<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[Banking Stack Notes]]></title><description><![CDATA[Banking Stack Notes]]></description><link>https://bankingstacknotes.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Fri, 20 Nov 2026 15:45:46 GMT</lastBuildDate><atom:link href="https://bankingstacknotes.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[What a bank must be able to prove about AI.]]></title><description><![CDATA[Vietnam's AI Law, Decree 142/2026 and the high-risk list: a control map for what a bank must prove, with dates and 90-day actions.]]></description><link>https://bankingstacknotes.hashnode.dev/what-a-bank-must-prove-about-ai</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/what-a-bank-must-prove-about-ai</guid><category><![CDATA[AI]]></category><category><![CDATA[Governance]]></category><category><![CDATA[banking]]></category><category><![CDATA[risk]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:50 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/c71bb8de-6210-4460-92be-369710c42bb2.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>The decision this quarter is not whether to keep using AI. It is whether to register the automated credit-decisioning engine, produce the file that defends it, and keep it in production — or stop automating decisions until someone can produce that file. The question that follows never reaches a slide: what must we put on the table, and by what date?</p>
<p>Vietnam answers part of that in law, and the answer is procedural: a classification, a file, a log, an override, a report inside a clock. Good news if you have run an audit. Bad news if you thought governance meant a policy PDF.</p>
<h2>What changed</h2>
<p><strong>1 March 2026 — the AI Law took effect.</strong> Law 134/2025/QH15 (passed 10 December 2025, eight chapters, 35 articles) covers development, provision, deployment and use, and Article 33 repeals Chapter IV of the digital technology industry law (71/2025/QH15). Article 4 says AI does not replace human authority and responsibility — further, the source material notes, than the EU model.</p>
<p><strong>15 August 2026 — the high-risk list began to bind, and banking is on it.</strong> Decision 33/2026/QĐ-TTg (issued 30 June 2026) covers six sectors. The banking examples are the two a consumer lender is likeliest to run: automated execution of electronic transactions, and automated credit-granting decisions. Press summaries count roughly 46 systems; read the annex.</p>
<p><strong>1 May 2026 — classification became the institution's liability.</strong> Decree 142/2026/NĐ-CP (issued 30 April 2026) requires self-classification before use, portal notification for medium and high risk with an automatic identifier, re-notification within 15 working days if risk rises, and incident reports in 72 hours, 5 working days or 15 days by severity.</p>
<p><strong>Since 1 March 2026 — the operator pays, and inspection asks for the file.</strong> Article 29 places compensation with the deployer where a high-risk system was lawfully used and damage still occurred, with recourse only where a contract provides it. Article 28 requires the technical file, trace logs and training data.</p>
<p><strong>1 October 2026 — no dedicated AI penalty decree, which is not the same as no enforcement.</strong> Decree 330/2026/NĐ-CP already fines personal-data processing in AI systems at VND 20–50 million; the draft figures — up to VND 2 billion for organisations, 1 billion for individuals — remain proposals.</p>
<p><strong>1 September 2027 — the transition ends for finance.</strong> Systems running before 1 March 2026 have 18 months in finance, health and education, and 12 months elsewhere.</p>
<h2>The decision framework</h2>
<p>Four roles exist — developer, provider, deployer, user — and one organisation can hold several. <em>Provider</em> includes software licensing, cloud services, APIs and embedded components, so a model API vendor is a provider and the consuming bank a deployer; fine-tune the model and you are both. I score seven criteria, weighted for a lender whose credit decisions are in scope.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Inventory and boundary</td>
<td>15%</td>
<td>Every system has an owner and a declared purpose</td>
</tr>
<tr>
<td>2</td>
<td>Classification file</td>
<td>15%</td>
<td>Reasoning recorded, mapped to the Decision 33 annex, signed by an officer</td>
</tr>
<tr>
<td>3</td>
<td>Human oversight you can test</td>
<td>20%</td>
<td>Logged override, measured intervention rate, a tested stop control</td>
</tr>
<tr>
<td>4</td>
<td>Technical file and traceability</td>
<td>15%</td>
<td>Any decision reproducible: inputs, model version, thresholds, timestamp</td>
</tr>
<tr>
<td>5</td>
<td>Conformity readiness</td>
<td>15%</td>
<td>Assessed before go-live; third-party cases identified; file maintained</td>
</tr>
<tr>
<td>6</td>
<td>Transparency and marking</td>
<td>10%</td>
<td>Customers know they face a system; content carries machine-readable marking</td>
</tr>
<tr>
<td>7</td>
<td>Contract and recourse</td>
<td>10%</td>
<td>Notice, audit, incident and indemnity duties; a Vietnamese contact point</td>
</tr>
</tbody></table>
<p>Two criteria are traps. On 3, a reviewer who approves 99% of outputs in two seconds is a signature, not oversight. On 5, the standards and the designated certifiers were still open points in November 2026 — you can prepare the file, not book the certificate.</p>
<h2>Options and trade-offs</h2>
<table>
<thead>
<tr>
<th>Option</th>
<th>Strength</th>
<th>Risk</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td><strong>Comply in place</strong> — register, classify, document what already runs</td>
<td>Fastest route to a register; no delivery freeze</td>
<td>Retrofit evidence; oversight claims may fail</td>
<td>Low in year one, steady run cost</td>
</tr>
<tr>
<td><strong>Build the evidence layer</strong> — registry, AI gate, standard logs</td>
<td>Compliance becomes a reusable control</td>
<td>Needs product, legal and data to agree; a slow gate gets bypassed</td>
<td>Moderate in year one, falling cost</td>
</tr>
<tr>
<td><strong>Replace or retire</strong> the high-risk decision systems</td>
<td>Removes the obligation at source</td>
<td>Expensive, loses decision speed already banked</td>
<td>High in year one, smallest surface after</td>
</tr>
</tbody></table>
<p>Most lenders need all three, applied per system.</p>
<h2>The numbers</h2>
<p>Effort is where estimates earn the label. A 2027 planning exercise I reviewed — anonymised, an estimate, not a quote — put the load at roughly 0.5 FTE in IT governance and architecture, 0.2 in legal and compliance, 0.2 in data: about one FTE across three functions, plus an optional VND 150–300 million for external legal review and conformity support. Certification was deferred a year, and in November 2026 nobody can price it: standards and certifiers are still pending.</p>
<p>What is law rather than estimate: 15 working days to notify a risk increase; 72 hours, 5 working days and 15 days for incidents; VND 20–50 million under Decree 330/2026/NĐ-CP; roughly 1 September 2027 for finance.</p>
<p>The measurable control targets: 100% of use cases registered with a risk level, zero unclassified production systems, 100% of high-risk systems assessed before go-live, 100% of IT staff trained.</p>
<h2>What I would do in 30, 60 and 90 days</h2>
<p><strong>First 30 days — count systems, not vendors.</strong></p>
<ul>
<li>Inventory every system, including embedded features and every model API in use. Shadow use is inventory.</li>
<li>One owner per system; one sentence of boundary — what it decides, for whom, with what data.</li>
<li>Start the classification file; the reasoning you record now is the evidence later.</li>
<li>Open a contract gap list with vendors: notice, audit, incident reporting, indemnity, a contact point in Vietnam.</li>
</ul>
<p><strong>Days 31–60 — test one control, not seven.</strong></p>
<ul>
<li>Override a decision on your riskiest system, then ask someone who did not build it to retrieve the override, the input, the model version and the time.</li>
<li>Build the technical-file template from that exercise. If it cannot be filled for one decision, it will not be filled for a thousand.</li>
<li>Table-top a Saturday incident: who reports what, to whom, inside 72 hours.</li>
</ul>
<p><strong>Days 61–90 — gate the lifecycle, price the rest.</strong></p>
<ul>
<li>Add an AI gate to the project process — role, risk class, evidence owner — and retrofit it to live systems.</li>
<li>Decide which systems you defend, rebuild or retire, in that order.</li>
</ul>
<h2>What I would not do</h2>
<ul>
<li><strong>I would not buy a governance platform before the evidence set is defined.</strong> The tool encodes whatever file you describe.</li>
<li><strong>I would not let a vendor's risk statement, or the Ministry's self-classification tool, become the classification.</strong> Both are inputs; liability stays with the institution.</li>
<li><strong>I would not keep the register outside delivery.</strong> If a use case can reach production without touching it, it is already wrong.</li>
<li><strong>I would not read the missing penalty decree as an absence of enforcement,</strong> or assume Article 27 stops at state agencies without a written opinion.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>The law does not ask whether your AI is good. It asks what you can show: that you classified the system and said why, that a person can overrule it and the overrule is recorded, that incidents are reported in days not months, and that the file survives an inspection. Fund one person across three functions, name an owner per system, and treat September 2027 as a delivery date.</p>
<h2>Sources</h2>
<ul>
<li><a href="https://vanban.chinhphu.vn/?docid=216334&amp;orggroupid=1&amp;pageid=27160">Law 134/2025/QH15 on Artificial Intelligence</a>, Cổng thông tin điện tử Chính phủ</li>
<li><a href="https://chinhphu.vn/?pageid=27160&amp;docid=218029">Decree 142/2026/NĐ-CP detailing the AI Law</a></li>
<li><a href="https://chinhphu.vn/?pageid=27160&amp;docid=218658">Decision 33/2026/QĐ-TTg — list of high-risk AI systems</a></li>
<li><a href="https://xaydungchinhsach.chinhphu.vn/nghi-dinh-330-2026-nd-cp-ve-xu-phat-vi-pham-hanh-chinh-trong-linh-vuc-an-ninh-mang-119260824172446407.htm">Decree 330/2026/NĐ-CP — administrative penalties in cybersecurity</a></li>
<li><a href="https://luatvietnam.vn/tin-van-ban-moi/danh-muc-he-thong-tri-tue-nhan-tao-co-rui-ro-cao-tu-15-8-2026-186-110045-article.html">"Danh mục hệ thống trí tuệ nhân tạo có rủi ro cao từ 15/8/2026"</a>, LuatVietnam</li>
</ul>
<hr />
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[The RFP that survives procurement.]]></title><description><![CDATA[How to write requirements that force evidence, score bids that survive negotiation, and avoid the RFP questions vendors answer wrongly in 2026.]]></description><link>https://bankingstacknotes.hashnode.dev/the-rfp-that-survives-procurement</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/the-rfp-that-survives-procurement</guid><category><![CDATA[procurement ]]></category><category><![CDATA[RFP]]></category><category><![CDATA[enterprise software]]></category><category><![CDATA[banking]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:48 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/c0b7dac2-a4b7-44a4-87a6-d6986c3ec367.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>On 12 October 2026 I issued an RFP with two lots, thirteen automatic exclusion grounds and a four-week proof of concept on anonymised production data. Shortlist on 13 November. Demos this week. The question is not which platform wins. It is: when a vendor answers "fully compliant" about a feature that does not exist, what is procurement actually buying?</p>
<h2>What changed</h2>
<ul>
<li><strong>29 February 2024.</strong> Zalo's OpenAPI stopped returning full user fields to applications and webhooks outside Vietnam; foreign IPs get HTTP 200 and the fields vanish silently. My 2 October 2026 review of 22 platforms found no international CMP with a first-party Zalo connector. A gate on "native Zalo" filters out honest bidders first.</li>
<li><strong>16 April 2024 and 5 November 2024.</strong> Zalo folded its purpose taxonomy into three tags, then replaced the after-sales quota formula. Templates in circulation still ask about four tag-5 messages a month.</li>
<li><strong>1 January 2026.</strong> Law 91/2025/QH15 took effect; Decree 356/2025 replaced Decree 13/2023. Consent is per purpose, silence is not consent, and the burden of proof is the controller's. A subscription centre is not a consent ledger.</li>
<li><strong>19 August 2026.</strong> Decree 330/2026 took effect: penalties for cybersecurity and personal-data violations, including 30–50 million VND for processing data collected without consent. It is quoted as an anti-spam decree. It is not. The three-message and one-call caps sit in Decree 91/2020, Article 13.</li>
<li><strong>9 September 2026.</strong> The Ministry of Public Security was still taking comments on a draft spam decree. It is not issued, yet bidders are asked to certify compliance with it.</li>
<li><strong>13 November 2026.</strong> Shortlist notified, demos 16–20 November, POC 23 November–18 December, contract by 15 January 2027, go-live in Q2 2027.</li>
</ul>
<h2>The decision framework</h2>
<p>Gates before scores. Thirteen conditions are pass/fail: frequency caps, consent and data-subject rights, statutory compliance, ISO/IEC 27001, a SOC 2 report, a POC on real anonymised data, holdouts, unrestricted export, a Vietnamese interface, a hard copy, a complete price schedule, no misrepresentation, acceptance of the bid terms. Fail one and you are out.</p>
<p>Then the weights — the published matrix, with the last four criteria merged into two rows.</p>
<table>
<thead>
<tr>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>Functional fit: journeys, dynamic segments, cross-channel caps</td>
<td>20</td>
<td>A mid-flight edit to a paused journey, done live, that does not corrupt customer state</td>
</tr>
<tr>
<td>Regulatory and consent architecture</td>
<td>20</td>
<td>Per-purpose consent ledger with timestamp, channel and text, exportable; withdrawal effective in minutes</td>
</tr>
<tr>
<td>Zalo channel and Vietnam IP gateway</td>
<td>12</td>
<td>Named egress and webhook IPs in Vietnam, plus a monitor for silently dropped fields</td>
</tr>
<tr>
<td>Core-system integration</td>
<td>10</td>
<td>LOS, core lending and collections each with method, direction, frequency and effort in days</td>
</tr>
<tr>
<td>Three-year TCO and price mechanics</td>
<td>12</td>
<td>Every line filled, unit declared, overage formula and exclusions named</td>
</tr>
<tr>
<td>Reliability and operability in Vietnamese</td>
<td>16</td>
<td>99.5% monthly uptime with credits; correct diacritics in segments; named in-country support</td>
</tr>
<tr>
<td>Vendor viability, references and exit</td>
<td>10</td>
<td>Three contactable financial-services references in Vietnam; 24-month written roadmap; free export</td>
</tr>
</tbody></table>
<p>Answers are coded full, partial, roadmap, none: 100, 50, 25, 0. Mandatory items count double. The threshold is 70/100, below which no bidder reaches the POC.</p>
<h2>Options and trade-offs</h2>
<p>Three ways to structure the buy; the product is the easy part.</p>
<table>
<thead>
<tr>
<th>Structure</th>
<th>Strengths</th>
<th>Risks</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td>A. One contract, Zalo delivery bundled through the platform's local partner</td>
<td>One SLA, one invoice</td>
<td>Consent split across parties you did not choose; no lever if the channel fails</td>
<td>Licence-led and flat, indexed to file size; per-message cost inside partner margin</td>
</tr>
<tr>
<td>B. Two lots, two contracts, a three-party data agreement</td>
<td>Each layer replaceable without the other; two suppliers at renewal</td>
<td>You own the integration contract and the incident bridge</td>
<td>Front-loaded year one, then licence plus usage; 200–300 VND per ZNS message</td>
</tr>
<tr>
<td>C. Platform only; keep the incumbent aggregator</td>
<td>Fastest to sign</td>
<td>You inherit the integration backlog; a partner change becomes your project</td>
<td>Lowest year-one services; maintenance sits in your budget</td>
</tr>
</tbody></table>
<p>Option B comes from a finding I did not expect. The only vendor with a long, documented official Zalo partnership is a messaging platform, not a CMP, and it has a Vietnamese legal entity. The one international CMP with Zalo workflows in its documentation routes them through that same platform. Two layers is not a compromise; it is the shape of this market.</p>
<h2>The numbers</h2>
<ul>
<li>ZNS lists at 200–300 VND per message before VAT. The first CTA button is free, later ones 100 VND, a permission request 400 VND. Timeouts (15 seconds for authentication, 7,200 for others) are not billed.</li>
<li>Public list prices that make bids comparable: $2,000 a month for 15,000 contacts on one enterprise suite, $100 a month for 5,000 profiles on a mid-market platform, overage $0.009 per profile.</li>
<li>Estimate, not quote: $30,000–80,000 a year for platform plus channel at the low end, $80,000–250,000 in the middle, $250,000 and above at the top.</li>
<li>Two billing models punish compliance work: one counts deleted profiles in a monthly high-water mark, so importing and then deleting to serve a data-subject request is billed twice; another bills profiles regardless of consent.</li>
<li>Cost carries 12% and scores as (lowest TCO ÷ bidder TCO) × 100; a bidder at twice the lowest TCO still takes 6 of 12 points. It breaks ties, it does not decide.</li>
<li>Terms worth more than a discount: a 7% cap on annual increases, 120-day validity, POC priced as a line, export at no charge.</li>
</ul>
<h2>What I would do in 30 / 60 / 90 days</h2>
<p><strong>30 days.</strong> Turn every requirement into an evidence form: which document proves it, what the POC tests, which answer is unacceptable. Search each bidder's full documentation index for the channel name — fourteen megabytes with zero hits is evidence, a partner logo is not. Correct the Zalo block to dated values. Keep the lots separate.</p>
<p><strong>60 days.</strong> Write the POC script before the demos. Two use cases: abandoned eKYC recovery, and collections reminders capped across Zalo, SMS and calls. Ten acceptance metrics, including consent withdrawal inside five minutes, full field completeness through a Vietnam-IP gateway, a working holdout. Normalise prices to one unit. Three reference calls per vendor, same six questions, and ask for the person who ran the migration.</p>
<p><strong>90 days.</strong> Sign a one-page decision record: criteria, weights, scores, rejected options and why. Then the terms that outlast price — a 12-month minimum at first signing, free exit and export, service credits, audit right, an increase cap, a three-party consent agreement for the Zalo flow. Go/no-go before the Q2 2027 date, not after.</p>
<h2>What I would not do</h2>
<ul>
<li>I would not make a native Zalo connector pass/fail. No international platform has one; that gate selects for confidence, not capability.</li>
<li>I would not score analyst quadrants or review counts. One vendor's public review total went from 50 to 232 in one year; reprints are sponsored by the vendors in them; a platform missing from both 2025 quadrants still runs in production at a Vietnamese lender.</li>
<li>I would not accept technographic evidence. BuiltWith-class tools show a script on a page, not a contract.</li>
<li>I would not score a three-year TCO with blank lines or an undeclared unit. A blank line means "not included", and that is a reason to stop reading.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>Two decisions are yours, not the team's. First, we buy in two contracts, so the messaging channel can be replaced without replacing the platform. Second, the cheapest bid will not win: cost is 12% of the score. What we are buying is evidence: four weeks on our own data, under NDA, before anyone signs.</p>
<h2>Sources</h2>
<ul>
<li><a href="https://developers.zalo.me/changelog/v240229-gioi-han-du-lieu-theo-dia-chi-ip-7667">Zalo IP restriction, 2024</a></li>
<li><a href="https://zalo.cloud/zns/pricing">Zalo ZNS price list</a></li>
<li><a href="https://zalo.solutions/news/dieu-chinh-ve-quyen-loi-gui-thong-bao-zns-tag-3-hau-mai-/oe18w3bi7qzrq1bkgqpjc9l9">Zalo after-sales quota, 2024</a></li>
<li><a href="https://zalo.solutions/business-message/partners">Zalo accredited partners</a></li>
<li><a href="https://xaydungchinhsach.chinhphu.vn/nghi-dinh-330-2026-nd-cp-ve-xu-phat-vi-pham-hanh-chinh-trong-linh-vuc-an-ninh-mang-119260824172446407.htm">Decree 330/2026</a></li>
<li><a href="https://lsvn.vn/goi-dien-quang-cao-sau-17-gio-bi-phat-den-30-trieu-dong-a65376.html">Decree 91/2020, Article 13</a></li>
<li><a href="https://xaydungchinhsach.chinhphu.vn/de-xuat-cac-bien-phap-chong-ngan-chan-tin-nhan-rac-thu-dien-tu-rac-cuoc-goi-rac-119260909163123265.htm">Draft spam decree, September 2026</a></li>
<li><a href="https://www.salesforce.com/marketing/engagement/pricing/">Salesforce MC list pricing</a></li>
<li><a href="https://docs.customer.io/accounts/billing/how-we-bill.md">Customer.io billing rules</a></li>
<li><a href="https://insiderone.com/news/gartner-customers-choice-multichannel-marketing-hubs-2026/">Gartner Peer Insights announcement, 2026, vendor-published</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[Productivity suite cost lives in the licences you forgot.]]></title><description><![CDATA[Where a productivity suite bill comes from: term mechanics, dead seats, add-ons, storage and the levers worth 42% before any migration.]]></description><link>https://bankingstacknotes.hashnode.dev/productivity-suite-tco</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/productivity-suite-tco</guid><category><![CDATA[tco]]></category><category><![CDATA[licensing]]></category><category><![CDATA[Productivity]]></category><category><![CDATA[banking]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:45 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/5b598411-3e75-4161-be8f-0e4e0d3823b0.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>In October 2026 I put a one-page decision to the board: renew the productivity suite for about 42% less a year than renewing as-is, without changing vendor and without migrating anyone. The question underneath was never whether to change suite. It was how much of the increase was price, and how much was licences nobody remembered buying?</p>
<h2>What changed</h2>
<ul>
<li><strong>1 July 2026.</strong> Microsoft's list prices moved: Office 365 E3 from $23 to $26, up 13.0%; Microsoft 365 E3 from $36 to $39, up 8.3%; Microsoft 365 E5 from $57 to $60, up 5.3%; Business Standard from $12.50 to $14, up 12%. Office 365 E1 stayed at $10, Business Premium at $22, the Copilot add-on at $30.</li>
<li><strong>1 May 2026.</strong> Microsoft 365 E7 reached general availability at $99 per user per month. Decline it unless agents are a programme, not a pilot.</li>
<li><strong>Q4 2026.</strong> Part of Copilot moves to usage-based billing. The $30 add-on stops being the whole cost, and an invoice without a spending cap moves on its own.</li>
<li><strong>1 January 2026.</strong> Law 91/2025/QH15 took effect and Decree 356/2025 replaced Decree 13/2023. The cross-border transfer impact file and the notification to the Ministry of Public Security remain an extra layer of obligation. No suite has a data region in Vietnam: Google defaults to Singapore and global, Microsoft's nearest is Southeast Asia.</li>
<li><strong>30 June.</strong> Microsoft's financial year closes. May and June negotiations have been worth 2–6% more than the same conversation in other months.</li>
</ul>
<p>The term mechanics underneath are less visible than the prices: monthly costs 20% more than annual, annual-paid-monthly 5% more, triennial 5–12% less. The cancellation window is 7 days on annual and 72 hours on triennial, and seats cannot be reduced mid-term.</p>
<h2>The decision framework</h2>
<p>My weighting, applied to a 300-seat model. It puts half the score on two things a price sheet cannot show: whether the licences match the people, and whether security controls survive the saving.</p>
<table>
<thead>
<tr>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>Licence-to-headcount accuracy</td>
<td>20</td>
<td>Every licence maps to a named active person or a documented shared function, reconciled monthly</td>
</tr>
<tr>
<td>SKU fit and capability tiering</td>
<td>20</td>
<td>Role matrix signed by the CISO; frontline and kiosk users on F-series or kiosk SKUs; E5 only where two or more exclusive controls are needed</td>
</tr>
<tr>
<td>Contract term and price mechanics</td>
<td>15</td>
<td>Volatile populations on monthly, stable ones on triennial; windows diarised; annual increase capped in writing</td>
</tr>
<tr>
<td>Security and compliance continuity</td>
<td>20</td>
<td>Conditional Access, DLP, eDiscovery, retention and residency unchanged or better, with legal sign-off on the transfer file</td>
</tr>
<tr>
<td>Migration and hidden-cost exposure</td>
<td>15</td>
<td>Inventory of VBA and Power Query models, SSO dependencies and e-signature add-ins before any comparison</td>
</tr>
<tr>
<td>Operability and support</td>
<td>10</td>
<td>In-country support, partner pass-through disclosed, helpdesk headroom for the change window</td>
</tr>
</tbody></table>
<h2>Options and trade-offs</h2>
<table>
<thead>
<tr>
<th>Option</th>
<th>Strengths</th>
<th>Risks</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td>A. Renew as-is, no cleanup</td>
<td>Nothing to explain; no execution risk</td>
<td>You keep paying for dead seats, monthly premiums and unused add-ons</td>
<td>$421,200 at annual; $505,440 if the subscription already rolled to monthly</td>
</tr>
<tr>
<td>B. Optimise in place</td>
<td>No migration, no new vendor, CISO controls intact</td>
<td>Needs seat and SKU discipline twice a year</td>
<td>$294,565, 42% below the monthly baseline; licence $98,188 a year</td>
</tr>
<tr>
<td>C. Replace with Google Workspace Business Standard</td>
<td>Licence down 72%; Gemini included</td>
<td>Loses Conditional Access, Purview, Intune and desktop Office; VBA models stop</td>
<td>$274,968, only 6.7% below B; negative in year one after migration and dual-running</td>
</tr>
</tbody></table>
<p>A fourth shape sits between B and C: keep Exchange Online and identity, move the lightest 60% to a document layer. It lands at $247,212 over three years, cheaper than full replacement. Replacement wins only if hidden cost stays under roughly $200 per user, which the analysis judged hard.</p>
<h2>The numbers</h2>
<ul>
<li>Per user per month at July 2026 list: Microsoft 365 E3 $39, Microsoft 365 E5 $60, Office 365 E3 $26, Office 365 E1 $10, F3 $10, Copilot $30. F1 at about $2.25 and the kiosk SKU at about $2 are unverified.</li>
<li>The gap between Microsoft 365 E3 and Office 365 E3 is $13 per user per month, $156 a year, roughly $78,000 a year at 500 users.</li>
<li>Copilot at $30 a month is $360 a year per user, more than an Office 365 E3 seat, and more than the usage report usually justifies.</li>
<li>The 300-seat illustrative model at July 2026 list, excluding VAT, with 30 recoverable seats: $168,480 of licence a year at the expiring monthly term, $98,188 after seat cleanup, tiering and a 5% commercial improvement. TCO falls from $505,440 to $294,565.</li>
<li>Cost shape: licence dominates. One-off migration is budgeted at $50 per user for hybrid and $200 per user for full replacement, plus $300 per user of hidden cost in the full-switch case.</li>
<li>Storage appears twice: inside the tier, with 50 GB of mail on Office 365 E1 against 2 GB on F1, F3 and kiosk, and inside retention. Legal hold and archive grow without a seat change. One mail alternative in the comparison meters storage at $0.08 per gigabyte on top of $4 per user, and leaving is days of export work rather than a per-gigabyte exit fee.</li>
<li>Currency and tax: 26,300 VND to the dollar and 10% VAT. Comparing a gross figure against a net one is how two business cases disagree by 10% and both get signed.</li>
<li>Seat hygiene: in real reviews, 8–18% of committed seats no longer match actual use. Partner pass-through has narrowed from 4–8% to 1–3%; co-investment credits run from $50,000 to $1 million on strategic SKUs.</li>
</ul>
<h2>What I would do in 30 / 60 / 90 days</h2>
<p><strong>30 days.</strong> Export licence assignment, active users and billing term. Reconcile against the HR list. Mark dead seats, functional mailboxes and Copilot seats with no usage report. Settle with Finance whether the overrun is price or prepaid amortisation, because the remedy differs.</p>
<p><strong>60 days.</strong> Get the role-to-SKU matrix signed by the CISO. Collect three partner quotes with pass-through disclosed. Move stable populations from monthly to annual, or to annual-paid-monthly at 5% if cash flow is the real constraint. Kill add-ons that duplicate entitlements already inside E5, such as Power BI Pro at $10 and Phone System at $8.</p>
<p><strong>90 days.</strong> Execute the tiering, sign with a capped annual increase, and start a 30–50 person hybrid pilot to price the alternative with real data. Book the main negotiation for May–June 2027. Review the transfer file and the outsourcing conditions for systems classified level 3 and above.</p>
<h2>What I would not do</h2>
<ul>
<li>I would not put triennial on Copilot or any SKU whose usage I cannot forecast. The window is 72 hours and seats cannot come down.</li>
<li>I would not drop Microsoft 365 E3 to Office 365 E3 for $13 without a written CISO decision. That gap is Windows, Intune, Defender P1 and Entra ID P1, and losing Entra P1 loses Conditional Access.</li>
<li>I would not migrate while VBA and Power Query models, Entra SSO and Purview retention are load-bearing, or before legal signs the transfer file.</li>
<li>I would not present a three-year TCO that omits dual-running, migration tooling, rewrite effort, and the cost of buying back MDM, DLP and eDiscovery.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>You are approving a renewal, not a migration. The renewal costs about 42% less per year than renewing as-is, with no vendor change and no new compliance exposure. Moving everyone to Google saves a further 6.7% over three years while putting audit, e-signature and device management at risk. We will test that option on 30–50 people first.</p>
<h2>Sources</h2>
<ul>
<li><a href="https://www.microsoft.com/en-us/microsoft-365/enterprise/microsoft-365-plans-and-pricing">Microsoft 365 enterprise plans and pricing</a></li>
<li><a href="https://exelegent.com/wp-content/uploads/2026/06/Exelegent_M365_Pricing_Grid_2026.pdf">Microsoft 365 pricing grid, July 2026</a></li>
<li><a href="https://www.epcgroup.net/blog/microsoft-365-e7-agent-365-enterprise-licensing-2026">E7, Agent 365 and the July 2026 increases</a></li>
<li><a href="https://atonementlicensing.com/blog/microsoft-nce-pricing-2026/">New Commerce Experience term mechanics</a></li>
<li><a href="https://microsoftnegotiations.com/blog/microsoft-365-kiosk-licensing">Frontline, F1, F3 and kiosk licensing</a></li>
<li><a href="https://workspace.google.com/pricing">Google Workspace pricing</a></li>
<li><a href="https://www.onlyoffice.com/docspace-prices">ONLYOFFICE DocSpace pricing</a></li>
<li><a href="https://vanban.chinhphu.vn/?classid=1&amp;docid=214590&amp;pageid=27160">Law 91/2025/QH15 on personal data protection</a></li>
<li><a href="https://vanban.chinhphu.vn/?docid=216387&amp;pageid=27160">Decree 356/2025/ND-CP</a></li>
<li><a href="https://thuvienphapluat.vn/van-ban/Tien-te-Ngan-hang/Thong-tu-09-2020-TT-NHNN-an-toan-he-thong-thong-tin-trong-hoat-dong-ngan-hang-455885.aspx">Circular 09/2020/TT-NHNN on IT safety in credit institutions</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[A PMO that measures decisions, not status.]]></title><description><![CDATA[A decision-first PMO for banks in 2026: decision latency, dependency age, cycle time and variance, and how to instrument them from data you already hold.]]></description><link>https://bankingstacknotes.hashnode.dev/pmo-measures-decisions</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/pmo-measures-decisions</guid><category><![CDATA[PMO]]></category><category><![CDATA[programme-management]]></category><category><![CDATA[portfolio]]></category><category><![CDATA[banking]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:43 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/2c82ba29-9d99-477d-b6c7-140dffbd8089.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>One question ends most portfolio reviews quickly: which decision is late, who owns it, and how long has it waited? The answer is usually a status colour. The decision in front of a PMO this quarter is whether to keep measuring activity or start measuring decisions. The test: if your pack cannot name the oldest open decision and its owner, what is the weekly meeting for?</p>
<h2>What the evidence says</h2>
<ul>
<li><strong>The portfolio is a list, not an instrument.</strong> The index carries 58 line items, each with a category, a system, a code, a mapping code, a budget owner and a project type. A line without a mapping code is invisible to whoever holds the money.</li>
<li><strong>The allocation logic is documented and checked against itself.</strong> The workbook reports 129 of 152 allocation lines (85%) reproduced by rule, with 23 exceptions. Recurring and one-off investment differ: a 15-year allocation produces an annual figure, a five-year contract divides by five, and the first year counts only its remaining months.</li>
<li><strong>Responsibility is already mapped.</strong> The task catalogue, version 16, dated 14 September 2026, lists 16 tasks across 8 people, with an R, S and K matrix, monthly time shares, a backup per person, renewal alerts at 90, 60 and 30 days, and minutes within 24 hours.</li>
<li><strong>The operating model assumes single owners.</strong> The squad playbook dated 17 October 2026 assigns product scope to the product owner, technical design to the tech lead, architecture principles to the enterprise architect, risk acceptance to the risk director and budget to the product council.</li>
<li><strong>The old model's timing is the problem being solved.</strong> Five to seven departments in sequence, 9 to 18 months to market, against a target of 30–50% faster — a target, not a measurement.</li>
<li><strong>Governance has moved from gates to guardrails.</strong> The playbook replaces stage-gate approval with architectural and risk guardrails, and asks for three metric layers: business outcome, delivery health including DORA measures, and risk.</li>
</ul>
<h2>The decision framework</h2>
<p>Pick measures that change behaviour. Weights assume a bank where a slow decision is paid for next quarter.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Decision latency</td>
<td>20%</td>
<td>Median days from "needed" to "decided"; the oldest open decision named, owned and dated weekly</td>
</tr>
<tr>
<td>2</td>
<td>Cycle time</td>
<td>20%</td>
<td>Lead time from funded idea to production, not phase-by-phase percent complete</td>
</tr>
<tr>
<td>3</td>
<td>Dependency age</td>
<td>15%</td>
<td>Every cross-team dependency has an owner, a promised date and an age; the oldest ten published</td>
</tr>
<tr>
<td>4</td>
<td>Variance to plan</td>
<td>15%</td>
<td>Variance explained by driver, re-forecast monthly, not at quarter end</td>
</tr>
<tr>
<td>5</td>
<td>Change and rework rate</td>
<td>10%</td>
<td>Incoming changes priced at intake; equal-size swaps cost nothing; larger ones go to committee</td>
</tr>
<tr>
<td>6</td>
<td>Outcome per decision</td>
<td>10%</td>
<td>Each funded item names the outcome it buys and the date that outcome is checked</td>
</tr>
<tr>
<td>7</td>
<td>Traceability</td>
<td>10%</td>
<td>Every number traces to a source line, with one named owner per number</td>
</tr>
</tbody></table>
<p>Two things make this work: a decision log with four fields only — what, who, due, made — and a chair with authority to close items. The playbook's test is whether a product owner can say no without penalty.</p>
<p>The benchmark is mundane. An initial credit rating runs to 11 steps, with a majority vote, a 48-hour appeal and publication within 24 hours. If a rating process can publish its clock, a PMO can publish one for a change request.</p>
<h2>Options and trade-offs</h2>
<table>
<thead>
<tr>
<th>Option</th>
<th>Strength</th>
<th>Risk</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td><strong>Status pack</strong> (today)</td>
<td>No new tooling; familiar to sponsors</td>
<td>Measures activity; decisions stay invisible; drift is found late</td>
<td>Low cash cost, high management time, rising cost of late decisions</td>
</tr>
<tr>
<td><strong>Decision log plus latency metrics</strong></td>
<td>Live within a quarter on existing tools; changes behaviour first</td>
<td>Decays if the chair cannot close items, or if the numbers are gamed</td>
<td>Mostly effort and discipline; the cheapest credible option</td>
</tr>
<tr>
<td><strong>Portfolio data platform</strong></td>
<td>One traceable number from index to allocation; the pack builds itself</td>
<td>Two to three quarters of build; automates politics if rights stay vague</td>
<td>Highest build, lowest manual reporting</td>
</tr>
</tbody></table>
<p>Run the second; start the third only where the same number is reconciled by hand monthly.</p>
<h2>The numbers</h2>
<ul>
<li><strong>Portfolio scale:</strong> 58 line items, 152 allocation lines, 129 reproduced by rule (85%), 23 exceptions. Counts from the workbook, not estimates.</li>
<li><strong>Catalogue scale:</strong> 16 tasks, 8 people, one matrix, one backup per person, three renewal alerts per contract.</li>
<li><strong>Team shape:</strong> squads of 5–9 people lasting 18–36 months; 4–6 squads for a mid-size bank, 8–15 for a large one; 10–20% of budget reallocated each quarter.</li>
<li><strong>Targets, not measurements:</strong> 30–50% faster time to market against a 9–18 month baseline. Hypotheses until a baseline exists.</li>
<li><strong>Effort:</strong> the sources carry no person-day estimate for a PMO rebuild. The first quarter is instrumentation on data already held: analyst work, not platform work.</li>
</ul>
<h2>What I would do in 30 / 60 / 90 days</h2>
<h3>Next 30 days — measure what you already have</h3>
<ul>
<li>Publish four numbers on existing data: decision latency, oldest dependency age, cycle time per value stream, variance with drivers. One page.</li>
<li>Open a decision log with four fields, backfilled one quarter so the first median is real.</li>
<li>Name the chair and quorum for each decision type in the playbook's list.</li>
</ul>
<h3>Days 30 to 60 — stop reporting activity</h3>
<ul>
<li>Retire percent complete. Replace it with dates: committed, forecast, and when the forecast last moved.</li>
<li>Put an owner and a promised date on every open dependency; publish the ten oldest.</li>
<li>Reconcile the task catalogue against squads of 5–9 people, so matrix and organisation agree.</li>
</ul>
<h3>Days 60 to 90 — move money on evidence</h3>
<ul>
<li>Move to a quarterly product review with 10–20% reallocation, showing what each decision bought.</li>
<li>Publish the decision log to the steering committee instead of the status narrative.</li>
<li>Keep the 23 exceptions visible. A rule that explains 85% of the lines is useful; one that is quietly overridden is not.</li>
</ul>
<h2>What I would not do</h2>
<ul>
<li><strong>Not measure percent complete.</strong> It moves when people talk, not when work finishes.</li>
<li><strong>Not count decisions as throughput.</strong> A busy log of trivial items is how expensive decisions get deferred.</li>
<li><strong>Not build the dashboard before the decision rights are written.</strong> You will automate an argument you have not settled.</li>
<li><strong>Not let a five-year total stand in for a plan.</strong> The allocation model is a reconciliation tool, not a commitment.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>The PMO's job is not to tell you a project is amber. It is to tell you which decision is late, who owns it, and what it delays. Ask three numbers: how long decisions wait, how old the oldest dependency is, how far the forecast has moved. If the pack cannot produce them, it is activity reporting with better typography.</p>
<h2>Sources</h2>
<ul>
<li>Internal project index and allocation workbook, 58 line items (no public URL).</li>
<li>Internal IT planning task catalogue, version 16, 14 September 2026 (no public URL).</li>
<li>Internal squad playbook, version 1.0, 17 October 2026 (no public URL).</li>
<li>Team Topologies, team types and interaction modes, <a href="https://teamtopologies.com/key-concepts">https://teamtopologies.com/key-concepts</a></li>
<li>McKinsey, ING's agile transformation, <a href="https://mckinsey.com">https://mckinsey.com</a></li>
<li>Spotify Engineering, squad, tribe and chapter model, <a href="https://engineering.atspotify.com">https://engineering.atspotify.com</a></li>
<li>FiinRatings, initial rating process with a 48-hour appeal, <a href="https://fiinratings.vn/Upload/Quy_trinh_XHTN_31_07_2025.pdf">https://fiinratings.vn/Upload/Quy_trinh_XHTN_31_07_2025.pdf</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[ISO 20022 is a data project, not an IT project.]]></title><description><![CDATA[The November 2026 SR was postponed; the data work was not. Field gaps, mapping failures and a 12-week ISO 20022 plan for banks.]]></description><link>https://bankingstacknotes.hashnode.dev/iso20022-is-a-data-project</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/iso20022-is-a-data-project</guid><category><![CDATA[#iso20022]]></category><category><![CDATA[payments]]></category><category><![CDATA[integration]]></category><category><![CDATA[banking]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:40 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/2191cb44-33d4-495e-947e-c763e4e21e8d.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>A payment operations manager sent me a six-week plan: map MT103 to pacs.008, switch the channel, declare cross-border done. I asked for it to be rewritten as a data plan, with the software line at the end. The decision is not whether to adopt ISO 20022 — CBPR+ coexistence ended in November 2025 — but how much work to finish before a deadline that, since 27 August 2026, has no date. Which fields can you populate, and who owns them?</p>
<h2>What changed</h2>
<p><strong>27 August 2026 — the release moved, the data did not.</strong> Swift postponed the entire November 2026 Standards Release after accepting a deferral of the structured address deadline; no new date was published. The Bank of England postponed its RTGS release to match.</p>
<p><strong>November 2025 — coexistence ended, for payments.</strong> CBPR+ cross-border payments moved fully to MX, and the deferral does not reopen MT. Trade finance did not move: MT700, MT707, MT734, MT754, MT756 and the MT760 family remain the working messages with no mandatory replacement.</p>
<p><strong>15 April 2026 — document rules are stable.</strong> The ICC Banking Commission's 2026–2027 action plan does not revise UCP 600 or ISBP 821: five business days to examine (article 14), one refusal notice (article 16).</p>
<p><strong>18 October 2026 — the map is eight rows.</strong> MT103 to pacs.008, MT202 to pacs.009, MT202 COV to pacs.009 plus underlying data, MT900 and MT910 to camt.054, MT940 and MT950 to camt.053, MT942 to camt.052 or camt.054, MT300 and MT320 to fxtr, and MT700 and MT760 with no mandatory equivalent.</p>
<h3>The field-level gap</h3>
<ul>
<li><strong>Address.</strong> <code>50K</code> free-format becomes structured <code>TwnNm</code>, <code>CtrySubDvsn</code> and <code>Ctry</code>. A record holding only <code>AdrLine</code> must be decomposed — a data-quality judgement, not a mapping rule.</li>
<li><strong>Charges.</strong> <code>71A</code> OUR, SHA and BEN map to <code>ChrgBr</code> DEBT, CRED and SHAR, with SLEV as the service variant. The mapping is trivial; agreeing which applies is not.</li>
<li><strong>References.</strong> <code>:20:</code> and <code>:21:</code> become <code>EndToEndId</code> and <code>InstrId</code>, and the trace rests on the 36-character <code>UETR</code> in block 3 field 121. Regenerate it mid-route and the chain breaks silently.</li>
<li><strong>Free text.</strong> <code>:72:</code> and <code>:79:</code> become <code>Purp</code> and <code>RmtInf</code>. What was stuffed into <code>:72:</code> is now a code from a maintained list, or a string that breaks reconciliation.</li>
<li><strong>Structure.</strong> A <code>pacs.008</code> is a GroupHeader plus transaction blocks. ISO 20022 itself mandates very few fields; the usage guideline mandates them, so two banks on different guidelines can both claim compliance and still fail each other.</li>
</ul>
<p>Five failures recur: address piled into <code>AdrLine</code>, a duplicated UETR, a charge bearer against the agreement, a purpose code outside the list, and a one-to-one mapping that runs but is not clean.</p>
<h2>The decision framework</h2>
<p>Score six criteria before the vendor conversation. The weights reflect a bank already running a payment hub; write yours down and hold them.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Field completeness at source</td>
<td>20%</td>
<td>Town, subdivision and country captured at origination, not parsed at send</td>
</tr>
<tr>
<td>2</td>
<td>Reference integrity</td>
<td>15%</td>
<td>One UETR per transaction from initiation, never regenerated, traceable on every hop</td>
</tr>
<tr>
<td>3</td>
<td>Reconciliation ownership</td>
<td>20%</td>
<td>camt.053 matches the ledger by reference and UETR same day, with a named owner</td>
</tr>
<tr>
<td>4</td>
<td>Code list discipline</td>
<td>15%</td>
<td>Purpose, charge bearer and reason codes validated at capture</td>
</tr>
<tr>
<td>5</td>
<td>Counterparty readiness</td>
<td>15%</td>
<td>Tested with real correspondents on real messages, not mock-ups</td>
</tr>
<tr>
<td>6</td>
<td>Change discipline</td>
<td>15%</td>
<td>A field change is a versioned mapping with tests, not a channel edit</td>
</tr>
</tbody></table>
<p>Criterion 3 carries most weight because nobody assigns it: the hub produces camt.053, the ledger the booking, and the gap is where a programme quietly fails.</p>
<h2>Options and trade-offs</h2>
<table>
<thead>
<tr>
<th>Option</th>
<th>Strength</th>
<th>Risk</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td><strong>Translate at the edge</strong> — map in the hub, leave sources alone</td>
<td>Fastest to demonstrate, no core release, contains blast radius</td>
<td>Every new field is another mapping in the same place; weak data returns</td>
<td>Front-loaded integration, then mapping maintenance; small licence change</td>
</tr>
<tr>
<td><strong>Fix at source</strong> — structured capture in channel, ERP and core</td>
<td>Removes the cause; one standard serves every message</td>
<td>Multi-team, no visible milestone for months, competes with revenue work</td>
<td>High people cost year one, falling repair cost after</td>
</tr>
<tr>
<td><strong>Rebuild the data spine</strong> — canonical party, account and instruction model</td>
<td>Best for reconciliation; makes the next standard cheaper</td>
<td>Longest path, hardest governance, easy to over-scope into a core replacement</td>
<td>Largest year one, lowest marginal cost per change</td>
</tr>
</tbody></table>
<p>Most banks do all three, in the wrong order: fix the fields reconciliation depends on, translate the rest, then argue about the canonical model.</p>
<h2>The numbers</h2>
<p>The sources carry no cost figure; I will not invent one. What they quantify is the shape of the work.</p>
<ul>
<li><strong>Five business days</strong> to examine under UCP 600 article 14; <strong>one notice</strong> to raise discrepancies under article 16. Timers that short are set by clean data, not harder work.</li>
<li><strong>Eight MT-to-MX rows</strong> make up the conversion map; <strong>ten trade message types</strong> stay on MT. That asymmetry is the biggest planning fact in the programme.</li>
<li><strong>Five recurring errors</strong> and <strong>six exception reason codes</strong> — invalid account, closed account, name mismatch, wrong BIC, insufficient funds, plus a sanctions hold — show where operating cost lands after go-live.</li>
</ul>
<p>My estimate: in a mid-size bank effort concentrates in address and name data, counterparty testing, and repair volume in the first two quarters. Licence is the smallest line. A business case built on licence delta will understate the programme.</p>
<h2>What I would do in 30 / 60 / 90 days</h2>
<p>Twelve weeks, independent of the new date.</p>
<p><strong>Weeks 1–4 — count, do not design.</strong> Run an address census: how many party records can produce town, subdivision and country today. Name one owner per field; two owners means none. Rank last quarter's repair reasons; that is the backlog. List the counterparties that must test with you.</p>
<p><strong>Weeks 5–8 — fix at source, one field at a time.</strong> Start with account identification, name and end-to-end reference. Put validation in front of the send, not behind the rejection. Write the exception playbook against the reason codes. Run one full round trip with a correspondent on real messages.</p>
<p><strong>Weeks 9–12 — prove it, then freeze it.</strong> Measure repair and straight-through rates before and after on the same volume. Freeze the mapping under version control with tests. Record the trade finance decision separately. Tell the sponsor what moved, what did not, and what the next Swift announcement triggers.</p>
<h2>What I would not do</h2>
<ul>
<li><strong>I would not read the deferral as extra time.</strong> It removed a date, not a work item; address data was already the long pole.</li>
<li><strong>I would not buy a new channel to fix a data problem.</strong> A cleaner front end on an unclean core produces the same exceptions faster.</li>
<li><strong>I would not run a one-to-one field mapping.</strong> That is how a programme ships a system that passes validation and still cannot reconcile.</li>
<li><strong>I would not let the UETR be generated anywhere except at initiation.</strong> Every downstream regeneration is a broken trace found later by operations, not tests.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>ISO 20022 is not a software upgrade; it is a change in the quality of the data the bank captures. The deadline moved in August 2026, but addresses, names, references and reconciliation cannot be compressed later. Twelve weeks now is cheaper than a rescue programme after a cutover nobody can date. Approve the data work before the software.</p>
<h2>Sources</h2>
<ul>
<li><a href="https://swift.com/standards/iso-20022">Swift — ISO 20022 and CBPR+</a></li>
<li><a href="https://swift.com/standards">Swift — standards and MyStandards</a></li>
<li><a href="https://bankofengland.co.uk">Bank of England — delay to the November 2026 RTGS standards release (27 August 2026)</a></li>
<li><a href="https://library.iccwbo.org">ICC — UCP 600, ISBP 821, URDG 758, URC 522, URR 725</a></li>
<li><a href="https://doccredit.world">Documentary Credit World — ICC Banking Commission 2026–2027 action plan: no UCP revision (15 April 2026)</a></li>
<li><a href="https://tradefinance.training">TradeFinance.training — ICC releases updated versions of eUCP and eURC (3 July 2023)</a></li>
<li><a href="https://iso20022.org">ISO 20022 — message repository and code lists</a></li>
<li><a href="https://swift.com/our-solutions/swift-gpi">Swift — gpi and UETR</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[Start here: six questions about banking integration]]></title><description><![CDATA[Six questions that decide a bank's integration spend in 2026, with the dates and sources behind each — and what to read next.]]></description><link>https://bankingstacknotes.hashnode.dev/integration-start-here</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/integration-start-here</guid><category><![CDATA[integration]]></category><category><![CDATA[architecture]]></category><category><![CDATA[banking]]></category><category><![CDATA[api]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:38 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/80e67abe-a5c5-4d0b-adce-59a3ecd23d9a.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>A vendor shortlist arrived before the inventory did. That order is why integration programmes fail politely: the decision is taken in a steering committee, and the evidence arrives six months later from operations. This series works the other way round. Six questions, in the order I ask them, each answered with a date and a source, and each ending in a trade-off rather than a pitch.</p>
<h2>The six questions</h2>
<h3>1. Which integration jobs do I still own?</h3>
<p>A bus and an API layer are not versions of each other; they encode opposite philosophies. A centralised bus puts routing, transformation, protocol mediation and orchestration in the middleware — smart pipes, dumb endpoints. API management keeps the logic in the service and the gateway thin — smart endpoints, dumb pipes. The jobs did not disappear when the products did. Mainframe and SAP adapters, long-running orchestration, B2B and EDI still need an owner, and usually a team.</p>
<h3>2. Which of my platforms has a date on it?</h3>
<p>The dates are public and close. Apache ServiceMix was retired in March 2026 and Apache Synapse sits in the Attic; Mule 3 Community Edition is end of life; Oracle Fusion Middleware 12c, the platform under Service Bus and SOA Suite, ends Premier Support in December 2026; SAP Process Orchestration 7.5 ends mainstream maintenance in 2027; BizTalk Server 2020 reaches end-of-sale on 31 March 2027; Red Hat 3scale ends maintenance on 30 June 2027. Ownership moves too: IBM completed its acquisition of Confluent in March 2026, on top of webMethods in 2024. Anything with a date inside three years belongs in this quarter's plan.</p>
<h3>3. Is ISO 20022 a format change or a data change?</h3>
<p>Format changes are projects; data changes are programmes. Cross-border coexistence ended in November 2025, and on 27 August 2026 Swift postponed the entire November 2026 Standards Release — the last date moved, not the work. What remains is address structure, reference integrity, charge bearer and reconciliation ownership, measured against a five-business-day examination window rather than a sprint calendar.</p>
<h3>4. Do I need a payment hub, or do I already have three?</h3>
<p>A payment hub is a layer that normalises and routes instructions across channels and banks. Many banks have that function spread across a core, a channel and a file-based integration, with no single routing table. The test is one question: can one place tell you which channel a payment will take, by currency and value, and what happens when it fails? Exception codes — invalid account, name mismatch, wrong BIC, insufficient funds — should be run by a written playbook, not by whoever is on shift.</p>
<h3>5. What is a corporate banking platform made of?</h3>
<p>Seven layers, from accounts and payments through trade finance, credit, FX and cash management. A mature bank runs four to six systems across them, and no single platform is best at all seven. That is why "replace the platform" is usually the wrong sentence. The question is which layer is missing, and whether it should be bought, rented or built.</p>
<h3>6. How do I price this when every enterprise price is quote-only?</h3>
<p>Price the metric, not the product. The category itself was estimated at USD 5.24 billion in 2025 and USD 6.92 billion in 2026, and per-call, per-unit, flat rate, credit consumption and per-core licensing behave completely differently as volume grows. Buyer-side estimates put large-enterprise discounts at 20–40% off list and first-year professional services at 20–40% of licence, and the hidden lines — environments, add-on analytics, connectors billed per environment, LLM tokens — are where budgets break. If the pricing metric does not grow roughly linearly with your volume, the business case is not finished.</p>
<h2>What to read next</h2>
<p>Start with <a href="/esb-isnt-dead">ESB isn't dead. It stopped being the default answer.</a> — the life-cycle dates, the three options, and the 90-day plan for an incumbent integration layer. Then read <a href="/iso20022-is-a-data-project">ISO 20022 is a data project</a> if payments are in scope this year, and <a href="/corporate-banking-build-buy-extend">Corporate banking platforms: build, buy or extend</a> if the decision is at layer level rather than interface level.</p>
<h2>Sources</h2>
<ul>
<li><a href="https://www.channelnews.fr/ibm-et-kong-dominent-le-magic-quadrant-de-la-gestion-des-api-159309">ChannelNews — Gartner Magic Quadrant for API Management 2026, report dated 6 October 2026</a></li>
<li><a href="https://www.silicon.fr/cloud-1370/gestion-api-magic-quadrant-2026-229527">Silicon.fr — API management market analysis (9 October 2026)</a></li>
<li><a href="https://access.redhat.com/articles/7141474">Red Hat — 3scale end of life (8 May 2026)</a></li>
<li><a href="https://attic.apache.org/projects/servicemix.html">Apache Software Foundation — ServiceMix in the Attic</a></li>
<li><a href="https://techcommunity.microsoft.com/blog/integrationsonazureblog/microsoft-biztalk-server-product-lifecycle-update/4478559">Microsoft — BizTalk Server product life-cycle update</a></li>
<li><a href="https://blogs.oracle.com/fusionmiddlewaresupport/end-of-premier-and-extended-support-for-oracle-fusion-middleware-12c">Oracle — end of Premier and Extended Support for Fusion Middleware 12c</a></li>
<li><a href="https://jp.newsroom.ibm.com/2026-03-18-ibm-completes-acquisition-of-confluent,-making-real-time-data-the-engine-of-enterprise-ai-and-agents">IBM — completion of the Confluent acquisition (March 2026)</a></li>
<li><a href="https://www.researchandmarkets.com/reports/5785642/api-management-market-report">Research and Markets — API Management market report 2026</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[Estimating a mobile programme by screen count is how you get late.]]></title><description><![CDATA[Why screen-count estimates fail on banking mobile programmes: integration links, native eKYC work, compliance evidence and the questions to put to a…]]></description><link>https://bankingstacknotes.hashnode.dev/estimating-mobile-programmes</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/estimating-mobile-programmes</guid><category><![CDATA[Estimation]]></category><category><![CDATA[mobile]]></category><category><![CDATA[programme-management]]></category><category><![CDATA[banking]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:35 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/1fcadf53-7b05-42e5-bf5c-e1d376db2738.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>On 4 December 2026 a consumer-finance lender reaches the confirmation gate on a two-app mobile programme. The decision is not React Native versus Flutter. It is whether to accept a supplier's fixed price built from a screen count. My question in that room: your own two versions of the baseline differ by 66 screens and 411 person-days, so what exactly is being bought, per screen?</p>
<h2>What the evidence says</h2>
<ul>
<li><strong>The screens were never counted.</strong> One version of the anonymised baseline carries 217 screens and 1,374 person-days; a later one 283 and 1,785. Both mark the count as assumed; counting is the supplier's first-month work.</li>
<li><strong>Most of the work is not a screen.</strong> The scope names 14 links to systems that stay, 13 services across two backends (7 customer, 6 sales), 32 security requirements in 12 groups and 9 shared packages at 205 person-days. Tests are 34% of development effort; security 11%.</li>
<li><strong>The unit does not reproduce.</strong> A hand-written screen takes 2, 4 or 7 days by complexity; a carried-over screen 22% of that, a new screen with AI tools 55%. Applied to 88 new screens, the inputs imply about 7.3 days each — above the document's 7-day maximum.</li>
<li><strong>The AI assumption is the widest lever.</strong> The base case assumes a 45% saving on new screens; the client's own review of published measurements ranges from no gain to about 30% on single tasks.</li>
<li><strong>The options differ by a fifth, not a factor.</strong> Reuse 1,785; Flutter all-new 2,155 (+20.7%) on the same calendar; hybrid 2,305 (+29.1%); native 3,125 (+75.1%). Stated range 1,580–2,310.</li>
<li><strong>The dates carry conditions.</strong> Proposal 16 November 2026; confirmation gate 4 December 2026; contract by 3 February 2027 or first release slips from 25 November 2027 to 10 February 2028; design gate 26 March 2027 is the last date to change technology without rework. Float: three weeks.</li>
</ul>
<h2>The decision framework</h2>
<p>Score the estimate, not the deck. Weights assume a regulated lender.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Productivity basis by screen class</td>
<td>20%</td>
<td>Days per class bottom-up, reuse and new separated, AI gain measured rather than assumed</td>
</tr>
<tr>
<td>2</td>
<td>Proof of reuse and build rights</td>
<td>15%</td>
<td>Build rights confirmed; eKYC vendor letter for the current framework and 16 KB pages</td>
</tr>
<tr>
<td>3</td>
<td>Integration ownership</td>
<td>15%</td>
<td>All 14 links owned and dated; status events (about 19 person-days) and no-chip route (about 25) priced</td>
</tr>
<tr>
<td>4</td>
<td>Native layer risk</td>
<td>15%</td>
<td>Named Swift and Kotlin engineers; chip reading on real devices by the design gate</td>
</tr>
<tr>
<td>5</td>
<td>Compliance evidence</td>
<td>15%</td>
<td>32 requirements mapped to evidence; Circular 77/2025; two penetration tests; AA and 5-second cold start tested</td>
</tr>
<tr>
<td>6</td>
<td>Calendar credibility</td>
<td>10%</td>
<td>1,785 person-days resourced inside the calendar, or own dates with reasoning; backends and database link dated</td>
</tr>
<tr>
<td>7</td>
<td>Run and exit</td>
<td>10%</td>
<td>Run effort in person-days, fix times, keys and store accounts with the client</td>
</tr>
</tbody></table>
<p>Three checks sit in front of the scoring: the eKYC kit supports the current framework and the 16 KB page size; shielding covers the chosen code and the Circular 77/2025 checks; the client holds source code and build rights. Any failure prices the work at the Flutter effort of 2,155. Circular 09/2020 adds the same demands from the regulator's side: provider assessment, contract clauses, source-code handover, no production data in tests.</p>
<h2>Options and trade-offs</h2>
<table>
<thead>
<tr>
<th>Option</th>
<th>Strength</th>
<th>Risk</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td><strong>React Native reuse</strong> (1,785)</td>
<td>Smallest effort; one code base for two apps</td>
<td>Three unverified checks and a 45% AI assumption</td>
<td>Build-heavy in year one, then support; cheapest if the checks hold</td>
</tr>
<tr>
<td><strong>Flutter, all new</strong> (2,155; +20.7%)</td>
<td>No dependence on existing code or build rights; same calendar</td>
<td>All 283 screens new; two skill pools to keep</td>
<td>A fifth more build; reuse question closed</td>
</tr>
<tr>
<td><strong>Native</strong> (3,125; +75.1%)</td>
<td>Closest to platform, no bridge to defend</td>
<td>Two code bases, slowest to staff, largest test surface</td>
<td>Highest build and run: two release trains</td>
</tr>
</tbody></table>
<p>The stated range spans 41% of the base case — wider than the 20.7% gap between the leading options.</p>
<h2>The numbers</h2>
<p>No rates, prices, budget or reserve appear in the sources, so any money figure attached to this programme is somebody's assumption. What is quantified:</p>
<ul>
<li><strong>Build:</strong> 1,374 or 1,785 person-days, against a stated range of 1,215–1,776 or 1,580–2,310. Client estimates, not quotes.</li>
<li><strong>Client side:</strong> 355 or 460 person-days, including a mobile technical lead and reviewers. One version counts 70 design-agency days.</li>
<li><strong>Excluded:</strong> two backends (13 services), the admin portal, the 14 links, platform capacity, data migration, legacy retirement, third-party licences, business-side effort.</li>
<li><strong>Run:</strong> the supplier must propose annual run effort in person-days; no figure is stated. An in-house alternative puts first release 7.1 months later at the same effort.</li>
</ul>
<h2>What I would do in 30 / 60 / 90 days</h2>
<h3>Next 30 days — to the 4 December 2026 gate</h3>
<ul>
<li>Count screens by class before anyone prices them.</li>
<li>Demand written evidence for the three checks: eKYC vendor letter, shielding documentation against Circular 77/2025, build rights.</li>
<li>Ask for days per screen class, reuse versus new, with the productivity basis. Refuse a blended rate per screen.</li>
</ul>
<h3>Days 30 to 60 — to contract</h3>
<ul>
<li>Decide 3 February 2027 deliberately: sign by then, or accept 10 February 2028 and rebuild the case.</li>
<li>Take the two-stage contract: foundation stage 18 January to 26 March 2027 on time and materials with a ceiling, break right at the design gate; then fixed per release from 29 March 2027, paid on acceptance.</li>
<li>Keep architecture, security, keys, store accounts and production access. A supplier holding the keys holds your release.</li>
</ul>
<h3>Days 60 to 90 — to the 26 March 2027 design gate</h3>
<ul>
<li>Verify the three checks on real devices. After this gate, changing technology means rewriting the app code.</li>
<li>Confirm the backends and database link have committed dates; the baseline calls this the largest dependency risk.</li>
<li>Re-derive the per-release gates; they do not map onto the four public releases.</li>
</ul>
<h2>What I would not do</h2>
<ul>
<li><strong>Not price per screen.</strong> The count moved 30% between versions of the same document, and the effort moved with it.</li>
<li><strong>Not plan on a 45% AI saving</strong> when the same document's review finds no gain to 30% on single tasks.</li>
<li><strong>Not start build before the backends and the national database link are committed.</strong> That buys an app that cannot go live on its date.</li>
<li><strong>Not treat three weeks of float as slack,</strong> nor the Tet freeze from 20 December 2027 to 9 February 2028, itself an assumption.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>A mobile app is priced per screen but paid for in integration, security evidence and testing. Here the screen count moved by 30% between two versions of the same document, and the effort moved with it. Approve the estimate only when the supplier shows days per screen class, states its productivity basis, and proves the three technical checks by 26 March 2027.</p>
<h2>Sources</h2>
<ul>
<li>Anonymised scope and effort baseline, two versions, 217 and 283 screens (internal, no public URL).</li>
<li>Team Topologies, team types and interaction modes, <a href="https://teamtopologies.com/key-concepts">https://teamtopologies.com/key-concepts</a></li>
<li>McKinsey, ING's agile transformation, on owning run, <a href="https://mckinsey.com">https://mckinsey.com</a></li>
<li>Spotify Engineering, squad, tribe and chapter model, <a href="https://engineering.atspotify.com">https://engineering.atspotify.com</a></li>
<li>Reserve Bank of India, outsourcing directions (12 August 2022), <a href="https://www.rbi.org.in/Scripts/NotificationUser.aspx?Id=12378&amp;Mode=0">https://www.rbi.org.in/Scripts/NotificationUser.aspx?Id=12378&amp;Mode=0</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[Start here: ECM in 2026]]></title><description><![CDATA[Where to start with content platforms in 2026: lifecycle cliffs, analyst reshuffles, open-source limits, archive rules and five-year TCO.]]></description><link>https://bankingstacknotes.hashnode.dev/ecm-start-here</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/ecm-start-here</guid><category><![CDATA[ECM ]]></category><category><![CDATA[Document Management,]]></category><category><![CDATA[architecture]]></category><category><![CDATA[banking]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:32 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/036d35c5-f3ef-4bab-9756-ef9064339089.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Every architecture series needs a first page that is not an opinion. This one is a reading order. Content platforms are in the middle of a generational change — brands renamed, support dates passed, analyst positions moved, a regulator's transition clock running — and most of the internal documents circulating about it are out of date. Six questions organise the rest of the series. Start with the one that matches the decision in front of you. Dates are as of November 2026.</p>
<h2>Six questions this series answers</h2>
<h3>1. Who is still building this software?</h3>
<p>IBM renamed FileNet to Content Cortex, GA June 2026, and FileNet 5.5.x left End of Usage Support on 30 September 2026, with Extended Support to 30 September 2030. That is a product transition, not an exit. The distinction decides whether your next move is a version upgrade or a platform replacement.</p>
<h3>2. Should an analyst demotion change a buying decision?</h3>
<p>Gartner retired the Content Services Platforms quadrant and published the Document Management edition on 28 April 2026: 16 vendors, six Leaders, with OpenText moved from Leader to Visionary a year after Forrester still called it a Leader. Those reports measure cloud document management for a broad market, not whether a repository survives inspection. Read them as roadmap risk, not as a score.</p>
<h3>3. Is open source a real alternative?</h3>
<p>Less than it was two years ago. OpenKM closed its community edition at 7.0 in March 2026; Alfresco Community Edition 26.2, released 30 July 2026, ships some connectors binary-only and dropped Solr; Apache Chemistry, the CMIS reference, is in the Attic. Mayan EDMS 4.12.2, released 17 September 2026, remains genuinely open. Licence cost was never the variable — project governance is.</p>
<h3>4. What does a renewal actually cost?</h3>
<p>In the five-year model used here, keeping and upgrading costs 4.98 times L, against 9.05 times L for replacing with a peer platform, where L is current annual licence and support. The gap is migration, dual-run, re-integration and retraining, not the licence line. An estimate from a model, not a vendor quote.</p>
<h3>5. Which obligations cannot be negotiated?</h3>
<p>The Law on Archives 2024 with Decree 113/2025 set a 24-month transition to build or upgrade digital archive systems. Circular 05/2025/TT-BNV fixes the archive data structure and the eight-step digitisation process. Decree 337/2025 made tamper-evident audit logs and trusted timestamps the standard for electronically signed contracts, and Law 91/2025 sanctions unlawful cross-border transfer at up to 5% of prior-year revenue. A platform that cannot produce evidence is not cheaper.</p>
<h3>6. What do the first 90 days look like?</h3>
<p>Baseline before architecture: users, documents, interfaces, annual cost, incidents. One quote template, issued to the incumbent and two alternatives. Three demos on your own data. A five-year TCO in L, a sensitivity test, and a contract with an escalation cap and a tested export clause. Nothing in that list requires a preference for any vendor.</p>
<h2>What to read next</h2>
<p>Two notes carry the detail.</p>
<p><a href="/ecm-market-split-2026">The 2026 ECM market split: FileNet becomes Content Cortex</a> is the market picture: what changed by date, the seven criteria I score, three options with three-year cost shapes, and what a renewal buyer should ask now.</p>
<p><a href="/defending-an-ecm-renewal">Defending a legacy ECM renewal: five questions before you sign</a> is the procurement mechanics: usage, egress, connectors, migration and exit, the TCO model logic, and the one page a board actually needs.</p>
<p>If you are mid-renewal, read the second first. If you are building a business case, read them in reverse order.</p>
<h2>Sources</h2>
<ul>
<li><a href="https://www.ibm.com/docs/en/content-cortex/26.0.0?topic=notes-whats-new">IBM, What's new in Content Cortex 26.0.0</a></li>
<li><a href="https://www.ibm.com/support/pages/ibm-filenet-content-manager55x">IBM, FileNet Content Manager 5.5.x lifecycle</a></li>
<li><a href="https://www.ibm.com/docs/en/announcements/filenet-content-manager-evolves-content-cortex-next-generation-agentic-content-services">IBM, FileNet evolves to Content Cortex</a></li>
<li><a href="https://www.silicon.fr/workspace-1376/magic-quadrant-gestion-documentaire-2026-227009/amp">Silicon.fr, secondary analysis of the 2026 Gartner MQ for Document Management</a></li>
<li><a href="https://www.hyland.com/en/resources/alfresco-product-support-status">Hyland, Alfresco product support status</a></li>
<li><a href="https://www.openkm.my/blog/openkm-community-edition-v70-a-new-stage-for-open-source-document-management.html">OpenKM, Community Edition 7.0 announcement</a></li>
<li><a href="https://attic.apache.org/projects/chemistry.html">Apache Attic, Apache Chemistry</a></li>
<li><a href="https://learn.microsoft.com/en-us/purview/retention-limits">Microsoft Learn, limits for retention policies and labels</a></li>
<li><a href="https://costbench.com/software/document-management/opentext-content-suite/calculator/">CostBench, OpenText Content Suite cost calculator</a></li>
<li><a href="https://vanban.chinhphu.vn/?pageid=27160&amp;docid=213822">Decree 113/2025/ND-CP</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[The 2026 ECM market split: FileNet becomes Content Cortex.]]></title><description><![CDATA[FileNet became Content Cortex in June 2026, 5.5.x support ended 30 Sep 2026, OpenText fell to Visionary: what a renewal buyer should ask.]]></description><link>https://bankingstacknotes.hashnode.dev/ecm-market-split-2026</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/ecm-market-split-2026</guid><category><![CDATA[ECM ]]></category><category><![CDATA[Document Management,]]></category><category><![CDATA[architecture]]></category><category><![CDATA[enterprise software]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:29 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/d56db3be-b995-4fdb-b60c-fa049a0e0a04.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Two things arrived in the same quarter for the person holding the renewal signature: the product was renamed, and the version in production left support. Neither changed a single document in the repository. Both changed the answer to a question nobody had asked — what exactly is being renewed: a repository, a brand, or a support contract that expired seven weeks ago?</p>
<p>What follows is dated November 2026.</p>
<h2>What changed</h2>
<p><strong>28 April 2026.</strong> Gartner published the Magic Quadrant for Document Management 2026 — 16 vendors, six Leaders: Box, Doxis (ex-SER), Hyland, Laserfiche, M-Files, Microsoft. The Content Services Platforms quadrant is retired.</p>
<p><strong>June 2026.</strong> IBM Content Cortex 26.0.0 reached GA. IBM's wording: <em>"Content Cortex is IBM's next-generation content services system, and the successor to FileNet."</em> It absorbs FileNet Content Manager, CMOD, Content Manager (CM8) and FileNet Image Services.</p>
<p><strong>30 September 2026.</strong> FileNet Content Manager 5.5.x, GA 26 June 2018, left End of Usage Support. Extended Support completes 30 September 2030. FileNet 5.7.x, GA 20 June 2025, is the interim release IBM recommends.</p>
<p><strong>September 2026.</strong> A paid Premium Edition went on sale. IBM partner ISR reports existing FileNet Content Manager and Cloud Pak licensees receive Essentials "at the same price point"; Premium adds LegalHold MCP, Redaction MCP, a Model Gateway and WDU Services.</p>
<p><strong>March 2026.</strong> Hyland released Alfresco Content Services v26 — full support, deprecation March 2028, end of maintenance March 2029. Nuxeo LTS 2025 runs to 28 April 2028.</p>
<p><strong>December 2025.</strong> OpenText made Content Manager 25.4 the LTS release, supported to 30 November 2028. Non-LTS windows ran about three months: 25.1 ended 28 February 2026, 25.2 31 May 2026, 25.3 31 August 2026.</p>
<p>The change with no clean date unsettles committees most. OpenText moved from Leader to Visionary, a year after Forrester's Q1 2025 Content Platforms Wave named it a Leader. Hyland and M-Files were promoted from Challengers. IBM lands last of 16 on both axes in secondary analysis of the reprint — a number to use internally, not to quote. OpenText also sold eDOCS to NetDocuments for approximately USD 160 million.</p>
<h2>The decision framework</h2>
<p>Seven criteria, weighted for a lender already replacing its core system. Write the weights down before the demo.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Evidence and audit capability</td>
<td>20%</td>
<td>Retention per record class, tested legal-hold release, tamper-evident audit trail, ten-year signatures</td>
</tr>
<tr>
<td>2</td>
<td>Programme risk</td>
<td>20%</td>
<td>No second replacement inside the core-banking window; dated milestones, named owner</td>
</tr>
<tr>
<td>3</td>
<td>Five-year TCO</td>
<td>20%</td>
<td>Six lines costed: licence, implementation, infrastructure, migration, dual-run, retraining</td>
</tr>
<tr>
<td>4</td>
<td>Integration and business fit</td>
<td>15%</td>
<td>Every live interface re-proven, including APIs that are themselves changing</td>
</tr>
<tr>
<td>5</td>
<td>Operational maturity</td>
<td>10%</td>
<td>Three years of availability, P1/P2 counts, DR tests, runbooks</td>
</tr>
<tr>
<td>6</td>
<td>Roadmap and AI capability</td>
<td>10%</td>
<td>Kubernetes 1.29+, OpenShift 4.16+, Helm 4.0+, MCP agent access, a dated release train</td>
</tr>
<tr>
<td>7</td>
<td>Vendor and lock-in risk</td>
<td>5%</td>
<td>Annual export test, escalation cap, ownership changes counted</td>
</tr>
</tbody></table>
<p>Two are pass/fail: a supported release path with a date and an owner, and a full export — content, metadata, audit trail — that has actually run on your data.</p>
<h2>Options and trade-offs</h2>
<p>Costs are multiples of <strong>L</strong>, current annual licence and support — five-year model estimates, not quotes.</p>
<table>
<thead>
<tr>
<th>Option</th>
<th>Strength</th>
<th>Risk</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td><strong>Upgrade in place</strong> — 5.5.x → 5.7 → Cortex 26.x</td>
<td>ISR reports back-end, data model, database and interfaces unchanged, migration not required</td>
<td>Brand reads "legacy" while delivery does not; last of 16; renewal price unpublished; scarce skills</td>
<td>0.73L, then 1.10L and 1.05L — about 2.88L, front-loaded</td>
</tr>
<tr>
<td><strong>Move to a peer platform</strong> — Alfresco 26 or Nuxeo LTS 2025</td>
<td>Open standards, CMIS endpoints, governance, rising analyst standing</td>
<td>~20% of Hyland staff cut; Gartner-named packaging complexity; ten classes of evidence to reproduce</td>
<td>3.50L, then 1.50L and 1.35L — about 6.35L, peak inside your core replacement</td>
</tr>
<tr>
<td><strong>Consolidate on Microsoft 365</strong> — SharePoint, Purview, Premium</td>
<td>Cheapest cash, usually already licensed, labels in Purview, Copilot</td>
<td>1,000 retention labels per tenant, 13 policies for all sites, 100 sites per policy; shallow file plan</td>
<td>1.85L, then 0.55L — about 2.95L, largest expected compliance damage</td>
</tr>
</tbody></table>
<p>There is no fourth option called "extend as we are". After 30 September 2026, 5.5.x is unsupported; a renewal that funds that state is a deferral with a signature.</p>
<h2>The numbers</h2>
<table>
<thead>
<tr>
<th>Scenario</th>
<th>Five-year cash</th>
<th>Expected cost with risk</th>
</tr>
</thead>
<tbody><tr>
<td>Keep and upgrade</td>
<td>4.98 L</td>
<td>5.06 L</td>
</tr>
<tr>
<td>Replace with a peer ECM</td>
<td>9.05 L</td>
<td>9.43 L</td>
</tr>
<tr>
<td>Microsoft 365</td>
<td>4.05 L</td>
<td>4.45 L</td>
</tr>
<tr>
<td>Do not upgrade (reference)</td>
<td>5.20 L</td>
<td>6.20 L</td>
</tr>
</tbody></table>
<p>Replacement costs about <strong>1.82 times</strong> the upgrade path over five years, more with risk priced: 25% probability of a major event at 1.5L damage, against 5% for the upgrade. Microsoft is cheaper in cash — 0.81 times — with the largest damage estimate, 2.0L at 20%, its exposure being regulatory.</p>
<ul>
<li><strong>Implementation and services</strong> run <strong>50–150%</strong> of the year-one licence on commercial ECM (CostBench, OpenText example). A quote below 50% has exclusions.</li>
<li><strong>Dual-run</strong> is <strong>6–12 months</strong>, both systems licensed and operated. No vendor quotes that line.</li>
<li><strong>Hypercare</strong> for a new platform is <strong>12–18 months</strong>, landing on the stressed half of a core programme.</li>
<li><strong>Migration scales with exceptions, not documents.</strong> Double-digit exception rates mean hundreds of thousands to millions of records by hand — an estimate, and the part migration contracts exclude.</li>
<li><strong>Public anchors:</strong> M-Files USD 65 per seat per month; Laserfiche USD 73, 93 and 42 by tier; SharePoint Premium about USD 0.05 per page prebuilt, 0.10 custom, 0.15 per document; Box near USD 36.5k a year by third-party estimate. IBM, OpenText and Hyland publish nothing: the incumbent's business case is a quote that has not arrived.</li>
</ul>
<h2>What I would do in 30 / 60 / 90 days</h2>
<p><strong>First 30 days.</strong> Put every content product's lifecycle on one page — version, GA, support end, owner — from vendor pages, not a slide. Baseline users, documents, live interfaces, annual cost, three years of incidents. Request a written Cortex upgrade quote.</p>
<p><strong>Days 31–60.</strong> Run three scenarios on your data: extract one customer's full record within two working days; place and release a legal hold on a disputed set; export three years of audit trail for one document. Take two alternative quotes on one template, five years, six cost lines. Test the exception cases.</p>
<p><strong>Days 61–90.</strong> Build the TCO in L, replace L with the real figure, run the sensitivity: compliance at 10%, TCO at 30%; if the ranking flips, it was political. Negotiate three to five years with an escalation cap, a price hold, an export right tested annually. Put upgrade dates in the minutes.</p>
<h2>What I would not do</h2>
<ul>
<li><strong>I would not sign a renewal that names a price but not a destination release and a date.</strong></li>
<li><strong>I would not let the quadrant decide, in either direction.</strong> It ranks cloud document management for the mass market, not survival under inspection.</li>
<li><strong>I would not buy Premium before a process owner names the LegalHold or Redaction dependency.</strong> Otherwise it is a subscription.</li>
<li><strong>I would not migrate the repository inside the core-banking window.</strong> Two moving variables in one release destroys attribution.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>FileNet has a new name and a new version; the old version stopped being supported on 30 September 2026. Staying is the cheapest and safest route, but only with a dated upgrade plan. Replacing costs roughly 1.8 times as much over five years and puts audit evidence at risk. The decision is a date, not a logo.</p>
<h2>Sources</h2>
<ul>
<li><a href="https://www.ibm.com/docs/en/content-cortex/26.0.0?topic=notes-whats-new">IBM, What's new in Content Cortex 26.0.0</a></li>
<li><a href="https://www.ibm.com/support/pages/ibm-filenet-content-manager55x">IBM, FileNet Content Manager 5.5.x lifecycle</a></li>
<li><a href="https://www.ibm.com/docs/en/announcements/filenet-content-manager-evolves-content-cortex-next-generation-agentic-content-services">IBM, FileNet evolves to Content Cortex</a></li>
<li><a href="https://isr.de/en/news/am-ibm-filenet-ibm-content-cortex/">ISR, IBM Content Cortex for existing FileNet customers</a></li>
<li><a href="https://www.silicon.fr/workspace-1376/magic-quadrant-gestion-documentaire-2026-227009/amp">Silicon.fr, secondary analysis of the 2026 Gartner MQ</a></li>
<li><a href="https://www.hyland.com/en/resources/alfresco-product-support-status">Hyland, Alfresco product support status</a></li>
<li><a href="https://learn.microsoft.com/en-us/purview/retention-limits">Microsoft Learn, retention policy and label limits</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[E-signature is an evidence problem, not a UI problem.]]></title><description><![CDATA[Why a digitally signed credit contract fails in court, what Vietnamese law actually requires, and how to wire signing into a lending workflow.]]></description><link>https://bankingstacknotes.hashnode.dev/e-signature-is-an-evidence-problem</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/e-signature-is-an-evidence-problem</guid><category><![CDATA[digital identity]]></category><category><![CDATA[esignature]]></category><category><![CDATA[banking]]></category><category><![CDATA[compliance ]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:27 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/8c606619-c461-4f82-be60-61f43f84523b.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>The decision this quarter is not which signing screen to buy. It is which file you would rather hand a judge in three years: a PDF with a picture of a signature and a database row, or a record that shows who signed, in what capacity, at what moment, and that nothing changed after. The interface is reversible. The evidence is not.</p>
<p>The law is more settled than the market thinks. What is unsettled is the plumbing: which document gets which signature level, who validates the certificate, what proves the timestamp, and who keeps the original. An architecture problem dressed as a legal one.</p>
<h2>What changed</h2>
<p><strong>1 July 2024 — three tiers of legal value, not one.</strong> Law on E-Transactions 20/2023/QH15 (in force 1 July 2024) makes a data message admissible, refuses to deny an e-signature value for being electronic, and reserves equivalence with a handwritten signature for a digital signature or a certified specialised e-signature. OTP, PIN and click-wrap can be evidence; they do not clear that bar.</p>
<p><strong>10 April 2025 — the certificate became a control point.</strong> Decree 23/2025/NĐ-CP (in force 10 April 2025) covers certificates, specialised signatures, certificate authorities and signing software. An officer's certificate must name the title and organisation, which makes signing authority checkable.</p>
<p><strong>1 July 2025 — electronic notarisation opened.</strong> Decree 104/2025/NĐ-CP (in force 1 July 2025) allows it, and some localities already run it for mortgages with online security registration. Readiness varies by locality, notary office and registry.</p>
<p><strong>3 July 2024 — a court showed how the file is read.</strong> Appellate judgment 79/2024/DS-PT (3 July 2024, a provincial people's court), a transaction of roughly VND 11.9 billion: username, password and OTP on a registered device were accepted as the customer's will unless the customer proved bank fault. That is the bar a digital signature has to beat.</p>
<p><strong>15 August 2026 — online lending got new limits.</strong> Circular 29/2026/TT-NHNN (in force 15 August 2026) sets a small-value loan at VND 200 million at people's credit funds and VND 400 million elsewhere, and requires each institution to set its own exposure limit for electronic-channel lending. Earlier contracts continue unless amended.</p>
<p><strong>1 January 2026 — the identity record became regulated data.</strong> Law 91/2025/QH15 (in force 1 January 2026) governs the eKYC and biometric evidence the signature rests on — two records, two retention rules.</p>
<h2>The decision framework</h2>
<p>The Article 22.3 test is worth memorising, because every demo assumes it: the signature confirms the signer, is uniquely linked to them, sits under their control, detects later change, and is backed by a certificate and a secure creation device.</p>
<p>The chain breaks in a predictable order: the framework agreement never mentions electronic execution, the signature level is chosen by what the screen allows, the timestamp is the application server's clock, the export is a PDF, the mortgage is on paper. Each gap survives alone; together they are the file you lose.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Identity proofing</td>
<td>20%</td>
<td>eKYC per Circular 09/2020/TT-NHNN, biometrics where Decision 2345/QĐ-NHNN applies</td>
</tr>
<tr>
<td>2</td>
<td>Signature level against value</td>
<td>20%</td>
<td>Digital signature above the band where OTP stops being credible</td>
</tr>
<tr>
<td>3</td>
<td>Integrity and timestamp</td>
<td>15%</td>
<td>Trusted timestamp, integrity service re-verifiable years later</td>
</tr>
<tr>
<td>4</td>
<td>Retention and export</td>
<td>15%</td>
<td>Original data, full context, exportable for litigation</td>
</tr>
<tr>
<td>5</td>
<td>Authority of the bank's signer</td>
<td>10%</td>
<td>Certificate names title and organisation; delegation current</td>
</tr>
<tr>
<td>6</td>
<td>Collateral continuity</td>
<td>10%</td>
<td>Mortgage on the same chain, or a documented bridge</td>
</tr>
<tr>
<td>7</td>
<td>Certificate lifecycle and exit</td>
<td>10%</td>
<td>Validity checked at signing, revocation stored, escrow on exit</td>
</tr>
</tbody></table>
<p>Criterion 3 is the quiet one: an application log proves what your system recorded, not that the document existed unchanged. Criterion 6 is where the money is — a signed contract attached to a paper mortgage is a file with two halves and no bridge.</p>
<h2>Options and trade-offs</h2>
<table>
<thead>
<tr>
<th>Option</th>
<th>Strength</th>
<th>Risk</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td><strong>Tier by value</strong> — eKYC and OTP below a band, digital signature above</td>
<td>Fast, fits existing flows, cheapest to start</td>
<td>The band is a policy line, not a safe harbour; a disputed large loan still fails</td>
<td>Low capex, per-signature fees growing with volume</td>
</tr>
<tr>
<td><strong>Sign the credit contract, keep collateral on paper</strong></td>
<td>Strong instrument, minimal process change</td>
<td>Split file; disputes over integrity and priority; reconciliation forever</td>
<td>Signing licences plus unchanged notary and registration cost</td>
</tr>
<tr>
<td><strong>Whole digital file, electronic notarisation where available</strong></td>
<td>Cleanest chain; one record from offer to registration</td>
<td>Availability varies by locality and asset type; a fallback is mandatory</td>
<td>Highest integration cost, lower handling cost later</td>
</tr>
</tbody></table>
<p>A consumer loan and a mortgage should not be the same project.</p>
<h2>The numbers</h2>
<p>Three figures are regulatory, not estimated. VND 200 million at people's credit funds and VND 400 million elsewhere define a small-value loan. VND 11.9 billion was the transaction in the 2024 judgment that showed how much turns on the log. VND 20–50 million is the Decree 330/2026/NĐ-CP fine for personal-data processing in AI systems; deepfake face authentication can be fined up to VND 50 million.</p>
<p>There is no effort baseline in the source material, so I will not manufacture one. Cost is per signature, per timestamp, per notarisation and per year of retention, so the expensive item is rarely the licence. It is the evidence you did not store, found in year three. The AI rules already require machine-readable marking for generated content — metadata or a digital signature: the same primitive an evidence chain needs.</p>
<h2>What I would do in 30, 60 and 90 days</h2>
<p><strong>First 30 days — map the file, not the flow.</strong></p>
<ul>
<li>List the documents in one credit product: framework agreement, offer, contract, collateral, disbursement instruction. For each, record the signature level and who signed.</li>
<li>Read the framework agreement for the clause permitting electronic execution. If absent, nothing downstream is protected.</li>
<li>Check whether the certificate validates at signing time and whether revocation is stored. Most implementations validate at issuance and never again.</li>
</ul>
<p><strong>Days 31–60 — fix authority and retention.</strong></p>
<ul>
<li>Set value bands with Legal and Risk, with the reason for each band written down.</li>
<li>Confirm every bank-side certificate carries title and organisation, and that delegation is current. A signature outside an officer's authority is worse than none.</li>
<li>Define retention: original data, full context, re-verification path, and an export format that survives the vendor.</li>
</ul>
<p><strong>Days 61–90 — run one product, then attack it.</strong></p>
<ul>
<li>Take one product end to end, including the collateral path where you actually lend.</li>
<li>Then the dry run that matters: hand the file to someone who did not build the process, and ask them to prove intent and integrity without the system.</li>
<li>Fix what they could not prove. That list is the roadmap.</li>
</ul>
<h2>What I would not do</h2>
<ul>
<li><strong>I would not treat "digitally signed" as "enforceable".</strong> The signature proves intent; notarisation and registration prove the collateral.</li>
<li><strong>I would not let OTP stand as the signature on a large secured loan.</strong> It may evidence intent; it is not the equivalence threshold.</li>
<li><strong>I would not keep screenshots as evidence,</strong> or hand a court a PDF with no verifiable original behind it.</li>
<li><strong>I would not promise a fully digital file before checking notarisation availability</strong> where the collateral sits.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>A signed PDF is not proof. What proves a loan is a chain: a verified identity, a signature at the right level, a trusted timestamp, the original record kept in full, and a notarised or registered collateral document connected to it. The cheap failure is the signing screen. The expensive one is finding, in year three, that the record cannot be shown to be unchanged.</p>
<h2>Sources</h2>
<ul>
<li>Law on Electronic Transactions 20/2023/QH15 — the legal basis for electronic signatures and electronic contracts in Vietnam.</li>
<li>Decree 23/2025/NĐ-CP and Decree 104/2025/NĐ-CP — guidance recorded in the working notes for this article.</li>
<li>Circular 29/2026/TT-NHNN — banking-sector requirements recorded in the same notes.</li>
</ul>
<p><em>Instrument numbers follow the working notes; deep links were not part of the source material and are therefore not asserted here.</em></p>
<hr />
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[Defending a legacy ECM renewal: five questions before you sign.]]></title><description><![CDATA[The renewal price is not the biggest cost. Five questions on usage, egress, connectors, migration and exit, plus a five-year TCO in one unit.]]></description><link>https://bankingstacknotes.hashnode.dev/defending-an-ecm-renewal</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/defending-an-ecm-renewal</guid><category><![CDATA[ECM ]]></category><category><![CDATA[tco]]></category><category><![CDATA[procurement ]]></category><category><![CDATA[banking]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:24 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/f2c3f192-0c80-4c21-bc70-080ea4a5f4a7.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>A renewal quote arrived with a small uplift, a three-year term and a discount that expires at quarter end. Fourteen days to sign. The committee rehearsed whether the uplift was fair. The question that mattered was what the same platform costs in 2029 if the licence metric changes, the connectors move to a separate SKU, and nobody has ever run the export. The quote looked small. That was the problem. Everything below is dated November 2026.</p>
<h2>What changed</h2>
<p><strong>30 September 2026.</strong> FileNet 5.5.x, GA 26 June 2018, left End of Usage Support; Extended Support ends 30 September 2030. Renewing 5.5.x renews an unsupported system.</p>
<p><strong>1 July 2025.</strong> The Law on Archives 2024 took effect; Decree 113/2025, dated 21 July 2025, allows 24 months to build or upgrade a digital archive system, 36 for the state archive database, 60 for redundant infrastructure.</p>
<p><strong>1 July 2025.</strong> Circular 05/2025/TT-BNV fixed the archive data structure and the eight-step digitisation process. No platform ships it; replacing means building it twice.</p>
<p><strong>1 January 2026.</strong> Law 91/2025 with Decree 356/2025 — data-subject requests in two working days, cross-border transfer sanctioned up to 5% of prior-year revenue. "Delete one person's records" is now a contract clause.</p>
<p><strong>1 July 2026.</strong> Decree 337/2025 set the bar for electronic signatures in employment contracts: identity at signing, trusted timestamps, tamper-evident audit log. Expect the same bar at an archive inspection.</p>
<p><strong>December 2025.</strong> OpenText's Content Manager 25.4 became the LTS release, supported to 30 November 2028; non-LTS windows lasted about three months — 25.1 ended 28 February 2026, 25.3 on 31 August 2026.</p>
<p>The metric moved too: consumption on top of seats — Laserfiche AI Units, M-Files Aino capacity, SharePoint Premium near USD 0.05 per page.</p>
<h2>The decision framework</h2>
<p>Five questions, in the order I would put them in writing. The weights are mine, tuned for a lender mid-core-replacement.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Question</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td><strong>Usage</strong> — what are we buying?</td>
<td>20%</td>
<td>Named user defined in the contract, the metering metric, consumption against entitlement, true-down rights</td>
</tr>
<tr>
<td>2</td>
<td><strong>Egress</strong> — what does data out cost?</td>
<td>20%</td>
<td>Export, API calls and page events priced in the contract, not a rate card the vendor can revise</td>
</tr>
<tr>
<td>3</td>
<td><strong>Connectors</strong> — what is in the licence?</td>
<td>15%</td>
<td>Every live interface named with a support status; nothing that exists only in a services SOW</td>
</tr>
<tr>
<td>4</td>
<td><strong>Migration</strong> — what must be rebuilt?</td>
<td>25%</td>
<td>The vendor names in writing what will not migrate; you count the exceptions yourself</td>
</tr>
<tr>
<td>5</td>
<td><strong>Exit</strong> — what can we prove?</td>
<td>20%</td>
<td>Content, metadata and audit log export in an open format, tested and minuted annually</td>
</tr>
</tbody></table>
<p>Two are pass/fail: a written named-user definition, and an export that has been executed and minuted. Price mechanics — escalation cap, price hold, re-compete right — wrap whatever the five answers produce.</p>
<h2>Options and trade-offs</h2>
<p>Costs are multiples of <strong>L</strong>, current annual licence and support; five-year model estimates, not quotes.</p>
<table>
<thead>
<tr>
<th>Option</th>
<th>Strength</th>
<th>Risk</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td><strong>Renew 3–5 years with caps</strong> and a dated upgrade</td>
<td>Cheapest five-year cash, near-zero migration, upgrade covered by Essentials at the same price point</td>
<td>Locks an unpublished pricing basis; if escalation is under-estimated you learn in year four</td>
<td>0.73L, then 1.10L and 1.05L — about 2.88L</td>
</tr>
<tr>
<td><strong>Bridge 12 months</strong>, then re-compete</td>
<td>Keeps optionality, forces comparison, no five-year bet on an unpublished metric</td>
<td>Uplift risk moves into next year; two procurements; a year comparing, not upgrading</td>
<td>1.0L, then the 2.88L or 6.35L curve</td>
</tr>
<tr>
<td><strong>Replace now</strong></td>
<td>Removes the incumbent's pricing leverage</td>
<td>3.50L in year one, dual-run, ten classes of evidence to reproduce, inside the core window</td>
<td>3.50L, 1.50L, 1.35L — about 6.35L; 9.05L over five years</td>
</tr>
</tbody></table>
<p>A bridge is honest only if the re-compete is funded. Otherwise it is a three-year renewal delayed by one.</p>
<h2>The numbers</h2>
<p>Normalised to <strong>L</strong>, current annual licence and support, so conclusions hold when L is real; probabilities and damages are subjective estimates.</p>
<table>
<thead>
<tr>
<th>Scenario</th>
<th>Five-year cash</th>
<th>Risk event</th>
<th>Damage</th>
<th>Expected cost</th>
</tr>
</thead>
<tbody><tr>
<td>Keep and upgrade</td>
<td>4.98 L</td>
<td>5%</td>
<td>1.5 L</td>
<td>5.06 L</td>
</tr>
<tr>
<td>Replace with a peer ECM</td>
<td>9.05 L</td>
<td>25%</td>
<td>1.5 L</td>
<td>9.43 L</td>
</tr>
<tr>
<td>Microsoft 365</td>
<td>4.05 L</td>
<td>20%</td>
<td>2.0 L</td>
<td>4.45 L</td>
</tr>
<tr>
<td>Do not upgrade</td>
<td>5.20 L</td>
<td>40%</td>
<td>2.5 L</td>
<td>6.20 L</td>
</tr>
</tbody></table>
<p>Replacement is about <strong>1.82 times</strong> the upgrade path before risk, <strong>1.86 times</strong> after. Microsoft is 0.81 times in cash with the largest damage estimate, its exposure being regulatory. The expected-cost gap between keeping and Microsoft is roughly 12% — that is where the argument lives.</p>
<p>What the quote will not show:</p>
<ul>
<li><strong>Implementation and services</strong> at <strong>50–150%</strong> of the year-one licence (CostBench, OpenText example; one cited case near USD 2.4 million over three years).</li>
<li><strong>Dual-run</strong> of <strong>6–12 months</strong> with both systems licensed, then <strong>12–18 months of hypercare</strong>.</li>
<li><strong>Ten classes of evidence</strong> to rebuild: checksums, classification, metadata, versions, retention, legal hold, entitlements, audit trail, signatures, deep links.</li>
<li><strong>Exception handling</strong>, priced per record. Double-digit rates in a large financial record store mean hundreds of thousands to millions of manual records — an estimate, and the exclusion in migration contracts.</li>
<li><strong>Public anchors:</strong> M-Files USD 65 per seat per month, Laserfiche USD 73 and 93 by tier, SharePoint Premium about USD 0.05, 0.10 and 0.15. IBM, OpenText and Hyland publish nothing.</li>
</ul>
<h3>The one page the board needs</h3>
<table>
<thead>
<tr>
<th>Line</th>
<th>The number</th>
</tr>
</thead>
<tbody><tr>
<td>Destination release and date</td>
<td>5.7 in six months, Cortex in 12–18</td>
</tr>
<tr>
<td>Users, documents, interfaces, cost each</td>
<td>From the baseline, not the vendor</td>
</tr>
<tr>
<td>Last successful export test</td>
<td>Dated and minuted</td>
</tr>
<tr>
<td>Alternative, same scope, five years</td>
<td>9.05L against 4.98L</td>
</tr>
<tr>
<td>Re-decision trigger</td>
<td>No quote in 60 days, uplift above 25% a year</td>
</tr>
</tbody></table>
<h2>What I would do in 30 / 60 / 90 days</h2>
<p><strong>Days 1–30.</strong> Take ECM out of the replacement programme's scope in writing. Baseline: users, documents, live interfaces, annual cost, three years of incidents. Inventory the ten evidence classes and rate their difficulty.</p>
<p><strong>Days 31–60.</strong> Send one quote template to the incumbent and two alternatives: five years, six cost lines, same scope. Run three demos on your data — extract one customer's full record in two working days, place and release a legal hold, export three years of audit trail. Minute pass or fail.</p>
<p><strong>Days 61–90.</strong> Rebuild the TCO with real inputs and run the sensitivity. Take it to architecture, compliance, legal and internal audit before the board. Negotiate three to five years with an escalation cap, a price hold, an annually tested export clause. Sign, with upgrade dates in the minutes.</p>
<h2>What I would not do</h2>
<ul>
<li><strong>I would not sign before an export test has been run and written down.</strong> An untested exit clause is a sentence, not a right.</li>
<li><strong>I would not accept "per user" without a definition.</strong> Named or concurrent, human or robot — the definition is the price.</li>
<li><strong>I would not let a discount decide the term.</strong> An uplift argument is noise beside a five-year bet on an unpublished metric.</li>
<li><strong>I would not replace the repository in the core-banking window.</strong> Two moving variables in one release destroys attribution.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>The renewal quote is the smallest number in this decision. The larger costs are finding and moving records the bank must keep, running two systems during the switch, and proving to an auditor that nothing was lost. Five questions — usage, egress, connectors, migration, exit — decide whether the three-year price is a price or an opening position.</p>
<h2>Sources</h2>
<ul>
<li><a href="https://www.ibm.com/support/pages/ibm-filenet-content-manager55x">IBM, FileNet Content Manager 5.5.x lifecycle</a></li>
<li><a href="https://www.ibm.com/docs/en/content-cortex/26.0.0?topic=notes-whats-new">IBM, What's new in Content Cortex 26.0.0</a></li>
<li><a href="https://isr.de/en/news/am-ibm-filenet-ibm-content-cortex/">ISR, IBM Content Cortex for existing FileNet customers</a></li>
<li><a href="https://costbench.com/software/document-management/opentext-content-suite/calculator/">CostBench, OpenText Content Suite cost calculator</a></li>
<li><a href="https://www.m-files.com/pricing/">M-Files, pricing</a></li>
<li><a href="https://www.laserfiche.com/products/pricing/">Laserfiche, pricing</a></li>
<li><a href="https://vanban.chinhphu.vn/?pageid=27160&amp;docid=213822">Decree 113/2025/ND-CP</a></li>
<li><a href="https://vbpl.vn/van-ban/chi-tiet/--177682">Circular 05/2025/TT-BNV</a></li>
<li><a href="https://vanban.chinhphu.vn/?docid=216387&amp;pageid=27160">Law 91/2025/QH15 with Decree 356/2025</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[Corporate banking platforms: build, buy or extend.]]></title><description><![CDATA[Suite, extension layer or build: choosing a corporate banking platform in 2026, with the vendor facts and the three-year cost shape.]]></description><link>https://bankingstacknotes.hashnode.dev/corporate-banking-build-buy-extend</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/corporate-banking-build-buy-extend</guid><category><![CDATA[banking]]></category><category><![CDATA[architecture]]></category><category><![CDATA[Core Banking]]></category><category><![CDATA[integration]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:21 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/5e7883d8-94ae-420d-8e2a-0c908d08474f.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>A steering committee approves "replace the corporate banking platform" and asks for a shortlist. Nobody in the room can say which of the seven layers is broken. That is the decision that goes wrong most often: a platform bought to fix a channel problem, or a channel built to hide a trade back-office problem. The question is not which vendor, but which layer you are missing, and whether it should be bought, rented or built.</p>
<h2>What changed</h2>
<p><strong>2 February 2026 — the counterparty changed legal entity.</strong> Apax completed its purchase of Finastra's Treasury &amp; Capital Markets business; Teciem launched with 1,300 specialists and the Kondor, Summit, Opics, Sophis and Arc products. For a bank running Kondor, vendor, roadmap and support now sit in a different company. Procurement should list Finastra and Teciem separately.</p>
<p><strong>2017 — the widest portfolio is a portfolio, not a suite.</strong> Finastra, formed from the Misys and D+H merger, sells four separately deployable components — Trade Innovation, Corporate Channels, Loan IQ, Global PAYplus — linked by APIs and a common data model. Buying "Finastra" is four purchases.</p>
<p><strong>September 2026 — the largest local footprint is a core footprint.</strong> Temenos states that around 30 domestic banks run its software, against more than 3,000 globally, and the named references are core-banking references. Running the core is not running corporate banking on it.</p>
<p><strong>30 August 2022 — one announced trade reference, four years later.</strong> The only announced Surecomp implementation in Vietnam went live on the bank's own infrastructure in Hanoi with a local integrator. A pure-play with 40 years and customers in 80-plus countries still had a single domestic reference after four years.</p>
<p><strong>30 April 2019 — the global standard, not the domestic one.</strong> CGI Trade360 was chosen by a global banking group as its worldwide trade platform; no domestic bank has been announced since. That absence reflects commercial model, not quality: it is sold as a multi-year group-wide transformation.</p>
<p><strong>31 July 2026 — the channel moved in-house.</strong> A large joint-stock bank launched a corporate digital channel with an integrator under a technology-transfer model in which the bank owns the IP, phased accounts and payments first, then credit, trade finance, collections, FX and investment.</p>
<p>A mature bank runs four to six systems across seven layers, and no platform here is best at all seven. Forcing one product to cover everything raises customisation cost and hands the roadmap to the vendor.</p>
<h2>The decision framework</h2>
<p>Score seven criteria before any demonstration. The weights assume an existing core the bank is not replacing this year.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Depth in the layer you are missing</td>
<td>20%</td>
<td>Import L/C with amendment, conditional guarantee, collection and one SCF deal end to end, unmodified</td>
</tr>
<tr>
<td>2</td>
<td>Integration surface to core and payments</td>
<td>20%</td>
<td>Tested against the real core and payment rails, not mock-ups</td>
</tr>
<tr>
<td>3</td>
<td>Local delivery capacity</td>
<td>15%</td>
<td>Two reference customers of similar size and region, plus a named partner with trade operations skills</td>
</tr>
<tr>
<td>4</td>
<td>Ownership and roadmap risk</td>
<td>10%</td>
<td>No ownership change in 24 months; ISO 20022, CBPR+ and MT798 roadmap in writing</td>
</tr>
<tr>
<td>5</td>
<td>Operating model and control</td>
<td>10%</td>
<td>Multi-level approval, branch authorisation and limits work as configured, not patched</td>
</tr>
<tr>
<td>6</td>
<td>Three-year cost including change requests</td>
<td>15%</td>
<td>Cost tracks the layers you use; change request rate bounded in the contract</td>
</tr>
<tr>
<td>7</td>
<td>Exit and data</td>
<td>10%</td>
<td>Data transfer, customisation ownership and termination terms agreed before signature</td>
</tr>
</tbody></table>
<p>Criterion 3 is the one committees discount and regret: a vendor with one local implementation has one delivery model, and it may not be yours.</p>
<h2>Options and trade-offs</h2>
<table>
<thead>
<tr>
<th>Option</th>
<th>Strength</th>
<th>Risk</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td><strong>Buy a suite</strong> — one vendor across several layers</td>
<td>Single contract, one data model</td>
<td>Widest coverage lands on layers you may already own; customisation grows where depth is thin</td>
<td>Highest licence, lowest integration; rises with every module</td>
</tr>
<tr>
<td><strong>Extend</strong> — keep the core, add a channel and trade workflow layer</td>
<td>Preserves the core; pairs a core platform with a documentary trade pure-play</td>
<td>Two vendors, two roadmaps, and the integration becomes yours</td>
<td>Moderate licence, higher integration and testing; most predictable if scope is fixed</td>
</tr>
<tr>
<td><strong>Build</strong> — technology transfer, bank owns the IP, phased by product</td>
<td>Roadmap control where the channel is the product</td>
<td>Needs internal product and engineering capacity plus a conversion budget</td>
<td>Low licence, highest people cost in years one and two, lowest later</td>
</tr>
</tbody></table>
<p>The pairing that recurs here is a core platform plus a documentary trade specialist. Not elegant, but it matches where each product is deep.</p>
<h2>The numbers</h2>
<p>The research carries no price for any of the four, and its one comparative statement is uncomfortable: cost is not proportional to coverage. The most expensive model covers only part of the portfolio, because you buy a multi-country operating model, not feature count.</p>
<p>What can be quantified is scope:</p>
<ul>
<li><strong>Seven layers</strong>, <strong>four to six systems</strong> in a mature bank. Below four, ask which layer is improvised.</li>
<li><strong>Around 30 domestic banks</strong> on the dominant core platform, against <strong>3,000-plus</strong> globally. Installed base buys skills, not depth.</li>
<li><strong>One announced domestic trade reference</strong> after four years for the pure-play; <strong>one global group reference</strong> for the SaaS platform, with no domestic bank announced.</li>
<li><strong>Three to five real transactions</strong> across four product types is the proof-of-concept scope that separates products; <strong>two reference interviews</strong> is the cheapest due diligence available.</li>
</ul>
<p>A three-year range built on licence alone misleads. My estimate: integration, data mapping and retraining outweigh licence in every option except the suite, where licence and customisation compete.</p>
<h2>What I would do in 30 / 60 / 90 days</h2>
<p><strong>First 30 days — name the missing layer.</strong> Map every corporate banking capability to the seven layers and mark the system serving it today. Interview the trade operations head, not only the CIO. Collect the public lifecycle and ownership facts on every incumbent, including the February 2026 spin-off. Write down the decision you are actually making.</p>
<p><strong>Next 30 days — test transactions, not features.</strong> Run three to five real deals: an import L/C with an amendment, a conditional guarantee, a documentary collection and one supply chain finance transaction. Test multi-level approval, branch authorisation and limits. Test integration against the live core. Require the business demo in Vietnamese, for the business team.</p>
<p><strong>Final 30 days — structure the deal.</strong> Separate the Finastra and Teciem lots. Put data transfer, customisation IP and termination on the table before pricing. Phase the channel with accounts and payments first. Take two reference calls and write the architecture decision record, including what you deliberately did not replace.</p>
<h2>What I would not do</h2>
<ul>
<li><strong>I would not buy a suite to fix a layer I already own.</strong> That is how a bank pays twice for the same capability and customises the second one.</li>
<li><strong>I would not put Finastra and Teciem into one tender.</strong> Different companies, different roadmaps, different support commitments since February 2026.</li>
<li><strong>I would not accept a demonstration as integration evidence.</strong> A mock-up proves the vendor's demo environment works.</li>
<li><strong>I would not sign a multi-year programme without data transfer, customisation ownership and termination terms.</strong> Those three clauses set the price of leaving.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>Corporate banking is seven capabilities, not one product. Before approving a replacement, ask which capability is failing, which systems serve it today, and whether the fix is a purchase, an extension of what you own, or a build. In my experience the cheapest answer is the narrowest one, and the licence is rarely the largest line.</p>
<h2>Sources</h2>
<ul>
<li><a href="https://www.finastra.com/solutions/corporate-banking">Finastra — lending and corporate banking</a></li>
<li><a href="https://www.finastra.com/lending/trade-supply-chain-finance">Finastra — trade and supply chain finance</a></li>
<li><a href="https://www.finastra.com/lending/solutions/trade-innovation">Finastra — Trade Innovation</a></li>
<li><a href="https://teciem.com/press/teciem-launch-announcement/">Teciem — launch announcement (2 February 2026)</a></li>
<li><a href="https://www.temenos.com/segments/corporate-and-commercial/">Temenos — corporate and commercial banking</a></li>
<li><a href="https://www.temenos.com/products/core-banking/core-for-corporate/">Temenos — core for corporate and commercial</a></li>
<li><a href="https://surecomp.com/solutions/for-financial-institutions/back-office-trade-finance-processing/">Surecomp — back-office trade finance</a></li>
<li><a href="https://www.cgi.com/uk/en-gb/CGI-Trade360">CGI — CGI Trade360</a></li>
<li><a href="https://vtv.vn/core-banking-cloud-ai-temenos-nhin-thay-gi-o-ngan-hang-viet-nam-100260909151119717.htm">VTV — core banking in Vietnam, around 30 banks (9 September 2026)</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[What a collections data programme actually costs.]]></title><description><![CDATA[What a collections data programme costs in 2026: contactability, contact-hours rules, audit evidence and the cost drivers behind AI collections.]]></description><link>https://bankingstacknotes.hashnode.dev/collections-data-programme</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/collections-data-programme</guid><category><![CDATA[Collections]]></category><category><![CDATA[data]]></category><category><![CDATA[compliance ]]></category><category><![CDATA[banking]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:19 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/fa897af4-2750-480d-a2ad-e03700754069.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>A consumer-finance lender is choosing between a collections data programme and a voice bot. Both arrive as one budget line and promise the same thing. The question: for any call made last month, can you show the lawful basis, the hour, the person contacted, the number dialled, and evidence the call was conducted properly?</p>
<h2>What the evidence says</h2>
<ul>
<li><strong>The conduct rules are now explicit.</strong> A debt-collection code of conduct for member banks, issued 31 March 2026 and effective immediately, runs to three chapters and eleven articles. Drafted with the IFC across eleven rounds with 48 organisations and more than 370 comments, it covers every type of debt, and responsibility survives outsourcing.</li>
<li><strong>The bucket structure is fixed, so your data must reproduce it.</strong> Circular 11/2021 sets specific provisions at 0, 5, 20, 50 and 100% for groups 1 to 5, plus 0.75% general. Treatment logic that cannot reconcile to those buckets is untestable.</li>
<li><strong>The history window is bounded.</strong> Circular 03/2013 limits the bad-debt history supplied by the credit bureau to five years, so a contact strategy assuming longer memory argues with the law.</li>
<li><strong>Scale explains the pressure.</strong> Non-performing debt across the system exceeded VND 227,000 billion, about USD 9 billion, in March 2025; consumer-finance companies averaged NPL above 10% in the first half of 2025.</li>
<li><strong>Contact hours are where regulators fine.</strong> India restricts recovery calls to 8am–7pm, with penalties of ₹1 crore in September 2024, ₹3.80 lakh on 10 February 2026 and ₹11.5 lakh on 27 February 2026. Indonesia allows 08:00–20:00, Monday to Saturday; the Philippines treats contact before 6am or after 10pm as unreasonable.</li>
<li><strong>Vietnam's enforcement levers have teeth.</strong> Decree 21/2021 requires at least 10 days' notice before disposing of movable collateral; Law 96/2025, effective 15 October 2025, codified seizure without a court in Article 198a where the agreement exists and the security interest has third-party effect.</li>
</ul>
<h2>The decision framework</h2>
<p>Score it on what an auditor would ask, not what a demo shows.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Lawful basis and consent per contact</td>
<td>20%</td>
<td>Every contact traces to a basis, a product and a date; withdrawal suppresses within one cycle</td>
</tr>
<tr>
<td>2</td>
<td>Data quality and contactability</td>
<td>20%</td>
<td>Right-party rates measured per portfolio and channel, wrong numbers retired, no duplicate identities</td>
</tr>
<tr>
<td>3</td>
<td>Contact strategy governance</td>
<td>15%</td>
<td>Hours, frequency and channel limits enforced by the system, not by agent memory</td>
</tr>
<tr>
<td>4</td>
<td>Audit trail and evidence</td>
<td>15%</td>
<td>Recording, disposition and quality check retrievable per account for the retention period</td>
</tr>
<tr>
<td>5</td>
<td>Treatment measurement</td>
<td>15%</td>
<td>Champion-challenger tests with same-age, same-bucket comparison; uplift attributed, not assumed</td>
</tr>
<tr>
<td>6</td>
<td>Third-party oversight</td>
<td>10%</td>
<td>Agency clauses with audit rights and penalties; outsourced conduct treated as own conduct</td>
</tr>
<tr>
<td>7</td>
<td>Integration with servicing facts</td>
<td>5%</td>
<td>Restructure, settlement and collateral status visible before a contact decision is made</td>
</tr>
</tbody></table>
<p>Criterion 5 is where programmes fail quietly: recovery rises when you call more; without controlled comparison the lift is volume, not intelligence.</p>
<h2>Options and trade-offs</h2>
<table>
<thead>
<tr>
<th>Option</th>
<th>Strength</th>
<th>Risks</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td><strong>Clean the data and govern contact</strong> inside the current call centre</td>
<td>Fastest compliance win; no new platform; fixes what an auditor asks for</td>
<td>Depends on core systems exposing account and collateral facts</td>
<td>Mostly people and data remediation; lowest build</td>
</tr>
<tr>
<td><strong>Add voice bots and AI quality checks</strong> on existing data</td>
<td>Cuts cost per contact; reviews far more calls than sampling</td>
<td>Automates bad data; conduct risk scales with volume; vendor cases are not your baseline</td>
<td>Front-loaded licence and integration, then per-minute cost that grows with volume</td>
</tr>
<tr>
<td><strong>Rebuild on a unified collections data model</strong></td>
<td>One customer, one exposure, one contact history; treatment becomes measurable</td>
<td>Longest path; needs business ownership, not only IT</td>
<td>Highest build over two years, lowest marginal cost after; the only option that makes criterion 5 real</td>
</tr>
</tbody></table>
<p>The second is not a substitute for the first: a bot dialling a wrong number faster is a bigger problem.</p>
<h2>The numbers</h2>
<p>The sources carry no person-day baseline, so I give none. What is documented is unit economics:</p>
<ul>
<li><strong>Cost per contact:</strong> vendor cases report cost per dollar recovered falling from 22¢ to 8¢ and cost per call down 75%. In one market an agent costs about ₹30,000 a month for 250 cases, with AI cited at 40–60% cheaper and 20 times the volume.</li>
<li><strong>Cost of poor treatment strategy:</strong> a card centre using champion-challenger reported a 30% cut in collections operating cost; a European bank reported 25% fewer collector minutes with the same money collected.</li>
<li><strong>Cost of scale:</strong> in Vietnam one lender moved from 200,000 to 5 million calls a month at halved operating cost; another ran 1,054,670 calls in eleven months, peaking at 52,152 a day.</li>
<li><strong>Cost drivers:</strong> data remediation and identity matching, consent reconstruction, contact-policy engine, recording and quality-check storage, integration to servicing, and the audit pack. No prices in the sources.</li>
</ul>
<h2>What I would do in 30 / 60 / 90 days</h2>
<h3>Next 30 days — measure the data, not the ambition</h3>
<ul>
<li>Baseline contactability: right-party rate, wrong-number rate, duplicates, accounts with no valid contact.</li>
<li>Inventory lawful basis and consent per product; find contacts traced to neither.</li>
<li>Audit last quarter's contacts against hours, frequency and channel limits; publish the exceptions internally.</li>
</ul>
<h3>Days 30 to 60 — make the rules executable</h3>
<ul>
<li>Put hours, frequency and channel limits into the system as controls with an exception log, not a policy PDF.</li>
<li>Run one controlled champion-challenger test inside a single bucket, same vintage, with a fixed contact ceiling.</li>
<li>Price every change request at intake, including the ones arriving as "just a new SMS template".</li>
</ul>
<h3>Days 60 to 90 — produce evidence, then scale</h3>
<ul>
<li>Assemble the audit pack: recording coverage, quality-check sampling, disposition codes, retention, and agency clauses with audit rights.</li>
<li>Decide the next increment on the first test's evidence, not a vendor deck.</li>
<li>Retire what you cannot justify: contacts without a basis, numbers without an owner, history beyond five years.</li>
</ul>
<h2>What I would not do</h2>
<ul>
<li><strong>Not buy a voice bot before contact data is trustworthy.</strong> That scales the error at a lower cost per error.</li>
<li><strong>Not treat outsourced collection as transferred responsibility.</strong> The expectation travels with the agent.</li>
<li><strong>Not use a vendor's recovery uplift as the business case.</strong> Without same-age comparison it measures contact volume.</li>
<li><strong>Not widen contact hours or frequency to hit a monthly target.</strong> It converts a data programme into an enforcement case.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>Collections spends money on contacts and reputation on how they are made. The programme is about knowing who to contact, when you may contact them, and proving it afterwards. Fund data and controls first, automation second: the bot is cheaper once the numbers are right, and expensive once they are not. Before any demo, ask for contactability, consent coverage and audit evidence.</p>
<h2>Sources</h2>
<ul>
<li>Vietnam debt-collection code of conduct, 31 March 2026, <a href="https://lsvn.vn/ban-hanh-bo-quy-tac-ung-xu-trong-hoat-dong-thu-hoi-no-trong-he-thong-ngan-hang-a172157.html">https://lsvn.vn/ban-hanh-bo-quy-tac-ung-xu-trong-hoat-dong-thu-hoi-no-trong-he-thong-ngan-hang-a172157.html</a></li>
<li>Circular 11/2021/TT-NHNN, provisioning at 0/5/20/50/100% plus 0.75%, <a href="https://english.luatvietnam.vn/circular-no-11-2021-tt-nhnn-dated-july-30-2021-of-the-state-bank-of-vietnam-providing-the-classification-of-assets-risk-provisioning-levels-and-met-206806-doc1.html">https://english.luatvietnam.vn/circular-no-11-2021-tt-nhnn-dated-july-30-2021-of-the-state-bank-of-vietnam-providing-the-classification-of-assets-risk-provisioning-levels-and-met-206806-doc1.html</a></li>
<li>Credit-bureau history capped at five years, <a href="https://chinhsachonline.chinhphu.vn/lich-su-no-xau-duoc-cung-cap-trong-toi-da-5-nam-75461.htm">https://chinhsachonline.chinhphu.vn/lich-su-no-xau-duoc-cung-cap-trong-toi-da-5-nam-75461.htm</a></li>
<li>Law 96/2025/QH15, Article 198a, seizure without a court, <a href="https://pbgdpl.camau.gov.vn/bai-viet/Ch%c3%adnh%20s%c3%a1ch%20m%e1%bb%9bi%20c%c3%b3%20hi%e1%bb%87u%20l%e1%bb%b1c%20t%e1%bb%ab%20th%c3%a1ng%206/quy-dinh-ve-quyen-thu-giu-tai-san-bao-dam.7142">https://pbgdpl.camau.gov.vn/bai-viet/Ch%c3%adnh%20s%c3%a1ch%20m%e1%bb%9bi%20c%c3%b3%20hi%e1%bb%87u%20l%e1%bb%b1c%20t%e1%bb%ab%20th%c3%a1ng%206/quy-dinh-ve-quyen-thu-giu-tai-san-bao-dam.7142</a></li>
<li>Decree 21/2021/ND-CP, notice periods for collateral, <a href="https://accgroup.vn/nghi-dinh-so-212021-nd-cp">https://accgroup.vn/nghi-dinh-so-212021-nd-cp</a></li>
<li>Reserve Bank of India, recovery-agent directions, <a href="https://www.rbi.org.in/Scripts/NotificationUser.aspx?Id=12378&amp;Mode=0">https://www.rbi.org.in/Scripts/NotificationUser.aspx?Id=12378&amp;Mode=0</a></li>
<li>Penalty for out-of-hours contact, September 2024, <a href="https://www.business-standard.com/amp/finance/personal-finance/rs-1-cr-penalty-on-hdfc-bank-how-to-protect-yourself-from-recovery-agents-124091700205_1.html">https://www.business-standard.com/amp/finance/personal-finance/rs-1-cr-penalty-on-hdfc-bank-how-to-protect-yourself-from-recovery-agents-124091700205_1.html</a></li>
<li>OJK on permitted collection hours, <a href="https://finansial.bisnis.com/read/20260507/89/1971988/ojk-tegaskan-aturan-penagihan-tidak-boleh-kasar-dan-harus-di-waktu-tertentu">https://finansial.bisnis.com/read/20260507/89/1971988/ojk-tegaskan-aturan-penagihan-tidak-boleh-kasar-dan-harus-di-waktu-tertentu</a></li>
<li>Champion-challenger in early collections, <a href="https://www.fico.com/en/newsroom/shanghai-pudong-development-bank-credit-card-centre-automates-early-collections">https://www.fico.com/en/newsroom/shanghai-pudong-development-bank-credit-card-centre-automates-early-collections</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[Start here: AI governance without blocking innovation]]></title><description><![CDATA[Six questions about AI governance in Vietnamese banking, answered in order: scope, accountability, registration, oversight, evidence and delivery.]]></description><link>https://bankingstacknotes.hashnode.dev/ai-governance-start-here</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/ai-governance-start-here</guid><category><![CDATA[AI]]></category><category><![CDATA[Governance]]></category><category><![CDATA[banking]]></category><category><![CDATA[compliance ]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:20:16 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/ed24b17c-8ee5-4bae-a5a2-ac3a4f7b209a.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Every AI governance programme I have seen fail, failed the same way: it started with a policy and ended with a register nobody maintained. The ones that hold up start with a narrower question — what will we have to show, to whom, and by when — and build backwards from that.</p>
<p>Vietnam has made the answer specific. Law 134/2025/QH15 took effect on 1 March 2026, Decree 142/2026/NĐ-CP took effect on 1 May 2026, and Decision 33/2026/QĐ-TTg put the high-risk list into force on 15 August 2026. This series is six questions, answered one at a time, without a maturity model.</p>
<h2>The six questions</h2>
<h3>1. Which systems are actually in scope?</h3>
<p>Classification is three-tiered and applies to complete systems, not to a model on its own. The list in Decision 33/2026/QĐ-TTg covers six sectors, and banking is one — with two examples that a consumer lender is likely to be running already: automated execution of electronic transactions, and automated credit-granting decisions. Embedded AI in purchased tools and model APIs are inside the boundary, not outside it.</p>
<h3>2. Who is accountable when a decision is automated?</h3>
<p>The law defines four roles — developer, provider, deployer, user — and one organisation can hold several. <em>Provider</em> includes cloud services, APIs and embedded components, so a vendor model is provided and the bank running it is deploying. Article 29 places compensation with the deployer where a high-risk system was lawfully used and damage still occurred, with recourse only where a contract provides it. An agent does not absorb liability; the institution that deployed it does.</p>
<h3>3. What must be registered, and by when?</h3>
<p>Medium and high-risk systems must be notified to the Ministry of Science and Technology through the one-stop portal before use, and the system receives an identifier automatically. If risk rises, notification is due within 15 working days. For systems running before 1 March 2026, obligations must be complete within 18 months in finance — around 1 September 2027 — and within 12 months elsewhere.</p>
<h3>4. What does human oversight look like when it has to survive an audit?</h3>
<p>Article 4 says AI does not replace human authority and responsibility, which is a stronger statement than requiring a reviewer. In practice, oversight means a logged override, a measured intervention rate, and a stop control that has been tested at least once. A committee that reviews monthly reports is governance theatre; an override nobody can retrieve afterwards is the same thing with better stationery.</p>
<h3>5. What evidence has to exist before an incident, not after?</h3>
<p>The clock starts when the incident does: a preliminary report within 72 hours for urgent incidents, 5 working days for other serious ones, and a formal report within 15 days, with data and system logs retained. Inspection under Article 28 can require the technical file, trace logs and training data. As of 1 October 2026 no dedicated AI penalty decree had been confirmed — but Decree 330/2026/NĐ-CP already fines personal-data processing in AI systems at VND 20–50 million, so the gap is not a holiday.</p>
<h3>6. How do you do this without stopping delivery?</h3>
<p>Put one AI gate in the project lifecycle: role, risk class, evidence owner. Classify at intake, reuse the same technical-file template, and let the sandbox handle genuinely novel solutions — Article 21 makes it voluntary, registered through the national public service portal, and its results are a basis for the authority to consider recognising conformity or reducing obligations. Completing a sandbox does not automatically earn either.</p>
<h2>What to read next</h2>
<p>The long-form piece in this series is <a href="/what-a-bank-must-prove-about-ai">What a bank must be able to prove about AI</a>: a seven-criterion control map, three delivery options with their three-year cost shape, and a 30/60/90 plan. If your next deadline is a signed credit contract rather than a model, start instead with <a href="/e-signature-is-an-evidence-problem">E-signature is an evidence problem</a>.</p>
<h2>Sources</h2>
<ul>
<li><a href="https://vanban.chinhphu.vn/?docid=216334&amp;orggroupid=1&amp;pageid=27160">Law 134/2025/QH15</a> — AI Law, in force 1 March 2026</li>
<li><a href="https://chinhphu.vn/?pageid=27160&amp;docid=218029">Decree 142/2026/NĐ-CP</a> — classification, notification, incidents</li>
<li><a href="https://chinhphu.vn/?pageid=27160&amp;docid=218658">Decision 33/2026/QĐ-TTg</a> — high-risk list, in force 15 August 2026</li>
<li><a href="https://xaydungchinhsach.chinhphu.vn/nghi-dinh-330-2026-nd-cp-ve-xu-phat-vi-pham-hanh-chinh-trong-linh-vuc-an-ninh-mang-119260824172446407.htm">Decree 330/2026/NĐ-CP</a> — penalties</li>
<li><a href="https://luatvietnam.vn/tin-van-ban-moi/danh-muc-he-thong-tri-tue-nhan-tao-co-rui-ro-cao-tu-15-8-2026-186-110045-article.html">LuatVietnam</a> — high-risk list summary</li>
</ul>
<hr />
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[ESB isn't dead. It stopped being the default answer.]]></title><description><![CDATA[A 2026 decision framework for ESB vs API management: dated lifecycle facts, TCO traps, and what a bank should do in the next 90 days.]]></description><link>https://bankingstacknotes.hashnode.dev/esb-isnt-dead</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/esb-isnt-dead</guid><category><![CDATA[integration]]></category><category><![CDATA[architecture]]></category><category><![CDATA[api]]></category><category><![CDATA[banking]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 10:07:24 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/856e5c03-8b59-417d-8c9f-4bf69b54f2e7.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Every budget cycle, the same sentence appears in a slide: <em>"ESB is a dead architecture, we should move to APIs."</em> It is usually followed by a quote that is either very large or very vague, and by a project plan that starts with a technology choice instead of an inventory.</p>
<p>The sentence is half right. Centralised ESB as the <em>default</em> integration model is finished. The product lines are being retired, the vendors are merging, and the analyst narrative has moved on. But the integration <em>jobs</em> that a bus was bought to do — protocol mediation, long-running orchestration, adapters to systems that will outlive all of us — did not disappear. They moved, and in many banks they moved nowhere at all.</p>
<p>This note is the decision framework I use before signing anything: what actually changed by November 2026, how to classify interfaces, and what I would do in the first 90 days.</p>
<h2>What changed</h2>
<p><strong>The products are being retired, and the dates are now close.</strong></p>
<ul>
<li>Apache ServiceMix was retired in <strong>March 2026</strong>; Apache Synapse is in the Attic; Mule 3 Community Edition is end-of-life; Red Hat Fuse became a build of Apache Camel; Microsoft states that <strong>BizTalk Server 2020 is the final release</strong>.</li>
<li><strong>Oracle Fusion Middleware 12c</strong> — the platform under Oracle Service Bus and SOA Suite — ends Premier Support in <strong>December 2026</strong>, with Extended Support ending <strong>December 2027</strong>.</li>
<li><strong>SAP Process Orchestration 7.5</strong> ends mainstream maintenance in <strong>2027</strong>.</li>
<li><strong>BizTalk Server 2020</strong> end-of-sale is <strong>31 March 2027</strong>, with mainstream support ending <strong>12 April 2028</strong>.</li>
<li><strong>Red Hat 3scale</strong> reaches end of maintenance on <strong>30 June 2027</strong> — it should not be selected for anything new.</li>
</ul>
<p><strong>The market has consolidated, which changes roadmap risk.</strong></p>
<p>IBM completed its acquisition of webMethods and StreamSets in <strong>July 2024</strong>, and its acquisition of Confluent (~USD 11 billion) on <strong>17 March 2026</strong>. Salesforce completed Informatica in <strong>November 2025</strong>. Every one of those deals moves a product you may depend on under a different roadmap owner.</p>
<p><strong>The analyst picture moved with it.</strong></p>
<ul>
<li>Gartner's <strong>API Management Magic Quadrant (28 September 2026)</strong> listed eight Leaders — Boomi, Google, Gravitee, IBM, Kong, Salesforce/MuleSoft, Sensedia and WSO2 — with Axway dropping to Challenger and Traefik entering the quadrant.</li>
<li>Gartner's <strong>iPaaS 2025</strong> quadrant demoted both Salesforce/MuleSoft and Oracle from Leader to Challenger, while TIBCO and Software AG no longer qualified.</li>
</ul>
<p><strong>A new integration consumer appeared: the AI agent.</strong></p>
<p>Within roughly twelve months, Apigee, Kong, Tyk, Azure API Management, Gravitee, IBM API Connect and WSO2 all shipped MCP proxy or governance features. A gateway without MCP support is now hard to shortlist. That is a real change in the <em>edge</em> layer — and it is another reason not to conflate "API management" with "the thing that replaces the ESB".</p>
<h2>The distinction people keep missing</h2>
<table>
<thead>
<tr>
<th></th>
<th>Centralised ESB</th>
<th>API management</th>
</tr>
</thead>
<tbody><tr>
<td>Philosophy</td>
<td>Smart pipes, dumb endpoints</td>
<td>Smart endpoints, dumb pipes</td>
</tr>
<tr>
<td>Traffic</td>
<td>East–west between internal systems</td>
<td>North–south from consumers to services</td>
</tr>
<tr>
<td>Communication</td>
<td>Mostly asynchronous: JMS, MQ, SOAP/WS-*, file, EDI</td>
<td>Mostly synchronous: REST, GraphQL, gRPC</td>
</tr>
<tr>
<td>Core job</td>
<td>Route, transform, mediate protocols, orchestrate</td>
<td>Authenticate, throttle, version, measure, monetise</td>
</tr>
<tr>
<td>Whose logic</td>
<td>The bus owns business rules</td>
<td>The service owns business rules</td>
</tr>
<tr>
<td>Scaling</td>
<td>Stateful, hard to scale, easy to make a single point of failure</td>
<td>Stateless, horizontal</td>
</tr>
</tbody></table>
<p>The overlap is narrow — routing, auth, some rate limiting — and it is exactly the overlap that makes commercial suites confusing. MuleSoft sells both an integration runtime and API management. IBM sells App Connect Enterprise, API Connect, webMethods and MQ. WSO2 sells API Manager and an Integrator. The question is therefore not <em>"API management or ESB?"</em> but:</p>
<ol>
<li>Do I need to govern <strong>consumption</strong> of APIs (partners, developers, agents, quotas, monetisation)?</li>
<li>Do I need to <strong>mediate and orchestrate</strong> internal systems, including legacy adapters and long-running transactions?</li>
<li>If both, do I buy one suite or compose specialised layers?</li>
</ol>
<h2>The decision framework</h2>
<p>I score seven criteria. Weights reflect a regulated financial institution; adjust them, but write them down before the vendor demo.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Criterion</th>
<th>Weight</th>
<th>What "high" looks like</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Legacy adapter depth (mainframe, MQ, SAP IDoc/RFC, EDI/HL7)</td>
<td>20%</td>
<td>Real systems need CICS/IMS/MQ adapters, not HTTP proxies</td>
</tr>
<tr>
<td>2</td>
<td>Transaction semantics (compensation, correlation, human tasks)</td>
<td>20%</td>
<td>Processes span days and must be recoverable</td>
</tr>
<tr>
<td>3</td>
<td>Consumer model (internal team, partner, developer, AI agent)</td>
<td>15%</td>
<td>External self-service, portal, quotas, MCP</td>
</tr>
<tr>
<td>4</td>
<td>Deployment and data residency</td>
<td>15%</td>
<td>On-prem or BYOC is mandatory, not preferred</td>
</tr>
<tr>
<td>5</td>
<td>Operating model and skills</td>
<td>10%</td>
<td>One central middleware team versus federated product teams</td>
</tr>
<tr>
<td>6</td>
<td>Three-year TCO including overage and professional services</td>
<td>15%</td>
<td>Pricing metric grows linearly with volume</td>
</tr>
<tr>
<td>7</td>
<td>Exit and lifecycle risk</td>
<td>5%</td>
<td>Config exportable, no EOL in the next three years</td>
</tr>
</tbody></table>
<p>Four go/no-go checks come first: does the candidate have deep SOAP/WS-* support, the adapters you actually use, a private deployment option, and no end-of-life inside the planning horizon? If any answer is no, stop — the rest of the score does not matter.</p>
<h2>Three options, not two</h2>
<table>
<thead>
<tr>
<th>Option</th>
<th>What it means</th>
<th>Strength</th>
<th>Risk</th>
<th>Three-year cost shape</th>
</tr>
</thead>
<tbody><tr>
<td><strong>Keep and strangle</strong></td>
<td>Keep the bus for what only it can do; put an API gateway in front; move interfaces out one at a time</td>
<td>Lowest disruption, preserves adapters, buys time</td>
<td>You keep the licence and the skills problem</td>
<td>Licence + integration work; savings are gradual</td>
</tr>
<tr>
<td><strong>Replace with an API-led suite</strong></td>
<td>MuleSoft, IBM, WSO2 or Boomi as the new centre</td>
<td>Strong governance, one vendor, faster onboarding</td>
<td>New lock-in, migration cost often exceeds licence savings, and the "suite" is itself consolidating</td>
<td>High licence + 25–50% professional services in year one</td>
</tr>
<tr>
<td><strong>Decompose</strong></td>
<td>API gateway (north–south) + event streaming (east–west) + iPaaS for SaaS/B2B + in-app connectors (Camel/Spring)</td>
<td>No single point of failure, best cost curve at scale</td>
<td>Requires four capabilities and a real architecture team</td>
<td>Higher initial complexity, lower marginal cost</td>
</tr>
</tbody></table>
<p>There is no universal winner. A bank with IBM MQ, CICS and SOAP interfaces has a different answer from a greenfield lender with five SaaS products.</p>
<h2>The numbers that decide it</h2>
<p>Licence size, not unit price, is usually the problem. In a published buyer-side example, a MuleSoft contract of 100 vCores at Platinum ran to roughly <strong>USD 800,000 per year</strong>, with <strong>29–41 committed vCores unused</strong> — between <strong>USD 232,000 and 328,000 per year</strong> of waste, or roughly <strong>USD 0.75–1.1 million over three years</strong>. Treat that as an indicative range from a buyer-side source, not a quote; the pattern it describes — committed capacity above real usage — is the one to check in your own contract.</p>
<p>For IBM App Connect Enterprise, the publicly listed SaaS entry points are around <strong>USD 8,484 per 10,000 VPC-hours per year</strong> and <strong>USD 2,544 per 100,000 flow runs per year</strong>. On-premise pricing is not published and is metered by PVU/VPC. Professional services in year one are commonly <strong>20–40%</strong> of licence for Boomi and <strong>25–50%</strong> for MuleSoft, again per third-party estimates.</p>
<p>The market context explains why vendors are repricing: third-party estimates put API management at <strong>USD 5.24 billion in 2025 growing to 6.92 billion in 2026</strong> (~32% CAGR), and iPaaS at <strong>13.88 billion in 2025, 18.72 billion in 2026, and 54.78 billion by 2030</strong>. Growth attracts consolidation, consolidation raises roadmap risk, and roadmap risk is now a line item in the decision.</p>
<h2>What I would do in 30, 60 and 90 days</h2>
<p><strong>First 30 days — inventory, not architecture.</strong></p>
<ul>
<li>Count interfaces, not applications: protocol, direction, payload size, messages per day, criticality.</li>
<li>List every end-of-life date that falls inside three years, with the owner's name against it.</li>
<li>Classify each interface into four buckets: retire, wrap (gateway in front of the existing service), re-platform, keep.</li>
<li>Stop all new development on interfaces classified "retire".</li>
</ul>
<p><strong>Next 30 days — contract and proof.</strong></p>
<ul>
<li>Open a licence workstream with four clauses on the table: true-down rights, escalation cap, price hold, and an exit/export clause.</li>
<li>Run a proof of concept on exactly two things: one partner-facing API through a gateway, and one event flow through streaming. Measure p95 latency with auth and rate limiting enabled — not on an empty policy set.</li>
<li>Decide the skills plan before the technology plan. A gateway nobody can operate is a renewal you cannot avoid.</li>
</ul>
<p><strong>Next 30 days — target state and a first migration.</strong></p>
<ul>
<li>Publish a two-tier target: an enterprise API layer at the edge, an event layer for east–west traffic, and a thin mediation layer for legacy adapters.</li>
<li>Migrate the three interfaces with the highest EOL risk and lowest coupling first. Not the most visible ones.</li>
<li>Write the architecture decision record, including what you deliberately did <strong>not</strong> replace.</li>
</ul>
<h2>What I would not do</h2>
<ul>
<li><strong>I would not buy a replacement bus to answer the sentence "ESB is dead".</strong> You would be buying the same centralisation with a newer invoice.</li>
<li><strong>I would not put business rules in the API gateway.</strong> That is how the original coupling problem returns, one policy at a time.</li>
<li><strong>I would not migrate by technology preference.</strong> Migrate by end-of-life date and coupling, in that order.</li>
<li><strong>I would not assume the gateway solves API security.</strong> OWASP's current API Top 10 (still the 2023 edition) puts broken object and function-level authorisation at the top; a gateway does not implement those checks — your application does.</li>
<li><strong>I would not let a three-year roadmap commitment outlive the vendor's own product commitment.</strong> If the product line has changed owner twice, cap the term.</li>
</ul>
<h2>For the non-technical sponsor</h2>
<p>An ESB is not a product category any more; it is a set of integration jobs. Some of those jobs are cheaper and safer to keep for now, some are urgent because the vendor's support ends soon, and some should move to an API layer or an event stream. The decision is which jobs move, in what order, and what the exit clauses say — not which logo replaces which.</p>
<h2>Sources</h2>
<ul>
<li>Apache Software Foundation, ServiceMix retirement (Attic), <a href="https://attic.apache.org/projects/servicemix.html">https://attic.apache.org/projects/servicemix.html</a></li>
<li>Microsoft, BizTalk Server migration overview, <a href="https://learn.microsoft.com/en-us/azure/logic-apps/biztalk-server-migration-overview">https://learn.microsoft.com/en-us/azure/logic-apps/biztalk-server-migration-overview</a></li>
<li>Oracle, Fusion Middleware lifetime support, <a href="https://docs.oracle.com/en/middleware/fusion-middleware/endsupport.html">https://docs.oracle.com/en/middleware/fusion-middleware/endsupport.html</a></li>
<li>IBM Newsroom, completion of webMethods and StreamSets acquisition (1 July 2024), <a href="https://newsroom.ibm.com/2024-07-01-IBM-Completes-Acquisition-of-StreamSets-and-webMethods,-Bolstering-its-Automation-Data-and-AI-Portfolios">https://newsroom.ibm.com/2024-07-01-IBM-Completes-Acquisition-of-StreamSets-and-webMethods,-Bolstering-its-Automation-Data-and-AI-Portfolios</a></li>
<li>IBM Newsroom, completion of Confluent acquisition (17 March 2026), <a href="https://jp.newsroom.ibm.com/2026-03-18-ibm-completes-acquisition-of-confluent,-making-real-time-data-the-engine-of-enterprise-ai-and-agents">https://jp.newsroom.ibm.com/2026-03-18-ibm-completes-acquisition-of-confluent,-making-real-time-data-the-engine-of-enterprise-ai-and-agents</a></li>
<li>MuleSoft, Q1 2026 product roadmap, <a href="https://blogs.mulesoft.com/news/mulesoft-q1-2026-product-roadmap/">https://blogs.mulesoft.com/news/mulesoft-q1-2026-product-roadmap/</a></li>
<li>Redress Compliance, MuleSoft pricing buyer guide (buyer-side estimate), <a href="https://redresscompliance.com/mulesoft-pricing-buyer-guide">https://redresscompliance.com/mulesoft-pricing-buyer-guide</a></li>
<li>The Business Research Company, iPaaS market report (February 2026), <a href="https://www.researchandmarkets.com/reports/5988426/integrated-platform-service-ipaas-market-report">https://www.researchandmarkets.com/reports/5988426/integrated-platform-service-ipaas-market-report</a></li>
<li>ZDNet, "Don't use an ESB unless you absolutely, positively need one", <a href="https://www.zdnet.com/article/dont-use-an-esb-unless-you-absolutely-positively-need-one-mule-cto-warns/">https://www.zdnet.com/article/dont-use-an-esb-unless-you-absolutely-positively-need-one-mule-cto-warns/</a></li>
<li>DZone, Gartner iPaaS 2025 summary, <a href="https://dzone.com/articles/gartner-ipaas-2025-ai-low-code">https://dzone.com/articles/gartner-ipaas-2025-ai-low-code</a></li>
</ul>
<p><em>Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.</em></p>
]]></content:encoded></item><item><title><![CDATA[Why I write about banking systems]]></title><description><![CDATA[A personal blog about technology decisions in banking: integration, content services, AI governance and the cost of getting them wrong.]]></description><link>https://bankingstacknotes.hashnode.dev/why-i-write-about-banking-systems</link><guid isPermaLink="true">https://bankingstacknotes.hashnode.dev/why-i-write-about-banking-systems</guid><category><![CDATA[banking]]></category><category><![CDATA[architecture]]></category><category><![CDATA[enterprise software]]></category><dc:creator><![CDATA[Chung VD]]></dc:creator><pubDate>Fri, 20 Nov 2026 09:56:42 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/gql/6afea8653a11ec043b1bc339/e8eb43f4-d7c1-4255-802c-44ee818097c3.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>For most of the last decade I have been the person who has to sign the paper that says <em>we chose this platform, and here is why</em>. Core banking, content management, API layers, identity, collections data, programme estimates. The work is rarely about technology being new. It is about a decision that has to survive a board, an auditor and a renewal cycle three years later.</p>
<p>This blog is where I make those decision patterns public — anonymised, with the numbers that can be cited and the trade-offs that usually stay inside a slide deck.</p>
<h2>Banking fails on decisions, not on ideas</h2>
<p>I have watched a bank renew a document platform because replacing it looked riskier than keeping it, and I have watched another replace the same class of platform because the renewal price had quietly become the second-largest line in the IT budget. Both were rational. Neither was obvious.</p>
<p>The difference was never the product comparison. It was whether someone had written down <strong>what would have to be true</strong> for each option to be the right one — and had the evidence to check it.</p>
<h2>What you will find here</h2>
<p>Six lines of work, each with its own series:</p>
<ol>
<li><strong>Integration and APIs</strong> — when an ESB is still the right answer, what ISO 20022 really costs to implement, and how payment hubs are chosen.</li>
<li><strong>Content services and ECM</strong> — renewal versus replacement for document platforms, and how to read a market that is consolidating.</li>
<li><strong>AI governance</strong> — what a regulated institution must be able to prove, not just promise, about models and agents.</li>
<li><strong>Identity and e-signature</strong> — the evidence chain behind a digital contract, and where implementations usually break it.</li>
<li><strong>Running programmes</strong> — estimates, PMOs and operating models that measure decisions instead of activity.</li>
<li><strong>Buying enterprise software</strong> — RFPs, scorecards and the questions that make a vendor answer honestly.</li>
</ol>
<p>Every article follows the same shape: what changed, a decision framework, options and trade-offs, the numbers, and what I would do in the first 30, 60 and 90 days. Where I do not know, I say so.</p>
<h2>What this blog is not</h2>
<p>It is not a review site. I do not rank products or take vendor briefings as fact. It is not investment or legal advice. And it is not a place to find client names: every case is anonymised, every figure is either public with a source or rounded beyond recognition.</p>
<h2>Who it is for</h2>
<p>Architects, engineering leads, CIOs and programme sponsors who have to choose between two reasonable options and explain the choice afterwards. If you are earlier in your career, the decision frameworks should still be useful — the constraints are the interesting part.</p>
<p>If that sounds like your week, subscribe to get the next note. I publish on Tuesdays, and the first series starts with integration: why ESB is not dead, and why it stopped being the default.</p>
<p><em>Personal views. Client and employer details are anonymised. Market figures are linked to public sources and dated.</em></p>
]]></content:encoded></item></channel></rss>