Skip to main content

ESB isn't dead. It stopped being the default answer.

A decision framework for the integration layer in 2026 — when a bus is still right, when it is a liability, and what replacing it actually costs.

Updated 10 min readMarkdown
ESB isn't dead. It stopped being the default answer.

Every budget cycle, the same sentence appears in a slide: "ESB is a dead architecture, we should move to APIs." 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.

The sentence is half right. Centralised ESB as the default integration model is finished. The product lines are being retired, the vendors are merging, and the analyst narrative has moved on. But the integration jobs 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.

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.

What changed

The products are being retired, and the dates are now close.

  • Apache ServiceMix was retired in March 2026; 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 BizTalk Server 2020 is the final release.
  • Oracle Fusion Middleware 12c — the platform under Oracle Service Bus and SOA Suite — ends Premier Support in December 2026, with Extended Support ending December 2027.
  • SAP Process Orchestration 7.5 ends mainstream maintenance in 2027.
  • BizTalk Server 2020 end-of-sale is 31 March 2027, with mainstream support ending 12 April 2028.
  • Red Hat 3scale reaches end of maintenance on 30 June 2027 — it should not be selected for anything new.

The market has consolidated, which changes roadmap risk.

IBM completed its acquisition of webMethods and StreamSets in July 2024, and its acquisition of Confluent (~USD 11 billion) on 17 March 2026. Salesforce completed Informatica in November 2025. Every one of those deals moves a product you may depend on under a different roadmap owner.

The analyst picture moved with it.

  • Gartner's API Management Magic Quadrant (28 September 2026) listed eight Leaders — Boomi, Google, Gravitee, IBM, Kong, Salesforce/MuleSoft, Sensedia and WSO2 — with Axway dropping to Challenger and Traefik entering the quadrant.
  • Gartner's iPaaS 2025 quadrant demoted both Salesforce/MuleSoft and Oracle from Leader to Challenger, while TIBCO and Software AG no longer qualified.

A new integration consumer appeared: the AI agent.

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 edge layer — and it is another reason not to conflate "API management" with "the thing that replaces the ESB".

The distinction people keep missing

Centralised ESB API management
Philosophy Smart pipes, dumb endpoints Smart endpoints, dumb pipes
Traffic East–west between internal systems North–south from consumers to services
Communication Mostly asynchronous: JMS, MQ, SOAP/WS-*, file, EDI Mostly synchronous: REST, GraphQL, gRPC
Core job Route, transform, mediate protocols, orchestrate Authenticate, throttle, version, measure, monetise
Whose logic The bus owns business rules The service owns business rules
Scaling Stateful, hard to scale, easy to make a single point of failure Stateless, horizontal

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 "API management or ESB?" but:

  1. Do I need to govern consumption of APIs (partners, developers, agents, quotas, monetisation)?
  2. Do I need to mediate and orchestrate internal systems, including legacy adapters and long-running transactions?
  3. If both, do I buy one suite or compose specialised layers?

The decision framework

I score seven criteria. Weights reflect a regulated financial institution; adjust them, but write them down before the vendor demo.

# Criterion Weight What "high" looks like
1 Legacy adapter depth (mainframe, MQ, SAP IDoc/RFC, EDI/HL7) 20% Real systems need CICS/IMS/MQ adapters, not HTTP proxies
2 Transaction semantics (compensation, correlation, human tasks) 20% Processes span days and must be recoverable
3 Consumer model (internal team, partner, developer, AI agent) 15% External self-service, portal, quotas, MCP
4 Deployment and data residency 15% On-prem or BYOC is mandatory, not preferred
5 Operating model and skills 10% One central middleware team versus federated product teams
6 Three-year TCO including overage and professional services 15% Pricing metric grows linearly with volume
7 Exit and lifecycle risk 5% Config exportable, no EOL in the next three years

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.

Three options, not two

Option What it means Strength Risk Three-year cost shape
Keep and strangle Keep the bus for what only it can do; put an API gateway in front; move interfaces out one at a time Lowest disruption, preserves adapters, buys time You keep the licence and the skills problem Licence + integration work; savings are gradual
Replace with an API-led suite MuleSoft, IBM, WSO2 or Boomi as the new centre Strong governance, one vendor, faster onboarding New lock-in, migration cost often exceeds licence savings, and the "suite" is itself consolidating High licence + 25–50% professional services in year one
Decompose API gateway (north–south) + event streaming (east–west) + iPaaS for SaaS/B2B + in-app connectors (Camel/Spring) No single point of failure, best cost curve at scale Requires four capabilities and a real architecture team Higher initial complexity, lower marginal cost

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.

The numbers that decide it

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 USD 800,000 per year, with 29–41 committed vCores unused — between USD 232,000 and 328,000 per year of waste, or roughly USD 0.75–1.1 million over three years. 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.

For IBM App Connect Enterprise, the publicly listed SaaS entry points are around USD 8,484 per 10,000 VPC-hours per year and USD 2,544 per 100,000 flow runs per year. On-premise pricing is not published and is metered by PVU/VPC. Professional services in year one are commonly 20–40% of licence for Boomi and 25–50% for MuleSoft, again per third-party estimates.

The market context explains why vendors are repricing: third-party estimates put API management at USD 5.24 billion in 2025 growing to 6.92 billion in 2026 (~32% CAGR), and iPaaS at 13.88 billion in 2025, 18.72 billion in 2026, and 54.78 billion by 2030. Growth attracts consolidation, consolidation raises roadmap risk, and roadmap risk is now a line item in the decision.

What I would do in 30, 60 and 90 days

First 30 days — inventory, not architecture.

  • Count interfaces, not applications: protocol, direction, payload size, messages per day, criticality.
  • List every end-of-life date that falls inside three years, with the owner's name against it.
  • Classify each interface into four buckets: retire, wrap (gateway in front of the existing service), re-platform, keep.
  • Stop all new development on interfaces classified "retire".

Next 30 days — contract and proof.

  • Open a licence workstream with four clauses on the table: true-down rights, escalation cap, price hold, and an exit/export clause.
  • 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.
  • Decide the skills plan before the technology plan. A gateway nobody can operate is a renewal you cannot avoid.

Next 30 days — target state and a first migration.

  • 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.
  • Migrate the three interfaces with the highest EOL risk and lowest coupling first. Not the most visible ones.
  • Write the architecture decision record, including what you deliberately did not replace.

What I would not do

  • I would not buy a replacement bus to answer the sentence "ESB is dead". You would be buying the same centralisation with a newer invoice.
  • I would not put business rules in the API gateway. That is how the original coupling problem returns, one policy at a time.
  • I would not migrate by technology preference. Migrate by end-of-life date and coupling, in that order.
  • I would not assume the gateway solves API security. 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.
  • I would not let a three-year roadmap commitment outlive the vendor's own product commitment. If the product line has changed owner twice, cap the term.

For the non-technical sponsor

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.

Sources

Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.

No comments yet. Be the first to comment.