AI9 min read
What is enterprise AI architecture?
The difference between a successful pilot and a production AI capability is rarely the model. It is the architecture around it. Here is what that architecture contains and which decisions matter.
Integration
SOA vs microservices: enterprise integration architecture that lasts
The debate is usually framed as old versus new. In practice, enterprises that integrate well take the principles of SOA, the autonomy of microservices and the governance of API management, and evolve toward them one interface at a time.
Published on 9 min readby Carlos Medger
Key takeaways
- SOA introduced reusable services with contracts. Its problems came from centralizing everything in a single bus, not from the principles.
- Microservices optimize team autonomy and deployability. They do not remove the need for governed integration between systems.
- API-led connectivity and event-driven architecture are the practical synthesis most enterprises converge on.
- Modernize ESB estates interface by interface behind stable contracts, with parallel runs, never as a big-bang rewrite.
What SOA got right, and where the ESB went wrong
Service-oriented architecture asked enterprises to expose business capabilities as reusable services with defined contracts, instead of integrating applications point to point. That idea is still correct. Systems of record such as ERPs, core banking and billing are stable; the capabilities they provide should be exposed once and reused by every channel.
The enterprise service bus was the implementation vehicle: one platform to route, transform, orchestrate and secure. It worked, and many businesses still run on IBM Integration Bus, Oracle Service Bus and similar platforms. The problems were organizational as much as technical: one central team became the bottleneck, business logic accumulated in the bus, and upgrades became multi-year events.
What microservices changed
Microservices moved the focus to autonomy: small services owned by teams, deployed independently, with their own data. This is a strong model for building new applications and digital products. It is a weak model for integrating a landscape of packaged systems that will never be decomposed.
Enterprises that tried to replace the ESB with a mesh of microservices often rebuilt the same integration logic in dozens of places, lost visibility and created a new kind of sprawl. The lesson: microservices are an application architecture, not an integration strategy.
The synthesis: API-led connectivity and events
API-led connectivity
Layered APIs make SOA principles practical again. System APIs expose systems of record with stable contracts. Process APIs compose them into business operations. Experience APIs adapt them for channels. API management provides security, versioning, catalog and analytics. Ownership is distributed by layer and domain, which removes the central bottleneck.
Event-driven architecture
Not every integration should be a synchronous call. Events published to Kafka or a message broker let systems react to changes without coupling, support real-time use cases and feed data platforms through change data capture. Messaging platforms such as IBM MQ remain the backbone of many critical flows, now combined with streaming.
Where the ESB fits
Modern integration platforms, whether IBM App Connect Enterprise, Oracle Integration Cloud, MuleSoft or cloud-native services, still perform transformation, orchestration and protocol mediation. What changes is that they are deployed as lighter, containerized runtimes, governed by API management and organized by domain rather than as one monolithic bus.
How to evolve without a rewrite
This strangler approach is how we modernize IBM and Oracle estates in practice. The business never depends on a single cutover, and every migrated interface improves observability and reuse.
- Inventory interfaces and rank them by business criticality, volume and change frequency.
- Put API management in front of existing services first. Stable contracts let you change the implementation later.
- Migrate interface by interface: build the new implementation, run it in parallel, compare outputs, switch consumers.
- Move business logic out of the bus into owned services or process APIs as you touch each flow.
- Introduce events where consumers need real-time data or where synchronous coupling causes failures.
- Retire point-to-point connections as their capabilities become available through APIs.
Why this matters for AI
AI agents need to read and change data in real systems. The safe way to give them that access is through governed APIs and events, with authentication, limits and audit. Enterprises with an API-led integration layer can put agents to work quickly. Those with point-to-point sprawl or a locked-down bus discover that the AI program starts with integration.
Frequently asked questions
Is SOA obsolete?
The principles are not: reusable capabilities with contracts remain the foundation of API-led architectures. The monolithic ESB as the only integration mechanism is what most enterprises evolve away from.
Should we migrate from IBM Integration Bus or Oracle SOA Suite?
Often yes, to current versions or cloud-hosted platforms such as IBM App Connect Enterprise or Oracle Integration Cloud, or to MuleSoft or cloud-native services, depending on skills and strategy. The migration should be incremental and behind stable API contracts.
Insights
Keep reading
AI11 min read
How to build an enterprise RAG architecture
RAG is simple to demo and hard to run. This guide walks through the components of a production retrieval-augmented generation architecture and the decisions that separate a reliable system from a chatbot that hallucinates politely.
AI8 min read
AI agents vs traditional automation: when to use each
Agents are not a replacement for every workflow engine and RPA bot. They are a new tool for a specific class of work. Knowing which class is the difference between a reliable operation and an expensive experiment.
Want to apply this to your architecture?
Talk to an architect about your systems, your data and what it would take to execute.