The excitement around Agentic AI is understandable. AI is beginning to move beyond generating answers towards taking actions and achieving outcomes.
But after many years working across integration, data and enterprise architecture, I have come to a simple conclusion:
AI transformation does not start with the AI.
It starts with whether an organisation’s data, services and business capabilities can be discovered, understood and safely consumed.
The complexity underneath AI
Large organisations inevitably accumulate complexity. Systems are built at different times, for different purposes and by different teams. APIs and integrations often follow the same path.
Over time, discovering what already exists, understanding what information means and knowing which capability can safely be reused becomes increasingly difficult.
My thinking around Domain APIs changed how I viewed this problem.
I initially saw Domain APIs as a way of organising capabilities around meaningful business domains, making them easier to discover, govern and reuse. But the emergence of Agentic AI gives that thinking another dimension.
A well-defined API is no longer simply an interface used by another application. Increasingly, it can become a tool through which an AI agent safely interacts with the enterprise.
Agents need more than access
Imagine an agent capable of understanding a customer’s objective and determining what needs to happen next.
The AI may be capable of reasoning about the problem. But that is only the beginning.
Where is the authoritative information? Which capability should it invoke? Is the agent authorised to perform that action? Which business rules apply? At what point should a human become involved?
These are not simply AI questions. They are also architecture, data, integration and governance questions.
This is why I believe integration will become increasingly important in the Agentic AI era. APIs expose trusted business capabilities. Events tell agents that something has changed. Integration connects information across boundaries. Business rules and policies help determine what an agent is permitted to do.
From “answer safety” to “action safety”
Much of the responsible-AI conversation rightly focuses on ethics, bias, hallucination and whether we can trust the answers generated by AI.
Agentic AI introduces another question:
Can we trust what AI does?
An agent should not be given unrestricted freedom simply because it can reason about a task. It needs boundaries around the data it can access, the capabilities it can invoke and the decisions it can make independently.
This is where architecture, data and integration become part of the guardrail.
Trusted data, governed APIs, identity, policies and business rules can create boundaries within which an agent is permitted to operate. We should not expect the AI model itself to understand and enforce every organisational constraint.
As AI moves from generating answers to taking actions, guardrails cannot exist only around the model. They must also exist around the actions the model is permitted to perform.
The foundation may matter more than the model
There is understandable excitement about increasingly capable AI models. But I suspect the organisations that succeed with Agentic AI will not necessarily be those with access to the most sophisticated model.
They may be those that have done the less glamorous work of making their data, services and business capabilities understandable, accessible, governed and safe for machines to consume.
AI can provide the intelligence, but it still needs trusted data to understand the enterprise, integration to reach and act across it, and architecture to define the boundaries within which those actions can safely take place.
The road to Agentic AI therefore starts long before the agent.
It starts with architecture, data and integration.
Read More Data & Decision Making