Data foundations of Agentic AI

Agentic AI is often described in terms of what it does: it observes, reasons, and takes action. That’s the part that captures attention. But underneath the behaviour sits something far more practical: an agent can only be as effective as the information it can access, interpret, and trust at the moment it makes a decision.

This is where many organisations get pulled into a familiar story. They hear “AI needs data,” and the conversation immediately collapses into data lakes, lakehouses, medallion layers, and enterprise-wide re-platforming. That story contains some truth, but it is incomplete. Agentic AI does depend on data, but it is not solely dependent on data platforms in the traditional sense. In many cases, agentic success depends more on real-time access, APIs, and the ability to structure unstructured information at the point of use than it does on having a perfect analytical architecture.

To understand why, it helps to go back in time—before the modern era of enterprise data architecture.

Before Data Platforms, There Were Control Loops

Long before anyone spoke about lakehouses, industry was already full of autonomous systems. In factories, aircraft, energy systems, and high-precision manufacturing environments, controllers were making decisions continuously. The logic wasn’t complicated in principle: read the signal, compute a response, apply an action. The reason it mattered was the constraint: the loop had to run in milliseconds, and it had to run close to the process.

That world never depended on centralised reporting layers. It depended on current state. It depended on data that was fresh enough to be meaningful. And it depended on interfaces that made the state available in a form the controller could use.

Agentic AI is not the first time we’ve built systems that act. What’s new is that the “signals” now include documents, workflows, digital operations, and enterprise knowledge—not just sensors and actuators. We are expanding decision loops into domains that were historically reserved for humans, because they relied on interpretation rather than measurement.

The Data Platform Era Delivered Visibility, Not Action

Over the last two decades, enterprises invested heavily in data architecture for good reasons. Those investments created cross-functional visibility, enabled regulatory reporting, supported forecasting and optimisation, and brought order to fragmented operational systems. Data warehouses, data lakes, and lakehouses became the backbone of modern analytics, and the value they delivered is real.

But most of that architecture evolved around a specific assumption: data is collected, transformed, and presented so that humans can interpret it and decide what to do next. Even when machine learning entered the picture, the typical pattern remained “predict, then hand off.” The enterprise platform became an engine for insight and oversight.

Agentic AI changes the centre of gravity. Not because analytics becomes irrelevant, but because the system is no longer producing insight for a human by default. The system is increasingly expected to use information to coordinate work.

That shift doesn’t make the data lake obsolete. It makes the data lake insufficient as the single foundation.

Agentic Systems Need Current State, Not Just Historical Truth

A lot of enterprise data architecture is optimised for correctness, consistency, and reconcilability. That’s the right design for reporting and strategic analytics. But agents often operate under a different constraint: the decision is only as good as the current state of the world the agent is acting in.

If an agent is scheduling work, routing a request, approving an exception, coordinating a response, or triggering a workflow, it needs to know what is true right now. It needs to query live inventories, current capacity, the present status of a batch, the latest lab result, the newest ticket update, the most recent policy revision, or the current risk posture.

This is why agentic systems frequently demand something enterprises underestimate: a reliable way to read and write operational state in real time or near real time. The most practical path to that isn’t a data lake. It’s usually an API. APIs are the nervous system of modern organisations. They are how systems expose truth, accept updates, and trigger downstream work. When you take agentic AI seriously, the question stops being “Do we have a lakehouse?” and becomes “Can the agent see and influence the system of record safely and correctly?”

Controlled Action in Agentic Systems

One important architectural question naturally follows from this discussion: if agents can observe enterprise systems, should they also be allowed to modify them? The answer in most enterprise environments is not through direct database access. Allowing an autonomous system to write directly to operational databases bypasses the business logic, validation rules, and audit controls that enterprise applications are designed to enforce.

Instead, agentic systems should interact with organisations through the same governed interfaces used by other software—typically APIs, service layers, or workflow tools that enforce permissions, validation, and logging. In this model, the agent does not manipulate data structures directly; it invokes approved operations such as creating a task, updating a record, or triggering a workflow. This approach preserves system integrity while still allowing agents to participate safely in operational decision loops. In regulated industries, this pattern also ensures that agentic decisions remain traceable, reviewable, and aligned with existing governance frameworks.

The Real Frontier Is Unstructured Information

The second major shift is not about speed. It’s about form. A huge proportion of what organisations rely on every day is not neatly structured in tables. It lives in standard operating procedures, PDFs, batch records, engineering notes, deviation narratives, emails, manuals, policies, contracts, and tickets. This is operational knowledge, and in many industries it is the difference between work that is repeatable and work that is fragile.

Traditional enterprise data architecture has never been good at turning this into something a system can reason over. That’s why so many automation programmes stall: they can integrate system-to-system transactions, but they can’t integrate the knowledge layer that tells people how to interpret exceptions, follow procedures, or navigate ambiguity. Agentic AI changes this equation because it can convert unstructured sources into structured representations on demand. It can extract entities, parameters, steps, constraints, and intent. It can produce a reliable intermediate form—often JSON or another structured schema—that downstream tools and workflows can consume. It can make documents actionable.

This is not a minor convenience. It effectively expands the data surface area of the enterprise. Suddenly, the “data the business runs on” includes the knowledge the business used to keep locked in text.

Structuring Data at the Moment of Use

This leads to one of the most important conceptual shifts in agentic design. Traditional data programmes tend to follow a “structure first, use later” mindset. You standardise, model, cleanse, and curate the data in advance. Then you build applications on top. That remains the right approach for many reporting and analytical needs, and it is often required in regulated contexts.

Agentic workflows introduce a complementary pattern: “structure dynamically, use immediately.” Instead of waiting for every document type or process variant to be fully modelled upstream, the agent can create structure when a task requires it, using the source material that is already available. The agent can interpret, extract, and normalise information inside the workflow itself. This does not eliminate the value of upstream modelling. It changes the boundary between what must be perfect in advance and what can be safely derived at runtime. In practical terms, it means organisations can begin to automate valuable work without waiting for multi-year data transformations to complete, while still evolving toward stronger governance over time.

Agentic Data Is Broader Than the Data Lake

If you zoom out, agentic systems typically draw from multiple kinds of data access simultaneously. They may use enterprise platforms for historical context, trends, and governed reference data. They may use operational APIs to get current state and execute actions. They may use event streams for responsiveness. They may use document stores and knowledge bases for procedural and interpretive context. They may use external data sources for market signals or third-party validation.

This is why agentic AI can succeed in organisations with very different levels of data maturity. The organisation with a world-class data platform has a deep asset the agent can leverage—especially for governed context, lineage, and analytics. But the organisation without that platform is not locked out. If it has clean operational systems, decent APIs, and well-managed knowledge artefacts, it can often move faster than expected. In fact, heavy “data architecture momentum” can become a double-edged sword. Large investments sometimes come with rigid operating models, long change cycles, and an assumption that value must wait for the platform to be complete. Agentic programmes reward a different capability: the ability to connect real work to real information and close the loop safely.

Challenging the Status Quo

So where does this leave the enterprise that has already spent millions on data platforms?

In a strong position—if it reframes what those platforms are for. Data lakes, lakehouses, and warehouses are excellent at what they were designed to do: provide reconciled truth, governed history, shared metrics, and analytical leverage. In an agentic world, those assets become part of the context layer. They help agents explain decisions, trace provenance, detect drift, and support auditability. They also provide the training and evaluation substrate that serious organisations will require as agentic systems scale. But those platforms are not, by themselves, the operational interface an agent needs to act. Agents also need safe pathways into the systems where work happens. They need reliable APIs, robust permissions, fine-grained controls, and feedback loops that confirm whether actions succeeded. They need mechanisms to observe outcomes and learn. Without those, the agent becomes an analyst looking through glass.

The bridge is simple: keep the enterprise data investment, and extend it with agent-ready interfaces and operational integration. The goal is not to replace the platform. The goal is to complete the loop.

The Practical Definition of “Good Data” for Agentic AI

In analytics, “good data” often means consistent, complete, and reconciled. In agentic systems, “good data” has additional properties.

It must be timely enough to represent current state. It must be accessible through interfaces that workflows can use. It must be interpretable, including unstructured sources. It must be reliable enough to justify action. And in regulated contexts, it must be traceable, with a defensible chain of evidence from source to decision to outcome.

The best way to think about it is this: agentic AI doesn’t just need data to exist. It needs data to be usable at the point of decision.

Data as the Enabler of Autonomy

When people say “agentic AI is coming,” what they really mean is that decision-making and coordination will move into software by default. That transition is not powered by model capability alone. It is powered by the information substrate that allows the model to understand what is happening, decide what to do, and act with confidence.

That’s why the data foundations of agentic AI extend far beyond the lakehouse. They include real-time state, APIs, operational tooling, and unstructured knowledge. They include the ability to create structure dynamically inside workflows, not only upstream in pipelines. They include the discipline to make action safe, traceable, and learnable.

For organisations with strong enterprise data architecture, the opportunity is to connect that investment into decision loops. For organisations without it, the opportunity is to move quickly by treating operational access and knowledge structuring as first-class capabilities. Either way, the destination is the same: systems that don’t merely report reality, but can increasingly participate in shaping it.