What are data foundations for AI?

Data foundations are the architecture, definitions, quality controls, access and governance that make an organisation's data trustworthy, accessible and consistent enough for AI to act on. They are the layer beneath any AI system, and for agentic AI, where agents take actions rather than just advise, they determine whether the AI creates value or causes operational damage.

Why does agentic AI need stronger data foundations?

Because agents act on data rather than just advising on it. With advisory AI, a human can catch a flawed recommendation. An agent takes the output and acts, often across multiple steps, with limited oversight. Weak or ambiguous data therefore produces wrong actions at speed, not just wrong answers, and in multi-step flows those errors compound into operational risk.

How do data foundations create business value?

Value comes from agents acting correctly and repeatedly on decisions that matter, inside processes designed to use them. That is only possible on a strong data foundation. On a weak one, either humans must check every action, removing the value, or nobody checks and incidents follow. Strong foundations are what let agents be trusted to act, which is where time and cost are saved.

How do you build data foundations without a multi-year project?

Sequence by use case rather than fixing the whole estate at once. Pick one valuable agentic use case, map the specific data it depends on, fix the architecture, definitions, quality, access and governance for that data, redesign the process, then let the agent act against a baseline. Each use case strengthens the foundation for the next, building capability instead of an endless cleanup.

Automation & AI

Data Foundations for Agentic AI: Turning Data into Business Value

27 June 2026·9 min read

Agentic AI is only as good as the data beneath it. Why data foundations decide business value, what they require, and how to build them before agents act.

Data foundations are the architecture, definitions, quality controls and governance that make an organisation's data trustworthy, accessible and consistent enough for AI to act on. For agentic AI, where autonomous agents take actions and make decisions rather than just suggest, these foundations stop being a nice-to-have and become the thing that decides whether the agents create value or cause damage. An agent that acts on weak data does not just give a bad answer. It takes a bad action, at speed, and often before anyone notices. MIT's 2025 research found around 95 per cent of generative AI pilots delivered no measurable value, with data and process readiness central to the divide, while Gartner forecasts over 40 per cent of agentic AI projects will be cancelled or scaled back by 2027. Behind most of those numbers is a data foundation that was never built.

Why do agents raise the stakes on data?

For most of AI's history in business, the model advised and a human decided. A weak data foundation produced a questionable recommendation that a person could catch. The human was the safety net. Agentic AI removes that net. An agent takes the output and acts on it, often across several steps, with limited human intervention. If the data underneath is wrong, ambiguous or undefined, the agent does not pause to wonder — it acts on the flawed input with full confidence, and in a multi-step or multi-agent flow that flawed action becomes the trusted input to the next step. This is why the same data weaknesses that produced disappointing dashboards now produce operational risk. The data did not get worse. The consequences of it being wrong did, because something is now acting on it without asking.

What do strong data foundations actually include?

A data foundation fit for agentic AI rests on five things, and an agent will expose a gap in any of them quickly. Architecture: data is structured and connected so that the systems and agents that need it can reach it reliably, rather than being trapped in silos that require manual extraction. Definition: everyone, and every agent, works from the same agreed meaning of each key field, because ambiguity that humans quietly work around becomes an error an agent acts on. Quality: accuracy, completeness and duplication are known and controlled, not assumed, since an agent inherits every flaw and amplifies it. Access and lineage: the right data is available to the right agent, and you can trace where any figure came from, essential when you need to understand why an agent did what it did. Governance: clear ownership of the data and rules for how it is used, so the foundation stays trustworthy as it feeds systems that act.

How do data foundations turn into business value?

Value does not come from agents existing. It comes from agents acting correctly, repeatedly, on decisions that matter, inside processes designed to use them. Every link in that chain depends on the data foundation. An agent on a strong foundation can be trusted to take real actions, which is where time and cost are actually saved. An agent on a weak foundation cannot be trusted, so either a human checks everything it does, which removes the value, or nobody checks and it causes incidents, which destroys it. There is no version where weak data and valuable agentic AI coexist. I have seen organisations spend heavily on agentic tooling while their data foundation went unfunded, then wonder why the agents never earned their keep. The agents were never the constraint. The foundation they stood on was.

How to build data foundations for agentic AI

Do not try to fix the entire data estate before doing anything, and do not skip the foundation and hope. Both fail. The workable path is to sequence by use case. Pick one valuable agentic use case. Map the specific data it depends on. Fix the architecture, definitions, quality, access and governance for that data — not the whole estate. Redesign the process the agent will act within. Then let the agent act, measured against a baseline. Each use case you do this way strengthens the foundation for the next, and you build real capability instead of a permanent cleanup project. This is slower than buying agents and switching them on. It is also the only version that produces value rather than incidents.