Why Good Data Governance Means Lift-Off for Agentic AI

I have the same conversation with technology leaders every few weeks and it always follows a familiar pattern. The board has approved the AI investment. The team has picked its models. The pilot works. Then somewhere between the prototype and production, everything slows down, and nobody can quite say why. 

In my experience it is almost never the AI. It is the data underneath it, and more precisely it is whether anyone can vouch for that data when an agent starts acting on it without asking permission first. 

The data nobody drew a map of 

Think about what is actually flowing through your organisation. Documents, emails, recorded calls, video, sensor feeds, chat threads, the notes someone typed into a customer record three years ago. Most of it is unstructured, it is growing faster than anyone can classify it, and it refuses to sit inside the taxonomies we built when data mostly meant rows in a database. 

Agentic systems change what that mess costs you. A traditional model worked with a static dataset: you trained it, drew a boundary around what it could reach and watched the outputs. An agent does not work that way. It ingests, interprets and acts continuously, reasoning across sources and making decisions with very little human involvement. The line between the systems that produce data and the systems that consume it dissolves. 

So, when an agent pulls from a document repository to draft a procurement contract, or queries a customer database to settle a dispute, it is acting on that data, and quietly assuming everything it touched was accurate, current and appropriate for the purpose. Let us not forget that an agent has no instinct for when something looks off. A person who opens a file and sees a figure that cannot possibly be right will stop and ask. An agent will use it. 

Policy documents do not stop agents 

The evidence backs this up. KPMG found that 62% of organisations see a lack of data governance as the main data challenge holding back their AI initiatives. That should give any CIO pause, because governance in most enterprises still means a set of policy documents, an approval workflow and a periodic review, all of it sitting some distance from the infrastructure that actually runs the business. 

That model assumed three things: that humans made the decisions, that data moved through defined pipelines, and that compliance could be checked after the event. Agentic AI breaks all three at once. Governance that lives in a document cannot enforce anything on a system making thousands of decisions an hour. 

This is where I see UK projects stalling. The model performs perfectly well. The leadership team simply cannot say, hand on heart, that they trust the data enough to hand over decision authority. So, the pilot stays a pilot, the investment sits there, and everyone waits for a confidence that never arrives on its own. 

Build it into the plumbing 

The organisations getting past that point treat governance as part of the data stack: lineage tracking, access policy and provenance built in, so every dataset an agent touches carries a chain of custody you can actually follow. Hitachi Vantara’s State of Data 2025 study found that 84% of organisations with strong data foundations report measurable AI ROI, against 48% of those in less mature environments. You do not close a gap that size by choosing a better model. 

The UK is in a genuinely good position here. Fifty-eight per cent of UK organisations have reached Managed or Optimised data infrastructure maturity, compared with 41% globally, and I think that discipline is why the AI success rate here runs ahead of the global average. It is a lead worth watching rather than resting on, though. Four in ten UK organisations have not reached that level, and that is precisely the population whose AI projects stall. 

What does good look like in practice? Policies that are enforceable rather than advisory. Lineage complete enough that an agent’s decision can be audited end to end, months later, by somebody who was not there. Duplicate datasets eliminated, because two versions of the truth is how an agent ends up confidently wrong. And a classification that reflects how data is used today rather than how somebody categorised it in 2019. Governance becomes a live property of the data itself. 

The question I would ask your board 

Agentic systems are already running in production across financial services, healthcare, logistics and the public sector, so this sits on the current agenda rather than a future one. The organisations delivering returns are the ones that sorted the foundations before they scaled, and the ones revisiting their AI roadmaps this year are mostly the ones that left governance until later. 

Knowing what you want from AI matters. Knowing that the data underneath it is accurate, traceable and governed as it moves is what makes any of it possible. So, the question I would put to any board is simple enough: if an agent made a decision in your business this morning that you had to explain to a regulator, a customer or a journalist tomorrow, could you show your working? 

Plenty of organisations cannot answer that yet. The ones that can will scale with confidence, and the gap between the two groups is going to widen considerably faster than anyone expects.