Governance, Skills and Data: What Is Holding UK AI Back

September is when the summer's experiments meet a board pack. The question is no longer whether the firm uses AI. In any business of real size, somebody already does. The question is whether the firm can explain what it built, who checks it, and whether the records underneath are fit to automate.

Those three constraints, governance, skills and data, are now doing more to decide the return than the choice of model. The tools got cheaper and more capable through 2026. Attention, clean records and a grown-up control did not.

The return follows the control

The State of UK AI Adoption Survey 2026, published on 20 June from 755 mid-market leaders, put a number on something operators already suspected. Firms reporting a measurable return climbed sharply with the maturity of governance, on much the same class of tools.

• No governance: 22.2 per cent report a measurable return. About 5 per cent of the market.
• Informal guidelines: 27.7 per cent. About 19 per cent of the market.
• Defined, applied inconsistently: 31.7 per cent. About 26 per cent of the market.
• Formal and organisation-wide: 59 per cent. About 29 per cent of the market.
• Mature and embedded: 85.3 per cent. About 21 per cent of the market.

Only about a fifth of that market is in the last group. The gap is not a footnote. It is the difference between a tool that pays for itself and a tool that produces activity.

Governance is a list, an owner and a stop

Boards do not need a novel policy before the autumn is over. They need to know what is already running, including the tools a team switched on without asking. Each workflow that touches a customer, a payment or a staff record needs a named owner. Anything customer-facing or money-related needs a review before it runs unattended. And someone needs to be able to halt it the same day, without raising a project.

That last point is the one agentic work makes urgent. A chatbot answers and stops. An agent takes the next step. The July piece on agentic AI in UK operations is about giving that agent a job. This one is about being able to take the job away.

The skill that matters is the check

A prompt contest will not decide this. The skill owners and managers need is knowing which work is safe to hand over, and how to tell a good result from a fluent one. People who cannot review the output end up in one of two places: they block everything, or they approve things they cannot see.

That is a leadership skill, and it is teachable. It is also why a short course for decision makers beats another all-staff licence launch. Our AI courses are built for that room: what to adopt, what to leave, and how to hold the work to account.

Data is the unglamorous constraint

An agent inherits the state of the records. Duplicate members, addresses that were never cleaned, a process that differs by office, a field nobody trusts. Automating that does not create intelligence. It creates a faster mess, delivered with confidence.

This is the constraint we see most often once the pilot excitement fades. The model was never the blockage. The customer record, the product file, the ownership of a step: those were. Firms that clean one process properly, and then automate it, outperform firms that connect a model to everything and hope. It is slower to explain and much faster to live with.

An autumn checklist

• Inventory what is already running, sanctioned or not.
• Name an owner for each workflow that touches a customer, a payment or a staff record.
• Write down how it is stopped.
• Decide what data must not leave the business.
• Train the people who check, not only the people who prompt.
• Do not widen an agent until the exception rate is boring.

If the inventory is embarrassing, that is the work. Hiding it until the next strategy day is how the gap in that table stays open. Where you are changing a real process, the August piece on where UK businesses are changing the work is the companion to this one: pick the process, then put the control around it.