Ben Hollebon

The layer nobody funds

In public services AI mostly lands on an unfunded integration layer, where a citizen's form is still re-keyed into legacy systems by hand.

A black and white photo of pipes on a brick wall

The thing blocking AI in public services is the layer the tools sit on, and nobody funds it.

A practitioner analysis published on 14 September 2026 put numbers on it. Forty-five per cent of public-sector AI is bolt-on, added on top of what already exists rather than built into it. The central estate was 28 per cent legacy in 2024, up from 26 per cent in 2023, so the ground it lands on is getting older, not newer. And the work of joining the new thing to the old, the integration layer, has no reliable funding line.

The clearest example in that analysis is Pension Credit. A citizen fills in an online form. A caseworker then re-keys what they typed into disconnected legacy systems, by hand. The form is the digital part. The re-keying is the actual system. From the outside it looks modernised. Inside, a person is retyping one screen into another because the two were never joined.

This is the part that goes missing when we talk about AI in public services. The argument is usually about the model, or about how it was bought, whether it’s safe enough and procured the right way. Both matter. Neither is the blocker. The blocker is the plumbing, and no one’s budget owns it.

A National Audit Office finding reported on 3 September 2026 names the same problem from another side. Legacy technology and fragmented data are the barrier. The same person is held under different identifiers in different places, an NI number here, a UTR there, an NHS number somewhere else, and those identifiers don’t join up across organisational boundaries. You can put the cleverest model you like on top of that. It still can’t tell that three records are one person.

When the barrier is the data and the joins, the fix rarely looks like a launch. It looks like correcting records across systems that were never meant to talk, one reconciliation at a time, until the same person is the same person everywhere. There’s nothing to demo in that. No new screen, no product, no before-and-after slide for a board meeting. Just a set of records that’s finally true. Work like that is hard to fund precisely because it’s hard to show.

That invisibility is why the veneer wins. A new front end is easy to demo, easy to fund and easy to put in front of a board. The join underneath is none of those things, so it gets deferred, and the demonstrable layer gets built on a base that can’t carry it. The outcome figures fit the pattern. Only 29 per cent of staff say their department is meeting most of what it promised on AI, and 75 per cent of UK adults can’t name a single government use of AI. Plenty has been announced. Much less has reached anyone’s daily life.

This is where AI adoption goes wrong before it’s started. A tool is bought and pointed at the estate on the assumption the estate is joined up. It isn’t. So one of two things happens. Either the tool produces confident answers from records that don’t reconcile, or someone keeps re-keying underneath it and the tool becomes one more layer of veneer. Either way the headline says modernised, and the caseworker’s Monday is unchanged.

The integration layer has no natural owner. The model has a vendor. Procurement has a process and a sign-off. The plumbing between old systems appears on no one’s roadmap.

I keep coming back to a plain test. Before an AI tool goes into a public service, and before anyone signs for it, ask what it assumes about the systems underneath, and who’s paying to make that assumption true. If the answer is that the data already joins up, that’s worth checking, because both of September’s reports say that in most places it doesn’t. Buying the model is the easy part. Making the ground underneath it true is the work, and it’s the work least likely to have a budget.

Has anyone seen the integration funded before the tool arrived? If so, where?

Ben

Photograph by Lucas van Oort.