The UK’s public sector AI problem was never the money

Whitehall has not been stingy with AI money. Since the AI Opportunities Action Plan launched in January 2025, the government has backed a Sovereign AI Unit worth up to £500 million, committed £1 billion to expand the AI Research Resource’s compute capacity twentyfold by 2030, and folded a further £2 billion of AI-focused funding into the 2025 Spending Review. This was not announcement theatre either: by its own count, government had met 38 of the plan’s 50 commitments within a year. If the complaint is that ministers talk about AI more than they fund it, the receipts say otherwise.

And yet adoption has not moved the way the money implies it should. The National Audit Office’s survey of departments found 70% of respondents still piloting and planning AI use cases rather than running them at scale. Research this year from FSP, in partnership with Microsoft, found a similar split: 65% of public sector staff experimenting with AI, but only 30% with it built into how they actually work. Roughly twice as many teams are trying AI as have put it to work. That gap, not the funding total, is the thing worth explaining.

The explanation already in wide circulation, including from the Public Accounts Committee itself, is legacy technology and poor data quality. The PAC’s March 2025 report on government’s use of AI concluded plainly that out-of-date legacy systems and poor data quality and data-sharing are putting AI adoption at risk. An estimated 28% of central government’s IT systems met the definition of legacy in 2024, and separately, the National Audit Office counted 228 legacy systems still in live use as of March 2024. Two different inquiries, run independently, landed on the same picture. That much is now well established and no longer counts as news.

What is worth asking is why legacy systems specifically block AI when they never stopped the previous two decades of government digital transformation. GOV.UK, digital tax accounts and online benefits claims all sit on mainframes and case-management systems every bit as old as today’s “legacy” estate. The trick that made those work was architectural: build a new front door and leave the back office alone. A citizen-facing form didn’t need to read or combine the underlying case data, only to collect it and hand it off down the same old pipes.

AI cannot take that shortcut. A model that is meant to triage a claim, summarise a case file or flag a pattern across cases has to reach into the data itself, often across several systems built decades apart with no shared reference for what a “case” or a “claimant” even is. Legacy systems display data to a person through a screen. They don’t release it to another system on demand. That is a narrower, more specific failure than “poor data quality” suggests, and it explains why the fix is not a data-cleaning exercise but an integration and interoperability one.

The pattern in what has actually worked backs this up. A third of NHS chest X-rays, roughly 2.4 million scans, are now AI-assisted, because medical imaging already sits in a standardised, interoperable format inside a single well-governed pipeline. The civil service’s own “Humphrey” tools, including Consult, Parlex and Redbox, work on consultation responses, parliamentary debate and meeting transcripts: unstructured text that is already digitised and self-contained, not scattered across incompatible case systems. Every credible example of AI running in UK government today sits on data that was already free to move before the model arrived. None of them had to fight a legacy back office to get there.

Set against that, the Public Accounts Committee found that of the 72 highest-risk “red-rated” legacy systems government promised to fix under its 2022-25 Digital and Data Roadmap, 21 still had no remediation funding as of March 2025, the deadline year. The government found £500 million to seed AI companies and a further £1 billion to expand compute twentyfold, but couldn’t find the money to finish a repair job it had already promised three years earlier. That is not a funding gap. It is a funding choice, and it consistently favours the visible layer over the plumbing beneath it. DSIT has told the committee that fixing legacy systems is now urgent. That claim runs straight into the fact that this is already the second deadline government has set itself for the same 72 systems.

None of this is contradicted by the UK’s strong showing in the OECD’s Digital Government Index. The UK ranked second among member states in 2019, third in 2023, and fourth in 2025, a steady slide across three editions and six years, even as its absolute score kept rising because every edition raises the bar. Other governments simply moved their own plumbing faster over the same period. Ambition and spend were never the constraint on the UK’s international standing either.

The verdict for anyone running a public sector transformation programme is straightforward: stop measuring AI readiness by budget or headcount, and start measuring it by how many of your core systems can hand data to another system without a person in the loop. Until that number moves, more compute and more pilots will keep landing on the same 30%.

Sources

• National Audit Office, “Use of artificial intelligence in government” (2024)
• National Audit Office, “Government Cyber Resilience” (January 2025)
• House of Commons Public Accounts Committee, “Government’s Use of Artificial Intelligence” (March 2025)
• GOV.UK, “AI Opportunities Action Plan: One Year On” (January 2026)
• GOV.UK, Sovereign AI Unit collection
• FSP (Future State Policy) and Microsoft, UK public sector AI research (2025)
• OECD, Digital Government Index, 2019, 2023 and 2025 editions
• Department for Science, Innovation and Technology, “Humphrey” AI tools announcement (January 2025)