TL;DR

Public sector AI keeps stalling on data rather than models, according to Nicola Furlong, who runs EMEA public sector at SAS. Her framing is unsentimental: attach a capable model to disorganised data and the result is confidently wrong, faster. Independent OECD numbers back the pattern — most catalogued European use cases have never reached production.

The evidence outside the vendor’s framing

This is a supplier making a case, so the third-party figures matter more than the quotes. Public Sector Tech Watch, the use-case registry run out of the European Commission, has catalogued nearly 1,500 of them; when the OECD went through them in 2025, 58% had got no further than planning, piloting or development. A separate OECD outlook the year after covered 36 countries and found all but six had stood up some institution to govern how the state uses AI — yet measuring what any of it achieved stayed a weak point throughout.

Those two findings together describe the actual problem. Governance structures exist on paper; almost nothing measures whether deployments deliver public value. Furlong’s version is that the gap sits between writing a policy and anyone in the building knowing what it means day to day.

Where it moves fast and where it should not

Fraud, waste and error teams are furthest ahead, helped by synthetic data that can model unusual event types without exposing real records — useful when, as Furlong notes, fraudsters are already generating synthetic identities at scale. SAS research from December found 40% of the UK fraud specialists it surveyed across the public sector already using AI, with near-universal expectation of adoption within two years.

Decisions about an individual’s entitlement move far more slowly, and she is clear that it should: benefits eligibility and immigration are places where an incorrect output has a person attached to it.

Looking forward

The most quotable line is also the most useful one for UK organisations: “Productivity is easy to put in a slide. Trust is the actual score to keep.”. Saving a caseworker two hours weekly is a loss if citizens stop believing the process is fair — and 96% of those same specialists thought fraud and error had already damaged public trust. Furlong expects the valuable applications to become less visible, catching problems early rather than performing tasks, and predicts councils will get access to the governed AI infrastructure currently reserved for central government sooner than expected. For UK suppliers, that widening of the buyer base is the commercial signal worth acting on.