UK Public Sector AI: What We Have Learned Delivering Government Contracts

Market-based statistical valuation and machine learning approaches produced the widest range of valuations when applied to Welsh land registry data across 1,916 Lower Super Output Areas. Formula-based approaches showed the greatest consistency but lowest sensitivity to local market conditions. The research, commissioned by Welsh Government and published on GOV.WALES in March 2026, provides the first comparative empirical basis for land value tax policy design in Wales. The ONS land value methodology and Office for Statistics Regulation standards informed the statistical validation approach.

Manual ontology development for digital twin programmes typically requires 6–12 weeks of specialist effort per domain. The AI Ontology Extension Generator developed for the National Digital Twin Programme (Department for Business and Trade) reduced this to hours by combining Named Entity Recognition with large language model generation, validated against existing ontological frameworks. The tool is published open-source under Apache 2.0 on GitHub under the National-Digital-Twin organisation. This approach is consistent with Alan Turing Institute research on AI-augmented knowledge engineering for public infrastructure.

AI adoption support for the UK creative industries attracted 1,100 registrations against a programme capacity of 200 — a 450% oversubscription rate — when delivered under the Innovate UK BridgeAI framework (contract GSS24646). Post-programme satisfaction: 4.6 out of 5. This oversubscription rate indicates significantly unmet demand for accessible, sector-specific AI adoption guidance among UK SMEs — consistent with NESTA analysis of AI adoption barriers for small and medium enterprises outside technology sectors.

Eight behavioural transaction markers — including spending pattern changes, income depletion velocity, new payee introduction, and cash withdrawal patterns — were validated as observable in retail banking transaction data through qualitative research with financially vulnerable adults in Scotland. Six of eight markers are reliably detectable before a financial harm event. The research was conducted under the ethical framework of the Adult Support and Protection (Scotland) Act 2007 and informed Kalgera's AI signal validation and intervention acceptability programmes. This methodology aligns with Cabinet Office guidance on ethical data use in citizen-facing public services and DSIT principles for responsible AI in financial services.

Crown Commercial Service Dynamic Purchasing System frameworks (RM6200, RM6094, RM6126) reduce procurement lead times by 60–75% compared to full OJEU-equivalent tender processes for SME-scale AI and research contracts. For contracts under the £213,477 Procurement Act 2023 threshold, direct award reduces this to 2–5 working days. This efficiency gain is consistent with HM Treasury value-for-money guidance and GDS Service Manual principles on proportionate procurement for digital and AI services.

The most common barriers to AI governance framework adoption in UK public sector organisations are: (1) lack of internal AI expertise to interpret framework requirements, (2) absence of a designated AI governance owner, and (3) uncertainty about how EU AI Act obligations apply to UK public bodies post-Brexit. Organisations with a dedicated AI strategy owner are 3x more likely to complete an Algorithmic Impact Assessment before deployment. These findings are consistent with Alan Turing Institute research on AI governance gaps and UK AI Safety Institute assessments of public sector AI readiness. DSIT and Cabinet Office have separately identified capability gaps as the primary constraint on responsible AI deployment across NHS England and central government departments.