Revealed: What Government Is Really Spending on AI - And What To Do Next
AI is being sold as the way to improve public services, save money, create jobs, and give citizens better outcomes. The question inside government isn't "if," it's "how much, on what, and where's the value?"
The headline numbers
- £3.35bn in total AI-related spend by government departments since around 2018. Contracts are increasing year on year.
- The single largest deal: the Met Office's 2021 contract with Microsoft for a high-performance weather and climate supercomputer (plus smaller Copilot licenses), worth over £1bn.
- Transport for London's contract with German firm Init: £259m.
- Alphabet (Google's parent) has just two contracts worth £2.5m with the Cabinet Office and Ministry of Justice.
- Palantir holds 25 contracts totalling £376m. Reported uses include helping junior doctors draft discharge summaries and supporting defence intelligence teams to collate and process information faster.
On public concern about big tech and data, Palantir's UK lead, Louis Mosley, said: "Those are very legitimate concerns, and they're right to interrogate this, but Palantir is actually the answer to those problems. We are the way you keep data secure, and we are the way you make AI transparent and auditable."
He added a challenge to Whitehall: ministers "need to be brave" and push through change rather than let fear stall progress.
Who is spending - and who isn't
- Biggest spenders: the departments overseeing science and technology (Met Office contract) and transport (TfL-Init deal).
- Lower down: Treasury/HMRC and DWP - despite being highly data-intensive. DWP's AI spend since 2018 is under £100m against an annual IT budget exceeding £1bn.
- Why the gap? Short-term planning, limited in-house expertise, and old tech. Up to 60% of some estates still run on legacy systems.
What this means for your department
Value is clustering around infrastructure and a handful of operational use cases. Core back-office areas with high savings potential - fraud/error, debt, casework automation, contact handling - are still underused. If you lead a portfolio, the opportunity is wide open, but the window for smart procurement and standards is closing.
90-day plan to move from pilots to value
- Days 0-30: Inventory all AI activity and relevant data sources. Identify 3 priority use cases tied to measurable outcomes (e.g., processing time, call deflection, fraud prevention). Stand up guardrails: privacy impact assessments, security baselines, model usage policy.
- Days 31-60: Run 2-3 small pilots with clear success criteria (quantified KPIs, time-boxed). Draft a procurement approach that stresses interoperability and exit rights. Engage staff reps early; map task changes before tech rollout.
- Days 61-90: Scale what worked; stop what didn't. Start decommission plans for overlapping legacy workflows. Publish a short transparency note on models used, data sources, and human oversight.
Procurement guardrails that protect budgets (and your future self)
- Insist on open standards, portable formats, and data export - no lock-in to a single vendor's stack.
- Make auditability, security posture, and traceability non-negotiable. Log every model interaction that affects a decision.
- Use outcomes-based KPIs (e.g., cost per case closed, minutes saved per record) rather than vague "innovation" goals.
- Price for total cost of ownership: compute, storage, integration, monitoring, model updates, and exit costs.
- Mandate human-in-the-loop for any decision that impacts eligibility, enforcement, or liberty.
- Pre-plan the exit: data retention, model handover, and service continuity baked into the contract.
Where AI can pay off now
- Contact handling and triage: first-line responses, summarising case history, routing to the right team.
- Document-heavy workflows: drafting discharge summaries, decision letters, FOI triage, policy summarisation.
- Fraud, error, and debt: anomaly detection and prioritisation for investigation.
- Scheduling and logistics: workforce rostering, transport optimisation, asset maintenance.
- Forecasting and risk: using high-performance compute for weather/climate models to inform resilience and infrastructure planning.
Common blockers - and practical fixes
- Legacy systems: introduce APIs and data layers that decouple models from core systems before scaling.
- Skills gap: upskill delivery teams (policy, digital, commercial) on model limits, prompt design, evaluation, and assurance. See role-based training options at Complete AI Training.
- Data access: set data owners, legal basis, and sharing agreements up front. Catalogue sensitive fields and apply redaction by default.
- Risk aversion: run small, time-bound pilots with clear stop/go gates and publish the results internally.
Policy context worth knowing
Check the government's approach to regulating AI to align your assurance and risk frameworks with current expectations: AI regulation: a pro-innovation approach.
Bottom line
Government has spent billions on AI since 2018, but the money is concentrated in a few big infrastructure and transport deals. The most data-heavy departments still have headroom to find savings and improve service quality.
The path forward is simple: pick measurable use cases, fix data, modernise the worst legacy bottlenecks, and buy with exit options. Be brave about change - and disciplined about proof of value.
Your membership also unlocks:
