The Tony Blair Institute for Global Change no longer argues that governments should use AI. Its paper of 21 September, Leading in the Age of AI: How to Build an AI-Enabled State, sets out how a government should be rebuilt around it, and it gives the lead role to one institution, which it calls the centre of government. “Only the centre can set priorities across government, determine how authority and resources should be distributed between institutions, resolve competing mandates,” the authors write. For the UK that is the weak point. In Scotland, health, education, local government and planning are devolved matters, according to the Scottish Parliament. Britain’s own review of its digital estate found that 70% of public sector leaders surveyed describe their data as poorly co-ordinated. And it is Holyrood, not Downing Street, that has voted to pause decisions on AI data centres in Scotland. The paper, written for any country, never engages with any of this.
Strategic Insight: TBI’s model fits a state whose executive office controls the services it wants to transform. Downing Street has a strong executive office, but in Scotland, health, education, local government and planning are devolved matters, run by the Scottish Government rather than Downing Street.
What does TBI actually propose?
The paper has three parts. First, prepare institutions: pick a small number of priorities the leader cares about, give the bodies delivering them multi-year money and clear decision rights, and let the centre step in when delivery stalls. Second, build digital foundations: common identity, lawful data exchange, a governed institutional memory that records decisions and the evidence behind them, and systems that can swap suppliers or models without collapsing. Third, deliver: a single pathway from pilot to scale, AI-powered strategic intelligence and simulation tools run from the centre, and “bounded” agents that track commitments and chase updates within defined limits.
Its organising phrase is to “automate intelligence to elevate judgement”: machines do the searching and reconciling, people keep the decisions. The authors are explicit that this is not a power grab. “But the objective is not greater centralisation,” the conclusion says, and departments are meant to keep responsibility for their own data and delivery.
The paper is by four TBI staff, Tone Langengen, Laura Britton, Ethan Dodds and Keegan McBride, with a foreword by Tony Blair. Its historical examples are South Korea’s Economic Planning Board, Singapore’s Housing & Development Board and Economic Development Board, and Estonia’s digital state.
Why does the UK reading matter?
TBI is not only publishing. In February the Guardian reported that the Treasury’s chief secretary, James Murray, chaired a meeting on public sector AI that included TBI’s director of AI, and that the advice would “feed into efficiency processes ahead of the next spending review”. Our news report covered that meeting. A framework from an organisation already advising the Treasury deserves to be tested against the UK specifically, even though it is written for a general audience.
| Measure | Figure | Source |
|---|---|---|
| Public sector leaders describing their data as poorly co-ordinated | 70% | State of digital government review, January 2025 |
| Leaders whose data lets them see operations in full | 27% | Same review |
| Share of central government systems estimated to be legacy | 28% in 2024, up from 26% in 2023 | Same review |
| Separate government accounts and identity-proofing routes | 44 in 2021 | Same review |
| Threshold for Scottish ministers to be notified of a data centre application | Over 50MW | The Scotsman |
| Representations on the Larbert data centre application | 7,284, of which 7,166 objections | The National |
Critical Context: The paper names the UK’s centre only once, as an example: “Whether organised around Downing Street, the White House, the Kantei”. A search of the full text finds no mention of Scotland, Wales, Northern Ireland, devolution, local government or the NHS.
Where does the model meet the devolved patchwork?
TBI’s mechanism depends on a centre that can reconfigure institutions around a priority, reallocate resources and retire bodies that get in the way. Its word for the last of these is “creative destruction”. In the UK, the Downing Street centre can do that for matters that are not devolved. It cannot do it for health services, education or councils in Scotland, because those sit with Holyrood. Gov.uk even carries civil service guidance whose whole purpose is to help UK officials allow for the devolved settlements in Wales, Scotland and Northern Ireland when they make policy or run services.
The digital foundations have already divided along the same line. The review counted more than 3 million GOV.UK One Login accounts by October 2024, used for a first tranche of 50 services run by central government. Scotland has ScotAccount, which UKAuthority reports has added HM Passport Office verification for access to digital public services. Scotland also has its own guidance on public sector AI: the Scottish Government has put out guidance on responsible AI use in its public sector through the Digital Scotland Service Manual.
None of that is a failure. It is what devolution is for. But TBI’s recommendation for common identity and shared data architecture “established once and made available across the state” does not say which state. In the UK, the realistic version is several centres agreeing on common standards. That is a negotiation between governments. It is slower and more political than the centre-led build the paper describes, and the paper offers no model for it.
Reality Check: A UK “common operating picture” of the kind TBI wants would, for health or education, have to be built by agreement with at least one other government that has its own priorities and its own ministers answering to its own parliament.
Does the data exist for the centre to see?
The second test is TBI’s own precondition. The paper says agents “should be introduced only where the underlying infrastructure, such as data and identity management, is in place and sufficiently mature”. That is a sensible gate. Measured against the UK’s own evidence, it is a long way from being met.
The review gathered views from more than 500 leaders in 120 public bodies. Only 27% said their data infrastructure let them see their operations or transactions in full. Legacy systems were estimated at 28% of central government systems in 2024, and the review’s authors noted there is “no comprehensive record of the scale of legacy IT across central government, let alone the entire public sector”. In a separate digital maturity survey of 76 councils, data use and management came out as the weakest area.
The review’s diagnosis of why is close to TBI’s, which is to TBI’s credit. “Fragmentation is a feature of the system,” it says. It describes bodies treating their data as their own, and says that at the extreme an organisation’s data “will not even be ‘shared back with them’”. It says data sharing agreements under existing law “are often laboriously agreed”. TBI likewise calls reform “primarily an organisational problem rather than a technological one”.
The gap is in sequencing. TBI’s strategic intelligence layer and simulation tools assume the centre can see a live picture of spending, delivery and risk. The UK’s own review says the centre cannot yet reliably count its legacy systems. We made a similar argument in September about why government AI is a plumbing problem, and earlier in the year about the thin soil under Whitehall’s AI ambitions.
Hidden Cost: The review found that half of respondents said legacy remediation budgets are frequently reallocated to other work. Any AI-enabled centre will be competing for the same money that already fails to reach the foundations.
Can the centre clear the consenting fights?
TBI wants the centre to step in when delivery stalls, including by “resolving regulatory or institutional barriers”. The physical side of the AI state, the data centres that host the compute, shows what those barriers look like in Britain.
Scottish ministers have issued a direction obliging planning authorities to tell them about any data centre application above 50MW within a week of it being validated, as we reported at the time. The Scottish Greens’ Mark Ruskell called it a first step, adding that “monitoring the pipeline of applications is not the same as managing it”. MSPs then rejected an outright moratorium but voted instead to hold planning and consent decisions for 12 months while the Scottish Government writes national planning guidance, according to The National. In a Green-led Holyrood debate reported by the BBC, Green MSP Patrick Harvie argued: “It’s simply impossible for local planning decisions made on a piecemeal basis to take into account the national strategic issues.”
At the same time, the UK government is backing capacity in Scotland. The Scotsman reports that DataVita secured a £300m debt facility for two data centres in North Lanarkshire’s AI growth zone, backed by a £202m guarantee from the UK’s National Wealth Fund. So UK money and Scottish planning control are working on the same infrastructure from different capitals.
The decisions themselves are often more local still. The National reports that Apatura’s proposed 300MW data centre at Larbert drew 7,284 representations, including 7,166 objections, and that the decision rests with a handful of Falkirk councillors. MSPs backed the pause the same night. In England, the BBC reports that Leeds City Council granted Microsoft permission to build at Skelton Grange in Stourton, once a power station, and that about 250 people protested against it.
Strategic Reality: For consenting, the UK has at least three layers that can stop or delay AI infrastructure: UK ministers, devolved ministers and parliaments, and local planning committees. Planning is devolved in Scotland, so a centre in Downing Street could convene these layers there but would not hold the planning decision.
Who carries the consequences?
| Stakeholder | What the TBI model asks of them | UK reality | Sign it is working |
|---|---|---|---|
| Downing Street and the Cabinet Office | Lead transformation, set standards, run strategic intelligence | Direct reach stops at devolved matters | Published cross-government data standards with devolved sign-up |
| Devolved governments | Not addressed | Separate ministers, parliaments, AI guidance and identity services | Joint standards agreed through intergovernmental machinery |
| Local authorities | Contribute data and delivery | Weakest data maturity; 2% of staff in digital and data jobs | Interoperable systems without bespoke deals per request |
| Suppliers | Portable models, audit rights, exit clauses | Procurement fragmented across bodies | Common contract terms reused across the sector |
| Communities near AI infrastructure | Not addressed | Mass objections and a Scottish pause on decisions | Consent decisions made against published national guidance |
What actually decides success
The paper is right that the scarce resources are political attention and authority, not compute. In Britain those resources are shared across governments by law. A centre-led plan that treats them as belonging to one executive office will stall at the first priority that crosses a devolved boundary.
Success Factor: The test of any UK version of this blueprint is whether it starts with an agreement between the UK and devolved governments on data and identity standards, rather than with a new unit in Downing Street.
What should UK public bodies and suppliers do with this paper?
💡 Implementation Framework: Reading TBI’s blueprint against a devolved state
Phase 1: Map jurisdiction (next month)
- For each AI use case, record which government holds the policy power and which body holds the data
- Flag any use case that relies on data crossing a devolved boundary
- Identify which identity service your users actually hold
Phase 2: Fix the foundations first (next two quarters)
- Count legacy systems and the data they hold before scoping agents
- Put data sharing agreements in place for the flows you need, using existing legal gateways
- Write portability and exit terms into new AI contracts
Phase 3: Build shared practice (next 12 months)
- Align with both GOV.UK and Digital Scotland service manual guidance where you operate across borders
- Record decisions and evidence in a form that survives reorganisations
- Pilot bounded agents only where the identity and audit trail already exist
Priority actions by starting point
For bodies just starting with AI
- Know your jurisdiction: Check whether your service is reserved, devolved or local before borrowing a UK-wide framework.
- Inventory your data: The national review suggests you may not have a full list of legacy systems. Build one.
- Pick one outcome: TBI’s advice to limit priorities is useful at any scale.
For bodies already deploying
- Test the handoffs: Find where your AI tools depend on data held by another organisation, and whether that flow is formally agreed.
- Audit the audit trail: TBI’s logging list, covering inputs, model versions, human decisions and overrides, is a good checklist for existing systems.
- Plan for model change: Check you could replace your model or supplier without rebuilding the service.
For suppliers and advanced adopters
- Build for several centres: Products that assume one national identity or data layer will need adapting for Scotland.
- Watch the consenting timetable: Scottish data centre decisions are paused pending guidance; hosting plans that depend on them need contingency.
- Offer portability as standard: TBI’s procurement recommendations call for data and model portability and exit terms.
Four problems the paper leaves open
Challenge 1: Whose centre?
The model gives one centre authority to reconfigure institutions. The UK has one at Westminster and others for devolved services.
Mitigation strategy: Treat cross-border AI work as intergovernmental from the outset, with shared standards agreed before shared systems are built.
Challenge 2: The precondition is not met
TBI’s own gate for agents, mature data and identity infrastructure, does not describe the estate the UK’s review found.
Mitigation strategy: Use the gate honestly. Fund the data and legacy work first, and let agent pilots follow where the foundations exist.
Challenge 3: Physical infrastructure sits outside the model
The paper does not discuss data centres or community consent, yet the compute behind an AI state has to be built somewhere, and in Scotland decisions on it are currently paused.
Mitigation strategy: Public bodies relying on UK-hosted compute should check where it will come from and on what timetable, rather than assuming capacity will arrive.
Challenge 4: Centralised visibility raises accountability stakes
TBI proposes logging, traceability and human ownership of agents, which is welcome. A strategic intelligence layer that spans governments raises the question of whose parliament scrutinises it.
Mitigation strategy: Settle scrutiny arrangements for any shared system before it is built, including which legislature can call its owners to account.
⚠️ Warning: A UK-wide AI layer designed in Downing Street and presented to devolved governments afterwards risks meeting the same response Holyrood has given data centre developers: a pause until its own guidance is written.
The takeaway: a good design for a state Britain is not
TBI’s paper is more careful than its headline suggests. It sequences agents behind foundations, insists that humans own decisions, and names fragmentation and organisational resistance as the real obstacles. Its failure is one of fit. It describes a strong centre reshaping a single state, and the UK is a union of governments with separate powers over health, education and planning, a data estate its own review says it cannot fully count, and a consenting system in which a Falkirk planning committee and the Scottish Parliament can decide what gets built.
Three things matter most:
- Devolution comes first: Any UK version has to begin with intergovernmental agreement on standards, not with a new unit at the centre.
- Foundations before agents: TBI’s own precondition, applied to the UK’s own data, says the plumbing work comes first.
- The compute has to be consented: An AI state depends on infrastructure that local and devolved decision-makers are currently contesting.
Measuring this differently
A useful measure of UK progress is not how many AI tools the centre deploys but how many cross-government data flows are formally agreed and working, including those across devolved boundaries. By that measure the UK’s own January 2025 review suggests a low starting point.
The paper’s best idea for Britain may be its least glamorous: a governed institutional memory that survives reorganisations. It works at any level of government and does not depend on resolving who the centre is.
Your next steps
Immediate actions (this week):
- Record whether each of your AI use cases sits in reserved, devolved or local territory
- Check which identity services your users hold
- Read TBI’s accountability logging list against your current systems
Strategic priorities (this quarter):
- Complete a legacy system inventory
- Formalise the data sharing agreements your AI tools rely on
- Add portability and exit terms to AI contracts at renewal
Long-term considerations (this year):
- Track whether the UK and devolved governments agree common data or identity standards
- Monitor Scotland’s national data centre planning guidance and any decisions that follow
- Watch whether TBI’s Treasury advice appears in the next spending review
Source: Leading in the Age of AI: How to Build an AI-Enabled State (Tony Blair Institute, 21 September 2026), by Tone Langengen, Laura Britton, Ethan Dodds and Keegan McBride, with a foreword by Tony Blair. UK context draws on the State of digital government review (January 2025), the Scottish Parliament’s page on devolved and reserved powers, UKAuthority on Scotland’s public sector AI guidance, The Scotsman on the data centre direction, the BBC on Holyrood’s data centre debate, The National on the Larbert hearing, the BBC on the Leeds data centre and the Guardian on the Treasury’s AI advisers.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether any UK plan for AI-enabled government begins with an agreement between the UK and devolved governments, and what Scotland’s data centre guidance says when it arrives. Read more analysis at Insights, or get in touch.