← Back to Technology

Technology: evidence and policy appraisals

Detailed research records, alternatives, costs and provisional assessments. These records preserve different scopes and historical reviews; they are not an additional ranked action plan.

Evidence → choices → outcomes

What could improve it?

AI that measurably improves services and productivity, with costs, dependency and continuity risks kept manageable.

Legacy systems leave services exposed

Resilience also depends on legacy systems and cyber controls. The NAO found significant control gaps and unfunded remediation for many vulnerable systems. AI adoption figures do not measure whether critical services can withstand an attack.

Evidence in the domain SWOT →

These options link to their national assessment, including funding, delivery constraints and remaining uncertainties.

  • Assessed option · provisional

    Make cloud exit and continuity work in practice

    Select a critical service and test usable backups, data export and a proportionate recovery route.

    8/10 criteria scored

    Coverage of the appraisal, not an impact score.

    Delivery: Amber

    Whole-life costs incomplete

    Evidence and delivery constraints

    Concentration establishes exposure; it does not establish the cost-effective resilience design.

    Run a costed service-level exercise before buying blanket duplication. No invented national AI-outage loss; retain service-specific risk thresholds.

    Resilience and sustainability appraisal queueEvidence and assessment

  • Assessed option · provisional

    Turn research into products and adoption

    7/10 criteria scored

    Coverage of the appraisal, not an impact score.

    Delivery: Amber

    Whole-life costs incomplete

    Evidence and delivery constraints

    The artifact offers weak support for its spinout/procurement package. Smart Grants evaluation is relevant to a different instrument and cannot validate the whole bundle.

    Separate spinout terms, procurement trials, grants and firm adoption; define market failure, beneficiaries and comparator.

    Growth and prosperity listEvidence and assessment

This domain in the UK SWOT →UK National Action Plan →

How we assessed these options

What would it take to improve this?

Detailed policy appraisal

The outcome we want: AI that measurably improves services and productivity, with costs, dependency and continuity risks kept manageable.

The main constraint: Cloud supply is concentrated and switching is rare, and adoption figures do not show whether AI helps — so the binding constraints are market structure and evidence, which organisations, buyers and the CMA must address workload by workload rather than by counting compute or provider nationality.

Constraint type: market structure, evidence.

The options on the table

What is funded now, and the changes assessed against it — the current baseline and the alternatives, before the position today and the actions.

Provisional single-reviewer assessments from the 6 October 2026 review; not approved business cases, and not government policy.

  • Make cloud exit and continuity work in practiceDelivery risk: Amber

    Funded now
    Not yet stated in the record.
    The change assessed
    Select a critical service and test usable backups, data export and a proportionate recovery route.

    Range if completed: 15.0–21.0 out of 30 (equal weights; not yet assessed: expected incremental benefit (Bn), incremental affordability (Co))

Assessed with another page as home

These options also bear on this topic. The full assessment sits on the home page named under each.

  • Turn research into products and adoptionDelivery risk: Amber

    Home: Innovation.

    Funded now
    Not yet stated in the record.
    The change assessed
    Not yet defined.

    Range if completed: 11.0–20.0 out of 30 (equal weights; not yet assessed: increment beyond baseline (Ad), expected incremental benefit (Bn), incremental affordability (Co))

Alternatives already on record: 21 options listed under the points in the position today (13 without an action record), and 3 earlier options under other options compared. They have not been scored in the review.

The position today

  • What we can build on (2)
  • What is holding us back (4)The NAO’s March 2024 audit identified value-for-money risks from unclear ownership and accountability for government AI adoption. This is a dated audit finding, not evidence that every gap remains unresolved in 2026; pilots and announcements alone do not establish service improvements. (March 2024 audit · UK central government)
  • What could improve things (1)Tested exit plans for critical workloads, and evaluated AI pilots compared with current processes, would show where AI helps and where dependency is a real risk.
  • What could make it worse (2)

Which quadrant a point sits in is our analytical judgement, not a statistical category. Each point says whether it is a measured fact, our synthesis or a risk.

What we can build on Strengths

What is holding us back Weaknesses

What could improve things Opportunities

What could make it worse Threats

The actions worth examining

Our shortlist for this page. Research quality, policy status, cost and funding are judged separately; none of these is government policy unless its status says so.Our shortlist. None is government policy unless its status says so.

  1. Make cloud exit and continuity work in practice — not yet costed / funding not identified
  2. Measure useful AI adoption before scaling it — not yet costed / funding not identified

  • Action 1A43Policy status: site-editorial action

    Make cloud exit and continuity work in practice

    Map critical dependencies, require proportionate portability and exit plans in contracts, and test that a critical workload can actually be restored elsewhere.

    The thread: issue, evidence, action

    1. Issue Weakness: Cloud infrastructure is concentrated: Microsoft and AWS each held 30–40% of UK cloud infrastructure and platform services in 2024 (CMA estimates), and fewer than 1% of customers switch provider in a year. Also: threat, opportunity
    2. Evidence Cloud customers switching provider in a year
    3. Action A43, this action

    Assessment (draft, 6 October review)

    Who decides:
    Technology owners and procurement teams in public and private organisations; the CMA and sector regulators (UK)
    Evidence:
    Expert or audit recommendation. A statutory market investigation documents the barriers; whether exit plans improve resilience needs workload-specific tests.
    Expected effect:
    Direction plausible; size not established.
    Cost and funding

    Not yet costed

    Funding: Not yet costed / funding not identified

    Why it is shortlisted here: Concentration and rare switching are the page’s clearest measured weakness; exit only counts if it has been tested.

    What it takes
    The problem

    Cloud supply is concentrated and switching is rare and costly, so organisations depend on providers they cannot easily leave if prices rise or a service fails.

    Who is affected, and where

    Public and private organisations running critical services in the cloud, and the people who use those services. Territory: UK. Competition regulation is UK-wide; public-sector technology policy is partly devolved.

    Policy status

    UK Facts proposal. Builds on regulatory action already under way: the CMA’s 2025 final decision, actions on egress fees and interoperability that Microsoft and Amazon set out in March 2026 (the CMA’s wording; not formal commitments), with a six-month review, and critical-third-party designations in financial services from July 2026. Tested exit plans for critical workloads are a UK Facts proposal.

    The evidence
    What the research found:
    The CMA found competition not working well: Microsoft and AWS each held 30–40% of UK cloud infrastructure and platform services combined in 2024 (Appendix D, Table D.3; for infrastructure alone AWS had 40–50%), and egress fees and licensing practices harm switching. In March 2026 the CMA accepted provider commitments on egress fees and interoperability instead of opening cloud SMS investigations, with a review after six months.
    Evaluated on:
    UK cloud customers and providers (CMA market investigation); United Kingdom, market data to 2024
    Our recommendation (separate from the research):
    Run a realistic exit and restore test on a suitable critical workload and use contract renewals to add proportionate portability terms.
    Strongest counter-evidence:
    Multi-cloud is not automatically cheaper or safer; poor implementation can weaken security; provider nationality does not guarantee resilience.
    Cost and funding

    Cost completeness: Not yet costed. Funding: Not yet costed / funding not identified.

    • Migration, egress, engineering, training and any duplicated capacity (administration; paid by central government, employers): Not yet costed. Multi-cloud is not automatically the cheapest solution; costs differ by workload.
    Who pays:
    Each organisation’s technology and resilience budget (migration, egress, engineering, training, duplicated capacity).
    Financing proposed:
    Each organisation’s own budget; regulator remedies do not pay migration costs.
    Gap remaining:
    Unknown.
    What delivery needs
    Skills:
    Cloud architects and engineers.
    Operational capacity:
    Time for tests outside live operations.
    Infrastructure:
    Alternative capacity for restore tests.
    Permissions:
    Data residency and security approvals.
    Law:
    None new.
    Public consent:
    Boards accepting the cost of tests.
    First step and stages

    First step: Map critical dependencies and run a realistic exit and restore test on one suitable workload. (owner: Organisation technology owners (in government, departmental CTOs))

    1. Map critical cloud dependencies — Technology owners (proposed stage, not an official timetable)
    2. Test exit and restore for one workload — Technology owners (proposed stage, not an official timetable)
    3. Add proportionate terms at renewal — Procurement teams (proposed stage, not an official timetable)

    Depends on: CMA review of provider commitments; Contract renewal cycles

    Trade-offs
    Who gains and who loses:
    Service users gain from continuity; organisations bear test and migration costs.
    What else the money or effort could do:
    Engineering time not spent on new services.
    Strongest counterargument:
    Large providers are more resilient than most in-house alternatives; duplicating capacity is costly.
    Risks:
    Complexity, weaker security and confusing UK ownership with resilience.
    How we would judge it
    Baseline:
    Fewer than 1% of UK cloud customers switch provider a year (CMA, 2024).
    Leading indicator:
    Critical workloads with a tested exit plan.
    Outcome:
    Tested recovery time; switching time and cost; material vendor dependencies.
    Continue, adapt or stop:
    Scale tests only where the first shows recovery is feasible at proportionate cost.
    Earlier option on this page

    This card replaces the option “Make switching between cloud providers easier” (operating: Microsoft case: proposed decision expected October 2026, statutory deadline 13 February 2027.). Its evidence: The CMA found that fewer than 1% of UK cloud customers switch provider a year, with egress fees and software licensing among the barriers. In March 2026 it chose not to open cloud strategic-market-status cases after Microsoft and AWS took steps on egress fees and interoperability, and to review progress; it opened a separate case on Microsoft’s business software, including cloud licensing, in May 2026. Whether these steps raise switching is not yet known. CMA, Cloud services market investigation: final decision · CMA, Microsoft’s business software ecosystem: strategic market status investigation

    Sources

    Source published July 2025 (CMA final decision); March 2026 (CMA announcement); evidence reviewed 2026-10-04. Reconsider if: The CMA Board’s review of provider commitments; The Microsoft business-software SMS decision (deadline 13 February 2027).

  • Action 2A44Policy status: site-editorial action

    Measure useful AI adoption before scaling it

    Pick low-risk service problems and compare an AI-assisted workflow with the current process — including human review and failure handling — before scaling.

    The thread: issue, evidence, action

    1. Issue Weakness: Adoption rates are not outcomes: the main evaluated adoption scheme found no statistically significant effect on productivity. Also: threat, opportunity, strength
    2. Evidence Firms with ten or more staff using AI
    3. Action A44, this action

    Assessment (draft, 6 October review)

    Not assessed in the 6 October review; this action stays in the topic plan with its own evidence.

    Who decides:
    Accountable service owners and technology teams, with independent evaluators (UK)
    Evidence:
    Untested proposal. An evaluation-led proposal: the evidence shows adoption rising and one flagship scheme without a measurable productivity effect, not what this approach would achieve.
    Expected effect:
    Direction plausible; size not established.
    Cost and funding

    Not yet costed

    Funding: Not yet costed / funding not identified

    Why it is shortlisted here: Adoption is rising fast; whether it improves services has to be measured before scaling.

    What it takes
    The problem

    AI use is rising fast, but adoption rates do not show whether services or productivity improve, and the main evaluated adoption programme found no significant productivity effect.

    Who is affected, and where

    Public-service users and staff; firms adopting AI. Territory: UK. Public services are partly devolved; pilots would sit with each government.

    Policy status

    UK Facts proposal. A UK Facts proposal following the Magenta Book; government AI programmes were not audited in this review.

    The evidence
    What the research found:
    ONS survey data show 34.9% of firms with ten or more staff used AI in June 2026, against 11.9% in September 2023. The Made Smarter Adoption evaluation’s more robust recent-beneficiary analysis found no statistically significant turnover, employment or productivity effects, alongside reported operational benefits.
    Evaluated on:
    UK firms (ONS survey); manufacturers supported by Made Smarter Adoption (evaluation); United Kingdom, 2023–2026
    Our recommendation (separate from the research):
    Define baseline quality, acceptable error and evaluation design first; pilot, evaluate independently, then expand or stop with published reasons.
    Strongest counter-evidence:
    Evaluation slows adoption that may be beneficial; generic AI tools change quickly, so trial results can date fast.
    Effects on record
    • Staff workload and service quality when AI outputs need checking: A mixed effect; size not established. plausible spillover, limited confidence; United Kingdom. This is why pilots must count review time and errors, not only throughput.
    Cost and funding

    Cost completeness: Not yet costed. Funding: Not yet costed / funding not identified.

    • Model and cloud use, integration, training, quality control, energy and exit (administration; paid by central government, councils): Not yet costed. Hidden human review work must be counted.
    Who pays:
    Service owners’ budgets (model and cloud use, integration, training, quality control, energy, exit costs) and evaluators.
    Financing proposed:
    A bounded pilot may use a named digital or innovation allocation; scaling needs a funded business case based on measured results.
    Gap remaining:
    Unknown.
    What delivery needs
    Skills:
    Service designers, data scientists, evaluators.
    Operational capacity:
    Staff time to review AI outputs.
    Infrastructure:
    Secure compute and data access.
    Permissions:
    Data protection impact assessments.
    Law:
    Existing data protection and equality law.
    Public consent:
    Staff and users informed; human oversight.
    First step and stages

    First step: Define baseline quality, user needs, acceptable errors and the evaluation design for one service problem. (owner: The accountable service owner)

    1. Set baseline and evaluation design — Service owner; evaluator (proposed stage, not an official timetable)
    2. Pilot with human review — Service team (proposed stage, not an official timetable)
    3. Expand or stop, with published reasons — Service owner (proposed stage, not an official timetable)

    Depends on: A43: exit plans for AI services; A42: trial and adoption route

    Trade-offs
    Who gains and who loses:
    Users may gain faster service; people less able to use digital channels may lose; staff workloads can rise from hidden review work.
    What else the money or effort could do:
    Evaluation time and money; other improvements not made.
    Strongest counterargument:
    Waiting for evaluations means falling behind organisations that adopt quickly.
    Risks:
    Automation bias, surveillance, unequal access, hidden review work and vendor dependency.
    How we would judge it
    Baseline:
    34.9% of firms with ten or more staff used AI (ONS, June 2026).
    Leading indicator:
    Pilots with a published baseline and evaluation design.
    Outcome:
    Net time and cost, quality, errors, accessibility and user outcomes.
    Continue, adapt or stop:
    Stop if quality or error rates are worse than the current process after review costs.
    Earlier option on this page

    This card replaces the option “Help organisations adopt AI well” (operating). Its evidence: Use is rising (11.9% to 38.1% of firms with ten or more staff since September 2023), but the main evaluated adoption programme, Made Smarter Adoption, found no statistically significant effect on productivity. ONS, Business insights and impact on the UK economy: 8 October 2026 · Made Smarter Adoption: impact and process evaluation (Ipsos)

    Sources

    Source published May 2026 (Magenta Book); July 2025 (CMA); evidence reviewed 2026-10-04. Reconsider if: Published evaluations of public-sector AI pilots; New ONS AI adoption data.

How the options are assessed

One shared record per option — the same one the national plan reads. Delivery risk, gate and list membership are separate judgements.

  • Make cloud exit and continuity work in practice — Amber delivery RAG; gate pending

Provisional single-reviewer assessments from the 6 October 2026 review; not approved business cases, and not government policy.

  • A43v2Delivery risk: AmberGate: pending

    Make cloud exit and continuity work in practice

    A new scope of action A43 (A43v2); A43 itself is unchanged and nothing is inherited from it.

    Provisional single-reviewer desk assessment, 6 October 2026.

    Who and where
    Select a critical service and test usable backups, data export and a proportionate recovery route. (from the 6 October review) Jurisdiction: not defined.
    Result so far
    Range if completed: 15.0–21.0 out of 30 (equal weights; not yet assessed: expected incremental benefit (Bn), incremental affordability (Co))
    Funded now (baseline)
    Funded baseline not yet stated
    The change assessed (marginal)
    Select a critical service and test usable backups, data export and a proportionate recovery route. (from the 6 October review)
    What the evidence supports—and the next decision

    A backup exists only as a resilience measure if data and services can actually be recovered within an acceptable time.

    AI-assisted analytical follow-up · 8 October 2026. Proposed appraisal design; the attributed scores below are unchanged. No funding or delivery agreement is implied.

    Define the change
    Select a critical service, define recovery needs and rehearse loss of its cloud provider or identity layer using a proportionate alternate route.
    Test whether it works
    Demonstrated recovery time, data loss, integrity and essential-service continuity; use independently observed exercises and record failed tests.
    Establish the cost
    Backup, export, alternate capacity, licences, staff exercises and continuing maintenance. Do not invent a national outage-loss estimate.
    Protect against harm
    Protect sensitive data and avoid a nominal second provider that shares the same critical dependency.

    Read with the outcome evidence and sources in the full assessment. Monitoring proposals are not existing baseline measurements. A numerical benefit or cost score needs evidence for the defined increment, not just the scale of the problem.

    Full assessment of Make cloud exit and continuity work in practice: ten criteria, scenarios, delivery, costs and monitoring
    The ten criteria (0–3)

    Each criterion is scored 0 to 3 by the reviewer. “Not yet assessed” means unknown: it is never counted as zero.

    1. Scale and severity (Sc)2 of 3 — Serious harm to a defined group or widespread moderate harm

      Reviewer’s reason: Problem scale judged from the linked domain evidence; not a population estimate of those helped.

    2. Increment beyond baseline (Ad)2 of 3 — Specific material change to existing delivery

      Reviewer’s reason: Select a critical service and test usable backups, data export and a proportionate recovery route. This is a provisional marginal-change judgement, to confirm against commissioned delivery.

    3. Confidence in the proposed action (Ev)0 of 3 — No effect evaluation

      Reviewer’s reason: Concentration establishes exposure; it does not establish the cost-effective resilience design.

    4. Expected incremental benefit (Bn)Not yet assessed

      Reviewer’s reason: Unknown: no sufficiently supported magnitude for this exact change.

    5. Residual harms (Hm)2 of 3 — Manageable harm with specified safeguards

      Reviewer’s reason: Duplicated cost, security complexity and a nominal exit plan that fails operationally.

    6. Incremental affordability (Co)Not yet assessed

      Reviewer’s reason: Unknown: the proposed increment has not been costed and matched to an approved affordable allocation in this review. Existing programme envelopes are not that allocation.

    7. Delivery capacity (Ca)2 of 3 — Stretched but plausibly obtainable

      Reviewer’s reason: No invented national AI-outage loss; retain service-specific risk thresholds.

    8. Distribution (Fa)2 of 3 — Mixed or broadly neutral with safeguards

      Reviewer’s reason: Distribution assessed against the target group and the burdens described here: Duplicated cost, security complexity and a nominal exit plan that fails operationally.

    9. Meaningful first outcome (Ti)3 of 3 — Within one year

      Reviewer’s reason: Timing judgement concerns the first meaningful outcome in: Recovery time, data loss and continuity under a rehearsed disruption. It does not imply all long-term benefits arrive then.

    10. Readiness (Rd)2 of 3 — Named commissioning/allocation/coordination step

      Reviewer’s reason: Run a costed service-level exercise before buying blanket duplication. Owner: Service owners, procurement teams and relevant regulators.

    Four weightings

    Normalised as 10 × weighted sum ÷ sum of weights, on a 0–30 scale. Where a criterion is not yet assessed, the range is the lowest and highest result if it were later scored 0 or 3 — not a score, a confidence interval or a rank.

    • Equal weights: Range if completed: 15.0–21.0 out of 30. Not yet assessed: expected incremental benefit (Bn), incremental affordability (Co). Funding gate: not passed.
    • Evidence doubled: Range if completed: 13.6–19.1 out of 30. Not yet assessed: expected incremental benefit (Bn), incremental affordability (Co). Funding gate: not passed.
    • Scale doubled: Range if completed: 15.5–20.9 out of 30. Not yet assessed: expected incremental benefit (Bn), incremental affordability (Co). Funding gate: not passed.
    • Cost ignored (the funding gate still applies): Range if completed: 16.7–20.0 out of 30. Not yet assessed: expected incremental benefit (Bn), incremental affordability (Co). Funding gate: not passed.

    Funding gate: Not passed — the cost is not yet assessed or no allocation is committed for this scope. Ignoring cost in a weighting never removes the gate.

    Delivery risk and gate

    Delivery risk shown: AmberReview’s claim: Amber. Best the recorded evidence allows: Amber.

    Review’s reason: Incremental lifetime costs and funding not established. No invented national AI-outage loss; retain service-specific risk thresholds.

    Why it is not Green yet

    • Scope (population, place and increment) not verified against the delivery baseline.
    • No verified owner with the powers in place.
    • Delivery capacity not verified.
    • Lifetime incremental costs not complete.
    • No credible funding (a committed allocation for the scope).
    • Safeguards not specified.
    • Monitoring lacks a metric, a data source or a baseline.

    Green needs verified scope, a verified owner with the powers, verified capacity, complete lifetime costs with credible funding, specified safeguards and monitoring with a metric, source and baseline. Amber means appraisal gaps remain; Red means a material delivery obstacle. Red does not mean unimportant.

    Gate: Pending — appraisal incomplete (review label: “Pending cost, funding, scope and capacity appraisal”).

    Why: Scores incomplete; Funding gate not passed (cost unknown or no credible funding).

    The gate is separate from national list membership: being listed does not make an option eligible, and an eligible option need not be listed.

    Cost and funding

    Lifetime costing: Not established.

    Still to cost: Incremental lifetime costs and funding not established. No invented national AI-outage loss; retain service-specific risk thresholds.

    Funding status: not established in review.

    Owner, powers and capacity

    Proposed owner: Service owners, procurement teams and relevant regulators. (proposed, unverified).

    Powers: needs a decision or allocation. Capacity: unverified.

    Powers and capacity read from the review’s Rd and Ca anchors; no proposed owner has agreed to act (the review: “Proposed owners have not agreed to act”).

    England:
    not assessed
    Scotland:
    not assessed
    Wales:
    not assessed
    Northern Ireland:
    not assessed

    Not defined in the review.

    Dependencies, safeguards and risks

    Depends on: No invented national AI-outage loss; retain service-specific risk thresholds.

    Safeguards

    None specified yet.

    Risks and residual harms: Duplicated cost, security complexity and a nominal exit plan that fails operationally.

    Monitoring and measured results

    Metric: Recovery time, data loss and continuity under a rehearsed disruption.

    Data source: Not yet named. Baseline: Not yet established.

    Measures already on the domain page

    A page measure is not necessarily the review’s metric; linking the two is an editorial decision still to be made.

    Evidence for the outcome

    Outcome: Recovery time, data loss and continuity under a rehearsed disruption.

    Concentration establishes exposure; it does not establish the cost-effective resilience design.

    Review record

    Assessment status: Provisional single-reviewer desk assessment, 6 October 2026.

    Reviewer: Original assessment author (name not supplied). Reviewed: 6 October 2026.

    Carried from the earlier domain review (C record): must be calibrated jointly with the G/W/E/NG/NW/NE records before any combined comparison.

    Decision route: Run a costed service-level exercise before buying blanket duplication.

    An editorial screening framework by one reviewer, not a formal appraisal or business case. Proposed owners have not agreed to act.

Decision briefs

  • Prove one critical government service can be restored on other infrastructure — by actually doing it
  • Measure one AI-assisted workflow against a control group before rolling it out

The leading option set out as a decision someone could take: who decides, who benefits, the best evidence, cost and funding, and the main trade-off. The full research record is one click away.

Decision briefPublic-policy proposalReviewed 2026-10-04

Prove one critical government service can be restored on other infrastructure — by actually doing it

The thread: issue, evidence, action

  1. Issue Weakness: Cloud infrastructure is concentrated: Microsoft and AWS each held 30–40% of UK cloud infrastructure and platform services in 2024 (CMA estimates), and fewer than 1% of customers switch provider in a year. Also: threat, opportunity
  2. Evidence Cloud customers switching provider in a year
  3. Action Make cloud exit and continuity work in practice
Decision required
The owner of one critical government service sets a recovery time objective, writes portability and exit terms into its cloud contract, and runs a real restore of the service on alternative infrastructure, publishing the measured recovery time and cost.
Who and where
Users of one critical government digital service. Applies in: UK.
Accountable body
The service-owning department (its CTO and senior responsible owner)
Intervention
Set a recovery objective, secure portability in the contract and run a live restore on a second platform.
Best evidence, and where it applies
The October 2025 AWS outage lasted about 14.5 hours and affected the Home Office, DVLA, DWP and HMRC; the CMA found fewer than 1% of cloud customers switch provider each year.Studied in: UK cloud infrastructure market (Market investigation, 2023–2025).How closely the comparison transfers to this decision: European Union — partial.
Extra cost of this change
Capital: none identified in the brief.Recurring: none identified in the brief.One-off or mixed: Migration, egress, engineering, training and any duplicated capacity — Not yet costed.
Funding
Not funded: no funding route identified yet.Committed to this change: none.Possible source, not an award: none identified.Unfunded: Not costed.
Savings and wider benefits
Not yet specified: no cashable saving or avoided cost has been costed for this change. Expected outcomes are listed one by one in the research appendix and are never added together or counted as money available.
First 90-day step
One department names the critical workload, publishes its recovery objectives and schedules a live restore testThe proposed first milestone; a 90-day date is not yet specified. Owner: The service-owning department. Proposed by UK Facts, not an official deadline.
Twelve-month milestone
Not yet specified.Time to a first useful result: The measured restore within a year.
Outcome, and when to scale or stop
Measure: Exit clauses agreed; test scheduled.Scale if: The service is restored within its objective at a cost the department judges proportionate; extend to the next critical workload.Stop if: Recovery on a second platform proves infeasible at proportionate cost — then invest in same-provider multi-region resilience instead and record why.
Main trade-off
A real restore test costs engineering time and duplicated capacity and can itself disrupt the service; a paper plan costs little but may not work.
Source and review date
Cloud services market investigation (case page; final decision 31 July 2025) (31 July 2025); Government cyber resilience (HC 546, Session 2024–25) (29 January 2025). Reviewed 2026-10-04.

An editorial proposal by UK Facts for discussion, not government policy. Records: A43 · costs C-A43-EXIT

Research appendix: study design, effect sizes and limits, alternatives, full costing, international comparisons and how we would know

1.The change, in plain words

Prove one critical government service can be restored on other infrastructure — by actually doing it

2.The decision requested

Decision
Select one critical citizen-facing service hosted with a single provider, set its recovery time and recovery point objectives, add portability and exit clauses at the next contract change, and run a live restore test on a second platform within the year, publishing the result.
Accountable institution
The service-owning department (its CTO and senior responsible owner)
Geography
UK
Powers
Existing departmental technology and procurement authority; GDS cloud and lock-in guidance; no legislation.
Delivery partners
Government Digital Service; Cabinet Office commercial function; Current and alternative cloud providers; National Cyber Security Centre; NAO (assurance)

3.Who it is for, and what changes beyond today

Target population
Users of one critical government digital service.
Eligibility
A citizen-facing service the department rates critical, currently dependent on one infrastructure provider, with data that can lawfully be held on the alternative platform.
Scale
One workload first; others only if the first restore is feasible at proportionate cost.
Baseline
AWS us-east-1 disruption on 19–20 October 2025 lasted about 14.5 hours; the government said the Home Office, DVLA, DWP and HMRC were affected and the cost was not yet known. The CMA estimated Microsoft and AWS each held 30–40% of UK cloud infrastructure and platform services combined in 2024 and fewer than 1% of customers switch a year. The NAO found 58 critical systems with significant gaps on independent review.

Source: Amazon Web Services; House of Commons, Summary of the Amazon DynamoDB service disruption in the Northern Virginia (US-EAST-1) Region, and written answer 84777 on UK government impacts (Opening paragraph; written answer 84777); Competition and Markets Authority, Cloud services market investigation (case page; final decision 31 July 2025) (Summary of final decision, paragraphs 5 and 24); National Audit Office, Government cyber resilience (HC 546, Session 2024–25) (Paragraph 19)

Change beyond existing law or funding
GDS guidance already asks business cases to include exit time and contingency plans to be tested; the finance sector has a critical-third-party regime. The change is one live, measured restore of a government service and contract terms that make it possible.

4.How it would work

Do: Set a recovery objective, secure portability in the contract and run a live restore on a second platform Changing: so that the department knows how long recovery takes and what blocks it before a real outage

  • Measured recovery time for the service against its objective

    Hypothesis to test

    The output of the test; whether continuity improves depends on fixing what the test finds.

  • Lower switching costs at the next contract renewal

    Hypothesis to test

    Depends on egress and licensing terms; the CMA accepted provider commitments on egress in March 2026 with a six-month review.

Outcomes are kept separate and never added into one impact figure.

5.Best evidence

  • Cloud services market investigation (case page; final decision 31 July 2025)

    Author or institution:
    Competition and Markets Authority
    Published:
    31 July 2025
    Evaluation period:
    Market investigation, 2023–2025
    Design:
    Statutory market investigation
    Sample or population:
    UK cloud infrastructure market
    Effect and uncertainty:
    Microsoft and AWS each 30–40% of UK cloud infrastructure and platform services combined in 2024 (CMA ranges; AWS 40–50% of infrastructure alone); fewer than 1% of customers switch each year; egress fees a key commercial barrier.
    Limits:
    A competition finding, not a resilience evaluation.

    Where: Summary of final decision, paragraphs 5, 24 and 26. Affiliation is provenance, not proof of quality.

  • Government cyber resilience (HC 546, Session 2024–25)

    Author or institution:
    National Audit Office
    Published:
    29 January 2025
    Evaluation period:
    Assessments to 2024
    Design:
    Value-for-money audit
    Sample or population:
    UK central government IT systems
    Effect and uncertainty:
    72 critical systems self-assessed, 58 independently reviewed with significant gaps; at least 228 legacy systems, 63 red-rated; the 2025 hardening aim will not be met.
    Limits:
    Cyber resilience, not provider-failure recovery specifically.

    Where: Paragraphs 19–20. Affiliation is provenance, not proof of quality.

6.Alternatives, counter-evidence and who loses

Closest feasible alternative
Multi-region deployment within the same provider, which protects against a regional failure but not against provider-wide problems or commercial lock-in.
Strongest counter-evidence
Multi-cloud is not automatically cheaper or safer; poor implementation can weaken security; provider nationality does not guarantee resilience.

Source: Competition and Markets Authority, Cloud services market investigation (case page; final decision 31 July 2025) (Summary)

Distributional effects
Users who rely most on digital-only services gain most from continuity; costs fall on the department’s technology budget.
Main downside
Duplicated capacity costs money every year if kept warm; a cold standby may be cheaper but slower.

The options compared, including current policy

  • The proposal

    Live restore test of one critical workload on a second platform, with exit terms in the contract

    Who is reached
    Users of one critical service
    Evidence
    CMA market findings; NAO cyber audit; 2025 outage
    Likely outcome and timing
    Measured recovery time within a year
    Cost and funding
    Not costed; departmental budget
    Main trade-off
    Engineering cost and test risk

  • Current policy

    Current practice: guidance asks for exit plans and tested contingency; no published live test

    Who is reached
    All services, on paper
    Evidence
    GDS lock-in guidance
    Likely outcome and timing
    No change
    Cost and funding
    Existing budgets
    Main trade-off
    Plans may not work when needed

  • Alternative

    Multi-region resilience within the current provider

    Who is reached
    Same service
    Evidence
    Protects against regional outages like October 2025
    Likely outcome and timing
    Faster to implement
    Cost and funding
    Extra capacity with one provider
    Main trade-off
    No protection against provider-wide or commercial lock-in

7.What other countries or the devolved nations show

Country lesson · European Union, Applies from 17 January 2025; first critical-provider designations November 2025

Mandatory tested exit plans for critical cloud services in EU finance

Evidence: Untested proposalPartial match: some conditions differ

What was changed
The Digital Operational Resilience Act requires financial firms to have exit strategies for technology services supporting critical functions, with exit plans “sufficiently tested and reviewed periodically”, and brings designated critical providers under EU oversight.
Study design and population
Legislation; no evaluation yet EU financial entities and 19 designated critical ICT providers
Measured outcome there
None measured yet: the rules are too recent for outcome data on tested exits or recovery times.
Cost status
Compliance costs not established.
What failed or remains uncertain
Whether tested exit plans would work in a real large-provider failure is unknown; testing may be paper-based.
Why the UK differs
The UK has its own critical third parties regime for finance (first designations from 13 July 2026) but no equivalent requirement for government’s own critical services.
First change required here
Require one critical government workload to complete a real restore on alternative infrastructure, and publish the measured recovery time.

Not a UK forecast: the effect above was measured in that place and population. Source: Official Journal of the European Union, Regulation (EU) 2022/2554 on digital operational resilience for the financial sector (DORA) (Articles 28(8), 31, 64). Reviewed 2026-10-04.

8.What delivery needs first

  • procurement — Portability, data-export and exit-assistance clauses at the next contract change. (Requirement; owner: The department with the Cabinet Office commercial function)
  • staff — Engineers able to run the service on a second platform. (Scarce resource shared with other priorities; owner: The department)
  • data — Lawful basis and security accreditation for holding the data on the alternative platform. (Genuine veto: nothing proceeds without it; owner: The department with NCSC)

9.What it would cost

How the cost is built up: Engineering time + egress fees + alternative-platform capacity (test, then standby choice) + accreditation; set against the cost of an outage, which the government says is not yet known.

  • Migration, egress, engineering, training and any duplicated capacity

    One critical workload

    Not yet costed (ledger C-A43-EXIT)

    mixed; price year not stated; Test year and ongoing standby; Not yet established

10.Who pays, and what is missing

Programme envelope (a possible route, not an award)
None identified
Allocation committed to this change
None
Remaining gap
Not costed.
Proposed financing
The department’s own technology and resilience budget; regulator remedies do not pay migration costs.
Responsible payer
central government
Budget displaced
Other work in the department’s technology programme.

11.First milestone and time to a useful result

One department names the critical workload, publishes its recovery objectives and schedules a live restore test (Owner: The service-owning department. Proposed by UK Facts, not an official deadline.)

Time to first useful result: The measured restore within a year.

What could delay it: Contract terms prevent data export without renegotiation Security accreditation of the second platform takes longer than expected

12.How we would know

Baseline
Untested exit plan; recovery time unknown.
Comparison
The recovery time objective set before the test.
Leading indicator
Exit clauses agreed; test scheduled.
Follow-up
Annual repeat tests; recovery time in any real incident.
Continue if
The service is restored within its objective at a cost the department judges proportionate; extend to the next critical workload.
Adapt if
The restore succeeds but slowly: fix the blockers (data transfer, licensing) and retest.
Stop if
Recovery on a second platform proves infeasible at proportionate cost — then invest in same-provider multi-region resilience instead and record why.
Thresholds
Not yet set: The recovery objective is set by the service owner from the service’s criticality before the test.

What is not yet known, and how to find out

  • Cost of the test and of keeping a standby (cost)Next step: The department to obtain quotes from the alternative provider and estimate egress.Scope: One workload · Data or quotes needed: Provider quotes; egress pricing; engineering days · Who can estimate it: The service-owning department
  • The cost of the October 2025 outage to government services (data)Next step: Ask DSIT for the assessment the written answer said was not yet known.
  • Whether tested exit plans improve recovery in real incidents (effect)Next step: Record recovery times in future incidents for tested and untested services.Defensible route: evaluation of existing delivery

Sources last checked 4 October 2026 · Owner: Evidence editor (James Harris) · Next check: the first trigger: A further major cloud outage affecting government services.

Decision briefPublic-policy proposalReviewed 2026-10-04

Measure one AI-assisted workflow against a control group before rolling it out

The thread: issue, evidence, action

  1. Issue Weakness: Adoption rates are not outcomes: the main evaluated adoption scheme found no statistically significant effect on productivity. Also: threat, opportunity, strength
  2. Evidence Firms with ten or more staff using AI
  3. Action Measure useful AI adoption before scaling it
Decision required
A service owner chooses one low-risk, high-volume workflow (for example drafting replies to routine correspondence), measures its quality, error rate and staff time — including review time — and compares teams using an AI tool with teams that do not, before any wider roll-out.
Who and where
Users and staff of one public-service workflow. Applies in: UK.
Accountable body
The accountable service owner in the chosen department
Intervention
Compare an AI-assisted workflow, including human review and failure handling, with the current process in matched teams.
Best evidence, and where it applies
Government trials so far rely on self-reports: UK civil servants estimated 26 minutes a day saved and Australian participants about an hour, with no control groups; an evaluation of the Made Smarter adoption programme found no significant average effect on firms’ value added.Studied in: 20,000 employees in 12 organisations (7,115 survey responses) (30 September to 31 December 2024).How closely the comparison transfers to this decision: Australia (federal public service) — partial.
Extra cost of this change
Capital: none identified in the brief.Recurring: none identified in the brief.One-off or mixed: Model and cloud use, integration, training, quality control, energy and exit — Not yet costed; Independent evaluation — not yet established.
Funding
Not funded: no funding route identified yet.Committed to this change: none.Possible source, not an award: none identified.Unfunded: Not costed.
Savings and wider benefits
Not yet specified: no cashable saving or avoided cost has been costed for this change. Expected outcomes are listed one by one in the research appendix and are never added together or counted as money available.
First 90-day step
The service owner publishes the workflow, baseline quality and time measures, and the evaluation design before the tool is usedThe proposed first milestone; a 90-day date is not yet specified. Owner: The accountable service owner. Proposed by UK Facts, not an official deadline.
Twelve-month milestone
Not yet specified.Time to a first useful result: Three to six months of comparison data.
Outcome, and when to scale or stop
Measure: Net staff time per case including review.Scale if: Quality and errors are no worse and net staff time falls after counting review work.Stop if: Quality or error rates are worse than the current process after review costs.
Main trade-off
A controlled comparison delays roll-out by months; scaling on self-reported savings risks paying for time that is spent checking outputs.
Source and review date
Microsoft 365 Copilot Experiment: Cross-Government Findings Report (2 June 2025); Made Smarter Adoption: impact and process evaluation (Ipsos) (30 July 2026). Reviewed 2026-10-04.

An editorial proposal by UK Facts for discussion, not government policy. Records: A44 · effects E-A44-REVIEW · costs C-A44-PILOT

Research appendix: study design, effect sizes and limits, alternatives, full costing, international comparisons and how we would know

1.The change, in plain words

Measure one AI-assisted workflow against a control group before rolling it out

2.The decision requested

Decision
Run a controlled comparison in one workflow — randomised or staggered teams — with pre-registered measures of quality, errors, accessibility and total staff time, and make further licences conditional on the result.
Accountable institution
The accountable service owner in the chosen department
Geography
UK
Powers
Existing departmental authority; Magenta Book evaluation guidance; data-protection impact assessment for the workflow.
Delivery partners
Independent evaluator; Departmental digital and data teams; Trade unions and staff representatives; The tool supplier (data access, not the evaluation)

3.Who it is for, and what changes beyond today

Target population
Users and staff of one public-service workflow.
Eligibility
A high-volume, low-risk workflow with a measurable quality standard and human review of every output.
Scale
Enough teams for a comparison, set by the evaluator’s power calculation.
Baseline
AI use among firms with ten or more staff rose from 11.9% (September 2023) to 34.9% (June 2026). In the UK Copilot experiment (20,000 staff, late 2024) participants self-estimated 26 minutes a day saved, with no control group; in Australia’s trial (about 5,765 licences) perceived savings were around an hour a day and up to 7% said it added time.

Source: Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026 (Firms with ten or more employees); Government Digital Service, Microsoft 365 Copilot Experiment: Cross-Government Findings Report (Findings and limitations); Digital Transformation Agency (Australian Government), Evaluation of the whole-of-government trial of Microsoft 365 Copilot: summary of evaluation findings (Summary of evaluation findings)

Change beyond existing law or funding
Trials have measured staff perceptions; none of those reviewed measured quality, errors and total time against a control. The change adds that comparison and ties scaling to it.

4.How it would work

Do: Compare an AI-assisted workflow, including human review and failure handling, with the current process in matched teams Changing: so that decisions to scale rest on measured quality and net time, not self-reported savings

  • Staff workload and service quality when AI outputs need checking

    Hypothesis to testEffect record E-A44-REVIEW

    This is why the comparison counts review time and errors, not only throughput.

  • Net staff time per case at equal or better quality

    Hypothesis to test

    Self-reported savings in the UK and Australian trials are not measured effects.

Outcomes are kept separate and never added into one impact figure.

5.Best evidence

  • Microsoft 365 Copilot Experiment: Cross-Government Findings Report

    Author or institution:
    Government Digital Service
    Published:
    2 June 2025
    Evaluation period:
    30 September to 31 December 2024
    Design:
    Survey and usage data; no control group
    Sample or population:
    20,000 employees in 12 organisations (7,115 survey responses)
    Effect and uncertainty:
    Self-estimated average saving of 26 minutes a day; how saved time was used could not be identified.
    Limits:
    Self-reports; licence delays; festive-period disruption; no control.

    Where: Findings and limitations. Affiliation is provenance, not proof of quality.

  • Made Smarter Adoption: impact and process evaluation (Ipsos)

    Author or institution:
    Ipsos, for the government
    Published:
    30 July 2026
    Evaluation period:
    Programme period covered by the evaluation
    Design:
    Impact and process evaluation
    Sample or population:
    Manufacturing firms in the Made Smarter Adoption pilot regions (England)
    Effect and uncertainty:
    The more robust recent-beneficiary analysis found no statistically significant turnover, employment or productivity effects; operational benefits were reported.
    Limits:
    Digital adoption in manufacturing, not public-service AI.

    Where: Impact findings. Affiliation is provenance, not proof of quality.

6.Alternatives, counter-evidence and who loses

Closest feasible alternative
Roll out licences widely and monitor usage and satisfaction.
Strongest counter-evidence
Evaluation slows adoption that may be beneficial; generic AI tools change quickly, so trial results can date fast.

Source: HM Treasury / Evaluation Task Force, The Magenta Book: Central Government guidance on evaluation (Chapter 1, sections 1.5–1.6)

Distributional effects
Users may gain faster service; people less able to use digital channels may lose; staff workloads can rise from hidden review work; administrative roles were flagged as most affected in Australia.
Main downside
A tool may improve between the trial and roll-out, so the result may understate or overstate later performance.

The options compared, including current policy

  • The proposal

    Controlled comparison in one workflow before scaling

    Who is reached
    Users and staff of one workflow
    Evidence
    Magenta Book; self-report limits of UK and Australian trials
    Likely outcome and timing
    Results in three to six months
    Cost and funding
    Not costed; departmental allocation
    Main trade-off
    Slower roll-out

  • Current policy

    Current practice: licence trials judged on self-reported time savings

    Who is reached
    Volunteer staff
    Evidence
    UK 26 minutes a day self-estimated; no control
    Likely outcome and timing
    Fast
    Cost and funding
    Licence costs
    Main trade-off
    Hidden review work and quality unknown

  • Alternative

    Wait for vendor or academic evidence before any use

    Who is reached
    None now
    Evidence
    Little public-service evidence exists
    Likely outcome and timing
    Indefinite
    Cost and funding
    None
    Main trade-off
    Forgoes possible gains

7.What other countries or the devolved nations show

Country lesson · Australia (federal public service), Six-month trial announced November 2023; evaluation 2024

A whole-of-government AI assistant trial in Australia

Evidence: Observational analysisPartial match: some conditions differ

What was changed
About 5,765 Microsoft 365 Copilot licences were issued across 56 agencies, with an independent evaluation of use, perceived productivity and risks.
Study design and population
Surveys, interviews and focus groups; self-nominated participants; no control group Australian Public Service staff who volunteered
Measured outcome there
Perceived time savings of around an hour a day on summarising, drafting and searching; 69% said it made them faster and 61% said it improved quality; only a third used it daily; up to 7% said it added time.
Cost status
Licence costs not reported in the summary; total trial cost not established.
What failed or remains uncertain
Self-reported savings from self-selected staff; inconsistent roll-out; concerns about jobs, bias and vendor lock-in. The UK’s own 2024 experiment (20,000 staff) also relied on self-estimates (26 minutes a day) with no control group.
Why the UK differs
Different public-service systems and licensing terms; both trials measured staff perceptions, not service quality or users’ outcomes.
First change required here
Choose one workflow, measure its quality, error rate and staff time before the tool, then compare with a control group doing the same work without it.

Not a UK forecast: the effect above was measured in that place and population. Source: Digital Transformation Agency (Australian Government), Evaluation of the whole-of-government trial of Microsoft 365 Copilot: summary of evaluation findings (Summary of evaluation findings); Government Digital Service, Microsoft 365 Copilot Experiment: Cross-Government Findings Report (Findings and limitations). Reviewed 2026-10-04.

8.What delivery needs first

  • data — A quality standard and error measure for the workflow, applied blind to whether AI was used. (Requirement; owner: The service owner)
  • law — A data-protection impact assessment and supplier terms that keep evaluation data with the department. (Genuine veto: nothing proceeds without it; owner: The department)
  • staff — An independent evaluator and staff agreement to take part. (Requirement; owner: The service owner)

9.What it would cost

How the cost is built up: Licences or usage for treated teams + integration + training + blind quality scoring + evaluation; savings counted only as measured net staff time at equal quality.

  • Model and cloud use, integration, training, quality control, energy and exit

    One low-risk workflow

    Not yet costed (ledger C-A44-PILOT)

    one-off revenue; price year not stated; Trial period; Not yet established

  • Independent evaluation

    Design, quality scoring and analysis

    Not yet established one evaluation

    one-off revenue; price year not stated; Trial period; Not yet established

10.Who pays, and what is missing

Programme envelope (a possible route, not an award)
None identified
Allocation committed to this change
None
Remaining gap
Not costed.
Proposed financing
A bounded pilot from a named departmental digital or innovation allocation; scaling needs a funded business case based on the measured result.
Responsible payer
central government
Budget displaced
Other digital projects in the department.

11.First milestone and time to a useful result

The service owner publishes the workflow, baseline quality and time measures, and the evaluation design before the tool is used (Owner: The accountable service owner. Proposed by UK Facts, not an official deadline.)

Time to first useful result: Three to six months of comparison data.

What could delay it: Supplier terms restrict evaluation data Staff decline to take part

12.How we would know

Baseline
Quality, error rate and staff time per case in the workflow before the trial.
Comparison
Randomised or staggered teams using the current process.
Leading indicator
Net staff time per case including review.
Follow-up
Quality and errors after three and six months; user complaints.
Continue if
Quality and errors are no worse and net staff time falls after counting review work.
Adapt if
Time falls but errors rise: add review steps and retest.
Stop if
Quality or error rates are worse than the current process after review costs.
Thresholds
Not yet set: Acceptable error margins come from the workflow’s quality standard and the evaluator’s power calculation.

What is not yet known, and how to find out

  • Pilot and evaluation cost (cost)Next step: The service owner to price licences, integration and an evaluator for the chosen workflow.Scope: One workflow, three to six months · Data or quotes needed: Supplier pricing; evaluator quotes · Who can estimate it: The service-owning department
  • Measured effect on quality and net time (effect)Next step: The controlled comparison.Defensible route: bounded pilot

Sources last checked 4 October 2026 · Owner: Evidence editor (James Harris) · Next check: the first trigger: Publication of a controlled evaluation of AI tools in government.

What it would take

The order things have to happen in, who pays, and the effects and appraisals on record.

The order things have to happen in

  1. First, the prerequisite

    Map critical cloud and AI dependencies; set baselines for one service problem

    Owner: Organisation technology owners; service owners

    No date: the timing depends on the steps before it.

  2. Then, the decision

    CMA Board reviews provider commitments on egress fees and interoperability

    Owner: Competition and Markets Authority

    Six months after 31 March 2026 (review timing as announced) — source

  3. Then, delivery

    Run an exit-and-restore test and an evaluated AI pilot

    Owner: Technology owners; service teams; evaluators

    No date: the timing depends on the steps before it.

  4. Finally, the measured outcome

    Tested recovery times; net time, cost and quality from pilots

    Owner: Organisations; independent evaluators

    No date: the timing depends on the steps before it.

Who pays, and what is still missing

Cost, committed funding and the remaining gap for the same scope and years. A programme envelope is not an award to a project; a proposal is not committed money. None of these costs has been established yet, so no total is shown.

Cost and funding ledger

  • Migration, egress, engineering, training and any duplicated capacity

    A43 · administration · not costed

    Amount
    Not yet costed
    Who pays
    central government, employers
    Committed
    None identified
    Proposed, not committed
    —
    Gap
    Not established

  • Model and cloud use, integration, training, quality control, energy and exit

    A44 · administration · not costed

    Amount
    Not yet costed
    Who pays
    central government, councils
    Committed
    None identified
    Proposed, not committed
    —
    Gap
    Not established

“Not yet costed” means no defensible cost has been found — not that the action is free. Lifetime benefits, benefit-cost ratios and forecast savings are never shown here as money available to spend.

Effects on record, with evidence ratings (1 record)

What the evidence says each action changes

Effects on record — these measure different outcomes, so they are not drawn on one scale
ActionOutcome and areaEffectCompared withEvidence
Measure useful AI adoption before scaling itStaff workload and service quality when AI outputs need checking Public service delivery · mixedSize not established—plausible spillover; limited confidence; United KingdomThis is why pilots must count review time and errors, not only throughput.

How we would know

Firms with ten or more staff using AI; Cloud customers switching provider in a year; Public AI Research Resource compute.

The outcome measures to watch. A baseline is where we start, not a target; where there is no official target we say so.

Measured results

What each option would be judged on, and the latest value of the measures already on this page. A measure moving is not proof that an option caused it.

Where the ideas come from

Research, counter-evidence and methods (10 sources; 1 claim held back)

Counter-evidence and criticism

  • Make cloud exit and continuity work in practice: Multi-cloud is not automatically cheaper or safer; poor implementation can weaken security; provider nationality does not guarantee resilience. (Competition and Markets Authority)
  • Measure useful AI adoption before scaling it: Evaluation slows adoption that may be beneficial; generic AI tools change quickly, so trial results can date fast. (HM Treasury / Evaluation Task Force)

Claims held back

  • Delivered public compute against the 2030 target — No official update on delivered AIRR capacity was found. Needed: A DSIT or UKRI progress report on AIRR capacity.

“Full text checked” means we read the report itself; “relevant passages” means the sections behind each figure; “summary or abstract only” means we could not read the full study, so no number from it is published here without another source.

Other options compared — not on the shortlist (3)

Other options compared

Options from earlier work on this page that are not on the shortlist. They remain valid choices; their evidence is shown as before.

  • capacityStatus: funded

    Expand publicly allocated compute (AIRR)

    Aims to:
    More researchers and smaller firms able to train and test models without buying commercial cloud.
    How:
    Government-funded supercomputers allocate GPU hours by application to eligible UK users.
    Who decides:
    BIST and UKRI (programme); host universities (operation).
    Main constraint:
    Chip supply, power and cooling; hardware ages quickly.
    Evidence:
    Not evaluated — a mechanism to test, not an evidenced fix
    See the trade-offs and evidence
    What the evidence shows
    Isambard-AI has been fully running since July 2025 and access routes are open, but there is no published evaluation of what the allocations have produced, and no official update on delivered capacity against the 2030 target.
    Policy status
    Compute Roadmap commitment of up to £2bn to 2030; over £1bn for AIRR.
    Resources
    Capital for hardware; electricity and grid connections; specialist operations staff.
    Consent and decisions needed
    Spending Review allocations; local planning for sites.
    Time
    Target date 2030; individual systems take one to two years to commission.
    Cost and yield
    Up to £2bn (cash, mainly capital) for the roadmap; running costs not separately published.
    Trade-offs and who pays
    Public money competes with other research spending; public systems may lag commercial hardware.
    How we would know it worked
    Leading indicator: GPU hours allocated, and share going to UK firms. Outcome: Research outputs and products built on AIRR. It would not be working if: Capacity rises but allocations stay unused or concentrated in a few users.
  • preventionStatus: operating

    Resilience rules for critical providers

    Aims to:
    Essential services keep running if a major cloud or AI provider fails.
    How:
    Regulators can require designated providers to test scenarios, report incidents and fix weaknesses; organisations plan exits and fall-backs.
    Who decides:
    HM Treasury; Bank of England, PRA and FCA (finance); sector regulators and organisations elsewhere.
    Main constraint:
    Covers financial services; most sectors have no equivalent; providers are based abroad.
    Evidence:
    Not evaluated — a mechanism to test, not an evidenced fix
    See the trade-offs and evidence
    What the evidence shows
    HM Treasury designated Microsoft, Google Cloud, Amazon Web Services and Oracle as critical third parties to the financial system from 13 July 2026, letting the Bank of England, PRA and FCA oversee their resilience directly. The regime is new and covers finance only.
    Policy status
    First designations effective 13 July 2026 (financial services only).
    Resources
    Regulatory supervision; organisations’ own continuity planning.
    Consent and decisions needed
    Extension to other sectors would need new law or designations.
    Time
    Supervision begins now; effects show in incident handling.
    Cost and yield
    Borne by providers and regulated firms; no public costing.
    Trade-offs and who pays
    Compliance costs may be passed on; duplicating systems for resilience is expensive.
    How we would know it worked
    Leading indicator: Scenario tests completed by designated providers. Outcome: Duration and impact of outages on critical services. It would not be working if: Major outages continue to stop critical services for long periods.
  • capacityStatus: operating

    Faster grid connections for data centres

    Aims to:
    Viable UK data centres connected and powered sooner.
    How:
    NESO’s connections reform reorders the queue by readiness; network build adds capacity.
    Who decides:
    NESO and Ofgem; network companies; DESNZ.
    Main constraint:
    Physical build times; competing large loads such as industry and electric heating.
    Evidence:
    Not evaluated — a mechanism to test, not an evidenced fix
    See the trade-offs and evidence
    What the evidence shows
    The reformed pipeline retained 283 GW (8 December 2025); there is not yet evidence that connection dates for data centres have shortened.
    Resources
    Network investment recovered through bills.
    Consent and decisions needed
    Ofgem price controls; planning for new lines.
    Time
    Queue reform now; new transmission lines take most of a decade.
    Cost and yield
    Recovered through electricity bills; no AI-specific costing.
    Trade-offs and who pays
    Grid capacity given to data centres is not available to other users.
    How we would know it worked
    Leading indicator: Time from connection offer to energisation for data centres. Outcome: UK-sited compute capacity in operation. It would not be working if: Pipeline shrinks on paper but connection dates do not move.
The options side by side
OptionResourcesTimeEvidenceOutcome to watch
Expand publicly allocated compute (AIRR)Capital for hardware; electricity and grid connections; specialist operations staff.Target date 2030; individual systems take one to two years to commission.not evaluatedResearch outputs and products built on AIRR.
Resilience rules for critical providersRegulatory supervision; organisations’ own continuity planning.Supervision begins now; effects show in incident handling.not evaluatedDuration and impact of outages on critical services.
Faster grid connections for data centresNetwork investment recovered through bills.Queue reform now; new transmission lines take most of a decade.not evaluatedUK-sited compute capacity in operation.

Plan reviewed 2026-10-04. How we judge evidence.