A plan from the top, and a week of evidence from below
Whitehall’s senior officials, most of them permanent secretaries, say government is coming out of the “foothills” of AI and needs a different way of organising itself to get more from it. That is the finding of The new AI operating model for government, a Global Government Forum (GGF) study led by former cabinet secretary Lord Gus O’Donnell and built on interviews with 13 permanent secretaries and other senior officials working on AI. Our news report covers the launch and the four priorities. This piece reads those priorities against three stories from the past week: the Public Accounts Committee chair’s annual report, complaints from Scottish Government staff about AI-drafted parliamentary questions, and the accuracy record of the Home Office’s asylum summarisation tool.
Read together, they suggest the four priorities are the right questions, and show how much of each is still unanswered.
Critical Context: This analysis draws on GGF’s published summary of the study and its report page. We were unable to access the full report text, so where we say the summary does not address something, that is a statement about the summary, not about the report.
What the permanent secretaries actually asked for
The study argues that from here, government AI should be judged on whether it transforms services, gives value for money and improves outcomes, not on how widely the technology has been taken up. One interviewee described the shift as moving towards “establishing the tooling to scale AI rather than just playing with it”.
The four areas the interviews produced are, in the report page’s own words:
- “Being clear about the ambition for AI, the expected outcomes and how this will be measured”
- “Finding better ways to work across government, sharing what works while preserving the flexibility to innovate”
- “Preparing for how AI could change the way people work, including roles, skills and career pathways”
- “Building the trust, accountability and approach to risk needed to move forward at pace”
Two details in GGF’s summary matter for what follows. First, it says deeper change to services and outcomes will not arrive of its own accord, so leaders have to choose where AI would do most good. Second, on trust, several participants warned that “the public may be less forgiving of mistakes made by AI than of human errors”.
The critical numbers
| Figure | Source | What it bears on |
|---|---|---|
| 13 interviewees | GGF | Permanent secretaries plus other senior officials in UK government departments |
| 4 priority areas | GGF | Ambition and measurement, joining up, workforce, trust and risk |
| 6% of the civil service in digital and data jobs | PAC, via The Register | Against a government target of 10% |
| 203 summaries spot-checked | Home Office documents, via Computer Weekly | From roughly 70,000 cases handled from April to July |
| 22% failed | Home Office documents, via Computer Weekly | 45 of the 203 summaries at the first stage of review |
| 23 minutes | Home Office ministers, via Computer Weekly | The time ministers say the tool saves per case |
Ambition and measurement: what does the asylum tool count?
The study asks leaders to find better ways to measure and demonstrate what AI is achieving. The Home Office asylum case shows why the choice of measure matters.
According to Computer Weekly, ministers say the Asylum Case Summarisation tool, introduced in April 2026, saves each caseworker 23 minutes per case. That is the benefit, and it is precisely stated. The quality side is measured far more thinly. Home Office specialists checked 203 summaries from roughly 70,000 cases handled between April and July, which Computer Weekly puts at 0.3%. Of those, 45 (22%) failed the first stage of review. And because relevance is not among the failure criteria, a summary missing material facts can still pass provided the content it does include is accurate.
Strategic Reality: A time saving is easy to count on every case. Accuracy, on this evidence, has been counted on a fraction of them, against criteria that do not catch omissions. If demonstrating impact means the first number without the second, the measurement priority will reward speed and leave quality to dip samples.
The study’s own framing allows for this. It credits wide adoption with efficiency and productivity gains, and treats better services as a separate goal that needs deliberate choices. The asylum tool is a live case where the efficiency figure exists and an outcome figure, such as whether decisions are as sound as before, does not appear in Computer Weekly’s reporting.
Trust and risk appetite: tested in public
The interviewees wanted a clear risk appetite, clear lines of accountability and more openness about how government uses AI. The asylum tool is an example of what an answer to those questions looks like in practice, and of the gaps.
Computer Weekly reports that ministers have confirmed applicants are not informed when AI has played a part in assessing their claims. The results of the dip testing have been “mainly discussed verbally” in a working group, an interim spot-testing report has reached an internal board but has not been published, and the department has told MPs no independent audit is planned. The tool went to all caseworkers even though the pilot evaluation found 9% of its summaries were inaccurate or incomplete and were taken out.
The Home Office’s position is that its AI tools support caseworkers and “do not make decisions on claims”, and that it monitors their performance alongside established quality-assurance processes. That is a statement of risk appetite. What the study’s trust priority would add is making it explicit and checkable: what error rate is acceptable, who is accountable when a summary is wrong, and what the people affected are told.
Warning ⚠️: The study’s warning that the public may judge AI errors more harshly than human ones applies with most force where the stakes are highest. Open Rights Group calls an asylum decision “one of the most serious and life-changing decisions that the state can make”, and telling applicants that AI was used is part of the transparency the interviewees asked for.
The Home Office was represented among the 13 interviewees by Christophe Prince, a director for data and technology policy. The study does not, in GGF’s summary, discuss individual departmental tools, so we cannot say how the interviews treated cases like this one.
Workforce: a plan without the specialists
The study argues that preparing the civil service for AI goes beyond building skills: leaders need to think about how AI could reshape jobs, professions and careers, and what follows for hiring and workforce planning.
The PAC chair’s report, reported the day before the study launched, describes the base that planning starts from. Sir Geoffrey Clifton-Brown writes that “AI is approaching government like a tsunami”, and that he remains sceptical government is prepared for it. According to The Register, just 6% of the civil service is in digital and data jobs against the government’s 10% target, with a particular shortage of senior specialists, and ministers have not said how or when they will close the skills gap.
Resource Reality: The study’s workforce priority is about redesigning work around AI. The PAC’s evidence is that government has fewer digital and data specialists than its own target calls for. Both are true, and the second constrains how fast the first can happen.
The PAC also points to legacy IT, quoting its report’s warning of “a worrying lack of urgency” over long-standing systems that mean government “cannot reap the benefits of new technologies”. Lord O’Donnell’s foreword makes a related point from the other direction: gains seldom appear unless workflows and organisations are redesigned and staff retrained.
Joining up: during a reorganisation
The study sees scope for more direction from the centre where common approaches would cut duplication, with flexibility where departments genuinely differ. It notes that leaders frequently have little view of activity beyond their own organisations, whilst departments invest separately in tools for needs many of them share.
That ambition meets a machinery-of-government change. The Register reports that Clifton-Brown fears setting up the Department for Business, Innovation, Science and Trade (DBIST) will make things worse. One of the study’s interviewees, Cat Little, is listed as the Cabinet Office’s permanent secretary and the incoming head of DBIST. GGF’s summary also notes that Andy Burnham, as prime minister, has set up an AI taskforce and moved the AI minister into Cabinet. Who sets the common approaches the study wants, and from which department, is a question the reorganisation makes more pressing.
Who is affected, and what the summary leaves out
The study is about how the UK government itself uses AI from here. The Scottish case is a reminder that AI also arrives from outside.
One Scottish Government staff member, posting on Saltire, a civil service message board, wrote: “In many respects, we’re now working against technology rather than benefiting from it.” The same post described questions that look AI-generated, “with MSPs lodging 20 or more at a time”. Another, in comments The Scotsman obtained through an FOI request, wanted senior leaders to fund AI tools able to search the system and draft briefing notes.
None of the 13 interviewees is from the Scottish Government, and the study is about the UK government, so this is not a gap in its sample. But the pattern is relevant to Whitehall too. GGF’s summary of the study does not mention parliamentary questions or freedom of information requests. If the public, campaigners and parliamentarians use AI to generate more correspondence, an operating model built only around government’s own adoption will be planning for one side of the workload.
Strategic Insight: At least one of the Scottish staff is asking for more AI, not less: AI to answer the demand that AI is helping to create. That is a version of the study’s ambition priority: deciding deliberately where AI makes the biggest difference, rather than letting the pressure decide.
| Group | How the four priorities bear on them |
|---|---|
| People subject to AI-assisted decisions | The trust and transparency priority determines whether they are told and what checks exist |
| Digital and data specialists | The workforce priority depends on a specialist base the PAC says is below target |
| Departmental leaders | The joining-up priority could mean more central direction on shared tools |
| Suppliers | The study’s knowledge partners see cloud infrastructure as the foundation for scaling |
The partners in the room
Readers should weigh who supported the work. GGF’s knowledge partners on the study were Cognizant and Amazon Web Services. Their joint foreword argues that real transformation depends on building AI into operating models, skills and “secure cloud-based infrastructure”. The launch took place at Public Service Data.AI, a conference GGF co-hosts with the UK government. None of this invalidates the interviews, which record what permanent secretaries said. It does mean the foreword’s account of what scaling requires comes from firms describing, in their words, “our own work with government”.
Lord O’Donnell strikes a more cautious note. His foreword says transformations of this kind take time to show an impact, and he said he hoped the report could help governments in “this daunting endeavour”.
What remains unresolved
Who owns the risk appetite?
The interviewees asked for clarity on risk appetite and accountability. GGF’s summary does not say who should set it. The asylum tool shows a department setting its own, without an independent audit, and without publishing its testing.
What counts as impact?
The ambition priority asks for better measurement. The figure ministers have given for the asylum tool is time saved; its accuracy figures came out through documents MPs and a campaign group obtained. Until outcome measures sit alongside efficiency ones, demonstrating impact can be met by the easier number.
Can the centre direct during a reshuffle?
The study wants more central direction on shared approaches. The PAC chair fears the creation of a new department will make things worse. Both can be true for some time.
What to watch
- The Home Office’s spot-testing report. An interim report has gone to an internal board but has not been published. Publication, with the error rates and the criteria used, would test the trust priority in a high-stakes case. Continued non-publication would suggest the transparency the interviewees asked for is not yet policy.
- The digital and data share of the civil service. The PAC puts it at 6% against a 10% target. Movement towards the target would show the workforce priority has a base to build on; a static figure would mean role redesign proceeds with the same thin specialist cover.
- The new DBIST. Watch where responsibility for shared AI tools and standards lands once it is formed, and whether that produces the central direction the study wants or the disruption the PAC chair fears.
- The Scottish Government’s change plan, which runs to 2031. It commits the organisation to “investing in skills and technology”. Whether that includes tools to handle AI-generated correspondence will show whether the demand side is being planned for anywhere.
The study gives Whitehall’s leaders a vocabulary for the next phase. This week’s evidence gives them the test cases. The asylum tool in particular is an early test of whether trust and pace can be held together, or whether pace is winning by default.
Source: GGF study with UK permanent secretaries outlines priorities for a new AI operating model for government (Global Government Forum, 15 October 2026) and the report page for The new AI operating model for government by Sarah Wray, GGF’s research editor. Comparison reporting from The Register on the Public Accounts Committee chair’s annual report, The Scotsman on Scottish Government staff, and Computer Weekly on the Home Office’s asylum summarisation tool.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether the Home Office publishes its spot-testing results, a public test of the trust the permanent secretaries say they need. Read more analysis at Insights, or get in touch.