In January last year, ministers promised to “mainline AI into the veins” of the nation. Central to that vision is using artificial intelligence to reshape the work of government itself and save tens of billions of pounds in productivity gains. Yet while ambition saturates the rhetoric, the picture on the ground is far patchier. And into the gap between political enthusiasm and operational reality has rushed a great deal of expensive technology, questionable procurement, and what one insider describes as a leadership tone of “let’s use AI” that is positive, but unfocused.
What’s happening on the ground?
The honest answer, according to both Whitehall-watchers interested in AI adoption and civil servants themselves, is: some things, in some places, unevenly. A survey of more than 2,000 civil service managers carried out by the FDA trade union last year found that nearly two in three have used AI personally. But of those, nine in 10 described using it only for basic administrative tasks like meeting management and communication support. Transformative it is not. Those based outside London, and in more junior grades, were less likely to have used AI at all.
David Black, a senior official who has spent the last few years running AI programmes in various government departments and who asked for his name to be changed for this feature, puts it bluntly. “AI is to some extent a solution looking for a problem,” he says. “A lot of it is driven by tech companies that need something to upsell.” He is particularly sceptical about the rollout of Microsoft Copilot, which was bundled into existing Office subscriptions at roughly a 6% price increase for UK government customers. “We’re lacking good evidence that the productivity benefits justify the cost,” he argues. “AI meeting summaries basically signal that nothing very important happened in that meeting. And free, off-the-shelf tools do largely the same job.”
“There is a lot of ‘AI is the future, get on with it’ from ministers who haven’t thought through what that means in practice”
David Black
But the government’s fondness for tools like Copilot is partly about data security – keeping sensitive material within controlled government environments rather than typed into a public chatbot. Black acknowledges this challenge but notes that, in practice, many civil servants simply use ChatGPT on their personal phones anyway.
Heloise Dunlop is a researcher in the Institute for Government’s civil service team who has been studying AI adoption and civil servants’ attitudes towards it. For her, the gap between what AI is doing now and what people expect it to do soon is striking.
The latest Civil Service People Survey found that only around 26% of civil servants say AI is currently saving them more than an hour a week. Yet 71% expect it to change their job significantly within five years. “There’s a mismatch there,” she tells CSW. “What do they think is going to change?”
She also points out that those departments with the biggest frontline and operational delivery roles are showing the lowest time savings from AI. Human contact and physical tasks are hard to automate, after all. “Where are you going to find the time savings if, as a work coach, for example, you’ve still got to have back-to-back, in-person appointments with clients in a job centre? Or if your role is to inspect vehicles – AI can’t do that for you.” Yet even staff in those kinds of roles still expect disruption ahead, Dunlop adds.
“So the big question for me is: is this just hype?”
Where it actually works
None of this is to say AI has no place in government. It absolutely does but, Black argues, it tends to work best in specific, bounded applications, rather than as a general-purpose productivity layer spread thinly across the whole machine. His favourite use-cases are those where AI does something humans simply cannot do at scale. Take the Department for Work and Pensions’ Whitemail tool, which scans thousands of incoming letters to flag people who may be at risk of harm. “The cost of a false positive is low,” he says. “And the benefit of catching someone in crisis is very high.” When Whitemail received negative coverage in the press, Black disagreed with the framing: in his view, it is exactly the kind of application for which AI is suited.
A second example is Extract, a planning data-extraction tool announced by the prime minister that converts planning documents – PDFs, maps, tree preservation orders, article 4 exclusion zones – into queryable data, freeing planning officers from hours of manual checking. “What I like about it is that you’re building a lasting data asset. And you can supervise the AI’s accuracy as you go,” Black says. “It’s AI as infrastructure rather than AI as shortcut, which is an important distinction.”
Justice Transcribe, developed under the Ministry of Justice’s AI Action Plan for Justice, is another application worth examining. The transcription tool replaces manual note-taking during probation meetings, allowing probation officers to focus on the person in front of them rather than their notepad or laptop. Between October 2025 and February 2026, it was used to summarise more than 150,000 probation meetings, saving over 25,000 hours of staff time. It is also a secure, internally managed system – not a commercial tool bolted on from outside.
The Government Communication Service’s Assist tool is, by common consent among those familiar with it, one of the civil service’s best AI case studies so far. Dunlop has studied it closely. “It works because it was built by a pre-existing, multidisciplinary team who understood the environment that they were in and the problems they were trying to solve. It was built specifically for their own function, with mandatory training and real attention to the hidden risks – not just hallucinations, but the human errors that creep in around any new tool,” she says. Black agrees: “They focused on adoption over speed of building. They embedded themselves with users. There were real feedback loops.”
In each of these success stories (and the number of others is growing by the week), AI is doing something people genuinely cannot – processing volume, extracting structure from unstructured data – and thereby freeing attention for tasks that require a human presence. Black likes to compare AI in government to “a really good, fast intern”: capable, quick and useful within limits. Not a replacement for judgement, not (yet) a wholesale transformation of the organisation, but valuable when given clarity about the problem it is solving.
The guidance and leadership gap
Clarity, however, is exactly what is missing in too many parts of government when it comes to AI. The FDA survey found that only 29% of civil servants have been consulted on AI at work. Many also described receiving contradictory signals. “The big first message from seniors was: ‘Do not use,’” recalled one focus group participant quoted in the survey. But when the messaging changed, the same official said, “it was not really clearly communicated”. Another respondent put it more directly: “I do not understand what my department wants to see AI being used for, or what outcomes it expects.”
Black says this is a familiar pattern. “There is a lot of ‘AI is the future, get on with it’ from ministers who haven’t thought through what that means in practice,” he says. “And beneath that, a lot of people are scrabbling around trying to find something they can point to.”
Dunlop’s view is that this reflects a structural absence rather than just poor communication. “We’re still waiting for the Strategic Workforce Plan from the civil service,” she says. “Which means there is no clarity on which skills to keep, which to build and which can be automated. You can’t sensibly deploy AI without having thought about that.” For her, initiatives like the government’s One Big Thing – an annual cross‑civil service training programme completed by hundreds of thousands of civil servants, which has most recently had AI as its focus – are fine as far as they go, but are a long way from the profession-by-profession, function-by-function thinking that meaningful transformation requires.
There is also a harder, less-discussed problem lurking beneath the surface: accountability. Generative AI, unlike more traditional algorithmic systems, does not produce outputs that are easily traceable or replicable. “Government needs to be able to explain how and why a decision was made,” Dunlop says. “Ministers and MPs have to be accountable. With generative AI, that is generally much harder to guarantee.” This is a constitutional concern, which is yet to be resolved.
Are we building this wrong?
Setting aside the guidance gap, there is a more fundamental question about whether the basic architecture of the government’s AI approach makes sense. Jack Perschke founded Great Wave AI in 2021 and has spent more than two decades in technology consulting, mostly with public sector clients.
His diagnosis of government’s current approach to AI is pointed. “It’s completely the wrong way round. Departments are going use-case by use-case, when actually what they should be doing is building shared infrastructure first and then deploying use-cases on top of it,” he says.
The consequence of the current approach, as he describes it, is a “procurement death spiral”. A department identifies a problem, writes a business case worth tens of millions of pounds, hands it to procurement, who then commission a major consultancy who, Perschke says, “frankly don’t know any more about implementing AI than the people on the ground, because the technology is so new”. The result, he argues, is that departments spend millions on procurement, burn through millions more in the first months of a contract to develop impressive-sounding proofs-of-concept – and then require a second funding-round and several years to develop one working use-case, which leaves many use-cases still to address and no budget left to meet them.
His alternative is what he calls an “agent orchestration” layer: shared infrastructure for supplying AI agents, connecting them together, evaluating their outputs, and logging everything. With that in place, deploying a new use-case is fast and relatively cheap. Without it, every use-case is a bespoke, expensive build from scratch.
Before the AI, fix the foundations
When it comes to data in the civil service, there is – in the words of one senior figure in the digital and data profession – “an absolute rat’s nest of practice across government”. Computer systems that are decades old sit alongside newer platforms that can’t talk to each other; data is siloed, poorly documented, and of wildly variable quality, and nobody has yet hit on the magic formula to align all these processes and technologies.
This is, according to Ruth Kelly, chief analyst at the National Audit Office, the single biggest barrier to AI adoption in government. “If you’ve got to manually extract data and you’re just building an AI on top, it can be very problematic,” she told CSW last year. The NAO’s core message, she added, is that investment in fixing legacy systems is more important than building AI applications on top of them.
Professor Sir Anthony Finkelstein, former chief scientific adviser for national security, has made the same point. Writing in the Heywood Quarterly earlier this year, he described how AI pilots can perform perfectly well in testing, only to hit a wall when departments discover that their data is spread across multiple systems with no authoritative source of truth, and their workflows are buried in legacy code. “To deploy the AI tool in live operations would have required changes across several interdependent systems,” Finkelstein wrote of one pilot. “No one could state with confidence what those changes would break, how long they would take or who ultimately owned the risk… AI revealed the limits of the organisation’s systems far more clearly than it delivered immediate benefits.”
There is, however, a real opportunity here. Fixing these data foundations is not just a precondition for transformative AI – the ambition of what AI could deliver is itself the best available argument for finally doing it.
Black broadly agrees with Perschke’s critique of big consultancy contracts. “The best ideas tend to come from grassroots staff,” he says. “AI is genuinely democratising in that sense.” But he adds another dimension: the UK has no shared framework for negotiating with commercial AI providers. “Every department has to do its own contracts, its own information governance carve-outs,” he says. “It’s wasteful duplication, and it leaves everyone exposed.”
Although they are coming at it from different angles, both Black and Perschke express the same concern: the UK’s growing dependence on large American AI providers, and what happens if that relationship shifts. Perschke flags the MoJ’s memorandum of understanding with OpenAI as an example of exactly the wrong approach. “The question isn’t just where the data sits,” he says. “It’s: can you change it? Can you replace it? What happens if the price spikes? What happens if geopolitics shift?”
His proposed solution is an orchestration layer that lets departments swap language models in and out based on cost and performance. Black’s is more fundamental: he points to France, which has its own sovereign AI model, Mistral, and a national government AI system, Albert, integrated with public sector IT. “The UK has shown very little interest in doing anything equivalent,” he says. “And that’s a real missed opportunity.”
What’s that coming over the hill?
Almost all of the debate about AI in government – this feature included, until now – is about the supply side: how civil servants use AI internally to work more efficiently and better serve the public. Black, however, thinks this framing is missing something important.
“Most of the conversation is about optimising existing government services,” he says. “But there’s a much bigger problem coming on the demand side, and almost nobody in Whitehall is talking about it.” His argument is as follows: government budgets have always, quietly, relied on a significant degree of citizen ignorance or inertia. Many people who are entitled to benefits don’t claim them. Many who have grounds to challenge a decision don’t. Many who could access a service don’t know it exists. AI could be about to change that – not because government will tell people what they’re missing out on, but because private companies will.
“Government needs to be able to explain how and why a decision was made. With generative AI, that is generally much harder”
Heloise Dunlop
He points to parking ticket dispute websites as an early version of this which is already happening: tools that identify grounds for appeal and generate the relevant correspondence automatically, at minimal cost to the user. “What’s potentially coming,” he says, “is the same thing for the entire benefits system. A startup will offer to get you every entitlement you’re eligible for, for a small cut. And government will suddenly be facing demand it never anticipated and never budgeted for.”
The same dynamic, he notes, applies to public consultations. Automation makes it easier for government to process responses at scale – but it also makes it incredibly easy for individuals and organisations to mass-generate them. The integrity of the consultation process, already under strain, could become harder to defend.
Black’s conclusion is stark: “Government needs to get ahead of this by proactively offering these kinds of services itself – tools that help citizens understand what they’re entitled to, rather than waiting for private intermediaries to do it first.” If it doesn’t, he argues, it will find itself in a reactive battle against commercial operators, while being simultaneously overwhelmed by a surge in demand it never saw coming.
It is, perhaps, the least comfortable argument to come out of all the conversations that have informed this feature. A civil service focused on using AI to save minutes per day may be preparing for entirely the wrong future.
What should government do?
A thread common to all these discussions is that AI isn’t wrong for government, but that the conditions for using it well are not yet in place. Those conditions, according to the experts and insiders CSW spoke to, include profession-by-profession skills planning with mandatory training for specific tools; a shared infrastructure put in place at the department level before any individual use cases are dreamed up; and a procurement process fit for a technology that moves faster than any tendering round can keep up with.
Finally, they call for leaders to get serious about fixing government’s data and legacy tech systems (see box), without which AI’s ability to be truly transformative will be severely hampered.
The technology is currently, as Black put it earlier, “a really good, fast intern”. If – at the risk of stretching the metaphor – it is going to be offered a permanent and prominent role in the UK civil service, it will need clear direction and leaders who understand the difference between enthusiasm and strategy.