AI and the UK state: what the numbers really show

This summer a chart did the rounds in British policy circles. It showed freedom of information requests to central government sitting more or less flat for ten years and then climbing steeply after November 2022, the month ChatGPT went public. The suggestion was obvious. People had found a machine that writes the letters for them, and Whitehall's postbag had filled up as a result.

Before writing a word of this I downloaded the Cabinet Office's own spreadsheet and added the columns up myself. On the face of it the chart is right. In 2022 the monitored bodies got 52,740 requests. In 2025 they got 94,526, and the bulletin says no year since 2005 has been higher. Go back to 2010 and you find about 44,000, and nothing much happened for the next eleven years.

Then I took one body out of the sum, and the picture changed. The National Archives had just under 8,000 requests in 2022 and 33,634 last year. The bulletin is quite open about why. Old Ministry of Defence service records have been moved there in bulk, and the requests came with them. Take the Archives away and everybody else in central government was getting roughly what they got before the pandemic, right up to the end of 2024. Only in 2025 does the line lift, to 60,892. That is a rise of a fifth in one year. It matters, but it is nothing like the near-doubling the chart suggests.

FOI requests to monitored central government bodies, 2010 to 2025

Show the data

Year	All monitored bodies	The National Archives	Excluding The National Archives
2010	43,921	3,194	40,727
2011	47,141	2,906	44,235
2012	49,464	2,856	46,608
2013	51,696	3,340	48,356
2014	46,806	3,236	43,570
2015	47,386	3,085	44,301
2016	45,415	2,922	42,493
2017	46,681	3,024	43,657
2018	49,961	3,395	46,566
2019	49,439	3,832	45,607
2020	44,195	2,243	41,952
2021	51,507	3,391	48,116
2022	52,740	7,983	44,757
2023	70,475	21,366	49,109
2024	83,041	32,575	50,466
2025	94,526	33,634	60,892

So most of that curve is military paperwork. Whether the smaller 2025 rise is AI, I genuinely cannot say. The Information Commissioner seems to think it plays a part, because in May 2026 it put out guidance for public bodies on AI-generated requests and warned of "higher volumes of requests". There was no figure attached, and I have looked for one.

Why open with a correction? Because I keep seeing AI offered as the cause every time a public service number moves, and I think we should be more careful than that. Sometimes it really is the cause. Often it sits alongside two or three other things. A department that plans for the wrong cause ends up paying to fix something that was never broken.

What the complaints data shows

The same video went on to complaints, and here the evidence is stronger, though it still needs reading with care.

Some figures check out exactly. Complaints and enquiries to the Local Government and Social Care Ombudsman went from 15,488 in 2022-23 to 27,625 in 2025-26, up 78%. Single claims to employment tribunals rose from 31,251 to 50,704 over the same years, which is the 62% quoted. Complaints to CCW, the water consumer body, more than doubled.

Look at what sits behind each one, though. The CCW figure counts complaints that escalate to the watchdog. Complaints made to the water companies themselves rose 56% in 2025-26, in a year when CCW says the average bill went up by 26%. On tribunals, claims were 39,615 in 2019-20, so against a pre-pandemic year the rise is 28%. The 2022-23 low was a dip, not a baseline. I would also remember that tribunal claims collapsed to around 17,000 a year while fees were charged, and recovered once the Supreme Court struck the fees down in 2017. Friction changes behaviour, and that lesson comes back later.

Rise in complaints and claims, 2022-23 to 2025-26

Show the data

Measure	2022-23	2025-26	Change
CCW complaints	6,197	15,115	+144%
DWP Universal Credit complaints	7,315	15,345	+110%
LGSCO complaints and enquiries	15,488	27,625	+78%
Employment tribunal single claims	31,251	50,704	+62%
All DWP complaints	22,060	31,610	+43%
GP practice written complaints	104,010	143,364	+38%
All NHS written complaints	229,458	305,264	+33%

Two of the video's numbers only work from a pandemic year. NHS England's tables show written complaints to GP practices at 72,087 in 2020-21 and 143,364 in 2025-26. Measured from 2022-23 the rise is 38%, which is still large, but it is well short of the figure that went round. At the Department for Work and Pensions I make the increase 43% from 2022-23, not 64%. Inside that total, though, complaints about Universal Credit more than doubled, from 7,315 to 15,345.

None of these publications puts the rise down to AI. The clearest statement I found comes from the Parliamentary and Health Service Ombudsman, whose latest annual report notes that tools like ChatGPT and Copilot "help complainants create and submit complaints", and that the results can be harder to follow.

Friction, fees and the Financial Ombudsman

The Financial Ombudsman Service is the most useful case I have found, because it measured something and then acted on it.

In March 2026 the service said that, in a small sample, up to a third of the replies it received to its initial assessments looked AI-generated. Some ran to more than 200 pages in answer to a six-page provisional decision. Note that this is a third of replies at one stage of a case, not a third of all complaints, which is how it has sometimes been repeated.

The other thing it did had nothing to do with AI. Since April 2025 claims firms and other paid representatives get ten free cases a year and then pay £250 for each one after that (£75 if they win). It showed up in the very first year. The service took in 305,700 complaints in 2024-25 and 214,600 the year after, and the share brought by representatives dropped from about half to 18%. The ombudsman's own report says the drop followed the new charge. Tribunals taught the same lesson a decade ago: put a price on the door and fewer people walk through it.

Financial Ombudsman Service: complaints before and after case fees for representatives

Show the data

Measure	2024-25	2025-26
Complaints received	305,700	214,600
Share brought by professional representatives	About 50%	18%

Tom Loosemore, one of the founders of the Government Digital Service, has been arguing for most of this year that this lever is about to stop working. In a February blog post he wrote that "many public services rely on friction to stay viable", and that he did not think governments were "remotely ready" for a world where software does the form-filling and the chasing for people. An academic paper presented this month, by Schmitz, Hammond and Chan, collects 84 cases of what they call "agentic flooding" across eleven jurisdictions.

The harder question is what the extra paperwork actually is. If it is mostly noise, a charge or a filter makes sense. If a good share of it is people who always had a case but could never write the letter, then a filter simply puts them back where they were. The DWP numbers point both ways. Complaints that went straight to the Independent Case Examiner before finishing the department's own process rose from 463 to 7,579, which looks like misdirected effort. The doubling of Universal Credit complaints is harder to wave away, and I would want to see how many were upheld before calling it noise.

How much AI government is really using

That is the pressure from outside. Inside government the picture is slower and patchier than the speeches suggest.

The best independent count is still the National Audit Office survey published in March 2024. Of the 87 government bodies that answered, 37% had deployed at least one AI tool, usually one or two. Across all of them that came to 74 live use cases. Seventy per cent were piloting or planning something, with a median of four ideas each. Only 21% had an AI strategy. And 70% said a lack of skills was holding them back.

I would like a newer figure, and I have looked. The NAO has since published guidance for audit committees and, in July this year, a report on workforce planning, but not a fresh census of AI use. The government's own State of Digital Government Review, from January 2025, adds one number I find more telling than any adoption rate: only 8% of AI projects could show measurable benefits.

AI in government bodies, NAO survey of 87 bodies (2024)

Show the data

Measure	Share of 87 bodies
Had deployed AI	37%
Piloting or planning AI	70%
Had an AI strategy	21%
Said skills were a barrier	70%

My own reading is that pilots are cheap and the government has lots of them. Getting a tool out of a pilot and onto the desks of people who process claims or answer letters all day is a much bigger job. You need data that is clean enough to trust, a security team that will sign it off, and training for thousands of staff. The NAO names skills, poor access to good data, and old systems as the main barriers, and I come back to the last two below.

The staff themselves are not waiting to be told. The NAO's July 2026 report quotes the 2025 Civil Service People Survey, where 76% said they expect AI to change the way they work within five years. So three in four civil servants think the change is coming, while fewer than one AI project in ten can yet show what it delivered. I would worry about that mismatch if I ran a department, because it is the kind that produces rushed rollouts.

Old systems and missing data

Whatever AI tool a department buys, it has to connect to the systems already there. In Whitehall a lot of those are old, and getting older.

The State of Digital Government Review asked departments how much of their estate they class as legacy, and the answer averaged 28% in 2024. A year earlier it was 26%, so the pile is growing, not shrinking. Some departments are at around 10% and others near 60%, which tells you the national average is not much use to anyone running one of the worse ones. The Public Accounts Committee went further in March 2025. Of the 72 systems the government itself rates as the riskiest, 21 had no funding in place to replace or repair them.

The data is the next problem down. Asked whether their data set-up gave them a complete picture of their own operations, only 27% of the review's respondents said yes. And 47% of central government services are not digital from end to end. Somewhere along the way there is still a paper form, a scanned letter or somebody on the phone typing notes, and an AI model can do very little with a step like that.

Sources: DSIT, State of digital government review (January 2025); Public Accounts Committee, Use of AI in Government (HC 356, March 2025).

Public sector technology spending, 2023 (about £26 billion)

Show the data

Part of spend	Amount
Contractors and consultants	About £14.5bn (55%)
Permanent staff	About £5bn (under 20%)
Remainder (derived)	About £6.5bn
Total	About £26bn

The money tells a similar story. The public sector spent about £26 billion on technology in 2023, which the review reckons is around 30% below the benchmark it uses. Of that, roughly £14.5 billion, 55%, went to contractors and consultants. Less than a fifth, about £5 billion, paid for permanent staff. I understand why it ends up that way, since contractors are easier to hire than civil servants on public pay scales. But it means a lot of what the state knows about its own systems walks out of the door when a contract ends.

My view is that this is the least glamorous part of the AI debate and the most decisive one. A department with clean, joined-up data and a modern case system can adopt a new tool in months. A department running a decades-old case system, with records split across several databases, will spend most of its AI budget on plumbing before anything useful reaches the public, and that spending rarely makes it into a press release.

The £45 billion question

The number everyone quotes is £45 billion. It comes from the January 2025 review, which put unrealised savings and productivity gains across the public sector at over £45 billion a year, or 4 to 7% of public spending. The companion blueprint gave a range of £45 to £87 billion. By July this year the NAO was describing £45 billion as the efficiency the government expects from digital and AI. That is a target the government set for itself, not an estimate an auditor has tested.

The academic work on what could be automated is worth reading in full. In 2024 researchers at the Alan Turing Institute went through 377 services across 57 organisations, which between them handle about a billion transactions a year. They picked out roughly 143 million that are complicated but follow a pattern, such as checking an application against a set of rules, and reckoned 84% of those could largely be automated. Their own illustration is that saving a single minute on each would be worth about 1,200 person-years every year. I think it is a careful paper, but it is a model of what might be possible. Nobody should read it as a record of what has happened.

What has actually been measured is more modest. Around 20,000 civil servants used Microsoft 365 Copilot between September and December 2024, and the government's report said they saved an average of 26 minutes a day. The figure was self-reported. When the Department for Business and Trade ran its own evaluation it "did not find evidence that time savings have led to improved productivity". HMRC's trial, published in July 2026, put the saving at about an hour a week, again by staff estimate.

The strongest evidence I found is narrower and more interesting. The government's Consult tool, part of the Humphrey suite built in-house, analysed more than 50,000 responses to the Independent Water Commission's consultation in about two hours, for £240, followed by 22 hours of checking by experts. That is a real, specific task done faster and cheaper, with a human still signing off the result. If I were setting targets, I would build them from cases like that one and work up, rather than from a national figure and work down.

Sources: DSIT blueprint; GDS Copilot findings; HMRC Copilot evaluation; DBT Copilot evaluation; DSIT on Consult.

Skills and people

If old systems are the first constraint, people are the second, and I suspect the second is harder to fix.

The government has made progress. Digital and data specialists were about 3% of the central government workforce in 2021 and about 6% by the time of the January 2025 review. The big operational departments, the ones that run benefits, tax and borders, averaged 5%. The review's own benchmarks were 6% for central governments elsewhere and 8 to 12% in regulated private-sector industries.

Hiring is where it stalls. The Public Accounts Committee reported that around half of the roles advertised in digital and data recruitment campaigns in 2024 went unfilled. For technical architects it found a pay gap of 35% against the market, roughly £30,000 a year. Anyone who has tried to recruit a senior developer knows how quickly a good candidate disappears when the offer is that far below the market.

Digital and data staff as a share of the workforce

Show the data

Group	Share
Central government, 2021	About 3%
Central government, 2025	About 6%
Large operational departments	5%
Benchmark: other central governments	6%
Benchmark: regulated private sector	8 to 12%
Target for 2030	10%

The digital roadmap published in January 2026 aims for one civil servant in ten working in digital and technology roles by 2030. It is a sensible ambition. Its own quarterly roadmap report for April to June 2026, though, says the cross-government digital workforce strategy "is not expected to be published to the previous timetable". That is a polite way of saying the plan for meeting the target is late.

The skills gap I worry about most gets less attention. A department does not only need engineers. It needs caseworkers and team leaders who understand what a model is doing well enough to spot when it is wrong, and managers who can judge a supplier's claims. The NAO's 70% skills figure from 2024 probably covers both kinds, although I could not find a breakdown. Training thousands of front-line staff to supervise a tool is slower and less visible than hiring a hundred data scientists, and it is the part I expect to be underfunded.

Trust and transparency

The last constraint is permission, in the broad sense. A state that uses AI to make or shape decisions about people needs those people to accept it, and the evidence on that is mixed.

Start with what the public can see. The UK has an Algorithmic Transparency Recording Standard, a template departments fill in to explain each algorithmic tool they use. When the Public Accounts Committee looked in early 2025, 33 records had been published, across the whole of government. The Treasury's response that summer aimed to publish records for every in-scope departmental tool by the end of 2025. I have not been able to confirm whether that happened.

Then what people think. The Ada Lovelace Institute and the Turing ran a large survey in 2024-25. Only 18% of respondents knew that AI is used to help assess eligibility for welfare. Concern about that use rose from 44% in the previous wave to 59%. And 72% said laws and regulation would make them more comfortable with AI, up from 62%. The government's own tracker is a little warmer: in its 2025-26 public engagement survey, published in July, 60% said they trust government departments a great deal or a fair amount to keep their data safe, although only 12% chose "a great deal".

Public attitudes to AI in public services

Show the data

Measure	Previous wave	Latest
Concerned about AI assessing welfare eligibility	44%	59%
Laws and regulation would increase comfort	62%	72%
Aware AI is used for welfare eligibility	n/a	18%
Trust government departments to keep data safe	n/a	60%

I read these numbers as conditional consent. People are not hostile to the state using AI, but they do not know where it is being used, their worry about its use in benefits is growing, and they want rules. That is a fragile base. One badly handled case, where a person loses money because of a model nobody can explain, could move those figures faster than any amount of good practice.

It also cuts the other way. In June 2025 the High Court dealt with a case, Ayinde v Haringey, where a claimant's legal team put five fictitious authorities before the court. The judges said public AI tools are "not capable of conducting reliable legal research". If the state is going to be judged on how it uses AI, it will also have to deal with citizens, and their advisers, who use it badly.

Three readings of the evidence

So what is really going on? I can see three serious readings of the evidence, and I do not think any one of them is enough on its own.

The first is the flood. Writing a complaint, an appeal or an FOI request used to take an evening; now it takes a minute, so more get written, and some of them are long, repetitive and hard to answer. The Financial Ombudsman's 200-page replies and the Parliamentary Ombudsman's comments on AI-drafted complaints fit this reading. On this view the answer is a filter of some kind.

The second is unmet need. Plenty of people had a fair grievance but lacked the time, the confidence or the English to put it in writing, and a chatbot has handed them a clerk. The doubling of Universal Credit complaints and the steady rise in council complaints are consistent with this as well. On this view a filter is the wrong answer, because it shuts out the people the system most needs to hear.

The third is the least exciting and gets the least attention: a lot of this has nothing to do with AI. Water bills went up by about a quarter in one year. Service records moved to Kew. Tribunal claims climbed back from an unusually low year. When I strip those out, the AI signal is still there in places, but it is smaller than the viral version and much harder to measure.

Where does that leave a department? Take a benefits office that suddenly gets far more letters than it can read. The quick fix is the Financial Ombudsman's: put a cost or an extra step in the way. The numbers fell there, at least in year one. But some good cases get lost with the bad ones, and Loosemore's point is that software does not get tired of extra steps the way people do.

The slower fix is to need fewer letters in the first place. If the office already holds a person's records, and can see that a payment was missed or a deduction was wrong, it can put that right before anyone has to write in. That is how I would want to see AI used in a benefits system, as much as for reading the post. It is also exactly the thing the earlier sections say is hard: it needs systems that talk to each other, data someone trusts, staff who can check the result, and a public that is willing to let the state use what it knows.

What I will be watching

I will keep an eye on a handful of things over the next twelve months. Any of them could change my mind.

Top of the list is a count. The Information Commissioner and the Parliamentary Ombudsman have both said they are seeing AI-drafted correspondence. Apart from the Financial Ombudsman's small sample, though, I have not found anybody who publishes a number. If an ombudsman told us roughly how many cases looked machine-written last quarter, and how many of those were upheld, we would learn more than any chart on social media can show.

Next is the DWP complaints series, and in particular whether the share of Universal Credit complaints that are upheld holds up as the volume grows. If it does, the unmet need reading gets much stronger. Then there is the digital workforce strategy that has now slipped. A date and a funded plan for the one-in-ten target would tell me the skills problem is being taken seriously. I would also like a fresh National Audit Office count of AI in government, because the 2024 survey is getting old and is still the best figure anyone has. And finally the transparency register, and whether it now covers the tools that actually touch people's money.

I do not think the British state is about to be broken by chatbots, and I do not think digital and AI will deliver £45 billion a year on the timetable ministers would like. My honest reading of the evidence is narrower than both. The paperwork is rising, partly because of AI and partly for older reasons, and the only lasting answer runs through the same unglamorous work on systems, data and people that departments have been putting off for years.

How I checked these figures: every number in this essay comes from the publication linked where it appears and listed in full below. Where a figure in wide circulation did not match its source, I have used the source and said so in the text. Percentage changes are my own calculations from the published counts.

Common questions

Sources

1. Cabinet Office (29 April 2026). Freedom of Information statistics: annual 2025 (bulletin and published data).
2. Information Commissioner's Office (6 May 2026). New guidance to support public authorities dealing with AI-generated FOI requests.
3. Local Government and Social Care Ombudsman (2026). Review of Local Government Complaints 2025-26, with data sheets.
4. Ministry of Justice (10 September 2026). Tribunal Statistics Quarterly: April to June 2026, main tables (employment tribunal single claims, table S_1).
5. UK Supreme Court (26 July 2017). R (on the application of UNISON) v Lord Chancellor [2017] UKSC 51.
6. CCW (Consumer Council for Water) (16 September 2026). Household Complaints Report 2025-26.
7. NHS England (15 October 2026). Data on Written Complaints in the NHS, 2025-26 (Tables 1 and 5).
8. Department for Work and Pensions (30 June 2026). DWP complaints statistics to March 2026 (official statistics in development).
9. Parliamentary and Health Service Ombudsman (2026). The Ombudsman's Annual Report and Accounts 2025 to 2026.
10. Financial Ombudsman Service (Marc Harris) (27 March 2026). Embracing AI's transformational impact on consumer complaints.
11. Financial Ombudsman Service (1 April 2025). Charges for professional representatives to refer cases to the Financial Ombudsman Service come into force today.
12. Financial Ombudsman Service (21 May 2026). Annual complaints data and insight 2025/26.
13. Loosemore, T. (19 February 2026). AI and the end of friction as a policy lever (blog post).
14. Schmitz, Hammond and Chan (October 2026). Characterizing Agentic Flooding of Government Services, Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society.
15. National Audit Office (15 March 2024). Use of artificial intelligence in government, HC 612, Session 2023-24.
16. National Audit Office (15 July 2026). Government workforce planning: lessons learned, HC 267, Session 2026-27 (cites the 2025 Civil Service People Survey).
17. Department for Science, Innovation and Technology (21 January 2025). State of digital government review, CP 1251.
18. Department for Science, Innovation and Technology (21 January 2025). A blueprint for modern digital government.
19. House of Commons Public Accounts Committee (26 March 2025). Use of AI in Government, Eighteenth Report of Session 2024-25, HC 356.
20. HM Treasury (June 2025). Treasury Minutes: Government response to the Committee of Public Accounts, CP 1341.
21. Straub, V. and others, Alan Turing Institute (March 2024). AI for bureaucratic productivity: measuring the potential of AI to help automate 143 million UK government transactions.
22. Government Digital Service (2 June 2025). Microsoft 365 Copilot experiment: cross-government findings report.
23. Department for Business and Trade (August 2025). Microsoft 365 Copilot evaluation.
24. HM Revenue and Customs (9 July 2026). Evaluating the impact of Microsoft Copilot in HMRC (phase 3 trial).
25. Department for Science, Innovation and Technology (16 October 2025). Ground-breaking use of AI saves taxpayers' money and delivers greater government efficiency (Consult and the Independent Water Commission).
26. Government Digital Service (20 January 2026). Our roadmap for modern digital government (blog post).
27. Government Digital Service (2026). Roadmap for modern digital government: quarterly report, April to June 2026.
28. Ada Lovelace Institute and Alan Turing Institute (2025). How do people feel about AI? 2025 findings.
29. Department for Science, Innovation and Technology (16 July 2026). DSIT Public Engagement Survey 2025/26.
30. High Court of Justice, King's Bench Division (Divisional Court) (6 June 2025). Ayinde v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin).

← More on governance & institutions  ·  All writing