Schools, councils and other public bodies have been telling the BBC that the complaint which used to fit on one side of A4 now arrives at 19 pages, or 27, dressed with case law and lines lifted from Acts of Parliament, some of them misquoted. Those lengths are what interviewees are describing, not a measured average. Somebody still has to read every one.
If you run operations, legal, HR or customer work, do not file this under government being government. An insurance claims file, an employee grievance, a supplier dispute: the same thing happens the moment drafting is nearly free and judgment is not. The letter gets longer. The citations get braver. The decision still sits with a person who is paid to be careful.
Deborah Evans, chief executive of Lawyers in Local Government, which represents 5, 600 local-authority lawyers, has called the rise in this kind of correspondence “massive.” Watch any system that still owes people an answer, and you can see what AI automation does to it.
The hour you saved is sitting in someone else’s queue
The sales story for AI in public services has been blunt: digitise the machinery of government and the money follows. On 25 April 2025, Parliament’s Science, Innovation and Technology Committee recorded that ministers and official documents had cited £45 billion a year from full digitisation of public-sector services. AI and Digital Government Minister Feryal Clark supplied a methodology note that shows how the number was built.
Of that £45 billion, £36 billion was put down to simplifying and automating delivery. The analysis behind it covered central government, the NHS and police personnel, then extrapolated the result across the whole public sector. It assumes a 15% lower-bound productivity saving, and that 100% of routine tasks and 10% of non-routine tasks can be automated. The note says the question is inherently hard to quantify. It gives no timeframe. Any savings are expected only over the long term.
Committee chair Chi Onwurah MP welcomed the transparency and still flagged the extrapolations, the assumptions, the missing clock, and the long record of public-sector technology programmes that never landed as promised. You are looking at a digitisation model. Nobody has booked it as an AI saving, and nobody can spend it next April.
Watch where the modelled hours actually go. A routine task marked automated on one desk reappears as a human on another, working out whether a 27-page letter is a crank, a template, or a carer who finally found the words. That is the productivity slide.
A record backlog, not a wild new intake
Employment tribunals in England and Wales are where the numbers get sharp, and where a sloppy summary does the most damage. Ministry of Justice figures, as set out by the Work Rights Centre, show 77, 594 open cases on 30 June 2026, the highest backlog since records began in 2009/10. That stock was up 8% in the quarter and 46% over 12 months, from 53, 320. The step from 53, 320 to 77, 594 is about 45.6%, so treat 46% as a rounded year-on-year comparison of those two stocks, not as a second, larger series.
Read the unit. Those 77, 594 are open cases (the pile waiting), not a count of new claims, and not the same measure as the 2.2 claims the Work Rights Centre finds, on average, inside a single claimant’s case. In the first quarter of 2026/27, 14, 459 cases were lodged. High for recent years, the Work Rights Centre notes, but closer to the average in the quarters before tribunal fees began in July 2013.
What has slowed is the exit. The Work Rights Centre, reading the same Ministry of Justice figures, describes disposals as unusually low, the lowest since the pandemic, and reports that some London hearings are listed into 2030. Capacity and low disposals built the stock. AI is a complexity multiplier that makes that stock harder to clear.
Money has not chased the pile. Real-terms Employment Tribunal funding, from Ministry of Justice figures the Work Rights Centre obtained under freedom of information, moved only from £73.83 million in 2022/23 to £79.75 million in 2025/26. In 2021/22, during COVID, it stood at £92.72 million. Judge and admin numbers have not kept pace with the open caseload, and London recruitment is especially weak.
Advice has thinned out too. Legal aid for employment matters in England and Wales is largely limited to discrimination, so people without lawyers reach for a chatbot. That is a staffing and advice failure. The AI subplot sits on top of it. The public did not suddenly find a cheat code, and you cannot wave the AI bit away either.
Where AI shows up in the files, it shows up as weight. The President of Employment Tribunals in England and Wales, along with other stakeholders, has linked rising case complexity to unrepresented claimants using AI to draft claim forms and manage cases, according to the Work Rights Centre’s reading of judiciary material. Acas data cited alongside that: in 2025/26, 58% of tribunal cases received for conciliation were open-track cases (Acas’s label for the more complex ones), up from 45% in 2022/23. Hearings get longer. Files get harder to unpick.
The cartoon version, claimants hurling every possible argument because a model told them to, is thinner than the anecdote mill suggests. The Work Rights Centre says single claimants have stayed steady at an average of 2.2 claims per case over two years. It also argues that chatbots are helping previously uninformed workers find rights they actually have, which it treats as good for the rule of law. A messier file can still be a real one. You can have both.
Freedom of information, now with extra clarification
Complaints and freedom-of-information teams are where the flood is happening now. The BBC reports that the Information Commissioner’s Office said in May 2026 that FOI teams were seeing higher volumes, greater complexity, misquoted legislation, and requests that need significant clarification. ICO guidance warns that without practical support:
this trend risks placing pressure on FOI teams and could lead to delays, errors or increased complaints.
Lynn Wyeth of the National Association of Data Protection Officers has called the result a “hidden workload”: longer correspondence, more clarification, more complex appeals. Her practical ask is permission for authorities to use AI on the way in, or the pile keeps growing. England’s Local Government and Social Care Ombudsman has told the BBC that AI should be used with “discretion.” The ombudsman has already seen more of it in legal challenges to its own decisions.
Schools are living the same pattern one inbox over. Pepe Di’Iasio of the Association of School and College Leaders has said complaints were already rising for several reasons. AI has added multiple, extensive parent complaints, and the strain on staff that comes with them. Evans’s counterpart to a spam filter is a form that limits length. Match the new volume by hiring, she has suggested, and you are looking at roughly doubling that complaint-handling workforce. That doubling figure is her estimate, not a published workforce study. The likely alternative, in her view, is slower replies. Critics quoted by the BBC hear a word cap and see a curb on accountability. They are not imagining the trade-off.
Guff, and the person inside it
It is tempting to treat every long AI letter as noise. Some of it is. These tools are often tuned to be agreeable, and bad at the sentence a school governor told the BBC they rarely produce: this is not worth pursuing. A parent, a claimant or a resident can be walked, politely, into a 20-page dead end, and a paid human has to walk back out of it.
Flatten that into “the public is abusing the system” and you will miss the other half of the same reporting. Dawn Harverson and others told the BBC that for some people a chatbot is the difference between complaining and staying silent, because the phone line has been replaced by a blank box and a word count. Neurodivergent people have said the same: the tool is what lets them challenge a decision at all. A teacher’s counter is blunter still. The failure may be that public services can no longer be dealt with face to face. One BBC listener put the question in a form every operator should steal: is the problem the model, or the fact that nobody can speak to a human anymore?
Email taught this once already. A channel sold as a replacement for meetings and paper became a permanent inbound pipe, because the cost of sending fell to almost nothing and the cost of being sent to did not. AI for business is running the same experiment at a higher word count. Cheap sending was always going to mean expensive receiving.
Both sides are already automating
Detecting machine text with more machine text is already happening. A BBC listener described cutting “50+ page” AI complaints down to a one-page AI reply. It looks like efficiency. It is also how a vulnerable person’s actual point gets summarised out of existence, and how a trivial grievance gets a polished answer it did not deserve. Wyeth is right that authorities need tools. That does not dissolve the duty to decide. Someone still owns the answer.
The Work Rights Centre’s preferred fixes are less glamorous than a new platform and more serious than a tollbooth. Emergency funding. Tighter case management. Deposit orders (a weak claim continues only if money is paid in) and strike-out for claims that are plainly hopeless. An ET1, the form that starts a claim, rebuilt so it uses AI to help a litigant in person present a coherent case and a schedule of loss, rather than a long account nobody can follow. Funded employment-advice services in the regions taking the worst hit. The President of Employment Tribunals has already issued guidance on remote hearings and widened what non-judicial staff can do. The system is improvising capacity. It is not waiting for a model to absolve it.
Fees are a different tool, and a legally loaded one. The Work Rights Centre reports calls for tribunal fees or a costs regime, and opposes both. It notes that the Supreme Court ruled the previous fee regime unlawful, and argues that a new toll would shut out low-income workers while weakening the deterrent against employer non-compliance. A congestion charge for justice picks who gets to be slow, and who gets to be silent.
Agents are the next turn, not this one
Keep AI agents in the future tense. What the evidence currently measures is assisted drafting: a person using a chatbot to write a complaint, an FOI request or a claim form, then sending it. No count in this picture covers fleets of agents filing, chasing and appealing at machine speed. Brief a board that agents are already overwhelming public services and you are ahead of the data. You will be easy to dismiss when the data is checked.
The direction is still obvious. An agent that can plan and act, not just draft, removes the last scrap of effort from the sender. Individual shortcuts become a bottleneck the moment everyone has one. Customer operations, supplier disputes, benefits-style appeals inside a large employer: the pattern does not stop at the council boundary.
The same applies to AI for sales. Sequences written and fired without a human pause train the other side to answer in kind, and they train your own inbox to stop distinguishing a real buyer from a spray. Once both sides generate at the same rate, sorting becomes the job.
Do not reach for a ban. Decide, before the volume arrives, which frictions you will actually keep. Length and structure on the way in. A duty to clarify before you refuse. A fast path to kill what is vexatious, and a slow path for what might be real. Permission for your own team to use AI, with a named human accountable for the answer.
Questions worth asking before the next efficiency slide
- What should a company do with this on Monday?
Count the work your AI automation moves, not only the work it deletes. If a tool makes it effortless to write to you, budget for reading, clarification and the occasional human conversation. A time saving that reappears in legal, compliance or customer operations was a relocation.
- Is £45 billion a year sitting there, ready for AI to collect?
No. The figure cited to Parliament’s Science, Innovation and Technology Committee on 25 April 2025 is a long-range digitisation estimate. £36 billion of it rests on automation assumptions (a 15% lower bound, every routine task automated, a tenth of non-routine work automated), with no delivery date. Treat it as a model with heroic premises. That money has not been saved.
- Are these long letters just abuse?
Many are ornate, error-strewn and trivial. They are not only that. For some people, including neurodivergent residents who can no longer get a human on the phone, a chatbot is what makes a complaint possible at all. A real grievance can still be buried in a bad draft.
- Should the answer be a fee, a word limit, or more AI?
Those are different tools. Length limits, strike-out of hopeless claims, and letting institutions use AI are all in play. None of them removes the duty to notice the genuine case. Fees are legally different: the Supreme Court struck down the last tribunal fee regime, and the Work Rights Centre argues a new toll would shut out low-income workers while weakening the check on employers who do not comply.
- Are AI agents already filing this at machine speed?
Not on the reporting and the tribunal figures now available. Today’s measured problem is assisted drafting and copy-paste volume. Agents that act without a person at the keyboard are the multiplier still ahead, which is why the design of friction should happen before they arrive, not after.
Extra words do not make anyone more accountable, and a faster draft still leaves you with the same case. And an hour removed from a routine task has a habit of popping up, slightly angrier, in the next department along. Build for that, or spend the next two years congratulating a model while your people drown in its output.