Tom Loosemore asked an AI agent whether the council tax band on his house was fair compared with his neighbours’. It came back at once: his band was higher than every house around him, with an offer to do something about it. He stopped it there, but he could see where it was heading: neighbours’ floor areas from the Gov.uk Energy Performance Certificate API, extensions measured off Ordnance Survey, Land Registry price history, then an appeal drafted and filed with the Valuation Office Agency, unsupervised.
Writing in Computer Weekly in April, Loosemore — a partner at Public Digital and a founder of the UK’s Government Digital Service — priced that operation at one click and about 12p, and noted it is the most expensive it will ever be. The friction that had quietly rationed council tax appeals for decades had evaporated.
That collapse in the cost of asking now has a literature. Characterizing Agentic Flooding of Government Services, posted to arXiv in August by Chris Schmitz, a PhD student at Berlin’s Centre for Digital Governance, with Lewis Hammond of the Cooperative AI Foundation and Alan Chan of GovAI, catalogues 84 cases across 11 jurisdictions and 13 service domains where submissions surged and officials or credible reporting blamed AI. TechCrunch covered the dataset on September 10; the paper is due at the AAAI/ACM Conference on AI, Ethics, and Society, October 12 to 14.
The curves look remarkably alike. Complaints to the UK housing ombudsman more than doubled after ChatGPT arrived, rising from 2,600 in 2022 to just over 7,000 last year. The US Consumer Financial Protection Bureau saw roughly 5x growth over the same period. Benefit appeals to the UK Department for Work and Pensions are up more than 60% since the first benefit-specific AI tools appeared in 2022. German parliamentary petitions, Brazilian judicial petitions and Dutch municipal valuation objections bend the same way. In most cases volumes were flat before 2022, then rose at an accelerating rate that has not yet slowed.
Schmitz separates two phenomena. Quantitative flooding, where the number of requests grows, appears in 50 cases. Qualitative flooding, where each request grows longer and more legally sophisticated, appears in 76, and 42 cases show both. In 87% of cases the mechanism is not an autonomous agent clicking through a portal but something more mundane: a language model generating competent legal prose for almost nothing.
“People are finding out that this is something one can do, and incrementally, it is just getting easier to do it,” Schmitz told TechCrunch. What once meant assembling context and prompting ChatGPT 3.5 precisely, he said, “may now be a question of just pasting or taking a photo of a letter with your Claude app and getting a pretty good response in one shot.”
The uncomfortable finding is who is sending all of it. “The vast majority of cases we find are people who are entitled to claim for something, claiming for that thing,” Schmitz said. The paper documents bad actors — AI-generated medical certificates filed for disability benefits, organised campaigns mass-filing records requests and voter roll challenges — but they are the exception. Mostly this is administrative burden lifting: people who were always eligible, and who used to give up between the criteria and the evidence requirements, now finishing the form. Two caveats belong here: the paper stops short of claiming AI caused the surges, and it calls the strain so far moderate.
Analysis: cheap demand meets fixed capacity
The structural problem is an asymmetry in marginal cost. Filing a well-argued appeal has gone from hours of unpaid, easily abandoned effort to the price of a few thousand tokens. Adjudicating one has not moved: it still takes a caseworker, a statutory timetable and sometimes a tribunal slot. One side of that ledger is falling toward zero on a technology curve; the other is pinned to a headcount set by a budget cycle. Five times the applicants on a flat budget is not a surge to absorb, it is a different service.
This is why legitimate automated volume is harder to handle than spam. Bug bounty programmes swamped by LLM-written reports last year had a clean exit: the submissions were worthless, so triage could get ruthless. An agency has no such exit. A valid claim creates a legal entitlement; a properly filed appeal starts a clock. Each submission must be read and answered on the merits, and because qualitative flooding makes every filing longer, the per-item cost rises with the item count. Screening that used to happen free at the point of discouragement now has to be paid for at the point of review.
Which leaves capacity or suppression, and suppression deploys faster. The paper counts 14 cases where agencies added friction — reinstating freedom of information fees, blocking IP ranges, batch-dismissing consultation responses — and finds those measures generally worked. They also screen out the poorest and least digitally literate claimants first, which is backwards if the point was widening access. Often they are not even lawful: charging a fee to apply for welfare is illegal in many jurisdictions. And friction does nothing about longer submissions, or adversaries who can pay it. Schmitz would rather governments treat this as a design brief than a denial-of-service attack. “A big part of making AI go well is being able to detail out what the good version of things looks like,” he said. “This could be the moment to say, we need to rethink pretty much everything about how this process looks.”
What to watch: whether any agency publishes an AI-era capacity plan rather than a fee increase; whether 2026 filings bend the curve further, given that consumer agent products only went mainstream late last year; and whether administrative law permits the automated triage that capacity plainly requires. The flooding so far came from people pasting letters into chatbots. The agents Loosemore was testing have barely arrived.
“The vast majority of cases we find are people who are entitled to claim for something, claiming for that thing.”— Chris Schmitz, PhD researcher, Centre for Digital Governance, Berlin