Characterizing Agentic Flooding of Government Services
Chris Schmitz, Lewis Hammond, Alan Chan
AAAI Conference on AI
cs.CY
2026-08-17
A scan of 12 countries yields 84 AI-attributed demand surges. 87% are cheap LLM text plus a human click. Fees work fast and ration access.
Talking to government is expensive in three ways: learning the rules, assembling the paperwork, and repeating yourself to officials. Public administration calls this administrative burden. Consumer LLMs already cut all three. They turn long statutes into plain language, check eligibility against a user's facts, and draft complaint letters that read like a lawyer wrote them. About 1% of Google Gemini requests are classified as help with government interaction. Accessibility going up is a public good.
Many windows were never sized for full take-up. They rely on friction to keep demand inside the budget. When submission costs fall, volume rises, and so does the length and tangle of each filing. The paper names this agentic flooding of government services: agents cheapen the interaction, then the window's volume or complexity outruns its capacity. Three questions follow. Is it happening. Which services break first. What can governments actually do.
The scan covers 12 countries and a fixed list of 13 domains, built from public-facing service catalogs in the UK, the US, and France. It includes benefits, tax, and courts, and also non-service channels such as consultations, regulatory complaints, and freedom-of-information requests. Collection is LLM-aided and human-gated. Models search country-domain pairs, fill a structured JSON template, and iterate. A human then keeps or drops the case. To limit hallucinations, agency names and counts are checked as fields, and cited URLs are taken from API metadata rather than generated in the model text.
A case must clear three tests: a plausible way AI lowers transaction costs for that service; a real shift in demand patterns, not a volume chart alone; and an official or a reputable third party attributing that shift to AI. Of 2,288 candidate services, 84 cases in 11 jurisdictions survived. The attribution rule was the tightest filter, passing fewer than one in twenty.
A risk matrix splits likelihood from severity. Likelihood tracks agent capability and price, whether the interface accepts open text, how hard a filing is, and how much a win is worth. Severity tracks processing effort per item, identity checks, room to scale, legal reply mandates, and whether the budget assumes low take-up. Near-term risk concentrates where a successful filing pays cash or triggers a legal duty, and where complexity, not design, used to keep people out. Tax returns and court claims sit in that cell.
Responses fall into two strategies. Suppress demand with fees, rate limits, in-person rules, bot blocks, or thinner entitlements. Raise capacity with staff, redesign, structured forms, or AI on the processing side.
In 87% of the 84 cases the mechanism is the same: an LLM cheaply writes legally sophisticated text, and a human clicks submit. Autonomous browser use and end-to-end agent pipelines do not show up in this set. Two reasons are on offer. Text generation is the mature, free capability. Officials can also spot AI-flavored prose in an open-text filing; an agent filling a short structured form looks like a person, so public attribution would be rarer.
Quantitative flooding (more filings) appears in 50 cases, 60%. Qualitative flooding (longer, denser filings) appears in 76 cases, 90%. Half show both. Justice and legal services lead with 19 cases, 23%, then regulatory complaints (10) and benefits (9). Government sources name AI in 69% (58 cases); the rest rest on third parties only. German social courts saw a 55% year-on-year caseload rise in 2025, and some courts tie a large share to AI-generated claims, including letters spanning more than 4,000 pages. Australia has considered restoring FOI fees. UK online money claims, Dutch WOZ valuation objections, and Korean electronic payment orders are in the table.
Governments respond in 56% of cases, usually with light-touch AI-use guidance. Friction appears in 14 cases (17%), including a Japanese IP block on a consultation portal. Redesign appears in 13 cases (15%), all small, such as verifying case numbers. Processing-side AI appears in 21 cases (25%), typically sentiment analysis of comments or detection of fake medical certificates: single-step tools, not autonomous case handling.
| Observation | Number |
| Cases kept / candidates | 84 / 2,288 |
| Mechanism is LLM text | 87% |
| Officials name AI | 69% (58) |
| Friction already used | 17% (14) |
| Processing-side AI tools | 25% (21) |
The authors refuse causal and prevalence claims. Some series in Figure 1 start rising before ChatGPT.
For people shipping civic, legal, or complaint agents, the product that is already in the wild is not a browser-using generalist. It is long-form drafting plus a human click. Open-text inlets (FOI, planning comments, pleadings) are the surface that floods first.
For governments the colder line is this. Operational collapse is not the near-term base case. Fees and caps deploy fast, and past demand surges show they work. The bill is equity. Friction hits poorer, less digital, unrepresented users first, and it also shuts down the accessibility gains agents could have delivered. Capacity-building avoids that trade-off and needs lead time that a live surge will not grant. The three near-term recommendations do not depend on a capability forecast: audit which windows are brittle, wire digital identity into the most exposed ones so per-person caps become automatic, and get a legal memo on which fees, caps, and processing-side AI uses are lawful now.
This is a policy diagnosis, not a new algorithm. The contribution is a checkable case file and a warning that the fastest tools are the ones that ration access.
Requiring public AI attribution writes undercounting into the method. Twelve countries were scanned and eleven produced cases; that is not a prevalence ranking. Only 25 cases have 2018-2025 volume series. The set is positives only, so 87% and 60% describe these 84 rows, not the service landscape.
LLM-aided collection has human review and URL extraction, but no inter-rater reliability and no human-vs-model discovery benchmark. The risk matrix and the response taxonomy come mostly from theory and older surges. The 84 cases are not plotted back onto the matrix, and the paper does not measure whether the responses already taken actually worked.