Agentic Flooding of Government Services: AI Does Not Remove Legal Procedure — It Changes Who Can Afford to Use It

Complex law. Clear action.

Reviewed by Oleksandr Sobovyi, Founder & CEO of CORVUS AI — editorial responsibility statement below.

Agentic Flooding of Government Services: AI Does Not Remove Legal Procedure — It Changes Who Can Afford to Use It

Most administrative and legal systems in Europe were built on an unspoken assumption: that preparing a coherent complaint, appeal, freedom-of-information request, or court filing costs the person making it something real — time to understand the rules, effort to locate the right authority, skill to write in the register the institution expects. That cost was never part of the legal test for whether a claim was valid. It was simply always there, absorbing part of the load that would otherwise reach the counter.

Generative and agentic AI is removing that cost, not by changing anyone's legal entitlement, but by making it cheap to exercise. A new study from researchers at the Centre for the Governance of AI and the Hertie School, Characterizing Agentic Flooding of Government Services (Schmitz, Hammond and Chan, August 2026), gives this phenomenon a name — "agentic flooding" — and a first empirical shape. CORVUS AI reads the paper as confirmation of something we have anticipated in our own regulatory work: the more interesting legal question is not how to stop AI-assisted submissions, but how public institutions should be redesigned once the transaction cost of exercising a legal right approaches zero.

What the Research Actually Found

The authors define agentic flooding as a surge in the volume or complexity of requests to a government service that (i) is caused by AI agents interacting with that service and (ii) substantially strains it. They separate two mechanisms that matter for very different legal reasons: quantitative flooding, where the number of submissions rises, and qualitative flooding, where individual submissions grow longer and more complex. A single AI-drafted filing running to several thousand pages, which the authors report arriving at a German social court, is qualitative flooding on its own — no volume increase is required to strain the recipient.

To test whether this is actually happening, the authors scanned public administrations in twelve countries across thirteen service domains and applied a deliberately conservative filter: a case was included only where a plausible cost-reduction mechanism existed, where measurable change in demand patterns was documented, and where a government body or a credible independent source explicitly attributed that change to AI use. Of roughly 2,300 candidate services surveyed, fewer than one in twenty cleared all three tests. The result is a dataset of 84 cases across eleven jurisdictions. In 69% of these, it is the government body itself that asserts AI involvement; the remainder rely on reputable secondary reporting.

It is important to state precisely what this dataset does and does not show. The authors are explicit that their methodology supports no causal or statistical claim about the prevalence of flooding, and that several of the underlying demand series were already rising before the 2022 release of ChatGPT. What the 84 cases do show is a mechanism, not a verdict: almost all of them — 87% — involve large language models cheaply generating legally sophisticated text, which a human then submits manually through an ordinary web form. Fully autonomous agents navigating government portals end-to-end are not yet a meaningful part of the pattern the authors observed. This is a finding about text generation lowering the cost of drafting, not yet about agentic autonomy replacing the citizen in the process.

The Deeper Issue: Administrative Friction as an Informal Gate

The paper's most useful conceptual move is to sort the costs AI is compressing into three categories, drawn from the public-administration literature on "administrative burden": learning costs (understanding what the rules require and whether one qualifies), compliance costs (drafting, form-filling, assembling evidence, tracking deadlines), and psychological costs (the toll of repeated, often dehumanising, contact with bureaucracy). AI agents address each of these directly — summarising regulation in plain language, drafting context-specific submissions, and maintaining case context across a multi-step process so a claimant no longer has to re-explain their situation at every contact.

Historically, these costs were not incidental. Some governments have knowingly tolerated — or deliberately maintained — administrative burdens that suppress take-up of entitlements they are nonetheless legally obliged to grant on request. Friction, in other words, has functioned as an informal capacity-control mechanism: it rationed access not by legal merit but by a claimant's literacy, patience, and access to professional help. AI does not touch the underlying statutory entitlement. What it changes is who can afford to invoke it. A person's right to challenge an incorrect benefit decision, object to a tax valuation, or file a freedom-of-information request is precisely what it always was. The practical cost of exercising that right has simply collapsed.

Access to Justice Versus Institutional Capacity

This is the tension CORVUS AI regards as the intellectual centre of the paper, and it does not resolve cleanly in either direction. On one side, AI genuinely improves access to justice: it helps individuals and SMEs who could not previously afford a lawyer identify the correct procedure, write a legally coherent submission, and overcome language barriers that would otherwise silently exclude them. This is not a marginal effect — the European Union has long treated effective access to administrative and judicial remedies as a foundational value, not an optional courtesy.

On the other side, courts, regulators, and administrative bodies remain bound by finite budgets, statutory response deadlines, and — in many jurisdictions — a legal duty to give reasoned consideration to every submission they receive, regardless of its origin. The same technological shift that lets a tenant challenge an unlawful eviction notice without paying for counsel also lets a claims-management operation multiply templated filings at near-zero marginal cost. The paper is careful — and CORVUS AI agrees this caution is warranted — not to treat these as separable populations. A single procedural channel routes both, and a response calibrated to one will inevitably strike the other.

Why Governments Cannot Simply Suppress AI-Assisted Claims

A claim that is legally well-founded does not become illegitimate because AI helped draft it. This follows directly from ordinary rule-of-law principles: procedural fairness, equality before the administration, proportionality, and the right to an effective remedy do not carry a carve-out for the tool used to prepare a submission. Within the EU legal order specifically, the Charter of Fundamental Rights is precise, and often misquoted, on where these guarantees actually bind. Article 41's right to good administration binds the institutions, bodies, offices and agencies of the Union itself — not, on the prevailing reading of the Court of Justice, the administrations of the Member States. Article 47's right to an effective remedy and to a fair trial has the broader reach: under Article 51(1) of the Charter, it binds Member States whenever they are implementing Union law, which captures a meaningful share of the administrative procedures the paper studies, including those tied to EU-derived entitlements and cross-border rights. National constitutional and administrative law fills the remainder. The point for counsel advising a public body is not to invoke the Charter reflexively, but to identify correctly which layer of obligation actually applies to the specific procedure at issue before recommending a response.

Because the underlying claims remain legally valid, suppressing volume by making the channel itself harder to use is not a neutral administrative choice — it is a decision that reallocates who can practically access a right. The paper's own data illustrates the shape of the trade-off: in the fourteen cases where governments have already responded with friction — reinstating fees, imposing digital-identity checks, blocking submissions by IP range — the measures were fast to deploy precisely because they ration by cost and access rather than by merit, and the authors note this predictably deters poorer and less digitally literate users first.

The Regulatory Paradox

This produces what CORVUS AI regards as the central regulatory paradox of agentic flooding: the fastest tools available to a strained institution are also the tools most likely to suppress legitimate access alongside any abuse. A reinstated fee for freedom-of-information requests, of the kind the Australian government has reportedly considered in response to AI-generated submissions, does not distinguish a vexatious mass filer from a citizen who simply could not previously afford to ask. Blocking submissions by IP address, as Japanese authorities did during a public-comment surge, is blunt by design.

The more durable answer the paper points toward is not friction but redesign: machine-readable procedures and structured digital submission formats that let straightforward cases proceed with minimal manual handling; identity verification that is proportionate to what is at stake rather than uniformly onerous; duplicate detection tuned to the specific abuse pattern rather than the channel as a whole; and AI-assisted triage on the institution's own side, so that processing capacity scales with demand rather than staying fixed while demand does not. None of these are free of legal risk of their own — identity verification requirements interact with data-protection law, and any triage system that meaningfully shapes an outcome for an individual raises the automated-decision-making and human-oversight questions addressed below. The point is that these risks are, in CORVUS AI's assessment, more proportionate and more defensible than blanket suppression, precisely because they can be targeted at the mechanism causing strain rather than at the channel through which legitimate claimants also pass.

The Coming Asymmetry

A related risk the paper flags, and which CORVUS AI sees as underappreciated in current governance planning, is institutional asymmetry. Citizens, claims-management firms, and advocacy organisations can adopt frontier AI tools essentially as fast as those tools are released. Public authorities cannot: they are constrained by procurement cycles, legacy IT that predates API-based integration, security clearance requirements, data-protection obligations attaching to any tool that touches personal data, and — since 2 August 2026 — the growing body of public-sector AI governance obligations under the AI Act itself. This is not a universal condition, and the paper is right not to overstate it, but it is a plausible and growing gap in specific service lines: the party submitting a request may, in some cases, already be working with materially more capable AI assistance than the official processing it.

The EU Regulatory Dimension

The AI Act is directly relevant to one half of this picture, and largely silent on the other. Where a public authority deploys AI to evaluate eligibility for public assistance benefits and services, Annex III, point 5(a) of Regulation (EU) 2024/1689 classifies that system as high-risk, triggering the full Article 8–15 obligations on risk management, data governance, technical documentation, human oversight, and accuracy. Point 8 does the same for AI used by or on behalf of a judicial authority to research or interpret facts and law. A narrow-procedural-task exemption exists under Article 6(3), but it does not extend to systems that meaningfully shape a decision affecting an individual's rights — which is exactly the category of triage and processing tool the paper suggests governments will increasingly deploy in response to flooding. Authorities building AI-assisted processing pipelines to cope with surging volume should assume Annex III applies unless a specific, documented exemption analysis says otherwise.

What the AI Act does not regulate is the shape of the administrative procedure itself — whether a fee is lawful, whether a rate limit is proportionate, whether a digital-only channel unlawfully excludes claimants without reliable internet access. Those questions sit in national administrative law, in GDPR where personal data processing is involved, and, where a decision is made "solely" by automated means and produces legal or similarly significant effects, in the safeguards of GDPR Article 22. Regulating the AI system that assists a submission and redesigning the procedure that receives it are two distinct exercises, and treating the AI Act as an answer to institutional capacity strain — which is the recurring error CORVUS AI sees in client thinking — misreads what the Act is for.

What Organisations Should Expect

Regulators, courts, municipalities, and regulated companies that receive statutory complaints or requests should treat AI-induced procedural volume as a planning variable in its own right, distinct from the AI-risk assessments most institutions have already built. This is what the paper's authors term auditing exposure: a service's vulnerability to flooding can be assessed today, without waiting to observe an actual surge, by asking whether it is financially consequential to the claimant, whether it has historically relied on complexity or professional knowledge to suppress demand, and whether its submission channel accepts open-ended free text. Tax administration, court filing systems, and benefits appeals score highly on all three factors and merit priority review. Legal and compliance functions inside regulated companies that receive high volumes of statutory requests — data subject access requests under GDPR are an obvious example — should apply the same audit internally, since the same cost-reduction dynamic applies wherever a statutory response obligation meets an open text channel. CORVUS AI would frame the planning question as one of procedural scalability: whether a given legal process can absorb a step-change in submission volume or complexity without either collapsing under backlog or resorting to blanket suppression that a court would later find disproportionate.

Conclusion

AI does not remove legal procedure. It changes who can afford to use it. The friction that used to ration access to complaints, appeals, and remedies was never a feature of the underlying law — it was an artefact of cost, and that artefact is now dissolving faster than most institutions can redesign around it. The response that repeats itself across the paper's dataset — fees, IP blocks, identity gates — is the one available on the shortest timeline, and it is also the one most likely to lock out exactly the claimants that lower-cost AI assistance was supposed to help. The more durable path is not to restore artificial friction but to redesign legal and administrative systems so that legitimate rights remain inexpensive to exercise, while institutions gain the machine-assisted capacity to identify, process, and resolve claims at the scale now arriving at their door.

Key Takeaways

Agentic flooding is a documented, if not yet statistically quantified, pattern: 84 cases across 11 jurisdictions, almost entirely driven by LLM-generated text rather than fully autonomous agents.

AI collapses the learning, compliance, and psychological costs that historically rationed access to administrative and legal remedies — without changing the underlying legal entitlement.

Friction-based responses (fees, ID checks, IP blocks) are fastest to deploy but fall hardest on the least-resourced claimants, creating a genuine access-to-justice risk.

Under the AI Act, public-sector AI deployed to process the resulting volume — benefits eligibility tools (Annex III, point 5(a)) and judicial-assistance tools (point 8) — is high-risk and subject to the full Article 8–15 regime; the Act does not itself regulate the fairness of the underlying procedure.

Organisations, public and private, that receive statutory requests should treat AI-induced procedural volume as a distinct planning risk and audit their most exposed channels now, before a surge forces a reactive response.

Disclaimer

This article has been prepared by CORVUS AI for general informational and educational purposes only. It is intended to make complex legal and regulatory developments easier to understand.

It does not constitute legal advice and does not create a professional adviser–client relationship. The information should not be relied upon as a substitute for advice based on the specific facts, circumstances and applicable law relevant to your organisation or project.

The article reflects our understanding of the law and regulatory framework as of the date of publication. Legislation, case law, regulatory guidance and administrative practice may subsequently change. While reasonable care has been taken in preparing this article, CORVUS AI does not warrant that the information is complete or remains current after the date of publication. We do not undertake to update this content.

To the fullest extent permitted by applicable law, CORVUS AI excludes liability for loss arising from reliance on this article. Nothing in this article constitutes an offer or solicitation to provide regulated legal services in any jurisdiction where doing so would be unlawful.

AI-assisted preparation: This article was prepared with the assistance of AI tools. Its legal analysis, conclusions and final text were subject to human review and editorial control and were reviewed and approved prior to publication by Oleksandr Sobovyi, Founder & CEO of CORVUS AI. CORVUS AI retains editorial responsibility for the published content.

For advice tailored to your organisation, project or specific circumstances, please contact CORVUS AI.

Key Sources

Schmitz, C., Hammond, L. and Chan, A., Characterizing Agentic Flooding of Government Services, arXiv:2608.16603 (17 August 2026)

Charter of Fundamental Rights of the European Union, Articles 41, 47, 51

Regulation (EU) 2024/1689 (AI Act), Article 6 and Annex III, points 5 and 8

Regulation (EU) 2016/679 (GDPR), Article 22

European Commission, Right to good administration

European Union Agency for Fundamental Rights, Article 41 — Right to good administration and Article 47 — Right to an effective remedy and to a fair trial

CORVUS AI, Rotterdam, the Netherlands — contact@corvusai.eu