Agentic Flooding: How Multimodal AI Is Breaking Legacy Public Service Intake

When consumer-facing generative AI tools lowered the barrier to generating structured text, developer platforms were the first to feel the strain. Code repositories and bug bounty programs absorbed waves of synthetic, low-yield submissions that drained review bandwidth. Now, that same dynamic has reached the public sector, creating a phenomenon termed "agentic flooding."

According to research conducted by Chris Schmitz, scheduled for presentation at the AI Ethics and Society conference, public sector intake pipelines across the globe are experiencing sharp, sustained spikes in formal submissions. Across 84 examined cases spanning 11 jurisdictions, the introduction and diffusion of consumer language models correlates directly with unprecedented application volumes.

The Empirical Footprint of Agentic Flooding

The statistical shifts documented across international intake systems show a structural break in historical filing baselines starting in late 2022:

• United Kingdom Housing Ombudsman: Official complaint volume rose from 2,600 cases in 2022 to more than 7,000 in the following year, more than doubling baseline load.
• United States Consumer Financial Protection Bureau (CFPB): Formal consumer complaints expanded by a factor of five over the same period.
• International Legislative and Judicial Pipelines: Parallel submission surges materialized in Brazilian judicial petitions and German parliamentary petition systems.

Prior to late 2022, intake trends across these 84 operational pipelines were broadly flat. After the public release of consumer LLMs, submission rates climbed rapidly and, crucially, show no signs of plateauing. While the study methodologically stops short of proving direct mechanical causality for every single filing, the synchronized inflection across disparate administrative domains points to a shared catalyst: the elimination of drafting friction.

The Mechanism: From Complex Prompting to Single-Shot Ingestion

The operational shift from text-only conversational interfaces to multimodal reasoning apps altered how citizens interact with bureaucracy. In the early phases of consumer generative AI, drafting an actionable dispute letter or housing appeal required complex prompt construction, contextual aggregation, and multi-turn refinement.

Modern multimodal tools require far less user effort. An individual can photograph a formal rejection notice or lease agreement using a mobile interface, pass it to an assistant model like Claude, and instantly receive a formatted, legally grounded submission ready for upload. By collapsing the cognitive and administrative effort required to interface with bureaucratic structures, consumer AI tools act as automated client-side proxies.

Legitimate Rights vs. Adversarial Noise

The core dilemma of agentic flooding differs fundamentally from corporate security triage. When bug bounty programs experienced LLM-generated submission spikes, the incoming traffic consisted overwhelmingly of hallucinated vulnerabilities and low-quality noise that wasted developer hours without surfacing real risks.

In public administration, the dynamic is reversed. Schmitz’s analysis indicates that the vast majority of incoming claims originate from real individuals filing for benefits, protections, or administrative reviews to which they are legally entitled. The previous suppression of these claims was not due to an absence of valid grievances, but rather the weight of administrative burden—the complex paperwork, time investment, and institutional jargon traditionally required to lodge a formal petition.

Consequently, public agencies cannot simply deploy blanket automated filters or adversarial bot mitigation without risking the unlawful disenfranchisement of valid constituents. Because the claims represent genuine civil and economic entitlements, processing agencies face an acute operational bottleneck: maintaining legal compliance and review fidelity while managing multi-fold increases in volume under fixed operational budgets.

Architectural Imbalance: Client-Side Agents vs. Legacy Backend Intake

The structural vulnerability exposed by agentic flooding is an asymmetric technical divide between client-side interfaces and backend intake pipelines:

• Client Side: Scalable, highly optimized multimodal language models that generate contextually complete, highly articulated legal appeals in seconds at negligible marginal cost per token.
• Server Side: Legacy relational databases, rigid web forms, manual document triage workflows, and human adjudicator pipelines constrained by operational capacity.

When intake systems are designed under the assumption that friction will naturally filter inbound volume, removing that friction overwhelms the backend. Applying basic CAPTCHAs or procedural hurdles merely introduces an adversarial race against increasingly capable vision-enabled models, doing nothing to solve the underlying processing deficit.

Designing Resilient Automated Public Infrastructure

Addressing the surge requires rebuilding public intake systems to accommodate an era where inbound traffic is agent-mediated. Several architectural adjustments are emerging as necessary responses:

1. Semantic Triage and Automated Ingestion: Public sector backends must integrate domain-specific language models capable of extracting structured claims, verifying jurisdictional requirements, and pre-populating evidentiary dossiers before human review occurs.
2. Decoupled Evidentiary Verification: Rather than relying on narrative length or argumentative complexity to establish claim seriousness, intake platforms must shift toward direct, authenticated data handshakes (such as verifiable digital credentials and programmatic database cross-referencing) to substantiate claims programmatically.
3. Standardized Machine-Readable Interfaces: As citizens increasingly use software agents to manage their administrative interactions, public agencies will need formalized, rate-limited APIs for petition filing rather than fragile web portals susceptible to unstructured automated submissions.

As AI tools become more integrated into consumer workflows, the volume of administrative, legal, and regulatory petitions will continue to climb. The challenge facing public institutions is not finding ways to block automated assistance, but re-engineering their intake architecture to ensure that increased accessibility does not result in systemic institutional failure.