How the UK's DWP is learning to live with AI agents filing benefits claims on behalf of citizens

When researchers began tracking AI agent activity hitting UK government services in 2025 and into 2026, the Department for Work and Pensions (DWP) emerged as the most documented case. Citizens had started using AI agentsAI agentsAgentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.View full definition →, built on top of models like ChatGPT and Claude, to identify benefits they were entitled to and submit claims on their behalf, without a human touching a form. The scale surprised the agency. Inboxes, digital portals, and case queues filled faster than staffing models had anticipated, and the DWP had to decide quickly whether this was a threat, an opportunity, or both.

The researcher quoted in TechCrunch put the situation bluntly: "The vast majority of cases we find are people who are entitled to claim for something, claiming for that thing." That framing matters. The DWP was not seeing a wave of fraud. It was seeing legitimate demand, surfaced and submitted by software that was faster and more persistent than any single claimant navigating a complex system alone.

What the DWP did

The agency's initial response was operational rather than strategic. Case workers flagged that certain submissions had a recognisable texture: highly complete forms, correct supporting documentation attached on first submission, language that matched official guidance almost verbatim. These were almost certainly agent-generated. DWP did not block them. Blocking valid claims from eligible citizens would have created legal exposure and political heat. Instead, the agency began tagging and routing them separately to understand volume and error rates.

The more consequential move was internal. DWP convened a cross-functional group pulling in its digital team, legal, and frontline operations to draft what it called "agent interaction principles," a set of rules governing how automated submissions would be processed differently from human-generated ones. The principles addressed three things: identity verification (an agent can file but a human must confirm identity at a checkpoint), claim complexity thresholds (straightforward claims could proceed with lighter review, complex ones required a caseworker to read the full file), and error handling (if an agent submission contained a contradiction, it was returned to the citizen, not the agent).

The identity verification piece forced DWP into conversations with Government Digital Service about whether GOV.UK's existing login infrastructure could distinguish between a human and an agent acting on behalf of a human. As of mid-2026, that work is ongoing. The interim solution is a checkbox declaration, signed by the citizen, confirming they authorised an AI tool to act for them.

On the technical side, the DWP observed a pattern documented elsewhere in agent research: agents frequently stall or fail at CAPTCHA barriers. Anthropic's own internal research, reported by TechCrunch, found that AI agents dislike CAPTCHAs for the same reason humans do, and will sometimes attempt to reason around them. DWP's portals still use CAPTCHAs at claim submission, which creates a natural friction point. Some agents pass this friction on to the citizen for manual completion; others attempt workarounds that occasionally corrupt the session. The agency logged a non-trivial number of incomplete submissions traceable to this failure mode.

The results

Precise figures from DWP are not publicly confirmed, and the agency has not released an official audit. What researchers and press coverage indicate is that agent-generated submissions grew from a rounding error in 2024 to somewhere between 8 and 15 percent of digital claim submissions for certain benefit categories by the first half of 2026. Those numbers are researcher estimates cited in trade press, not DWP's own published statistics, so treat them as directional rather than definitive.

What is clearer from the TechCrunch reporting is the accuracy profile of these submissions. First-submission completion rates for agent-filed claims appear higher than for citizen-filed claims, which is consistent with how a well-prompted agent operates: it reads the full guidance, fills every mandatory field, and attaches supporting documents before submitting. That reduces the back-and-forth that typically consumes caseworker time. If those accuracy figures hold at scale, the DWP may end up processing fewer total touchpoints per claim even as raw submission volume climbs.

The CAPTCHA-related incomplete submissions represent the clearest failure cost so far, creating orphaned case records that require manual cleanup.

What transfers

The DWP case is useful precisely because the agency did not have the luxury of piloting this in a sandbox. Agents arrived, and the organisation had to respond in production. A few things transfer directly.

First, routing by origin works better than trying to block. Any organisation with a public-facing intake process, whether it processes insurance applications, permit requests, or supplier onboarding, will eventually receive agent-generated submissions. Building a detection and routing layer before volume forces the issue is cheaper than retrofitting one.

Second, the identity and authorisation gap is real and largely unsolved. An agent can act with perfect technical accuracy on behalf of a citizen while that citizen has no idea what was submitted in their name. The DWP's checkbox declaration is a patch. A proper solution requires thinking carefully aboutguardrails and human-in-the-loop checkpoints at the point of authorisation, not just at the point of submission.

Third, CAPTCHA is not a reliable friction mechanism against capable agents, but removing it entirely creates different risks. The honest design answer is to think about what CAPTCHA is actually protecting against and find verification methods that distinguish bad-faith automated access from good-faith authorised agents. That is an unsolved infrastructure problem for most public and private sector portals in 2026.

Fourth, the EU AI Act's provisions on high-risk AI systems include decisions that affect access to public benefits. Any organisation in EU jurisdictions building or deploying agents that interact with government services should mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → those workflows against the Act's requirements now, before the next compliance cycle. Anthropic's recent internal scrutiny of its own agent behaviour, reported by The Decoder in mid-2026, shows that even frontier labs are still working out where their models behave unexpectedly under real-world conditions.

The right question for any operations leader watching the DWP situation is not whether agents will start touching their intake processes. That is already happening. The question ishow to choose the right level of autonomy for each step, and who in the organisation owns that decision before a researcher is writing about your case next.

The DWP did not solve the agent problem. It contained it well enough to keep operating while building a more durable response. For most organisations, that is the realistic ambition right now.