The UK government is testing whether an AI assistant can do more than point citizens toward another webpage. A new Department for Science, Innovation, and Technology pilot with Anthropic will use Claude to guide people through employment services, training opportunities, and available support.
The project follows a February 2025 memorandum and moves toward an agentic model that retains context and helps users progress through a process. Government portals usually contain abundant information; the harder problem is matching complex rules and services to a person who may not know the official vocabulary.
Success will depend less on conversational fluency than on safe routing, data control, and a reliable route to a human.
Employment Services Test Stateful Public AI
Job seeking is not one transaction. A person may return repeatedly, change goals, complete training, update circumstances, or need support from several agencies. Remembering relevant context could prevent users from entering the same information again and make the next recommendation more useful.
The assistant is intended to help people find work, identify training, and understand support mechanisms. That makes it an intelligent routing layer between an individual situation and existing public services rather than a source of new policy.
Statefulness also creates responsibility. The system must distinguish current information from old assumptions and show users what it remembers. A mistaken detail about availability, eligibility, or location could repeatedly send someone down the wrong route if it silently persists.
Citizens need to correct records, control memory, and continue without consenting to unnecessary retention. The source says users will be able to opt out and determine what the system remembers, with personal information handled under UK data-protection law.
Scan, Pilot, Scale Makes Restraint Part of Delivery
Public services cannot treat a polished demonstration as evidence of readiness. The project follows a “Scan, Pilot, Scale” framework that requires staged testing before wider deployment.
This creates space to evaluate whether the assistant identifies the right service, explains uncertainty, protects personal data, and escalates difficult cases. The UK AI Safety Institute will participate in testing so safeguards can be assessed before reach expands.
Evaluation should include more than user satisfaction. Completion rates, incorrect routing, repeated contacts, accessibility, demographic differences, human escalations, and the outcomes of recommended steps all matter. An assistant that reduces calls by discouraging users would look efficient while weakening the service.
Employment support can involve vulnerability, disability, immigration status, finances, and other sensitive circumstances. The system should ask only for information necessary to the task and avoid presenting probabilistic guidance as a determination of eligibility.
Knowledge Transfer Limits Long-Term Dependency
Anthropic engineers will work alongside civil servants and Government Digital Service developers. The explicit aim is to build internal capability so the government can maintain the service after the initial engagement.
That co-working model addresses a familiar outsourcing risk. When an external provider becomes the only party able to operate a public service, contract renewal and model changes can constrain policy and budgets. Skills transfer makes the government a more informed buyer and operator.
Technical independence requires more than training. The team needs access to evaluation sets, prompts, workflow definitions, incident records, and architecture documentation. It should understand which components can move to another model and which depend specifically on Claude.
The system also needs clear accountability. Anthropic can provide technology, but the government remains responsible for the service citizens receive. A person must be able to challenge harmful guidance and reach an official who can resolve the underlying issue.
This pilot reflects a mature ambition: move from answering generic questions toward helping people complete useful journeys. Yet action raises the stakes. A search result can be ignored; a personalised assistant may be trusted as an authority.
The safest public AI will not imitate certainty. It will preserve user control, identify the limits of its role, and make the government easier—not harder—to reach when circumstances do not fit the standard path.
