Summary

• DWP's Work Hub uses Claude Sonnet 4.6 through Amazon Bedrock alongside structured employment-support workflows.
• Several thousand people have accessed the service through private beta, with wider availability planned.
• The system provides voluntary guidance rather than benefit decisions, with DWP retaining responsibility for its operation and assurance.

The Department for Work and Pensions (DWP) is operating an AI-powered employment assistant with several thousand private-beta users, using Anthropic’s Claude Sonnet 4.6 to provide CV improvements, skills analysis, and personalised career guidance.

Known as Work Hub, the service forms part of DWP’s Jobs and Careers Service and combines conversational AI with structured processes for assessing work experience, identifying transferable skills, and suggesting jobs or training. The department intends to expand access nationally, although it has not specified a general launch date.

A government algorithmic transparency record published on 30 October identifies Amazon Bedrock as the managed service through which the language model processes text. DWP operates the surrounding application and data storage within its approved technical environment, with Amazon Web Services providing the external model-processing service.

Work Hub is voluntary and its recommendations are advisory. The department expressly excludes decisions about benefit entitlement, enforcement, compliance, and other statutory outcomes from its intended function, separating employment support from processes that affect legal rights.

Career guidance built around individual circumstances

Users interact with Work Hub through an online interface that gathers information about their employment goals, experience, skills, and preferences. The assistant can examine a CV, identify relevant capabilities, and recommend occupations or training opportunities that may fit the information provided.

Someone seeking work in a different sector might use the service to identify skills gained through previous employment and explore roles requiring comparable experience. The system can also generate suggested improvements to a CV or create an action plan covering steps towards a particular career objective.

These activities combine structured software with conversational responses. CV parsing and skills mapping help organise the information supplied by a user, while the language model produces explanations and guidance adapted to the conversation.

Previously, personalised employment support was primarily available through work coaches and other manual services, alongside online job boards and general career information. Work Hub provides another route to tailored guidance without requiring a separate appointment for every interaction.

The system can be used repeatedly as an individual’s circumstances change. A person may revise their CV, investigate alternative occupations, or explore different training options, with the service using the information made available during the relevant interaction.

Its effectiveness depends on the accuracy of that information and the quality of the recommendations. Job requirements can vary between employers, while an individual’s circumstances may involve qualifications, caring responsibilities, accessibility needs, or other conditions that cannot always be interpreted reliably from a short conversation.

Claude operates within a managed government service

DWP uses Claude Sonnet 4.6 through Amazon Bedrock rather than relying on a user accessing a general consumer chatbot. The department specifies the model in its transparency record as the version operating at the time of publication, while acknowledging that the configuration may change as Work Hub develops.

Information submitted by users can include CVs, employment histories, skills, preferences, and answers to questions posed by the assistant. Where a user is logged in, the system may also use other information available through the relevant account arrangements.

Some of these details can be sensitive. A user may disclose information about health, disability, previous employment difficulties, or personal circumstances while seeking appropriate work. DWP therefore applies data minimisation and protective controls intended to limit the personal information processed externally.

The department says the external model-processing service handles text transiently and does not retain it or use it for model training. Its own systems remain responsible for storing information required by the application under the department’s approved hosting and security arrangements.

These controls affect how the service is operated, but they do not remove the need for careful governance. Employment recommendations are generated from personal circumstances, making accuracy, confidentiality, and access to information relevant throughout the service’s lifecycle.

Testing checks accuracy and unequal outcomes

Generative AI can produce plausible text that is incomplete or incorrect, creating risks when a user relies on advice about qualifications, vacancies, or potential careers. DWP identifies misleading outputs among the operational risks associated with Work Hub.

The department uses testing, evaluation, user information, and quality assurance to manage those risks. Users remain responsible for deciding whether to act on the guidance, while DWP monitors the service and updates its operation as development continues.

Fairness testing forms another part of that process. The department describes using synthetic and paired demographic profiles to examine whether equivalent users receive materially different recommendations when demographic characteristics or related signals change.

Such tests can help identify inconsistent treatment, particularly where advice might reproduce patterns found in historical labour market information. The value of the results depends on the range of situations tested and whether the system behaves consistently when people use it under less controlled conditions.

DWP completed a data protection impact assessment and equality assessment in May 2026. The transparency record identifies these exercises as part of its work on information handling, fairness, and appropriate operational safeguards.

Individual Work Hub interactions do not require mandatory human review because the outputs are advisory rather than statutory decisions. Human oversight instead operates through testing, quality assurance, and continuing assessment of system performance.

That model places importance on explaining the nature of the service to users. An individual seeking employment advice needs to understand that a generated recommendation is a suggestion to consider, not a verified job offer, official determination, or guarantee that a particular occupation will be suitable.

Private beta provides the next source of evidence

Although the transparency record identifies Work Hub as being in production, access remains limited to several thousand users participating in private beta. The software is therefore in operational use without being generally available to everyone seeking employment support.

Expanding beyond the initial users will increase the range of questions, employment backgrounds, and accessibility requirements the service must handle. It will also require the surrounding infrastructure and support processes to accommodate greater demand.

Digital careers guidance can improve availability, particularly for people who want to explore opportunities outside conventional appointment times. However, individuals with limited digital skills or more complex support needs may still require help from work coaches or other established services.

DWP has not published measured outcomes showing that Work Hub increases job placements, improves employment retention, or reduces the cost of delivering employment services. Nor has it provided comparative results demonstrating that its AI recommendations are more effective than other forms of advice.

The department’s immediate evidence will therefore come from the continuing beta programme and its evaluations. Those findings can inform improvements to guidance, safeguards, and the service’s wider operating arrangements before access is expanded nationally.