A caseworker's decision is often the only point where an organisation looks at a person's circumstances as a whole. AI agents can check an application against criteria and draft an assessment. The caseworker decides, uses discretion where rules do not fit, and explains the outcome to someone who may appeal. That makes caseworkers the real control.
What does a caseworker's work look like with an agent?
Caseworkers assess applications and decide on cases: benefit claims, permits, grants, support packages, licences. The work combines reading evidence, applying criteria, using discretion and communicating the decision. Much of the time goes on the first two.
A governed agent takes on much of that reading and checking. At depth 1 it reads the application, supporting documents and the organisation's own records. At depth 2 it drafts an assessment against the criteria, notes missing evidence and prepares the letter. It does not decide. Approvals, refusals and anything involving discretion are held for a named caseworker.
The caseworker therefore starts each case with a prepared assessment. That saves time and changes the risk: the assessment looks complete, and the caseworker has to judge whether it is.
Why is the caseworker the deciding control?
Three reasons, each particular to casework.
Discretion. Criteria rarely capture every situation. An applicant just outside an income threshold because of a one-off payment, or a missing document because of circumstances outside their control: these need someone to decide whether the rule applies as written. An agent applies criteria; a caseworker decides when the criteria are not the whole story.
Significant effects on individuals. Casework decisions often have legal or similarly significant effects on people. Under GDPR Article 22, people have rights in relation to decisions based solely on automated processing with such effects. Meaningful human involvement takes a decision outside that category; a caseworker who approves the agent's draft without considering the case does not provide it.
Explanation and appeal. Public bodies face additional transparency and fairness duties, and under the EU AI Act some public uses are high-risk, with human oversight expected. An applicant is entitled to understand a decision. If the caseworker cannot explain it without pointing at the system, the decision is not defensible on appeal.
What must caseworkers be trained to do?
Most of it builds on skills good caseworkers already have. What is new is reading an agent's work:
- Check the evidence base. Which facts in the assessment are sealed by the organisation's own records, which are signed by an identifiable party, and which are declared by the applicant?
- Find what the agent did not weigh. Circumstances mentioned in a letter but not in a form field. Context from a previous case.
- Recognise where discretion applies, and record the reasoning when it is used.
- Write explanations in plain words that would hold up if the applicant appealed.
- Spot vulnerability in the application and route the case appropriately.
The guide on how caseworkers review an agent's decisions sets out a review method in detail.
Does the country change what the control looks like?
The principles are shared across Europe; the institutions differ. In Italy, for example, caseworkers in the public sector work under GDPR and the EU AI Act, with the Garante as the data protection authority, and with administrative law expectations about reasoned decisions. AI agent training for caseworkers in the public sector in Italy covers what that means for training and oversight.
Will agents replace caseworkers?
No, though they will change the job. Agents will take more of the checking, and caseworkers will spend more of their time on discretion, explanation and complex cases. The argument is made in full in whether agents will replace caseworkers. For organisations, the important point is that the caseworker's judgement becomes more concentrated, not less important.
How does this compare with other roles?
Paralegals and compliance officers are in a similar position. Paralegals review agent-prepared documents where a missed clause can change a matter's outcome. Compliance officers sign off findings an agent drafted. In each, the agent narrows the work and the person's judgement decides it. What sets caseworkers apart is that the person affected is usually a member of the public with fewer resources to challenge a poor decision.
What is the trade-off?
Treating caseworkers as the control means investing in training and keeping caseloads manageable. Public bodies are under pressure to clear backlogs, and an agent that drafts assessments looks like a way to do that with fewer people. If caseloads rise to match the agent's speed, caseworkers will have less time per case and rubber-stamping becomes likely.
There is also the risk that organisations treat the agent's assessment as the decision and the caseworker as a formality. That would save time and expose the organisation to challenges it cannot answer.
The right balance is to let the agent take the checking and keep caseloads at a level where caseworkers can use real judgement on what remains. Savings come from not having to gather evidence by hand, not from reducing the time spent deciding.
What should a service director do differently?
Name the caseworker as the decision-maker in every process where an agent prepares assessments. Train caseworkers on reading agent drafts, provenance and discretion. Keep caseloads at a level that allows real review. Sample decisions for quality, not just speed. And make sure every decision letter can be explained without reference to the system. The guide to AI agent skills by role sets out what each role needs.
Where the Academy fits
Pryme Intelligence Academy's Operator track prepares people who review a live workspace every day. Module tests are set in the learner's own industry and role, so a caseworker's questions involve applications, criteria and discretion. Certification includes a two-week practicum of real review decisions signed off by a Pryme assessor, and the certificate can be checked at source by anyone.
Questions people ask
Can an agent refuse an application on its own?
In most public-sector deployments, no. Refusals have significant effects on people and must be explainable, so they are held for a caseworker.
How should caseworkers record discretion?
In their own words, with the verdict, explaining which circumstance led them to depart from or apply the criteria. That record is what supports the decision on appeal.
Do applicants need to be told an agent was involved?
Transparency duties vary, but under the EU AI Act people must be told when they interact with an AI system. Many public bodies go further and explain how agents are used in assessments.