Two weeks is enough to onboard a new caseworker if the plan teaches the agent's limits before the caseworker's discretion. Week one covers what the agent gathers, drafts and checks, and how to read its sources. Week two puts real cases in front of them with a senior colleague checking each decision. Entitlement decisions stay with the caseworker.
What does a caseworker do alongside the agent?
In a public service or advice organisation that already runs agents, the casework agent typically assembles the case file from records, checks eligibility rules against the facts provided, drafts letters and requests for information, and schedules follow-ups. It works at depth 1 and 2, sometimes depth 3 for scheduling within set rules. Decisions on eligibility, sanctions, support and anything a person could challenge stay with caseworkers.
That leaves the caseworker with the parts that need judgement: incomplete or conflicting information, people in difficult circumstances, discretion the rules allow, and explaining a decision so the person can understand and challenge it. How this reshapes the day is described in what changes in a caseworker's day when AI agents arrive. The casework team manager owns onboarding; an experienced caseworker coaches.
Why do public bodies need extra care?
Public bodies face additional transparency and fairness duties, and people affected by their decisions often have a right to an explanation and a review. Under GDPR, Article 22 restricts decisions based solely on automated processing that have legal or similarly significant effects, and the UK has an equivalent. Under the EU AI Act, some public uses of AI are classed as high-risk. A new caseworker does not need the law in detail, but they must understand why the agent drafts and they decide, and why their reasons must be written so a person could challenge them.
What does the two-week plan look like?
- Day 1. Person (team manager): introduces the agent's role and bounds, using the team's new joiner induction template so nothing is missed.
- Day 1 to 2. Person (new caseworker): learns to read sources. Records from the organisation's own systems are sealed. Documents from another agency are signed. What the person says on the phone is declared. The agent's eligibility check is only as good as the facts beneath it.
- Day 2 to 4. Agent, depth 2: drafts letters and eligibility checks on live cases, while the new caseworker sits with the coach, who explains each decision and where they departed from the draft.
- Day 5. Person (new caseworker): practises in a sandbox workspace on closed cases, making decisions from the agent's drafts and writing the reasons, then comparing with what was decided.
- Week 2. Person (new caseworker): decides live cases, each checked by the coach before letters go.
- Week 2. Person (new caseworker): records reasons wherever they change the agent's draft.
- End of week 2. Person (team manager): signs off solo decisions on routine cases, keeping complex or high-impact cases under review.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Assemble file | Pulls records with provenance | 1 | Asking for missing evidence |
| Eligibility | Checks rules against facts | 2 | The decision |
| Letters | Drafts decisions and requests | 2 | Sending, and the reasons |
| Scheduling | Books follow-ups within rules | 3 | Exceptions |
| Discretion | Nothing | none | All discretionary choices |
What must the new caseworker get right before sign-off?
Writing reasons a person could follow. "You do not meet criterion 3" is not enough. The person needs to know which fact led there and what they could provide. Caseworkers must rewrite agent drafts where reasons are thin.
Spotting vulnerability. People in crisis often mention it in passing. The agent should hand such cases to a person; the caseworker must notice when it did not.
Using discretion openly. Where rules allow discretion, the caseworker exercises it and records why. They must never ask the agent to apply discretion as a rule.
How should they record corrections?
Every change to an agent draft gets a short reason: missing fact, wrong rule, reasons unclear, vulnerability missed, or discretion applied. The last code is important: it tells the agent owner that the change was a choice, not an error, so it is never banked into the agent's brief. Corrections without personal details go to the agent owner; corrections with them stay in the case file.
Which controls and evidence should onboarding leave?
- Checked decisions. Evidence: each week-two decision shows caseworker and coach.
- Written reasons. Evidence: a sample reviewed for clarity by the manager.
- Staged rights. Evidence: dated access for routine decisions, later for complex ones.
- Sandbox practice. Evidence: closed cases attempted and compared.
What should the manager measure?
Decisions overturned by the coach in week two, by reason. Clarity of written reasons on a sample. Vulnerability cases the new caseworker caught that the agent missed. After sign-off: reviews and complaints on the caseworker's decisions over the first quarter.
Where does it usually go wrong?
The draft becomes the decision. A heavy caseload pushes new caseworkers to send the agent's letter unchanged. The supervised week must make clear that sending is deciding.
Learning the system, not the rules. A caseworker who can navigate the workspace but does not understand the eligibility rules cannot spot when the agent applied them wrongly.
No link to IT or compliance. New caseworkers need to know who to call when the agent behaves oddly; plans for a new IT administrator and a new compliance officer cover the other side of that call. The wider training approach is in training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy's Foundations track gives a new caseworker the core ideas in about four hours: why agents must be governed, the decision chain, and what the record shows. Caseworkers who will review agent drafts daily continue to the Operator track. Lessons can be read or listened to as audio, which helps people learning between appointments.
Questions people ask
Can a caseworker rely on the agent's eligibility check?
As a starting point. They must check the facts it used and the rule it applied, and they own the decision.
What if a person asks whether an AI was involved?
Be open about it. People should know when they interact with an AI system, and public bodies carry extra transparency duties.
How soon can new caseworkers handle complex cases?
After a period of accurate routine decisions, usually several weeks or more, depending on the team and the stakes.