Train by role, not by tool. Approvers learn to read a held item and reject with confidence; reviewers learn to sample and classify errors; exception handlers learn what the agent passes back and why; everyone learns to report problems and take work back. Teach with real cases from shadowing, including the agent's mistakes, assess before go-live and refresh regularly.
What do staff actually need to learn?
Not how the model works. Staff need to know what this agent does in their process, how far it may act on each task, what they are responsible for, and what to do when it is wrong. That is practical knowledge about one agent in one team, and it changes when the agent's depth changes.
The team leader owns the training, because they know the work and the people. The agent owner supplies the content about the agent. Learning and development can help with format and records. In the EU, the AI Act's AI literacy provision, Article 4, expects deployers to ensure their staff have sufficient understanding of the AI systems they work with, which gives this training a regulatory as well as practical purpose.
Training sits inside a wider change effort. How you tell staff about the agent in the first place is covered in communicating an AI agent rollout to staff, and the plan itself in the staff communication plan template.
What are the steps?
- Team leader: lists the roles people will play around the agent: approver of held items, quality reviewer, handler of refused and exception work, member of the fallback rota, and everyone else who receives its output.
- Agent owner: writes what each role must know: what the agent does at what depth, the four verdicts (allowed, held, refused, withheld), how provenance labels on facts work, what to check, when to reject, how to report a problem.
- Agent, depth 1: assembles training cases from shadow and early live runs: clean cases, cases where the agent was wrong, cases where it held correctly, cases with weak sources.
- Team leader: runs sessions by role, working through real cases, not slides. Approvers practise on held items with known flaws.
- Team leader: assesses each person before they take on their role: can they spot the flawed item, explain why a case was held, and say how they would take work back?
- Team leader: records completion and assessment. Evidence: training records per person and role.
- Agent owner: updates the content whenever the agent's depth, scope or sources change materially.
- Team leader: runs refreshers, at least every certificate cycle, using the latest cases.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Roles | Nothing | none | Team leader |
| Content | Nothing | none | Agent owner |
| Cases | Assembles from records | 1 | Choosing which to use |
| Sessions | Nothing | none | Team leader |
| Assessment | Nothing | none | Team leader |
| Records | Nothing | none | Team leader |
| Updates | Flags changes to depth or scope | 1 | Rewriting content |
| Refreshers | Nothing | none | Team leader |
Why teach with the agent's mistakes?
Because the main risk is not that staff distrust the agent; it is that they trust it too much. An agent that is right most of the time trains people to stop looking. Showing them real cases where it was wrong, and asking them to find the error, builds the habit of checking. It also shows that finding errors is the job, not a failure of the agent or the team.
Use cases where the error was subtle: a correct-looking letter with the wrong rate, a claim registered with a loss date that the claimant later corrected, a supplier change that came from a lookalike address. Obvious errors teach little.
Which controls make this safe?
Role-based assessment before go-live. No one approves held items until they have passed. Evidence: assessment results.
Training linked to depth. When a capability's depth changes, the people affected are retrained before it takes effect. Evidence: training dates against depth change dates.
A clear route to report problems. Staff know how and to whom, and that reports are welcome. Evidence: problem reports logged, with responses.
A practised fallback rota. Some staff keep doing the work by hand on a rota, so the team can take work back. Evidence: rota and practice records.
Approver capacity planned. Training is pointless if approvers have no time. How to staff that work is covered in how to staff the queue of actions AI agents hold.
What are the signs it is going wrong?
Approvers never reject. Across hundreds of held items, nothing is changed. Either the agent is perfect or no one is reading.
Staff redo the agent's work quietly. They do not trust it, but do not say so, and check everything twice. Effort doubles and no one learns why.
No one reports problems. Silence usually means people do not know how, or fear blame.
Staff cannot do the work by hand. When the agent drops to depth 1, the team struggles. The fallback rota has lapsed.
People ask who is responsible. "If I approve it and it's wrong, is that my fault or the agent's?" This question is the most common source of unease, and the reason staff resistance is usually about accountability, not jobs. Answer it in training, plainly: the approver answers for what they approve, on the information shown to them; the agent owner answers for what the agent shows.
How do you get staff to accept the agent?
Training helps acceptance more than persuasion does. Staff who understand exactly what the agent does, what it may not do, and what they are accountable for are more willing to work with it. The broader approach is in how to get staff to accept AI agents. In training, be specific: which tasks change, which stay, and how their judgement is still needed.
Where Pryme fits
In Pryme, held items show the facts behind the agent's proposal with their provenance stamps, so approvers can see what came from the system of record and what came from outside. That makes the training content concrete: staff learn to read the stamps, not to trust the agent.
Questions people ask
How long does training take?
For most roles, a few hours of case-based sessions before go-live, then shorter refreshers. Approvers of high-value items need more practice.
Should the agent's builders run the training?
They should help write it, but the team leader should run it. Staff need to hear it from someone who does the work. The steps in train, shadow, certify, widen only hold if the team owns them.
Do new joiners need the same training?
Yes, before they take on any role around the agent, including the assessment.