Train people before the agent acts, not after. Teach everyone the shared vocabulary of depth, verdicts and the record. Train reviewers and team leads first, because they meet held actions first; then builders, owners and risk. Prove readiness with assessed decisions on real cases and certificates anyone can check, not with attendance lists.
Why do teams need training before an agent acts?
A governed agent does not remove people from the work. It moves them to the points where it stops: the held action waiting for a named person, the refusal someone must explain, the lapsed certificate that sends routine work back to the queue. Those points are where the risk concentrates, and they arrive on the first day the agent drafts anything.
Untrained people fail at those points in two predictable ways. Some approve whatever the agent drafts, because it looks tidy and the queue is long. Others block everything, because they cannot tell a sound draft from a weak one. The first group creates losses; the second makes the agent pointless. The case for firms that rely on judgement is made in accounting practices will not scale AI agents without trained reviewers, and it applies anywhere an agent's output passes through a person.
There is also an obligation. In the EU, Article 4 of the AI Act asks providers and deployers to ensure a sufficient level of AI literacy among the people who operate AI systems on their behalf. Elsewhere, supervisors expect named individuals to be competent for what they approve.
What should each group learn?
Start from the decisions each group makes, then teach what those decisions need.
| Group | What they must learn | How you know they learned it |
|---|---|---|
| Everyone who touches an agent | Depth 1 to 5, the four verdicts, provenance classes, lapse, the record | A short test set in their own industry |
| Reviewers | Reading a held action, checking sources, edit and approve, refusing with a reason, red lines | A run of real review decisions, marked |
| Team leads | Cover, calibration, drift, what to do when a certificate lapses | Calibration results and a drill |
| Builders | Writing the brief, choosing sources and bounds, reading Attest results | An agent certified on the business's own cases |
| Agent owners | What each depth means for their process, accepting risk, handover | A signed depth decision they can explain |
| Risk and compliance | Reconstructing a decision from the record, testing controls | An audit sample traced end to end |
| Board | The questions to ask and the answers to expect | A briefing and a recorded discussion |
Two points sit underneath the table. Everyone learns that withheld actions are a platform decision: the action exists and works, but no agent may take it at any depth, and nobody in the business can switch that off. And everyone learns that depth is earned, not configured, so a promise of more autonomy is a promise of more evidence first.
Who should be trained first?
Reviewers and team leads, before the agent goes live. They meet the agent's output first, and their decisions are what the agent's record will show to an auditor.
Can work start before everyone is trained? Partly. An agent at depth 1 only reads, and at depth 2 it drafts while a person sends, so a business can run the agent in shadow while reviewers train on its drafts. The detailed answer for one sector is in whether accountancy firms can use AI agents before staff are trained. The rule that travels: no agent should reach depth 3 or above on a process until the people handling its held actions are certified.
Builders and owners come next, because their decisions shape what reviewers will see. Risk follows, once there is a record to test. The full sequence is set out in deploying an AI agent: train, shadow, certify, widen.
How should the training be delivered?
Mostly in the work. The 70-20-10 model holds that most workplace learning comes from doing, some from colleagues and a smaller share from formal courses. Agent work fits that shape well. The formal part teaches the vocabulary; the real learning happens in the held queue, where each case is a lesson with a known right answer.
Three methods carry most of the weight:
- Shadow review. While the agent runs at depth 1 or 2, reviewers decide its drafts and compare their decisions with an experienced colleague's. Disagreements become the next session's material.
- Calibration. Every few weeks, reviewers decide the same small set of cases separately, then discuss where they differed. This keeps a team's judgement consistent as people join and leave.
- Drills. Walk through an incident before one happens: a lapsed certificate, a supplier fraud attempt, a refusal a customer complains about. The AI agent incident drill script template gives a ready script with roles, injects and what good looks like.
Keep formal modules short and set in the learner's own industry. A claims handler learns faster from a held claim than from a generic example about invoices.
How do you prove your people are ready?
With evidence that would satisfy someone who was not in the room. Attendance shows people were present; it does not show they can decide a held payment correctly.
Readiness evidence has three layers. First, a test of the shared vocabulary, scored by concept so you can see where the team is weak. Second, a run of real review decisions for each reviewer, marked for correctness and for any missed red line. Third, a certificate per person that names what was assessed, carries an expiry date and can be checked at source with the issuer, so a supervisor, auditor or client does not have to trust a PDF.
Keep it all in one register next to the agents' own certificates. A team lead should see, for each agent, whether its certificate is current and whether the people approving its held actions are current too. A business can adopt the practice of tying approval rights to a current certificate, depth by depth, so the register also decides who may approve what.
How do you keep training current?
Agents change and so does the work. An agent's certificate lapses after 90 days; a run that clears no depth drops it to depth 1 the same day. Sources change, bounds narrow, new capabilities arrive. People trained on last quarter's agent are partly trained for this quarter's.
Build three habits. Treat every refusal and every overturned approval as material for the next calibration. Run a short update whenever an agent gains a capability or depth. And recertify people on a fixed cycle, usually yearly, sooner when they change role. New joiners need a defined path too, because they arrive into a team where the agent already handles the routine work they would once have learned on.
How do you get the board to back it?
Frame training as a control, not a perk. The board approves the risk appetite for agents; training is how the business keeps people able to hold that line. Show which agents act at depth 3 or above, who approves their held actions, and whether those people are certified and current.
The guide on how to brief a board on AI agent training gives the structure: the agents, the decisions people still make, the evidence of competence, and the gaps with a date to close each. Boards rarely refuse a control they can see working.
Where the Academy fits
Pryme Intelligence Academy follows this order. Foundations gives everyone the vocabulary for free. Operator certifies reviewers and team leads through an exam and a two-week practicum of real review decisions signed off by a Pryme assessor. Module tests are set in each learner's industry and role, and a free readiness assessment for teams of up to 25 shows scores by concept, the gaps, and the first agent to start with.
Questions people ask
How long does it take to train a team before an agent goes live?
It depends on the depth you want and how many held actions the team will handle. The shared vocabulary takes hours; reviewer competence takes weeks of real decisions. Plan the agent's shadow period to match, so reviewers practise on its drafts before it acts.
Should training be run by learning and development or by operations?
Both, with operations owning the content. Learning teams can run scheduling, tracking and delivery, but only the people who own the process know what a correct decision on a held case looks like. Ask the process owner to sign off every assessment.
What happens to training records when someone leaves?
Keep them with the record of the decisions they made. If a past approval is questioned, you will need to show the person was certified when they made it. Remove their approval rights the day they leave, and update the register.