Start with reviewers who already decide well, not with keen volunteers. Certify them as reviewers first, then teach them to teach: explaining why the agent held a case, running calibration, and coaching a new reviewer through shadowing. Observe each trainer deliver twice, sign them off against a written standard, and recheck them whenever the agent changes.
Why train trainers at all?
One central trainer cannot keep up once agents spread across teams. Each new agent brings reviewers who need lessons, practice, calibration and sign-off, and every change to an agent's depth or bounds means retraining. A small group of internal trainers in each function solves the capacity problem, and puts the teaching close to the work.
The risk is dilution. Each layer of retelling loses detail, and a trainer who half-understands why the agent refuses an action will pass on the half. A train-the-trainer programme exists to stop that: it qualifies people to teach, and it checks them.
The head of learning and development usually owns the programme. Each function's operations lead nominates trainers and owns their time.
What are the steps?
- Operations lead: nominates candidates from reviewers with a strong sampling record, not from whoever has free time.
- Candidate: completes reviewer certification for the agent they will teach. A trainer must be able to do the job before teaching it.
- Programme owner: teaches the trainer content. How a governed agent decides, the five bounds, the four verdicts, what depth means for this agent, and the common misconceptions new reviewers bring.
- Agent, depth 1: assembles a teaching case library. The trainer asks for held, refused and corrected cases with verdicts and the deciding bound, then selects and anonymises them.
- Candidate: co-delivers a lesson and a calibration session with an experienced trainer.
- Candidate: delivers alone, observed twice. The observer scores against the trainer standard.
- Programme owner: signs off the trainer for named agents and records it.
- Trainer: re-qualifies when the agent's depth, sources or bounds change, through a short briefing and one observed session.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Nominate | Nothing | none | Operations lead |
| Certify as reviewer | Nothing | none | Candidate, assessor |
| Trainer content | Nothing | none | Programme owner |
| Case library | Lists held and refused cases | 1 | Trainer selects and anonymises |
| Co-deliver and observed delivery | Nothing | none | Candidate, observer |
| Sign-off | Nothing | none | Programme owner |
| Re-qualify | Lists cases affected by the change | 1 | Trainer, programme owner |
What should trainers be able to do?
Write the trainer standard down. Five abilities cover most of it.
- Explain a verdict from the record. Given any held item, the trainer can show which bound decided it and why, without guessing.
- Distinguish refused from withheld. Refused means outside this agent's bounds, which the customer can change. Withheld is a platform decision for every agent at every depth. New reviewers mix them up constantly.
- Run calibration. Collect independent decisions, show the spread, steer discussion to reasons, and record a gold standard without imposing their own view.
- Coach a shadow reviewer. Sit alongside, ask "what would you check next?", and resist taking over.
- Spot when the agent has changed. If drafts start looking different, the trainer asks the agent owner before teaching the old behaviour.
Notice what is missing from that list: presentation polish. A trainer who speaks well but cannot trace a verdict to its bound will mislead confidently. A quieter trainer who can open the record and show exactly why the agent stopped will produce better reviewers.
The train-the-trainer guide template for AI agents turns these into an observation form.
Which controls keep quality from drifting?
Reviewer certification before trainer status. Evidence: the certificate, checked at source and current.
Observed delivery. Evidence: two observation forms per trainer.
Scope by agent. A trainer signed off for the invoice agent is not automatically signed off for the claims agent. Evidence: the sign-off record lists agents.
Re-qualification after change. Evidence: a dated briefing and observation for each change.
Expiry tracked. Reviewer certificates lapse after 12 months unless renewed. A trainer whose own certificate lapses stops training until it is renewed. Evidence: the expiry date in the training register.
What should you measure?
- Results of reviewers each trainer taught: practicum outcomes, early sampling accuracy.
- Calibration agreement in sessions each trainer facilitates, over time.
- Questions escalated to the programme owner, which show where the trainer content is thin.
- Trainer capacity versus demand per function, so you can nominate before a backlog forms.
Where does it usually go wrong?
Trainers chosen for enthusiasm. Keenness is welcome, but a trainer must decide well first.
No observation. Programmes that skip observed delivery cannot tell good trainers from confident ones.
Old behaviour taught. The agent's brief changed, nobody told the trainers, and new reviewers learn last month's standard.
Trainers become the agent's owner by default. Trainers teach and calibrate; changes to the agent still go through its owner and certification.
One-off thinking. Treating training as an event rather than a rhythm starves the trainer network. The case is made in is AI agent training a one-off or a programme, and how to build an AI agent training programme from scratch shows where trainers fit. For a sector example, see what an AI agent training programme for lenders should include. The wider picture is in training teams for AI agents.
Some organisations use an external certificate as the entry gate. Pryme Intelligence Academy's Certified Pryme Operator is one option: a written exam of 40 questions, a pass mark of 32, and a two-week practicum signed off by a Pryme assessor, checkable at source by number or QR code.
Questions people ask
How many trainers does a function need?
Enough that no trainer teaches more than they can observe properly. In practice, one or two per team of reviewers, with at least one cover for absence.
Should trainers be paid more or given a title?
Recognition helps retention. At minimum, protect their time in workload planning; unprotected trainers stop training when the queue rises.
Can a trainer teach a team they do not work in?
Only after learning that team's agent and its held cases. Teaching skills transfer; knowledge of the specific agent and process does not.