A programme. Only the shared vocabulary can be taught once. Everything else changes underneath your people: agents gain capabilities and depth, new agents arrive, certificates lapse, staff join and move, and the product itself moves on. Each of those changes needs a defined training response, so the work never finishes; it settles into a rhythm.
Why can it not be a one-off?
Because the thing people supervise is not fixed. A governed agent's authority is granted per capability, proven on the business's own cases, and it expires. Its certificate lapses after 90 days and is renewed on fresh evidence; a run that clears no depth drops it to depth 1 the same day. Sources are added and retired. Limits are narrowed after an incident or widened after a good quarter. A reviewer trained on the agent's first version is supervising a different agent six months later.
People change too. In most operations, a meaningful share of the people approving held actions next year are not the people approving them today. Every new joiner arrives into a team where the agent already handles the routine work they would once have learned on, so they need training to understand what they are now responsible for.
What does the programme have to cover?
Five recurring triggers, each with a defined response:
| Trigger | Training response | Who |
|---|---|---|
| A new agent is deployed | Reviewer training on that agent's capabilities, sources and red lines, with practice during its shadow period | People who will approve its held actions |
| An agent gains a capability or steps up in depth | A short update before the change goes live | People whose queue changes |
| A person's certificate approaches expiry | Renewal: an update and a short test | Every certified person |
| A new joiner or role change | The shared vocabulary, then the role's path, before approval rights are granted | The individual |
| An incident, overturned approval or missed red line | A walkthrough of the record, then a calibration case | The team, then everyone |
The only part that can be close to one-off is the vocabulary: depth, the four verdicts, provenance classes, lapse, the record. Even that needs a refresh when the product changes materially.
What stays the same, and what keeps moving?
The method stays the same: practice on real held decisions, calibration across reviewers, feedback with reasons, assessment by someone who knows the process. The content keeps moving with each agent.
That is why sector programmes look different in their content but alike in their shape. A hospital's programme is built around patient identity and clinical boundaries, as described in what an AI agent training programme for hospitals should include. An insurer's is built around authority levels and claims evidence, as in what a programme for insurers should include. Both renew, both expand with each new agent, and both have an owner.
An example over eighteen months
Take a hypothetical insurer. In the first quarter, 20 claims handlers are certified for an intake agent. In the second, the agent starts settling glass claims at depth 3, and the same 20 take a one-hour update. In the third, a second agent arrives for complaints, and 12 complaints handlers start the reviewer path. In the fourth, 5 of the original 20 have left; their replacements need the full path before they approve anything. By month 13, the first certificates are due for renewal. By month 18, a missed red line on a fraud indicator has produced a new calibration case used across both teams.
None of that was a one-off. All of it was predictable, which is why it can be planned.
How do you run it as a programme?
Give it an owner, a register and a calendar. The owner is usually the process owner, with learning and development running delivery. The register shows every agent's certificate and depth next to every reviewer's certificate and expiry. The calendar shows renewals, planned depth changes and new agents for the next two quarters.
Build internal capacity as well. Team leaders who can coach and calibrate keep the programme running between formal courses; how to run a train-the-trainer programme for AI agents covers how to develop them. If you are starting with nothing, how to build an AI agent training programme from scratch gives the first steps. Pryme Intelligence Academy's certificates carry fixed validity, 12 months for Operator and Builder and 24 for Architect, which gives the calendar its fixed points.
Questions people ask
Can we buy the programme rather than build it?
You can buy the courses and the certification. You cannot buy the knowledge of your own red lines, cases and limits, which is the core of the programme. The argument is made in training is the part of AI adoption you cannot buy off the shelf.
How much ongoing effort does it take?
Most of it happens in the work: calibration sessions, short updates and practicum supervision. The formal parts are renewals and new-agent training.
Who should own it?
The person who owns the process the agent works in. The guide to training teams for AI agents explains why operations owns the content and learning teams run delivery.