Build an AI agent training programme by starting from the agents, not the curriculum. List each agent, its depth and who approves its held actions; map each role to the training it needs; train highest-risk approvers first; prove skill on real review decisions; and plan renewals before the first certificates lapse. The programme runs as long as the agents do.
What is the programme for, and who owns it?
The programme exists so that every person who oversees, approves or builds an agent is competent to do it, and so the organisation can prove it. That is a narrower and more useful goal than "AI upskilling". It ties training to authority: who may approve a held payment, who may change an agent's brief, who answers for a multi-agent deployment.
Ownership usually sits with learning and development for delivery, with an operations or risk executive as the accountable sponsor, because the programme is a control. Whether this is a one-off push or a standing programme is discussed in is AI agent training a one-off or a programme; the short answer is a programme, because agents change, people move, and certificates expire. The wider picture of what to teach and who first is in training teams for AI agents.
What are the steps?
- Agent, depth 1: reads the agent register and the record, listing each agent, its depth, and the people who approved, corrected or overrode its actions in the last quarter.
- Person (programme owner): runs the needs analysis. From that list, group people into four kinds: those who need awareness only; those who review and approve daily; those who design and certify agents; those who lead regulated or multi-agent deployments.
- Person (programme owner): builds a skills matrix showing each role, the level required, and who holds it now.
- Agent, depth 2: drafts a schedule that trains approvers of the highest-depth agents first, for the owner to adjust around operational peaks.
- Person (team leads): release people for learning and for supervised practice on real work.
- Person (assessor or senior reviewer): signs off competence on real decisions, not just tests.
- Person (programme owner): plans renewals and reports to the sponsor quarterly.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Inventory | Reads register and record | 1 | Deciding which agents matter most |
| Needs analysis | Nothing | none | Grouping people by need |
| Skills matrix | Nothing | none | Setting required levels |
| Schedule | Drafts training order | 2 | Fitting it to operations |
| Assessment | Nothing | none | Signing off competence |
| Renewals | Lists expiries | 1 | Acting on them |
How do you decide who needs what?
Use depth. The higher an agent's depth, the more harm a wrong approval can do, and the more skill the approver needs. People who approve depth 4 actions, such as payments or hard-to-reverse changes, need to understand the decision chain thoroughly: the five bounds every action meets, why a verdict was held, refused or withheld, and how to read the record. People who only read depth 1 outputs need much less.
A skills matrix makes this visible. Rows are roles, columns are the skills, and cells show required level against current level. Building a skills matrix for an AI agent team walks through it. Different sectors weight the columns differently; a fintech, for instance, puts payment holds and provenance near the top, as what an AI agent training programme for fintechs should include describes.
How do you prove skill, not attendance?
Attendance proves nothing about competence. A credible programme has three kinds of evidence.
A test on concepts, set in the learner's own industry and role, so a claims handler is tested on claims cases.
Supervised practice on real work. Two weeks of real review decisions, checked by a senior reviewer, shows whether someone can actually approve, correct and escalate. The bar should be high, and a missed red line should fail the practice period on its own.
A certificate that can be checked. If a supervisor or auditor asks whether a named person was competent on a given date, the answer should be verifiable at source, and a lapsed certificate should show as lapsed.
How do you scale beyond the first team?
Train trainers. Senior reviewers who have been through the programme and done well become the people who supervise practice for the next group. Running a train-the-trainer programme for AI agents covers how to pick and support them. Keep the trainers' own certificates current; a trainer whose certificate has lapsed should not be signing off others.
What controls and evidence should the programme leave?
- A skills matrix kept current. Evidence: the matrix with a date and owner.
- Approval rights tied to competence. A business can make approval rights depend on a current certificate, depth by depth. Evidence: the access list compared with certificates.
- Supervised practice records. Evidence: the decisions reviewed and the sign-off.
- Renewal tracking. People's certificates typically run 12 or 24 months depending on level. Evidence: expiry dates and actions taken before them.
What should you measure?
Measure coverage: approvers of each agent who hold a current certificate. Measure quality from the record: accuracy of review decisions, overturned approvals, missed red lines, before and after training. Measure speed to depth: how long it takes a new agent to reach its target depth, since trained reviewers produce cleaner evidence in Attest. And report all of it in a consistent format; the training impact report template for AI agents gives one.
Where does it usually go wrong?
Starting with content. Teams buy or build courses before knowing which agents and roles matter, and train everyone the same.
Training once. The first wave is certified, nobody tracks expiry, and within a year a large share of approvers are lapsed without anyone noticing.
Measuring completions. A dashboard of completed modules tells the sponsor nothing about whether approvals got better.
Leaving out agent owners. The people who change briefs and run certification need deeper skills than reviewers, and are often the last to be trained.
Questions people ask
How long does it take to set up a programme?
The needs analysis and skills matrix can be done in weeks if the agent register is in order. The first certified approvers take longer, because supervised practice runs on real work.
Should everyone in the organisation be trained?
Everyone who touches agents should have awareness. Deeper training should follow authority: the more someone can approve or change, the more they need.
Who should pay for it: L&D or risk?
Delivery often sits with L&D, but the case is a risk case, because training is a control on agent actions. Many organisations fund it from both.