Yes, for anyone who approves an agent's held actions or designs agents, and the requirement should be competence, not attendance. For everyone else who works beside agents, a short shared course should be expected but need not gate their job. The strongest form of "mandatory" is simple: no current certificate, no approval rights.
Why make it mandatory for approvers?
Because their decisions are the control. A governed agent stops at defined points: it holds an action for a named person, refuses what is outside its bounds, or is withheld from acting at the platform floor. Every held action is a decision that a person makes on the business's behalf. If that person cannot read the agent's sources, recognise a red line or give a reason for declining, the hold protects nothing.
There is also an external case. 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. In the UK, the Senior Managers and Certification Regime puts named individuals behind regulated activities, and they will want to show the people approving agent work are competent. Neither is satisfied by an optional course that most people skip.
Who should it be mandatory for?
| Group | Mandatory? | What it should mean |
|---|---|---|
| Reviewers who approve held actions | Yes | An assessed certificate, current, before approval rights are granted |
| Team leads over approval queues | Yes | The same certificate, plus calibration duties |
| Builders who design and certify agents | Yes | A certificate showing they built an agent that passed certification |
| Agent owners who sign depth decisions | Yes | Enough training to explain each depth they approve |
| Staff who work beside agents but approve nothing | Expected | The shared vocabulary, ideally within their first weeks |
| Executives and board members | Expected | A briefing and the vocabulary, so they can ask the right questions |
"Mandatory" in the first four rows should mean a pass on an assessment that tests decisions, not a completion tick. The idea of mandatory training in most organisations is attendance-based; for agent approvers it needs to be evidence-based.
How should approval rights depend on it?
The cleanest practice is to link rights to depth. A business can adopt a rule like this:
- Depth 1 and 2 work (the agent reads or drafts and a person sends): anyone who has completed the shared vocabulary may handle drafts, under supervision until certified.
- Held exceptions on depth 3 capabilities (the agent acts within bounds and holds exceptions): only certified, current reviewers may approve.
- Depth 4 approvals (money moves or actions are hard to undo): only certified, current reviewers with the role's authority level.
The rule enforces itself if the approval queue checks the certificate. When a certificate lapses, the person drops off the queue that day, the same way an agent drops to depth 1 the day its certificate lapses. Nobody needs to remember to remove them.
How long does mandatory training take?
It depends on the role and the volume of real decisions available for practice. Claims handlers and AML investigators, for example, face very different red lines and case volumes; the timings are compared in how long it takes to train claims handlers to supervise AI agents and how long it takes to train AML investigators. Plan the mandatory period into the agent's shadow phase so reviewers practise on real drafts.
An example
Take a hypothetical bank's onboarding team of 30. Before go-live, the bank makes reviewer certification mandatory for anyone approving held account openings. Twenty-two pass in the first round; eight need another attempt. The agent goes live at depth 3 for low-risk personal accounts, and only the 22 see held cases. The other eight keep working on depth 2 drafts under supervision until they pass. Methods for that team are covered in how to train the onboarding team to review AI agent decisions.
Where the Academy fits
Pryme Intelligence Academy's professional certificates expire, 12 months for Operator and Builder and 24 for Architect, and a lapsed certificate shows as lapsed when checked at source. That makes them usable as the gate for approval rights. The Foundations course is free, which makes the expected level easy to require.
Questions people ask
Does mandatory training make people resent it?
Attendance-based training often does. Training that decides whether someone may approve their own team's work tends to be taken seriously, because it is tied to the job.
What happens if someone keeps failing?
They keep working at drafting level under supervision. Repeated failure usually points to a gap in coaching, which the team lead should address. The finance version is set out in how to train the finance team to review AI agent decisions.
Who decides who is mandatory?
The process owner, recorded in the training register. The guide to training teams for AI agents explains how that register works.