A table that starts from decisions, not courses. A training needs analysis for AI agents lists each agent and what it may do, each role that touches it, the decisions that role makes, the skills those decisions need, where people are today, and the gap. Only then does it name a method. Done well, it fits on a few pages.
Why start from decisions?
Because the agent changes what people decide, and training should follow the decisions. A claims handler who used to assess every claim now approves the claims the agent holds. A finance clerk who used to key invoices now checks the agent's matches. A team lead who used to allocate work now calibrates reviewers and watches the agent for drift. Each of those is a new decision with a new skill behind it.
Starting from a course catalogue gets this backwards. You end up sending everyone on the same general AI course, which teaches nobody the decisions they actually face. The general idea of a training needs analysis is the same everywhere; what is particular to agents is that the decisions change with each agent's depth.
What are the steps?
- List the agents and their capabilities. For each capability, note its depth now and the depth you plan: read, draft, act within bounds, move money.
- List the roles that touch each capability. Reviewers, team leads, builders, owners, risk, audit, and people who simply work beside the agent.
- Write each role's decisions. "Approves held refunds above the limit." "Signs the decision to raise invoice matching to depth 3." "Traces a sample of verdicts each quarter."
- Name the skills behind each decision. Reading a held action, checking provenance, recognising a red line, giving a reason, running calibration, reading certification results.
- Assess the current level. A concept test for the team, observation of real decisions, and the team lead's view.
- Record the gap and its priority. Gaps on capabilities that will reach depth 3 or 4 come first.
- Choose the method. A course for concepts, a practicum for judgement, calibration to keep it, a drill for rare events.
What does it look like on the page?
| Role | Decision | Skills needed | Current level | Gap | Method |
|---|---|---|---|---|---|
| AP clerk | Approves held invoice matches | Provenance; red lines on bank detail changes | Knows the invoice process, not provenance | High | Course, then practicum on the live queue |
| AP team lead | Calibrates clerks; investigates drift | Running calibration; reading the workspace | Strong on process, new to agents | Medium | Reviewer path plus calibration training |
| Finance director | Signs depth decisions | What each depth means; reading the record | Has the vocabulary | Low | Briefing and one guided trace |
| Systems accountant | Designs the matching agent | Brief, sources, bounds, certification results | Strong systems skill | Medium | Builder path |
| Internal audit | Traces verdicts | Reading the record | None | Medium | Short course and a guided trace |
A ready layout is in the training needs analysis template for AI agents, and the drafting steps are in how to write a training needs analysis for AI agents.
What makes a good one?
Three tests. Every decision with a held action behind it has a named role. Every role that approves at depth 3 or above has a method that includes real decisions, not just a course. And the analysis has a date to be redone, because the next agent or the next step in depth will change the table.
A good analysis also notices where the work itself can teach. Many gaps close fastest in the review queue, with a lead discussing hard cases, as the review queue is a classroom, if you run it like one argues. Name that as a method in the table, with time allowed for it.
How does it change for a small business?
It shrinks, but the columns stay. A business of ten may have one row per person. The analysis still needs the decision column, and it still needs at least two people able to approve each held action. The lean version is described in what good AI agent training looks like for a small business.
Where the Academy fits
Pryme Intelligence Academy's tracks map onto the columns: Foundations for the shared vocabulary, Operator for reviewers and team leads, Builder for agent designers, Architect for people leading regulated multi-agent deployments. Its free readiness assessment for teams of up to 25 fills the current-level column by concept.
Questions people ask
Who should write the analysis?
The process owner with learning and development. Operations knows the decisions; learning teams know the methods.
How detailed should the skills column be?
Specific enough to test. "Understands AI" cannot be tested; "can tell a declared fact from a sealed one in a held draft" can.
When should it be redone?
Before each new agent, each step up in depth, and at least yearly. The guide to training teams for AI agents covers the wider cycle.