A training needs analysis for AI agents works role by role: write down the decisions each role makes about agents, list the skills those decisions need, measure where people stand today, and name the track that closes each gap, with a priority and deadline tied to go-live. The template below is a ready-to-copy table plus a skills list.
What is a training needs analysis for, with agents?
A training needs analysis compares the skills a job requires with the skills people have, then decides what training closes the difference. It stops two common mistakes: training everyone in everything, and training nobody properly because the budget went on a general course.
Agents change the skills a role needs without always changing the job title. An accounts assistant who used to prepare reconciliations now reviews the agent's reconciliations. The work looks similar on an organisation chart, but the skill is different: judging someone else's work, checking its sources, knowing which bound held it. The analysis has to start from what people will decide once the agent is live, not from what they do today. A worked illustration of the finished output is in what a training needs analysis for AI agents looks like.
How do you fill it in?
- List the agent's capabilities and their depth. This tells you which decisions will reach people.
- List the roles that meet those decisions. Include approvers, team leads, builders, owners, risk and audit.
- For each role, pick the skills it needs from the skills list below.
- Measure the current level. Use a short test set in the team's own work, scored by concept, plus a learning needs survey for self-reported confidence. Trust the test over the survey where they disagree.
- Name the gap and the track. One track per role, at the lowest level that closes the gap.
- Set priority and deadline against the agent's timeline: approvers of depth 3 and 4 actions first.
The full method is in how to write a training needs analysis for AI agents.
The template
Skills list (choose from these)
- S1 Vocabulary: depth 1 to 5, the four verdicts, provenance classes, lapse
- S2 Reading a held action and the bound that held it
- S3 Checking sources: sealed, signed, declared, external
- S4 Applying the team's red lines
- S5 Refusing or rejecting with a recorded reason
- S6 Handling a lapsed agent certificate or a drop in depth
- S7 Running calibration and cover
- S8 Writing an agent brief, choosing sources and bounds
- S9 Reading certification results and making a depth decision
- S10 Reconstructing a decision from the record
Analysis table
| Role | Decisions about agents | Skills needed | Current level (test and survey) | Gap | Track to close it | Priority | Deadline |
|---|---|---|---|---|---|---|---|
| [Role] | [What they approve, design, own or audit] | [S codes] | [Scores by concept] | [S codes below standard] | [Foundation / reviewer / builder / architect] | [High / medium / low] | [Relative to go-live] |
Worked example
An illustrative logistics company plans an agent for its management accounts team. The agent will prepare monthly variance analysis at depth 2, drafting commentary for accountants to check, and post routine accruals at depth 3. Whether an agent can do that work well is discussed in can an AI agent do variance analysis; this analysis is about the people.
| Role | Decisions | Skills needed | Current level | Gap | Track | Priority | Deadline |
|---|---|---|---|---|---|---|---|
| Management accountants (6) | Approve variance commentary; approve accruals over threshold | S1 to S6 | Strong on S4; weak on S2, S3 and S6 | S2, S3, S6 | Certified reviewer | High | Before shadow period ends |
| Finance manager (1) | Calibration, cover, escalations | S1 to S7 | Weak on S6, S7 | S6, S7 | Certified reviewer plus lead duties | High | Before shadow period ends |
| Systems accountant (1) | Builds and maintains the agent's brief | S1 to S9 | No exposure to S8, S9 | S8, S9 | Certified builder | High | Before certification run |
| Financial controller (1) | Signs depth decisions | S1, S9 | Weak on S9 | S9 | Foundation plus depth decision briefing | Medium | Before depth 3 |
| Internal audit (2) | Samples decisions | S1, S10 | Weak on S10 | S10 | Foundation plus record tracing | Low | Within two months of go-live |
The test showed something the survey did not: the accountants rated themselves confident on sources, yet most treated a budget holder's emailed explanation of a variance as verified. That is why S3 is a gap despite high self-ratings.
What usually goes wrong with these analyses?
Three things. Analysing job titles instead of decisions, so a role whose work changed completely shows no gap. Relying on self-assessment alone, which overstates skills people have never had to use. And forgetting the people around the queue: leads, owners and auditors, who need different skills from reviewers. The guide to training teams for AI agents sets out each group's needs.
Questions people ask
How long does a training needs analysis take?
For one team and one agent, a few days: a day to map decisions and skills, a short test and survey, and a day to agree the table with the agent owner.
Who should sign off the analysis?
The agent owner, because the analysis sets who will be competent to approve the agent's work. The learning lead runs it.
When should it be repeated?
Whenever the agent gains a capability or a depth, and for each new team. Re-test a sample of people yearly to see whether the gaps closed.