Training evaluation is the systematic assessment of whether a training programme achieved what it set out to do, from how participants found it through to what changed in their work. For teams overseeing AI agents, the question it answers is concrete: are people making better decisions on the agent's held actions, drafts and refusals because of the training?
Why does training evaluation matter?
Agent oversight training is a control. Like any control, it should be tested. Without evaluation, an organisation knows it spent time and money on training and has no evidence it reduced risk. If an auditor asks how you know reviewers are competent, attendance lists are not an answer.
Evaluation also improves the training. It shows which parts land and which do not, so the next cohort gets a better version. Programmes that are never evaluated tend to stay as first drafted.
How does it work in practice?
A common approach looks at four levels, each harder to measure and more meaningful than the last:
| Level | Question | Evidence for agent oversight |
|---|---|---|
| Reaction | Did people find it useful? | Short survey after each session |
| Learning | Did they learn the concepts? | Test scores by topic: verdicts, sources, limits, red lines |
| Behaviour | Do they apply it at work? | The agent's record: what they approved, sent back, escalated |
| Results | Did outcomes improve? | Fewer wrong approvals, fewer complaints, faster safe resolution |
The third level is where agent work makes evaluation easier than usual. Every verdict and every human decision on a held action is filed in an append-only record. You can see whether a trained reviewer sends back drafts that rely on external sources, rather than guessing.
Findings belong in a lessons learned log so the next version of the programme uses them.
An example
A logistics company evaluates its training for dispatchers who review an agent's carrier bookings. Reaction scores are high. Test scores are good on limits, weak on provenance. In the record, trained dispatchers still accept carrier rates from forwarded emails in cases where the contract system disagrees. The training is revised to add scenarios on provenance; a month later, the record shows those acceptances have largely stopped. The content of such programmes is set out in what an agent training programme for logistics companies should include, with a parallel piece for contact centres.
How is training evaluation different from AI evaluation?
AI evaluation tests the agent: whether its answers are grounded, whether it refuses what it should, whether it stays in scope. Training evaluation tests the people and the programme that prepared them. Both feed the same decision about how much depth the agent can be trusted with, but they use different evidence and different owners.
Frameworks for testing AI systems, such as the NIST AI RMF's approach to testing, evaluation and validation, cover the system side. Training evaluation covers the human side.
Which terms are related?
- Training transfer: whether learning shows up at work, the behaviour level.
- Training completion rate: a process measure, not an evaluation.
- Summative assessment: one source of evidence at the learning level.
For the full approach, read training teams for AI agents.
Questions people ask
When should training be evaluated?
At each level in turn: reaction straight after, learning at the end, behaviour after a few weeks on the live queue, results over a few months.
Who should evaluate the training?
Someone other than the person who designed it, using the record and test results rather than opinion.