What is a worked example?

A worked example is a problem shown together with its full solution, step by step, so the learner can follow the reasoning before trying similar problems alone. In training for AI agent oversight, it is usually a single case from the queue, with each check the reviewer made and why, ending in the correct decision.

Why do worked examples matter?

Newcomers to agent review do not yet know what to look at. Given a held payment and asked to decide, they often stare at the amount and miss the source. A worked example shows the order of attention: the verdict, the bound that decided it, the source class of each document, the limit, then the decision. Seeing that sequence a few times builds the habit.

They also reduce early errors. Learners who start by solving cases from scratch tend to practise their mistakes. Learners who first study solved cases, then attempt similar ones, make fewer errors on the live queue, because they have seen what good reasoning looks like.

How is a good worked example built?

A worked example for agent review should show reasoning, not just the answer:

1. The case: what arrived and what the agent did.
2. The verdict: allowed, held, refused or withheld, and which bound decided.
3. The sources: each one named with its class: sealed, signed, declared or external.
4. The checks: what the reviewer verified, and where.
5. The decision: approve, send back or escalate.
6. The reason: one or two sentences tying the decision to a concept.

Then fade the support. The next example leaves out the reason; the one after leaves out the checks; eventually the learner solves a fresh case alone.

An example

An accounts payable team teaches reviewers with this worked example. The agent drafted a supplier payment and held it: the amount exceeds the team's limit. Sources: a purchase order from the procurement system (sealed); a goods receipt signed by the warehouse (signed); an invoice by email (external). Checks: the invoice amount matches the purchase order; the supplier's bank details match the supplier record and have not changed. Decision: approve. Reason: the payment is supported by sealed and signed sources, and the external invoice only confirms them. The next example changes one fact, the bank details changed last week by phone, and asks the learner to decide. The guide on training an accounts payable team to review agent decisions uses this pattern throughout.

How is a worked example different from a scenario?

A worked example shows the solution. A scenario asks the learner to decide first and then shows the consequence. Worked examples suit the start of learning, when people need to see good reasoning. Scenarios suit later stages, when they need to practise it. Most good programmes move from one to the other.

A worked example also differs from a walkthrough, which is a live, guided demonstration. A worked example is usually written and studied alone.

Which terms are related?

• Scenario-based learning: practice that follows worked examples.
• Counterexample: a case that looks right but is not, often paired with a worked example.
• Resistance to change: often eased when people see clear reasoning; see what resistance to change is.

To check whether examples actually improved decisions, see how to measure whether agent training worked. Fintechs moving quickly should read whether fintechs can use agents before staff are trained. For the full approach, read training teams for AI agents.