Log every practice decision with the same fields: the case, what the agent proposed, your verdict, your reason, the source you relied on, whether a red line was involved, and what you would check next time. Review the log weekly for patterns. Fifty decisions with clear reasons beat five hundred clicks with none.
Why keep a decision log while you practise?
Reviewing agent decisions is a judgement skill, and judgement improves when you can see your own reasoning later. A log does that. It also turns practice into evidence: when an interviewer asks how you would handle a borderline case, you can describe one you actually decided, with the reason you wrote at the time.
It mirrors how the real work is recorded. In a governed system, every verdict is filed in an append-only record, so any decision can be reconstructed. Learning to write your reasons as if someone will read them later is the habit the job needs. For the basics, see what it means to review an AI agent's decision and what a decision record is and why job seekers should know it.
How do you write a good reason?
A good decision rationale answers three questions in one or two sentences: what did you rely on, what rule or limit applied, and why does that lead to your verdict? "Looks fine" is not a reason. "Order record and delivery photo both support the claim; refund is within policy and limit" is.
Name the source and its strength. A fact from the system of record is stronger than a customer's statement, and an emailed attachment is weaker still. When you rely on a weak source, say so and say why you accepted it, or why you did not.
The decision log template
| Field | What to write |
|---|---|
| 1. Date | When you made the decision |
| 2. Case ref | Your own reference (never real customer details) |
| 3. Case type | One line: refund request, supplier payment, address change, etc. |
| 4. Agent proposed | What the agent drafted or wanted to do, and its verdict (allowed, held, refused) |
| 5. Your decision | Approve, amend, reject or escalate |
| 6. Reason | One or two sentences: source relied on, rule or limit, why |
| 7. Source strength | System record, named party, customer's statement, or arrived from outside |
| 8. Red line? | Yes or no; if yes, which one |
| 9. Confidence | High, medium or low |
| 10. Check next time | What you would look for first in a similar case |
| 11. Later review | Filled in at your weekly review: still right? If not, why |
Fields 9 and 11 are the ones people skip and the ones that teach most. Low-confidence decisions show where to study. Decisions you would change a week later show where your judgement is moving.
A filled sample
Three entries from an invented learner, Lena, practising on a simulated retail refunds queue:
| Field | Entry A | Entry B | Entry C |
|---|---|---|---|
| Case type | Refund, item not received | Refund, damaged item | Address change before dispatch |
| Agent proposed | Held: third non-receipt claim on the account | Allowed: within limit | Held: change made by a new device |
| Your decision | Escalate to fraud review | Approve | Reject until verified |
| Reason | Tracking shows delivery with photo; customer's statement conflicts with carrier record; pattern of repeated claims | Delivery photo and customer photo both show damage; within limit | Change came from an unrecognised device with no confirmation from the email on file |
| Source strength | Carrier record versus customer's statement | Two photos, one from carrier | Arrived from outside |
| Red line? | No | No | Yes: possible account takeover |
| Confidence | Medium | High | High |
| Check next time | Carrier photo first | — | Device and confirmation history first |
| Later review | Still right | Still right | Still right |
Entry A was medium confidence, and Lena noted why: the photo was clear, but the customer's story was plausible. Escalating rather than rejecting was the right call when she was unsure; a fraud team has more context.
How should you review the log each week?
Spend fifteen minutes on it. Count decisions by verdict: if you rejected almost nothing, you may be too trusting; if you rejected most things, you may be hunting for errors. Read every low-confidence entry again. Look for repeated "check next time" notes; three of the same means a habit to build.
Measure decision quality, not volume. The guide on how to learn to read an AI agent decision record shows how to compare your reasoning with the record the agent leaves.
Questions people ask
What should I practise on if I have no access to agents?
Use a simulated queue in a free workspace if one is available, or build cases from your own work experience with the details changed. Practise with realistic cases, not puzzles.
Should I use a spreadsheet or a notebook?
A spreadsheet, so you can sort by verdict, confidence and red line during weekly reviews.
Can I show the log to employers?
Show a short anonymised extract and a summary, not the whole log. The AI skills for job seekers guide explains how evidence like this fits into an application.