Read a decision record in four passes. First the verdict: allowed, held, refused or withheld. Then the deciding bound, which names the limit that set the outcome. Then the basis: the sources the agent relied on and where each came from. Finally the person who answered, if a person was involved. Practise on real records until each pass takes seconds.
Why is reading the record a core skill?
Because every verdict is filed in an append-only record, the record is how anyone reconstructs what happened: your manager after a complaint, an auditor months later, you at the start of a shift. A reviewer who cannot read it quickly will either approve on trust or waste time hunting for context. A reviewer who can read it well sees in seconds why an action was held and what needs checking.
It is also what interviewers test. "Here is a record; what happened and what would you do?" is a common scenario question for oversight roles, because it shows whether you understand the system or just the screen. The guide to AI skills for job seekers puts this skill near the top of the list.
What does each part of the record tell you?
| Part | What it tells you | Question to ask | Common misreading |
|---|---|---|---|
| Verdict | Allowed, held, refused or withheld | Is this the outcome I expected for this action? | Treating held as a failure rather than a request for a person |
| Deciding bound | Which of the five bounds set the outcome | Is this the bound I would expect to decide? | Assuming the agent's ability decided, when it was the customer's limit |
| Basis | The sources relied on, with provenance | Is any key fact only declared or external? | Trusting a fact because it looks official |
| Person who answered | Who approved, held or escalated | Was it the named person for this kind of action? | Missing that the wrong person approved |
What are the four passes, step by step?
- Read the verdict. Allowed means the action went ahead within all bounds. Held means it was drafted and is waiting for a named person. Refused means it was outside the agent's bounds. Withheld means it was refused on purpose at the platform floor for every agent at every depth, such as reversing a settled payment on the platform's own rail. Withheld is a platform decision, not a customer setting.
- Read the deciding bound. Every action meets five bounds at once: role scope, certificate, sources, the customer's own limits and the platform guardrails. The lowest one wins, and the verdict names it. If the bound was "certificate", ask whether the agent's certificate had lapsed or its depth had dropped after a failed run. If it was "sources", the evidence was not strong enough.
- Read the basis. List the facts the agent relied on and their provenance: sealed (from the system of record, tamper-evident), signed (a named party answers for it), declared (stated by a person, unverified) or external (arrived from outside, such as an email). An external claim cannot authorise a payment, so any payment resting on one should never have been allowed.
- Read who answered. For held actions, the record shows the named person who decided. Check that it was the right person for that kind of action and that a reason was recorded.
How do you practise?
Use a free workspace where you can generate records safely. Three drills work well.
Predict, then read. Before opening a record, read only the proposed action and predict the verdict and deciding bound. Then check. When you are wrong, find out why.
Find the weakest fact. For each record, identify the least trustworthy source in the basis. If it is declared or external and the action depended on it, ask whether the outcome was right.
Reconstruct a sequence. Take several records about one customer or case and put them in order. What happened first, what changed, who decided at each step? The guide on how to reconstruct an AI agent's decision from the record shows this in detail.
Pryme Intelligence Academy's Operator track covers how a governed agent decides and how to read the workspace, with labs in a free Build workspace, if you want a structured version of these drills.
What does a worked example look like?
A sample record, with invented values: an agent proposed a payment of 6,000 to a supplier. Verdict: held. Deciding bound: sources. Basis: invoice from email (external); supplier bank details from the same email (external); supplier master record (sealed) showing different bank details. Person who answered: the accounts payable lead, who kept it held and recorded "bank details differ from sealed record; contacted supplier via registered number".
Read in four passes: the action was held, not refused, so a person had to decide. The bound was sources, because the only evidence for the new account was external. The basis shows a direct conflict with a sealed record. The right person answered, with a clear reason. A good record, and a correct decision.
What should you measure?
How long each pass takes you, and how often your prediction of verdict and bound is right. Track separately how often you spot the weakest fact in the basis. As you improve, the four passes should take well under a minute for a routine record.
Where do people usually go wrong?
They read the agent's draft and skip the record, so they never see why it was held. They confuse refused with withheld. They trust a fact because it appears in a tidy document, without checking its provenance. And they forget the last pass: a correct verdict approved by the wrong person is still a problem.
If you lead a team, the guide on how you train staff to read an AI agent's decision record adapts these drills for groups. Reading individual records pairs naturally with learning to read AI agent workspace numbers, which shows patterns across many records.
Questions people ask
Can a decision record be edited after the fact?
The record is append-only, so a decision is never overwritten. A later decision is added as a new entry, which is what lets anyone reconstruct the full history.
Why does the record name the deciding bound?
So that anyone can see which limit set the outcome without guessing. It turns "the agent would not do it" into a specific, checkable reason.
How is this different from judging decision quality?
Reading the record tells you what happened and why. Judging quality asks whether it was the right call. You need the first to do the second; the page on what decision quality is covers the second.