A quick reference guide is a short document, usually one page, that a person keeps beside them while doing a task, so they can check a rule or a step without opening the full procedure. For people who review AI agents, it holds the few things they must get right on every decision, such as what each verdict means.
Why does a quick reference guide matter?
Training fades fastest on the details people meet rarely. A reviewer may handle hundreds of routine drafts before the first withheld action or the first payment backed only by an external document. On that day, they will not reopen a 30-page procedure. A one-page guide puts the decision in front of them at the moment it matters.
It also standardises judgement. When ten reviewers each remember the rules slightly differently, the agent's held actions get ten different treatments. A shared page narrows that spread.
What belongs on one for agent review?
A good guide is selective. It covers the decisions where a mistake is costly, not everything a reviewer could know. A typical page for a team reviewing an agent's work holds:
| Section | What it says |
|---|---|
| The four verdicts | Allowed, held, refused, withheld, and what the reviewer does with each |
| Source classes | Sealed, signed, declared, external; which may support which action |
| Before you approve | Three checks: the source, the amount against the limit, the bound that decided |
| Red lines | The actions no reviewer may approve, whoever asks |
| Who to call | The agent's owner, the team lead, the escalation route |
It should fit on one side and use the same words as the workspace. If the screen says "held", the guide says "held", not "pending".
An example
A shared services team runs an agent that matches supplier invoices and drafts payments for approval. Their guide has one line that does most of the work: "An emailed invoice is external. It cannot authorise a payment. Check the supplier record before approving." Reviewers pin it beside the queue. When a new starter meets a held payment with a changed bank account, the guide tells them to stop and call, not approve.
Writing the page is a good test for a team's trainers. If they cannot agree what goes on it, the training itself is not settled; the train-the-trainer guide template helps here.
How is it different from a standard operating procedure?
A standard operating procedure is the full, owned description of how a process runs: scope, roles, steps, exceptions, records. It is the authority. A quick reference guide is a summary of the parts people need at the point of work, and it defers to the procedure. If the two disagree, the procedure wins and the guide is wrong.
A guide is also not training. Handing out the page without teaching the reasoning behind it produces people who follow the line until a case does not fit it. That gap is one reason AI agent pilots fail when nobody is trained.
Which terms are related?
- Standard operating procedure: the full document the guide summarises.
- Runbook: step-by-step instructions for a specific event, such as an incident.
- Capability building: the wider effort a guide supports; see what capability building is.
For how guides fit into a training programme, read training teams for AI agents.
Questions people ask
How often should a quick reference guide change?
Whenever the agent gains or loses a capability, or a limit changes. Date and version the page so reviewers know they have the current one.
Should the guide be printed or on screen?
Either works if it is current. Many teams keep it beside the queue on screen and retire printed copies when the version changes.