How to run an AI agent's first week in production

Start narrow and look at everything. On day one the agent works one queue, in staffed hours, under a low ceiling, and a reviewer checks every action after the fact. A short daily review looks at held, refused and allowed work and decides whether to continue, widen volume or stop. By day five the process owner signs a week-one record.

What is the first week for?

Certification shows the agent can do the job on past cases. The first week shows it can do the job on today's cases, in your live systems, with your people around it. Things surface in production that replay cannot catch: a connector that times out at the morning peak, a case type that has become common since the certification set was drawn, an approver who does not know the held queue exists.

The process owner runs the week. The decision to go live at all should already have been made on evidence, using the criteria in how to know an AI agent is ready for production. The owner is in operations, for the reasons argued in operations, not IT, should own agents in production: they are the ones who can judge whether the work was done right.

What happens each day, and who acts?

1. Day 0, process owner: confirms go-live. Certificate in date, approvers named and briefed, fallback team available, stop rule written down. Evidence: the signed go-live record.
2. Day 1, process owner: sets the slice. One queue or case type, staffed hours only, and a ceiling on volume and value below what the certificate allows. How to choose that first limit is covered in setting the first ceiling for a new agent. Evidence: the slice and ceiling, recorded in the agent's bounds.
3. Days 1 to 5, agent: works within the slice at its certified depth. Every action passes the minimum chain and is filed with its verdict. Evidence: the verdict records.
4. Days 1 to 2, reviewer: checks every allowed action after the fact against the procedure. Not a sample; all of them. Evidence: the review log, item by item.
5. Days 1 to 5, approvers: work the held queue at least twice a day. Evidence: approval times and outcomes.
6. Each day, process owner and reviewer: hold a short review. Figures from the agent's records, every refused item, every error found, every complaint. Decide: continue, widen volume within the ceiling, narrow, or stop. Evidence: the decision and the reason.
7. Days 3 to 5, reviewer: moves to a large sample if days one and two were clean.
8. Day 5, process owner: signs the week-one record. Evidence: the record, with the decision on what happens in week two.

What is the stop rule?

Write it before day one, so no one has to decide under pressure. A typical stop rule says: any error that reached a customer and could cause them harm, any action outside the agent's bounds that was not refused, or any failure of the verdict record, stops the agent at depth 1 that day. Lesser errors are reviewed at the daily meeting.

When the stop rule fires, the fallback team takes the work, the agent drafts only, and the problem is investigated. The broader sequence, including what to tell customers and how to restore depth, is in what happens when an AI agent fails in production.

Why review every action on days one and two?

Because the first days carry the most surprises and the fewest data points. A sample of ten from a hundred actions can easily miss the one case type that is going wrong. Reviewing all of them costs a reviewer's day or two, and in return you either find the problem early or gain real confidence. Once the first two days are clean, a large sample is enough.

What should you watch?

• Errors found in review, by severity and case type.
• Held queue age. Items waiting hours mean approvers are not ready.
• Refusals by bound. Many refusals by one bound often mean the slice includes work the agent's certificate does not cover.
• Connector failures and timeouts, especially at peaks.
• Staff reactions. Questions, workarounds, and anyone quietly redoing the agent's work.
• Customer contacts about anything the agent did.

What does the week-one record contain?

The go-live record; the slice and ceiling; daily figures; every error and what was done about it; every stop rule event; the review decisions; and the process owner's decision for week two, which is one of: keep the same slice, widen volume, add a case type, narrow, or return to shadow. This record becomes the first entry in the 90-day evidence, and follows the sequence set out in train, shadow, certify, widen.

Where does it usually go wrong?

The slice is too wide. Going live on everything at once means any problem affects everything. Choose a slice where an error is cheap, ideally the kind of work set out in choosing the first process to give an AI agent.

Approvers were not told. Held items sit unread because the approvers did not know they were approvers.

Out-of-hours work starts on day one. The agent runs overnight with no one watching. Keep it to staffed hours in the first week.

Clean days lead to fast widening. Two good days are not evidence for a large change. Widen volume within the ceiling; widen depth only with new certification.

The review becomes a celebration. Everyone wants the launch to go well. Keep the reviewer independent of the team that built the agent.