Your first month of operating an AI workflow

A bounded release gives a team something valuable: real work to learn from. It also creates a new operating obligation. Someone must watch what happens after the demonstration ends, distinguish recurring problems from isolated cases, and decide which changes deserve another release.

Consider a hypothetical facilities team using an agent to prepare maintenance visit briefs. The agreed scope is narrow: assemble approved site information, list outstanding questions, and produce a draft for a coordinator. Scheduling visits and contacting contractors remain human decisions. Here is a practical plan for the first month of operating that workflow.

Week one: watch complete cases

Follow cases from arrival through acceptance. Keep the release scope stable so the team can see how the agreed process behaves under ordinary conditions. Record the source material available, the draft produced, the coordinator's corrections, and the final disposition.

In this workflow, a polished brief might still omit a site's access restrictions. Another might contain everything needed but arrive after the coordinator has prepared the visit manually. Those cases reveal different operating problems. Reading only the generated text would miss the second one.

Ask the coordinator to leave a short reason when work is changed, rejected, or completed outside the workflow. Preserve the relevant evidence under the existing access rules. Observing a process does not justify collecting unrelated private conversations or copying restricted site records into a wider team folder.

The operating owner should also check that someone notices stalled work. A case with no visible error can still be waiting indefinitely for information that nobody has been asked to provide.

Week two: classify the friction

Review the observed cases together. Separate incomplete input, outdated source material, unclear instructions, incorrect output, missing permission, and delays in human handling. Give each category a practical consequence and a person who can address it.

These distinctions prevent an easy mistake: changing the agent whenever the workflow disappoints. If the approved site record is stale, the source owner needs to correct it. If a coordinator cannot tell whether a brief is ready, the completion rule needs attention. If necessary information is restricted, the case must wait for an authorized decision.

Keep a small set of representative examples, including an ordinary success, a correction, and a stopped case. Select material the reviewers are permitted to see. An approved, redacted example may be more useful for wider learning than a full operational record.

Choose the most consequential recurring issue that the team understands well enough to change. Leave the other findings visible rather than bundling them into an ambitious redesign.

Week three: change one thing and replay

Make one controlled change with a clear expected effect. For the facilities team, that might be an explicit requirement to identify the approved source for access instructions and flag its absence. The change should have an owner, a version, and a way to return to the previous behavior.

Replay the saved examples using only the information and permissions available in each case. Check whether the intended problem improves and whether previously acceptable work deteriorates. Include a missing-source case so a better draft does not conceal a weaker stopping rule.

The person reviewing the change should examine the evidence, including unresolved cases. An agent's own assessment cannot authorize its release, approve its consequential output, or enlarge its access.

If the change works in replay, introduce it within the existing release boundary and observe again. Keep a record of when the new version starts handling work. Otherwise, later comparisons can mix behaviors and leave the team arguing about which change caused what.

Week four: decide the next boundary

Bring together the case record, correction reasons, unresolved exceptions, and operating effort. The decision can be to continue at the same scope, adjust the process, pause it, or propose a specific extension. Continued narrow operation is a legitimate outcome when the evidence remains incomplete.

An extension needs its own description. Adding another site type, using another source, and sending a message are separate changes with different consequences. Name the new task, responsible person, required access, and evidence that would justify continuing. Existing permissions do not expand because the month went smoothly.

Check whether people can sustain the review and recovery work. A workflow that depends on the coordinator remembering undocumented fixes still needs operational attention, even when its drafts look good.

Before the next month begins, write a short operating decision: what stays in scope, what changes, what remains unresolved, and when the owner will review it again. Put that decision where the people running the workflow can find it before the next case arrives.