Duty desk note: AI drafts for incident updates

During an incident, the duty engineer may be checking dashboards, coordinating responders, and answering customer questions at the same time. Writing a clear status update becomes another urgent task. AI can help turn verified notes into readable prose, but it should not decide what happened or what customers need to know.

Treat the model as a drafting assistant, not an incident commander. Your team supplies the facts, uncertainty, audience, and communication deadline. The model proposes wording. A responsible human checks and publishes it.

Start with the facts the desk can defend

A useful incident update explains the current customer impact, what responders are doing, and when readers should expect another message. It does not need to explain every technical detail.

Before asking AI to write anything, collect a small, approved fact sheet:

• Record the incident start time and its time zone, distinguishing the first observed symptom from the confirmed start of impact.
• Describe the affected service and the user-visible symptoms, including any confirmed limits on scope.
• Separate confirmed findings from hypotheses that responders are still testing.
• Record the actions underway and the next communication time approved by the incident lead.

Do not paste raw logs, credentials, customer records, or private responder conversations into a model. Use the minimum necessary information and follow your organization's rules for approved AI services and data handling.

Give the model boundaries before asking for prose

An open-ended request such as “write an outage update” invites the model to fill gaps. A constrained request makes those gaps visible instead. Specify the audience, length, permitted facts, and things the draft must not claim.

A reusable prompt could read:

“Draft a public incident status update of no more than 90 words using only the facts below. Describe customer impact in plain English. Preserve uncertainty. Do not infer a cause, expand the affected scope, promise a recovery time, or describe the incident as resolved. Include the supplied next-update time. If required information is missing, list it separately for the reviewer rather than inventing it.”

Supply the fact sheet after those instructions. Keep internal review questions separate from the text intended for publication so they cannot accidentally reach the status page.

Watch a short fact sheet become a status update

Consider a fictional incident. Responders have confirmed elevated errors on the reporting API beginning at 14:10 UTC. Some report downloads are failing. The cause is unknown, the team is investigating, and the next update is scheduled for 14:45 UTC.

A suitable draft would be: “We are investigating elevated errors affecting the reporting API since 14:10 UTC. Some customers may be unable to download reports. We have not yet confirmed the cause. We will provide another update by 14:45 UTC.”

Notice what is absent: no invented regional scope, no claim that customer data is safe, and no recovery estimate. The next-update time is a communication commitment, not a prediction that service will recover by then.

Review every claim before the update leaves the desk

Fluent wording can hide an unsupported claim. Check each sentence against the approved facts, especially statements about scope, root cause, recovery, and data integrity. Reject phrases such as “fully restored” unless responders have actually confirmed recovery through the team's established checks.

For later updates, give the model the previous published message and newly confirmed facts. Ask it to preserve the timeline, remove stale actions, and distinguish mitigation from resolution. Keep publication behind human approval rather than connecting model output directly to a status-page publishing action.

Choose a drafting route your team can govern

If your team wants to try this workflow across model families, Relay365 is a pay-as-you-go relay desk that can fit the drafting step. It works with Claude Code, Codex, Gemini CLI, and OpenAI SDK clients. Create an account to test sanitized incident notes, provided the service meets your organization's approval requirements.

Start with a tabletop exercise, not a live outage. Compare drafts against the same fact sheet and record unsupported additions. The goal is straightforward: clearer communication without surrendering control of the facts.