Your service is unavailable. Customers are asking whether their work is lost, and the engineering team is still investigating. The social inbox is filling up faster than anyone can answer it.
AI can help turn a short set of confirmed facts into a readable update. It cannot establish the cause, declare customer data safe, or decide which promises your organization can keep. The useful starting point is a shared incident note, not a prompt asking the model to “manage the crisis.”
Put confirmed facts in one place
Before drafting, ask the incident owner for four things: the affected service, what customers can do now, what remains unknown, and when the team will provide another update. Include the time zone. A vague “soon” creates another question for every person reading it.
Keep reports separate from confirmation. “Three customers reported missing files” and “files have been deleted” are different claims. Neither a sentiment dashboard nor a generated summary closes that gap.
The CDC's Crisis and Emergency Risk Communication resources provide an established framework for emergency communication. For a business incident, the practical lesson is to give people usable information while communicating uncertainty clearly. Health and safety incidents require the appropriate qualified responders and organization-specific procedures.
Draft a holding statement that answers something
Hypothetical example: a booking service has confirmed that users cannot complete reservations. The cause and restoration time are unknown.
Some customers cannot complete bookings right now. We are investigating the interruption. Please avoid repeating payment attempts while we check transaction status. We will post another update here at 15:00 UTC, even if the investigation is still underway.
The payment instruction belongs in that statement only if the incident team has approved it. Remove it if the team cannot confirm that advice.
A useful drafting instruction would be: “Use only the facts below. Describe the customer impact in plain language. Do not add a cause, restoration estimate, or claim about data safety. Include the approved next-update time.” This constrains the draft without asking it to manufacture reassurance.
For preparation before an incident, use the separate guide to AI crisis communication strategies. During the incident, the immediate task is a credible update.
Answer the question behind each comment
Customers may all sound angry, but they need different information. One person needs a workaround. Another needs to know whether they were charged. A third is pointing out that your status page contradicts your social update.
- Public service questions: link to the current update and give the approved action. Repeating the answer is appropriate when the question repeats.
- Account-specific problems: offer a secure support route. Do not ask people to publish transaction identifiers or personal details in the thread.
- New evidence: send it to the incident owner. A report that changes the understanding of the incident should update the shared note before it changes public messaging.
A reply draft can acknowledge the actual problem: “You need to know whether the booking completed before trying again.” That is more useful than “We understand your frustration” followed by no answer. Our guide to customer interaction with AI copilots covers this distinction in routine support.
Keep the update promise
If the investigation is unresolved at the promised time, say that. A short update explaining what is still being checked is preferable to silence or an invented estimate.
When a fact changes, correct it visibly. State what the earlier message said, what is now confirmed, and which action customers should take. Update the source page and tell the people answering comments. Do not let an older generated draft continue circulating after its facts expire.
Measure the communication by unanswered urgent questions, conflicting messages, and whether promised updates occurred. A rising comment count is not success during an outage.
Give drafting tools a narrow role
ReplyPilot can assist with a reply draft through its Chrome extension and /rp workflow. The writer reviews and manually posts the result. It is not an incident monitoring service or a source of verified operational facts.
Before the next incident, draft one holding statement from a fictional scenario, identify who must approve its facts, and time a rehearsal. The exercise should expose missing ownership and unclear instructions before customers have to expose them for you.