The hardest time to decide who can speak for your business is when people are already demanding an explanation. An AI assistant may produce a statement in seconds. It cannot establish what happened, authorize a remedy, or decide which uncertain detail is safe to publish.
Use AI during preparation to expose gaps in your communication plan. During an incident, use it within that plan. The distinction matters when a confident sentence can turn an unconfirmed theory into a public claim.
Name the decisions before choosing the software
A practical crisis plan identifies an incident owner, a communications lead, a subject expert, and a final approver. It also identifies backups and a channel for reaching them outside normal hours.
Write down the events that trigger this group. Several complaints about the same fault deserve investigation; an angry comment alone does not establish a systemic failure. Set thresholds appropriate to your business and leave room for a person to raise a concern that the rules miss.
For communication principles, the CDC's Crisis and Emergency Risk Communication manual provides a useful framework grounded in major emergencies. Apply its emphasis on credible information and audience needs thoughtfully to your own setting. A business complaint and a public health emergency have different stakes.
Build a facts sheet that can survive pressure
Use one maintained document for confirmed facts. Include the source, the person who verified it, and the time of the latest check. Put open questions in a separate area so nobody mistakes them for agreed answers.
For a hypothetical service outage, the facts sheet might establish that checkout is failing for some customers, engineers are investigating, and the next public update is due at 3 p.m. It might explicitly leave the cause and restoration time unknown.
The AI drafting input should come from those confirmed fields. Ask it to identify missing information and produce a draft that preserves uncertainty. Do not ask it to fill gaps with the most likely explanation.
Rehearse situations, including awkward ones
AI can help create fictional prompts for a tabletop exercise. Label them clearly and use synthetic information. A good rehearsal includes a customer asking for a refund, a reporter requesting comment, a misleading screenshot, and two internal teams describing the same event differently.
Ask the team to work through four questions:
- Verification: What would establish whether the allegation is true?
- Responsibility: Who owns the operational response and the public statement?
- Action: What can affected people do now, and through which channel?
- Update: When will the next confirmed information be available?
Rehearsal should test the approval chain as well as the wording. A beautifully drafted statement is useless if its approver cannot be reached.
Write a holding statement people can use
A hypothetical first update could read:
Some customers are unable to complete checkout. Our team is investigating. We do not yet have a confirmed restoration time. We will post another update here by 3 p.m. For an existing order, contact our support team through the help page.
Only use commitments the team can keep. Do not promise compensation, assert that personal data is safe, or announce a cause without the appropriate confirmation. A tone edit must not soften a fact until the meaning changes.
AI-assisted customer support needs the same separation between an acknowledgment and an actual resolution. ReplyPilot can assist with a reviewed reply draft; it is not a crisis monitoring or incident management system.
Review what failed after the incident
Record outdated reference material, slow approvals, contradictory replies, and questions you could not answer. Judge the process by accuracy, useful updates, and follow-through, rather than how much content the assistant produced.
Generative models can present false information confidently, a risk addressed in NIST's generative AI profile. Keep factual approval with the people who can verify the event. For wider rules on truthful content and data handling, use the ethical AI marketing guidelines.
Begin with one rehearsal this month. Choose a plausible failure, run it through your existing contacts, and repair the first broken handoff before writing more templates.