Crisis Communication Plan Template: Draft Holding Statements With AI

A PR agency can use AI to draft a crisis communication plan, scenario list and holding statements from a short fact sheet, then rehearse them with a media Q&A drill. A named client decision-maker and, where liability is involved, a lawyer must approve every word before release. AI drafts; people confirm facts and decide what is said.

The reason to prepare drafts in advance is speed under pressure. In the first hour of an incident, the delay is rarely writing, it is finding out who can approve and what is confirmed. A plan with templates, named approvers and a fact sheet removes that delay. The plan below is filled in for one illustrative client and marks each point where AI stops and a named person approves.

Follow me on Instagram@sagnikteaches

What the plan holds, in the order a team will need it

Keep the plan short enough to read in a stressful moment. Eight parts cover it, and each one can be drafted with AI from your client interview notes.

Connect on LinkedInSagnik Bhattacharya

1. The scenario list: what might go wrong for this client, ranked.
2. The crisis team: who decides, who writes, who speaks, who checks.
3. The fact sheet: a one-page form filled in as facts are confirmed.
4. Holding statements: one template per scenario, with gaps to fill.
5. The approval chain: the order of sign-off and a time limit for each.
6. Channels and audiences: staff, customers, media, partners, regulators if relevant.
7. The Q&A bank: likely questions and approved answers.
8. The log: timestamps of what was said, to whom, and who approved it.

The main example: a specialist food importer

This is an illustrative client: a specialist food importer with 25 staff, selling olive oils, cured meats and preserves to restaurants and independent shops. The agency has a retainer for general PR and the importer has asked for a crisis plan. The agency interviews the managing director and the quality manager for an hour and records the notes. Those notes, with names and supplier details removed, are what goes to the AI tool.

Subscribe on YouTube@sagnikteaches

Step one: building the scenario list

Ask for scenarios from the client's own situation, not a generic list. This prompt uses the interview notes as input.

You are helping a PR agency prepare a crisis plan for a specialist food importer
(25 staff; sells to restaurants and independent shops). From the notes below, list
eight realistic crisis scenarios. For each give: a one-line description, who is
affected, likelihood (low/medium/high), impact (low/medium/high) and the first
audience we would need to tell. Base every scenario on something in the notes.
Mark any scenario you added from general knowledge rather than the notes.

[paste anonymised interview notes]

An illustrative result, trimmed:

The team then rewrites the ranking. The model rated the delivery failure "medium" impact, but the client says a week without stock loses a restaurant account for good, so it moves up. AI proposes; the people who know the client's customers decide. Keep the top four to six for full templates and park the rest in a list to revisit.

Step two: the fact sheet that every statement draws from

The fact sheet is the single most useful page in the plan. Every statement is built only from what is on it, and each line carries a status. Draft the empty template once, then fill a copy per incident.

INCIDENT FACT SHEET - version 3 - updated 14:20
What happened: [one sentence]            Status: CONFIRMED / UNCONFIRMED
When it was found: [time, date]          Status:
Who is affected: [group, number]         Status:
Products / services involved: [list]     Status:
What we have done so far: [list]         Status:
What we are doing next: [list]           Status:
What customers should do: [instruction]  Status:
Who is the spokesperson: [name, role]    Status:
Next update promised at: [time]          Status:
Not to be said: [cause, blame, numbers not confirmed]

The "not to be said" line is where the lawyer and the client leader record anything that must stay out of public statements until confirmed. The rule for AI is mechanical: it may use CONFIRMED lines only. Put that in the prompt, and check the output for any fact it could not have taken from a confirmed line.

Here is the same sheet filled in for the importer at 09:45 on the morning of a recall, as an illustration. Much of it is still unconfirmed, and a holding statement can go out anyway.

Two lines decide what the holding statement may say. "Batch 2241 only" stays unconfirmed until the neighbouring batches are tested, so any sentence claiming no other products are affected comes out. And the supplier is not named, because the legal position on naming them has not been agreed. A model given this sheet and told to use CONFIRMED lines only should produce a statement that names the batch, tells customers what to do and promises an update at 13:00, and nothing more. If its draft says "no other products are affected", it has used an unconfirmed line, and that is the first thing the reviewer checks.

Step three: holding statement templates, with the gaps showing

A good holding statement has five parts: what happened, who is affected, what you are doing, what people should do, and when you will say more. Ask the model to leave square brackets wherever a fact is unknown, so nothing is filled in by guesswork.

Write a holding statement template for the scenario "product recall after a supplier
contamination finding" for a specialist food importer. Use this structure: what has
happened; who is affected and what they should do; what we are doing; when we will
update. 80-120 words. Plain, calm, no blame, no speculation about cause, no promises
of outcome. Use [square brackets] for every fact that is not yet known.

An illustrative result, and the edits a person would make:

We are recalling [product name], batch [number], after [a supplier's test result /
our own checks] found [issue]. If you bought this product, please do not eat it
[or: do not use it] and return it to [place] for a full refund. No other products
are affected [CONFIRM BEFORE SENDING]. We have stopped sales of the batch and are
working with [authority/lab] to establish the cause. We will update this page at
[time] and every [interval] until the matter is resolved. For questions, contact
[contact].

Edits at review: "please do not eat it" is right for food but the sentence about "no other products are affected" must come off unless the quality manager has confirmed it in writing, because a wrong reassurance is worse than silence. The phrase "our own checks" is wrong if a supplier alerted the company, and getting that wrong looks like concealment. The reviewer also removes any apology wording that counsel has not cleared, a topic covered at the legal checkpoint below.

Step four: the approval chain with time limits

An approval chain with no clock fails in the first hour. Write it as a table with a name per role, a deputy for each, and a maximum time for each step.

That adds up to an hour in the worst case, so for a fast-moving incident agree in advance that steps two and three run in parallel on the same draft. Write the deputy's name beside every role. Chains break on holidays and flights, not on drafting.

The legal review step: what to put in front of counsel

Do not send a lawyer a polished statement and ask "is this OK?". Send the fact sheet, the draft and a short list of questions:

• Does any sentence admit fault or liability, and is that what we intend?
• Are there any statutory notification duties triggered by these facts, and by when?
• Does anything conflict with insurance policy terms on communications about a claim?
• Is any named person or company at risk of a defamation complaint?
• Is a statement to staff required before the public release?

Ask AI to prepare that list per scenario in advance, then have counsel correct it once during plan preparation. In a live incident the lawyer then reviews against a list they have already shaped. None of this is legal advice; your counsel decides what applies to the client.

Step five: the media Q&A bank and a drill

Journalists will ask the questions the statement leaves out. Prepare answers in advance, in three groups: questions you can answer, questions you cannot answer yet and what you will say instead, and questions you will not answer, with the reason.

Using the fact sheet and the holding statement below, list the 12 hardest questions a
journalist would ask a spokesperson, in order of difficulty. For each, draft an answer
using only CONFIRMED facts. Where the answer is not yet known, draft a response that
says so, says what is being done to find out, and says when we will update.
Do not draft any answer that speculates on cause or blames a third party.

Illustrative lines from the bank:

Now run a drill. One agency person plays the journalist with the question bank but is told to improvise follow-ups. The spokesperson answers aloud from the templates and a timer runs. Afterwards note every question that had no prepared answer, every sentence that drifted into speculation, and every time someone reached for a fact that was not confirmed. Feed those notes back to the model to extend the bank, then have a person approve the additions.

Four more scenarios and how the template bends

The structure stays; the emphasis changes.

• Allergen mislabelling. The instruction to customers is the first line, ahead of any explanation. The statement lists the product and batch exactly, and the spokesperson answers only on what the label said and what is being done about it.
• Delivery failure. The primary audience is trade customers, so a direct email and a call list come before any public post. The statement offers a specific remedy and a date. Over-apologising and under-specifying is the usual error.
• Data breach. The statement describes what kind of data was involved only when confirmed, tells people what to do, and may trigger legal notification deadlines. Counsel leads, and the AI draft is a starting point only. Data-protection law in many places sets short deadlines for notifying the regulator and the people affected, so the lawyer checks those before anything is published.
• Viral accusation on social media. Speed matters less than accuracy. A short acknowledgement and a promise to look into it beats a long rebuttal. Do not argue in the comments and do not delete critical ones. For monitoring and labelling see how to label AI-generated images and video on social media if altered images are involved.

The first four hours, walked through with clock times

A timeline makes the approval chain concrete. This is an illustrative run of the recall scenario, using the templates above.

Notice that AI is not in the critical path for any decision. It saves minutes at the drafting and reformatting steps, and the 85 minutes from the first report to release are mostly checking and approving. If the plan has no drill history, expect the real timeline to take longer, which is why the time limits in the approval table should be tested rather than assumed.

The staff message that goes out first

Staff often hear about an incident from social media before they hear from their employer, and they are then asked about it by customers. A short internal message, issued at the same time as the public statement or before it, prevents improvised answers. Ask the model for one version in plain language and one line telling staff where to send media enquiries.

Subject: Product recall: what you need to know and what to say

Team, we are recalling [product], batch [number]. The public statement goes out at [time].
If a customer or journalist asks you about it, please:
1. Share the statement link and nothing beyond it.
2. Pass media calls to [name, number] straight away.
3. Do not post about this on any personal account.
If you are unsure of the answer, say: "I don't have more detail yet, but I will pass your
question on." Next update to the team at [time].

A reviewer would add one thing the model rarely includes: a sentence acknowledging that this is stressful and that staff can come to a named manager with concerns. The template should also be checked for anything that reads as an instruction to hide information, which can look bad if the message is ever published.

A social reply, before and after review

Social replies are where drafts go wrong quickest, because tone matters more and review time is shortest. Here is an illustrative pair for a post alleging that a jar was contaminated.

The first draft apologises for a feeling, asserts safety it cannot yet prove, and moves nothing forward. The reviewed reply asks for the one piece of information the team needs and gives a specific action. Add the rule "no claims of safety or quality in a live incident" to the social prompt, and keep every reply in the log.

The review after the incident

Within a week of the incident closing, hold a one-hour review with the agency and client teams. The aim is to improve the plan, not to apportion blame. Use the log of timestamps and ask five questions: how long did each approval take against its limit, which statements needed rewriting and why, which questions came up that the Q&A bank lacked, what did staff and customers say they needed that we did not provide, and what should be added to the scenario list.

AI is useful here as a summariser. Give it the anonymised log and ask for a timeline, the longest delays and the questions without answers. Then a person decides which changes to make to the plan. Update the templates the same week, while the lessons are fresh, and note the version and date in the plan header so that anyone can see it is current.

What to charge for, and what to promise the client

For an agency, a crisis plan is a deliverable you can scope: the interview, the scenario list, the templates, the approval chain, one drill and a review at six months. AI shortens drafting, so the value you sell is the judgement in the fact sheet, the legal checkpoint and the rehearsal. Be careful in the proposal not to promise that the plan prevents a crisis or guarantees outcome in the media. What it can promise is speed and consistency, with the right people approving the right words. Show the client the drill results, since they demonstrate better than any document that the plan works.

What AI gets wrong when the clock is running

• Reassurance it cannot support. "There is no risk to the public" appears in drafts unprompted. Strike it unless the fact sheet confirms it.
• Corporate fog. "We take this matter very seriously" says nothing. Replace with the action taken.
• Premature apology or blame. Both have legal consequences. Apology wording is a counsel decision.
• Stale facts. A draft built on version 2 of the fact sheet is still on someone's screen when version 4 exists. Put the version number and time in the header of every draft.

Keeping the plan alive between incidents

Store the plan where the whole team, including out-of-hours deputies, can open it without a password hunt, and print the one-page version of the contacts and approval chain. Review the scenario list every six months, rerun the drill after any change in the client's products or leadership, and update the Q&A bank with real questions from every minor incident. A short AI risk register is a useful companion if the client also wants to track AI-related risks in its own communications.

For wider agency workflows, AI press release drafts and client approvals uses the same principle: AI drafts, the named person approves. If you would rather build the plan with someone, that is what my AI implementation consultation covers.

Questions PR teams ask about AI in a crisis

Can AI write the statement during a live crisis?

It can produce a first draft in seconds from a fact sheet, which is useful when time is short. A named person must still approve the final text, because the facts change by the hour and the model does not know which ones are confirmed. Never let it publish on its own.

What is a holding statement?

A short, factual public message issued early, before the full picture is known. It confirms what has happened, says what you are doing, and says when you will say more. It does not speculate on cause, blame anyone or promise outcomes.

Should we put client facts into an AI tool in a crisis?

Only into a tool your agency has approved for client work under a business plan, and only the facts needed. Remove personal details of victims, staff and customers. Agree this in the plan beforehand, because nobody decides it well at 2 a.m.

How often should we rehearse the plan?

At least twice a year, and whenever the client's leadership or product changes. A one-hour drill, where one person plays a journalist and the team uses the templates against the clock, finds more gaps than rereading the document.

Further reads

• AI Media List Tools: Muck Rack vs Semrush AI PR Toolkit vs ChatGPT — Media list tools compared for building a response list.
• How to Write a Press Release With AI That Journalists Will Read — Writing releases that journalists actually open.
• AI Content Workflow for Agencies: From Brief to Approved Draft — A brief-to-approved-draft workflow for agency content.
• AI Media Monitoring for Small PR Agencies: Alerts, Sentiment, Reports — Alert set-up in 45 minutes a client, a sentiment spot-check that catches the tool's errors, a coverage log and report, and what each option costs.
• AI PR Pitch Emails: Personalise for Journalists With a Review Step — A five-step pitch workflow for small PR agencies: beat note, hook line, three-line pitch, manager sign-off and a follow-up cadence that stops on time.
• PR Coverage Report: Build One for Each Client With AI — A coverage report structure, the clippings log columns, AI prompts for the summary, and a sample monthly report with the checks a person must do.
• AI Agents for PR: What They Do and How an Agency Supervises Them — A PR agency supervision method for media lists, monitoring and pitch drafts, with a strict approval-before-send gate.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general crisis communication practice; no vendor pricing or product claims are used in this tutorial. Have counsel review any statement that touches legal liability.