The First Hour: AI Prompts for Rapid Crisis Assessment
Before you write a single word, you need to know what you’re actually dealing with.
Do these five things before you read anything else
- Is anyone hurt or at risk? If yes, stop reading. Safety response comes first and comms follows it.
- Write down what you actually know — confirmed only. Not what you’ve been told. Not what you assume. Three bullet points is fine.
- Call your legal counsel. Not email. A regulatory clock may already be running, and they need to tell you which.
- Say nothing externally yet beyond an acknowledgement that you’re aware and looking into it. Do not explain, speculate, or characterise.
- Screenshot and timestamp everything. Posts, articles, messages. They will be edited or deleted, and you’ll need the record.
Then come back and run the triage prompt below. It takes about four minutes and tells you who needs to be called.
The first hour, compressed
- Verify before you classify. Separate confirmed from told from assumed. Conflating them is how retractions happen.
- Classify before you notify. Severity determines who gets called, not how the statement is worded.
- Check the clock. GDPR gives 72 hours from awareness — not from full understanding. SEC materiality disclosure is four business days.
- Acknowledge, don’t explain. A short “we’re aware and looking into it” buys verification time without committing you to anything.
- AI structures the assessment. Humans and lawyers decide. An invented detail in a crisis statement is a legal event, not a typo.
Something has just broken, you’ve had a near miss and want a protocol, or you’re the person who’d be called at 11pm and you’d rather know the sequence now.
You need the statement itself — that’s the holding statement guide — or you’re dealing with a fabricated video, which has its own verification-first sequence.
On this page
The mistake: drafting before classifying
When something breaks, almost everyone’s first instinct is to open a document and start writing. It feels like action. It’s the wrong first move, and it causes two specific failures.
| Failure | What it looks like |
|---|---|
| Disproportionate response | A full corporate statement on a matter that would have died in a day — which converts a minor complaint into a documented incident, and creates a searchable record where none existed. |
| Premature commitment | A statement built on what someone told you in the first ten minutes. When the facts change, you retract — and the retraction becomes the story, because it is now about your credibility rather than the original issue. |
Both come from the same root: writing before knowing. Classification takes about ten minutes and prevents both.
The four-step sequence
The whole sequence runs in about twenty minutes with the prompts below, which leaves forty minutes of the first hour for the things only humans can do — making calls, checking records, and deciding.
Prompt 1: What do we actually know?
This is the foundational one, and it’s the step most often skipped. In the first hour people conflate three very different categories of information, and statements built on the wrong one are the single most common cause of a retraction.
## THE SITUATION What has happened, as I currently understand it: [BRAIN-DUMP IT. Messy is fine. Include what you've been told, what you've seen, what you're inferring — don't try to sort it yourself, that's the job.] Where this information came from: [Who told you what, and when. Include "saw it on social", "a colleague said", "the article claims", "I assumed".] What I have personally seen or verified: [Documents, system records, screenshots, direct observation] ## YOUR TASK Sort every single claim into exactly one category: **CONFIRMED** — independently verified against a record, system, document or direct observation. State what verifies it. **REPORTED** — someone has said it, but it hasn't been checked. Name who said it and note that it's unverified. **ASSUMED** — inference filling a gap. Nobody said this; we're joining dots. Flag every instance, including ones that feel obvious. Then: 1. THE CRITICAL UNKNOWNS — what would change our response most if we knew it? Rank by decision-impact, not by how hard it is to find out. 2. WHO CAN VERIFY EACH ONE — name the function or role that holds the answer, so we know who to call. 3. WHAT WE MUST NOT SAY YET — anything that depends on REPORTED or ASSUMED information. Be explicit: list the specific claims that are not yet safe to make publicly. 4. THE DANGEROUS ASSUMPTION — of everything in the ASSUMED column, which one are we most likely to state as fact without noticing? Quote it back to me. RULES: - Do not add any information. If something is unclear, put it in ASSUMED and say why. - Do not speculate about cause, motive, or who is at fault. - Do not draft any statement. This step is assessment only.
Prompt 2: The severity triage
Severity determines who gets called and how fast — not how the statement is worded. Run this immediately after the verification prompt, feeding in its output.
## ORGANISATION CONTEXT Sector: [AND WHETHER REGULATED] Size / profile: [HEADCOUNT, PUBLIC OR PRIVATE, MEDIA PROFILE] Jurisdictions we operate in: [FOR REGULATORY EXPOSURE] Who can authorise an external statement: [ROLES] ## THE INCIDENT [PASTE THE CONFIRMED / REPORTED / ASSUMED OUTPUT FROM PROMPT 1] Where it surfaced: [Social, press enquiry, employee report, customer complaint, regulator contact, internal discovery] Current visibility: [Who knows so far, and roughly how many] Movement in the last hour: [Growing / static / unclear] ## CLASSIFY ACROSS SIX DIMENSIONS Score each and explain the score in one line: 1. HARM — has anyone been physically, financially or psychologically harmed, or is anyone at risk? (This one overrides the others. If yes, say so first.) 2. VERIFICATION — is the underlying claim confirmed, alleged, disputed, or apparently fabricated? 3. VELOCITY — how fast is it moving, and is that accelerating? 4. EXPOSURE — which stakeholder groups are affected or will ask: employees, customers, regulators, investors, partners, press, community? 5. REGULATORY — does this plausibly trigger a notification obligation? Flag for legal; do not attempt to determine it. 6. REVERSIBILITY — if we do nothing for four hours, does this get better, stay the same, or get materially worse? ## THEN OUTPUT A. SEVERITY LEVEL — 1 (critical) to 4 (monitor), with the single dimension that drove the classification. B. WHO MUST BE TOLD IN THE NEXT 60 MINUTES — by role, in priority order, with the contact method for each (call / message / email) based on severity. C. WHO SHOULD NOT BE TOLD YET — and why. Premature internal circulation is a common route to a leak. D. THE DECISION THAT CANNOT WAIT — the one thing that must be decided in this hour, and by whom. E. WHAT CHANGES THE CLASSIFICATION — the two developments that would escalate or de-escalate this, so we know what to watch for. F. THE HOUR-TWO QUESTION — what we should be finding out right now to be ready for the next assessment. RULES: - Do not draft a statement. - Do not determine legal or regulatory obligations — flag them for counsel. - If information is insufficient to classify a dimension, say so rather than estimating. - Err toward higher severity when uncertain, and say that you have.
The severity levels and who they trigger
Three to five levels is standard practice, each mapped to a defined notification chain. Here’s a four-level model adapted for communications rather than IT incidents — the dimensions differ because comms severity turns on harm, verification and audience rather than system availability.
Triggers: harm to people · existential business threat · confirmed regulatory breach · executive implicated.
Notification: CEO, General Counsel, Comms lead, Board chair — by phone, with confirmed receipt. Within minutes.
Assume: a legal clock is running until counsel tells you otherwise.
Triggers: verified, spreading, material business impact, multiple stakeholder groups affected.
Notification: Comms lead, GC, affected function head, CEO informed — phone or direct message, receipt confirmed. Within the hour.
Watch: velocity. Level 2 becomes Level 1 when it stops being containable.
Triggers: limited audience, single stakeholder group, no regulatory dimension, reversible.
Notification: Comms team and relevant function head. Channel message is adequate.
Risk: over-responding. A Level 3 handled as Level 1 creates a record that outlives the issue.
Triggers: unverified, low visibility, no harm, not moving.
Notification: log it, assign an owner, set a check-in time.
Do: screenshot and timestamp anyway. Level 4s occasionally become Level 1s, and you’ll want the early record.
Prompt 3: Stakeholder exposure map
THE INCIDENT: [TRIAGE OUTPUT FROM PROMPT 2] OUR STAKEHOLDERS: [List the groups that actually apply: employees, customers, regulators, investors, partners, suppliers, community, press, unions, franchisees...] For each stakeholder group: 1. ARE THEY AFFECTED — directly, indirectly, or not at all? 2. WILL THEY HEAR ABOUT IT — and from whom? From us, from press, from each other, or from social? 3. WHAT WILL THEY ASK FIRST — the actual first question, phrased the way they'd phrase it. 4. WHAT THEY NEED FROM US — reassurance, information, instruction, or an apology. These are different and confusing them is a common error. 5. SEQUENCE — what order should groups hear from us, and what's the risk if that order is wrong? Then: A. WHO MUST HEAR FROM US BEFORE THEY HEAR ELSEWHERE — the groups where being told second is itself damaging. Employees are almost always on this list. B. THE CONTRADICTION RISK — where might different groups receive versions that appear inconsistent? Anything said internally should be assumed to reach outside. C. THE FORGOTTEN GROUP — based on our sector, which stakeholder do organisations like ours typically overlook in the first hour? Do not draft messages. This is mapping only.
Point A, in practice, almost always means employees. Staff learning about their own organisation’s crisis from a news alert is a reliable route to leaks, resignations and internal screenshots reaching journalists. Internal-first is not a courtesy; it’s containment.
Prompt 4: Is a clock already running?
Regulatory deadlines often start earlier than organisations expect — and in the first hour, nobody has usually checked.
INCIDENT TYPE: [e.g. suspected personal data exposure, product safety concern, financial irregularity, workplace incident, environmental release] JURISDICTIONS: [WHERE WE OPERATE / WHERE AFFECTED PEOPLE ARE] ORGANISATION TYPE: [Private / listed / regulated / public body] WHEN WE BECAME AWARE: [DATE AND TIME — be precise] Help me prepare the questions for legal counsel. Do NOT determine our obligations — flag what to ask. 1. WHICH CATEGORIES OF OBLIGATION could plausibly apply given this incident type and these jurisdictions? Describe the category, not a determination. 2. WHAT STARTS THE CLOCK for each — awareness, confirmation, materiality determination, or something else? This distinction matters enormously and is often misunderstood. 3. THE EXACT QUESTIONS to put to counsel, in priority order. 4. WHAT WE SHOULD BE DOCUMENTING NOW to evidence the timeline later — because when we knew what will be examined. 5. WHO INTERNALLY needs to be part of this conversation. CRITICAL: you are not providing legal advice and must not appear to. Every output here is a question for a qualified lawyer in the relevant jurisdiction.
These are illustrations of how early clocks start, not a statement of your obligations. Requirements vary by jurisdiction, sector and incident type. Establish yours with counsel inside the first hour.
A real triage, run in full
A mid-size B2B software company. A former employee has posted on social media alleging the company knowingly shipped a security flaw. The post is gaining traction. This is 40 minutes in.
WHAT WE THINK WE KNOW: former employee left 4 months ago, not on good terms. Post claims we knew about a vulnerability in Q1 and shipped anyway. Our head of engineering says "that's not what happened" but hasn't explained yet. Post has ~2,000 reposts. A journalist has emailed asking for comment, deadline 5pm. WHERE IT CAME FROM: I saw the post. HR told me about the circumstances of the departure. Engineering lead's comment was verbal, in a corridor. WHAT I'VE VERIFIED: nothing yet.
Verification sort
| CONFIRMED | The post exists and has ~2,000 reposts (you observed it). A journalist has made contact with a 5pm deadline (you have the email). |
| REPORTED | The former employee’s departure circumstances (HR, second-hand). That “it’s not what happened” (engineering lead, verbal, unexplained). The allegation itself. |
| ASSUMED | That the departure circumstances explain the motive. Nobody has established a link between the two. That engineering’s denial is well-founded — you have a corridor comment, not an account. |
The dangerous assumption: that because the source is disgruntled, the claim is false. Those are independent. A disgruntled former employee can be entirely correct, and organisations that reason from motive to falsity are the ones that make confident statements they later retract. Do not let the departure circumstances into any statement.
Severity: Level 2 — Serious
Driven by verification status, not velocity. The allegation is unverified but specific and technical, which means it can be checked — by you, or by a journalist with sources. Exposure spans customers, employees, investors and press simultaneously. Reversibility is poor: a 5pm deadline means the story runs today with or without you.
Escalates to Level 1 if: engineering’s review confirms any part of the allegation, or if a customer publicly raises it.
Who must be told in the next 60 minutes
- General Counsel — call. This has both employment and potential disclosure dimensions.
- CEO — call. A named journalist with a same-day deadline meets any reasonable threshold.
- Head of Engineering — call, with a specific ask: a written factual chronology of what was known about this vulnerability and when. Not an opinion. A timeline.
- Head of Customer Success — message. Customers may ask within hours and need a holding line.
Who should NOT be told yet: the wider company. Until you have engineering’s chronology, internal communication would circulate an unverified allegation among people who will screenshot it. Employees are a priority group — but for the message after verification, not this hour.
The decision that cannot wait: whether you respond to the journalist before 5pm, and if so with what. That’s a CEO and GC decision, and it needs the engineering chronology to be made properly. Which means the chronology is the critical path — request it now with an explicit deadline.
The hour-two question: what does the engineering record actually show? Not what the team remembers — what the ticketing system, the commit history and the release notes show. That’s the difference between a defensible statement and another assumption.
Note what the triage did: it identified that the reassuring assumption — “disgruntled ex-employee, therefore false” — was the most dangerous thing in the room, and it made the engineering chronology the critical path rather than the statement. Forty minutes in, the organisation now knows what it’s waiting on and who’s deciding.
Level-up: what does our response create?
This is the prompt almost nobody runs, and it’s what separates handled crises from reactive ones. Every statement generates the next cycle — new questions, new headlines, new obligations, and a permanent record. Modelling that before you speak is the difference between choosing your second day and having it chosen for you.
We are considering the response below. Before it goes anywhere, model what it creates. THE SITUATION: [TRIAGE OUTPUT] THE DRAFT RESPONSE: [PASTE IT] WHO IT GOES TO: [Press / customers / employees / all] Seven sections: A. THE NEXT QUESTIONS — what does this response invite someone to ask next? List the five most likely follow-ups from a competent journalist, and note which we can currently answer. B. WHAT WE'VE COMMITTED TO — every implicit or explicit promise in this draft. Investigations promised, timelines implied, standards asserted. Each becomes a future obligation we can fail. C. THE HOSTAGE TO FORTUNE — the sentence most likely to be quoted back to us in three months if the facts develop unfavourably. Quote it exactly. D. WHO ELSE THIS ACTIVATES — which stakeholder or party might now feel compelled to respond, comment or escalate because we said this? Regulators, competitors, former employees, plaintiffs' lawyers. E. THE PERMANENT RECORD — this statement will be indexed and retrieved by AI search engines for the foreseeable future. Is it accurate, specific and self-contained enough to be quoted out of context without misleading? If it would read badly stripped of surrounding context, rewrite it. F. THE SILENCE ALTERNATIVE — what happens if we say nothing for another four hours? Sometimes that's worse and sometimes it isn't. Argue both sides honestly. G. THE ONE CHANGE — the single edit that most reduces downside without weakening the response. RULES: - Do not soften the analysis. We need the objections now, not from a journalist. - Do not rewrite the statement unless asked — this is assessment. - Flag anything requiring legal review before release.
Section E is the one this industry hasn’t caught up with. A crisis statement is no longer a document that circulates and fades — it becomes a primary source that AI engines retrieve and quote for a long time afterwards. Which means it needs to survive being pulled out of context: specific, accurate, and complete enough to stand alone. That’s a drafting standard crisis comms hasn’t traditionally applied, and it now matters more than the press release format.
The first 72 hours no longer just shape coverage. They build the citation record that AI engines retrieve for the next 18 months.
Silence isn’t neutral any more. If you publish nothing while others publish plenty, theirs becomes the source of record — and it stays that way long after the news cycle ends.
What AI must never do in a crisis
| Never | Why |
|---|---|
| Fill a gap in the facts | An invented detail in a crisis statement is a legal event, not an inaccuracy. Every prompt here instructs the model to flag rather than infer — keep it that way. |
| Determine legal or regulatory obligations | Jurisdiction-specific, fact-specific, and consequential. AI prepares the questions; counsel answers them. |
| Decide whether to apologise | Admission of liability has legal and insurance consequences. That’s a decision for counsel and leadership. |
| Release anything unreviewed | Every output is a draft. No prompt on this page produces a sendable statement. |
| Hold privileged material | Don’t paste legal-privileged communications, regulatory correspondence or investigation findings into a consumer tool. Use an enterprise tier with proper data terms, and check first. |
| Assess whether an allegation is true | It cannot know. It can only structure what you’ve told it. Verification is human work against records. |
Which model, and one setup note
Use a reasoning-tier model for the triage and the second-order prompt — both reward careful analysis over speed, and the few extra seconds are irrelevant against a 60-minute window. Use a large-context model if you’re pasting long threads, article text or message histories.
The setup note that matters more than model choice: decide now, before an incident, which tool your team will use and whether it has appropriate data terms. Nobody should be evaluating AI privacy policies at 11pm during a live crisis. Put it in the crisis plan alongside the phone list.
This category moves quickly — detection capability, platform policy and disclosure rules all shift. We re-verify on each review cycle; confirm current obligations with counsel regardless.
Common questions
What should you do in the first hour of a crisis?
Verify, classify, notify, and only then draft. The most common first-hour error is beginning to write a statement before establishing what is actually known and how serious the incident is. Verification means separating confirmed facts from reported claims and assumptions. Classification determines who must be told and how urgently. Drafting comes last, and no external statement should be released without legal review.
How do you assess the severity of a communications crisis?
Across six dimensions: whether anyone has been harmed, whether the claim is verified or alleged, how fast it is spreading, which stakeholders are exposed, whether a regulatory obligation is triggered, and whether the situation is reversible. Harm to people overrides all other considerations. Most organisations classify into three to five severity levels, each mapped to a defined notification chain and escalation timeframe.
Should you respond immediately when a story breaks?
Respond quickly, but classify first. Speed matters more than it used to, since manufactured crises can spread within minutes rather than hours, but a fast statement built on unverified information frequently requires retraction, and a retraction is a second crisis larger than the first. The practical approach is a short acknowledgement that commits to nothing while verification proceeds.
Can AI help during a live crisis?
For structuring assessment, mapping stakeholders, anticipating questions and drafting variants, yes. For deciding what to do, no. AI cannot verify facts it has not been given, cannot assess legal exposure, and must never be permitted to fill gaps in a crisis, because an invented detail in a crisis statement is a legal event rather than an inaccuracy. Every output is a draft for human and legal review.
What regulatory deadlines apply when a crisis breaks?
It depends on jurisdiction and incident type, and the clocks often start earlier than organisations expect. Under GDPR, personal data breaches must be reported to the supervisory authority within 72 hours of becoming aware, and awareness rather than full technical understanding starts that clock. Listed companies in the United States must disclose material cybersecurity incidents within four business days of determining materiality. Establish which obligations apply within the first hour, with counsel.
Who needs to be told in the first hour of a crisis?
That depends on severity, which is why classification precedes notification. At the highest severity, that typically includes the chief executive, legal counsel, the communications lead and any function directly implicated, contacted by phone with confirmed receipt rather than by message. Lower severity incidents may require only the communications team and the relevant function. Establishing decision rights before an incident is what makes this fast during one.
What is the difference between what you know and what you have been told?
Confirmed information has been independently verified against a record, a system or a document. Reported information has been stated by someone but not yet checked. Assumed information is inference filling a gap. In the first hour these three categories are routinely conflated, and statements built on the second and third are the most common cause of later retraction. Every crisis assessment should label each item explicitly.
How fast do crises spread now compared with previously?
Considerably faster, to the point that practitioners increasingly describe the response window in minutes rather than the traditional golden hour. Manufactured incidents including fabricated video and documents can reach very large audiences before verification is complete, and the material published during the first seventy-two hours forms the primary source record that AI search engines subsequently retrieve. Speed of classification therefore matters more than it once did.
Download: The First-Hour Crisis Triage Card
A single printable page — the four-step sequence, the six severity dimensions, the four-level classification with notification chains, and the questions for counsel. Designed to be pinned up, not read.
Agencies: supplied unbranded and free to white-label for client crisis plans.
Enter your email and we’ll send the card plus occasional crisis comms updates. Unsubscribe anytime.
Next in this series
About this guide
Narracomm is a communications and content strategy team. We build and test prompt systems inside live client work and revise them as models and conditions change. [REQUIRED BEFORE PUBLISHING: named reviewer with genuine crisis experience — a former agency crisis lead, in-house comms director, or a lawyer who has run incident response — with credential and review date shown. This page may be read by someone acting under pressure, and that makes review non-negotiable rather than advisable.]
Sources & further reading
- GDPR data breach notification requirements (2026)
- GDPR Article 33 — the 72-hour breach notification rule
- Everything-PR — the 72-hour AI crisis playbook
- Death of the golden hour: navigating AI-manufactured crises
- 2026 Global Crisis Comms Report — AI risks up, protocols missing
- Incident severity levels: classification and escalation
Scope: this guide is general information about communications process. It is not legal advice, does not determine your regulatory obligations, and must not delay safety response or professional counsel. Regulatory references are illustrative and jurisdiction-specific — verify yours with a qualified lawyer. Last reviewed: July 25, 2026 · Next review due within 14 days.
