Can AI Handle Crisis Communications? Where It Helps and Fails

AI can handle parts of crisis communications, but not the crisis. It's genuinely useful for monitoring mentions at volume, summarising what's being said, and drafting statement variants fast. It fails at the decisions that determine how a crisis ends: what you know versus what you're guessing, which stakeholder hears first, when to speak and when to wait, and what the company is willing to lose to stay credible.

I've worked through enough market shocks in crypto to say this plainly. The tools got faster. The hard part didn't move.

Why is crisis communications different from routine PR work?

Routine PR rewards preparation. Crisis PR rewards judgment under incomplete information.

In a launch, you know the facts, you control the timing and you can polish every line. In a crisis, the facts change while you're writing, someone else often controls the timing, and the first statement you publish will be compared against everything you learn later. A hack, an exploit, a model producing harmful output, a regulator's notice, a founder's bad post: each one starts with a team that doesn't yet know the full picture.

AI tools are trained to produce a confident, complete answer. A crisis rarely has one. That mismatch explains most of what follows.

Where does AI genuinely help in a crisis?

It's fair to start with what works, because refusing useful tools during an incident is its own mistake.

• Monitoring at volume. Pulling mentions from X, Telegram, Discord, Reddit and news into one stream, and flagging spikes. No person reads that fast.
• Summarising the conversation. Turning thousands of posts into "the three things people are angry about" so the team can decide what to address.
• Drafting variants. Producing three versions of a holding statement at different lengths and tones in minutes, so humans can choose and edit.
• Translation drafts. First-pass versions for regional communities, which a native speaker then checks before anything is posted.
• Q&A preparation. Generating the hostile questions a reporter or community member is likely to ask, so the spokesperson isn't surprised.
• Timeline building. Organising internal notes into a clean sequence of events for the post-incident write-up.

Every item on that list is about speed and volume. None of them is a decision.

Where does AI fail when the story is still unfolding?

It fails in four places, and each has a cost you can name.

It can't tell known from assumed. Ask a model to draft an exploit statement and it will often state a cause, an amount and a fix, because that's what complete statements contain. If your engineers haven't confirmed the cause, that sentence becomes the thing you retract later.

It can't read the room. A statement that would be fine on a normal Tuesday can land badly if the market is already falling or a competitor was hacked last week. The model doesn't know what your community went through yesterday.

It can't weigh stakeholders against each other. Users want reassurance, investors want facts, regulators want accuracy, lawyers want caution, and the community wants it all now. Each of those groups needs a different message at a different moment. Deciding the order is a trade-off between real interests, and a model has no stake in any of them. That's why I've long argued a company's communicator is a competitive asset, not a support function:

if your company has a good external/internal communicator, you’re half way ahead of your competitors. Shilika Jain on X, 29 Aug 2023

It can't push back on the founder. In a crisis, founders often want to fight. Their instinct is to defend, correct and counter-attack, sometimes in public, sometimes within minutes. I once summed it up in one line:

founders are naturally born FUD-fighters. Shilika Jain on X, 27 Feb 2023

That instinct is useful when the criticism is false and you can prove it. It's dangerous when the facts aren't in yet. A model will happily draft the counter-attack the founder asks for. A comms lead's job is sometimes to say "not yet" and hold that line.

The Crisis Clock: who does what in each phase

This is how I'd split the work between tools and people across the life of an incident. The phases overlap in practice, but the split holds.

The "first statement" row is where most damage happens. A good holding statement says what you know, says what you're doing, and says when people will hear from you next. It doesn't guess. My holding statement templates are built around that structure, and the human work is filling them with only confirmed facts.

Three questions a model can't answer for you

When I'm in the first hour of an incident with a client, these are the questions that matter, and none of them can be delegated to a tool:

1. What are we certain of, and are we willing to say what we don't know yet? Admitting uncertainty early buys credibility later. Models tend to fill the gap instead.
2. Who needs to hear this before the public does? Exchanges, partners, investors, affected users, sometimes regulators. The order is a judgment about relationships and obligations.
3. What are we willing to lose to stay credible? Sometimes the honest statement costs short-term price, users or a partnership. Deciding to accept that cost is leadership, not drafting.

All three answers come from the same place. A crisis puts enormous pressure on a team to say something, anything, fast. I wrote a one-line rule for that moment:

PR should follow purpose, not pressure! Shilika Jain on X, 21 Apr 2025

A model responds to the pressure of the prompt. A person can decide the purpose first: protect affected users, stay credible with reporters, keep the facts straight. Being transparent under pressure is a choice someone makes. A tool can format that choice. It can't make it.

What new risks does AI add during a crisis?

Using AI in an incident brings its own hazards, and they're worth planning for before you need them.

• Confidentiality. Pasting incident details, wallet addresses, legal advice or internal logs into a third-party tool may breach your own policies or create discovery issues. Agree with counsel in advance what can be shared with which tools.
• Speed without review. AI makes it possible to post within minutes. That's a risk when the statement hasn't been checked by engineering and legal.
• Uniform tone. A crisis statement that sounds generated reads as evasive. People want to hear a person who's accountable.
• Model-caused crises. For AI companies, the crisis itself may be the model's output. Using the same class of tool to write the apology needs extra care. I cover that case in the AI hallucination crisis playbook.

How should a small team set this up before anything goes wrong?

Most startups don't have a crisis team. They have a founder, a few engineers, a community lead and maybe a fractional comms person. Set up the split before the incident:

• Pick one decision-maker for public statements and one backup.
• Pre-approve which AI tools can be used on incident data, and which can't.
• Prepare holding statement templates with blanks for confirmed facts only.
• Agree the review chain: engineering confirms facts, legal checks wording, comms owns tone and timing.
• Decide your update cadence in advance, so nobody improvises it at 3am.
• Keep a stakeholder list with contacts, in notification order.

If you're in Web3, my crypto crisis communications playbook walks through this in more depth. If you'd rather have an experienced person in the room when it happens, crisis work is part of my Web3 PR campaigns engagements.

When can you lean harder on AI?

Not every incident is a crisis. For low-stakes issues, such as a short outage with a known cause, a minor bug, or a delayed feature, AI-drafted updates with a quick human check are fine. The facts are clear, the audience is mostly users, and the cost of a slightly generic update is small.

The line I use: if the statement could be quoted in a news story or a legal filing, a person writes the final version. If it's an operational status update, AI can draft it and a person approves it.

FAQ

Should AI tools monitor social media around the clock for early crisis signals?

Yes, if you have someone who looks at the alerts. Monitoring is one of the best uses of AI in comms, but an alert nobody reads is worthless. Set thresholds so the tool flags genuine spikes, and route them to a named person with authority to start the response.

Can an AI chatbot answer user questions during an incident?

Only with tight limits. A support bot can point users to the official status page and the latest statement. It should not improvise answers about causes, losses or compensation, because anything it says will be screenshotted and treated as the company's position.

Is it acceptable to use AI to draft a public apology?

As a starting structure, yes. But an apology has to sound like the person signing it and has to name specific failures and specific fixes. If it reads like a template, people will treat it as one. Expect to rewrite most of it.

What's the biggest mistake teams make with AI in a crisis?

Publishing faster than they can verify. The tool removes the delay that used to force a second look. Put that delay back on purpose: no statement goes out until engineering has confirmed every fact in it.

Want your crisis plan stress-tested before you need it? Grab 30 minutes with me.