Rehearse a crisis before you are in one
Test a holding statement, an apology, or a silence against simulated press, customers, and staff — and see which response de-escalates and which one feeds the story.
MiroFish simulates how a crisis unfolds under each response you are considering. You describe what happened, what is already public, and the statement you are drafting; the engine models journalists, affected customers, staff, and opportunistic competitors as interacting agents, then runs the story forward over several news cycles. The report gives the most-likely trajectory with an estimated probability, which specific phrase in your draft is most likely to be quoted against you, whether the story escalates or decays, and the response variant that performs best. It is rehearsal, not prophecy — it cannot know what a journalist has that you do not.
How MiroFish simulates it
A crisis simulation is run under time pressure, so the pipeline is tuned for speed and for comparing response options:
- 1
State the facts and the exposure
What happened, what is already public, what is provably true, and what you fear comes out next. The gap between those last two is usually what determines the trajectory, so name it honestly — the simulation is only as good as the exposure you admit to it.
- 2
Model the actors, including the hostile ones
Reporters on the beat, affected customers, your own employees, regulators if relevant, and competitors with an incentive to amplify. Employees are the actor most teams forget, and frequently the source of the second-day story.
- 3
Run each response variant separately
Simulate the full apology, the narrow acknowledgement, and saying nothing as three separate runs. Comparative output is far sturdier than one absolute prediction, and the differences between variants are the actual decision.
- 4
Trajectory and quotable-line report
Peak-attention timing, whether the story decays or escalates by cycle, the phrase in your draft most likely to be lifted as a headline, second-day angles to prepare for, and the variant with the best estimated outcome.
A worked example
A fintech app suffers a 14-hour outage during a payday weekend. Customers could not access funds. A local paper has asked for comment and a viral thread has 40k views.
What you give it
- Incident
- 14h outage, payday weekend
- Already public
- Viral thread, 40k views
- Draft response
- Technical explanation, no compensation
- Feared next
- Root cause was a skipped upgrade
What the report estimates
- Most-likely trajectory
- Escalates one cycle, then decays
- Riskiest draft phrase
- "a small number of users"
- Second-day angle
- Why the upgrade was deferred
- Best variant
- Acknowledge + goodwill credit, same day
Illustrative output. Crisis trajectories are highly sensitive to facts not yet public — treat this as rehearsal, not forecast.
What it can’t do (honest limits)
- —It cannot know what a journalist already has. If there is a document you have not disclosed to the simulation, the real trajectory can diverge sharply.
- —Crisis prediction is the least certain thing this engine does — confidence figures here are lower than on pricing or launch work, and should be.
- —It models a plausible press and public reaction, not named journalists or a specific outlet’s editorial line.
- —It is not legal advice. Where liability, regulatory disclosure, or employment law is involved, the simulation is an input to counsel, never a substitute.
Questions people ask
Can AI predict how a PR crisis will unfold?
It can estimate a trajectory rather than predict events. A multi-agent simulation models the actors — press, customers, staff, competitors — and plays their reactions forward over several news cycles, returning the most-likely path with an estimated probability alongside alternatives. Crisis is the hardest category to forecast because the decisive variable is often information you do not yet know is coming, so treat the output as structured rehearsal rather than a forecast you can bank.
How do I test a crisis statement before publishing it?
Run each version as a separate simulation and compare the reports. MiroFish flags which specific phrase in your draft is most likely to be quoted back at you — usually a minimiser like "a small number of users" or "an isolated incident" — and estimates whether that framing escalates or defuses the story. Comparing two or three variants is more reliable than judging one in isolation.
Is it better to respond immediately or wait?
Simulate both. Speed usually reduces escalation risk when the facts are established and you can concede something concrete; waiting tends to perform better only when a fast statement would have to be walked back. The report estimates the trajectory under each timing choice, which is more useful than a general rule, because the answer depends on how much you can honestly say today.
What is a second-day story and can it be anticipated?
The second-day story is the follow-up angle that appears once the initial facts are reported — often about why the failure was possible rather than the failure itself. The simulation surfaces the most likely second-day angles from the exposure you describe, which is why naming what you fear comes out next materially improves the output.
Should employees be considered in a crisis simulation?
Yes, and they are the most commonly omitted actor. Staff read the same statement your customers do, and a response that feels dishonest internally frequently produces the leak or the anonymous quote that drives the second cycle. MiroFish models employees as their own agent group for this reason.
Can it tell me whether to apologise?
It estimates the outcome of apologising versus acknowledging versus staying silent, given what is public and what is true. What it cannot weigh is legal exposure — an apology that reads well publicly may carry admissions that counsel would strike. Use the simulation for the reputational half of that decision and a lawyer for the other half.
How accurate are crisis predictions?
Less accurate than the other things this engine does, and the reports say so — confidence estimates on crisis work run materially lower than on pricing or launch simulations. The value is in surfacing the angles and phrasings you had not considered while there is still time to change them, not in the probability figure itself.
Can I simulate a crisis that has not happened yet?
Yes, and it is the better use. Running your three most plausible failure scenarios before they occur gives you drafted, tested responses and a sense of which failures escalate fastest. That is ordinary crisis preparedness work, done in minutes rather than in a workshop.
Go deeper on the method: read the full guide on the blog →
Run it on your own scenario
No install, no API keys. Monthly plans from $2.99 — results in about 30 seconds, cancel anytime.
Start predicting