Skip to content
World Model Readiness
Engraved governance instrument

For your company

Governance & Compliance

Module · The night the model gets it wrong

The AI Incident Preparedness Check

Most AI failures are not dramatic; they are quiet, wrong outputs that ship for hours before anyone notices. This module tests the machinery that turns a misfire into a contained event: whether you would detect it, who acts, how you communicate, and whether anyone has ever rehearsed it.

Question 1 of 5 · Misfires get detected

Would you detect an AI system producing wrong outputs before your customers did?

Most AI failures are silent: plausible outputs that are simply wrong. Without monitoring on the output, not just on uptime, the first alarm is a customer complaint or a regulator's letter.

Question 2 of 5 · A playbook exists

Is there a written playbook for what to do when an AI system misfires?

A playbook names the first moves: how to pause the system, who to call, what evidence to preserve. Instinct is slower than a checklist, and incidents rarely happen at a convenient hour.

Question 3 of 5 · Roles are named

When an AI incident hits, does everyone know who decides to pause the system?

The costly minutes of an incident are spent finding out who is allowed to act. One named person with authority to pull a system offline beats a chain of approvals every time.

Question 4 of 5 · Communication is ready

Do you know what you would tell customers, staff, and regulators if an AI decision caused harm?

Disclosure duties come with clocks: the GDPR and the AI Act both impose notification timelines. The messages you draft under pressure and legal review are the ones that cost you.

Question 5 of 5 · It has been rehearsed

Have you ever run a drill for an AI system failing?

A plan on paper is a hypothesis. A tabletop exercise, walking a plausible misfire through with the people who would respond, is the cheapest way to find out where it breaks.

For the statistics · one click each

Three questions for the public picture

These do not affect your score. They feed the anonymised, aggregated statistics; groups under 8 respondents are never shown.

How would you most likely first learn that an AI system had misfired?

Do you have a written plan for AI incidents?

Have you ever rehearsed an AI system failing?

Your context

Used to calibrate the report. Company size and sector remain in the anonymized dataset; your email does not.

What the five levels look like

Every dimension in this assessment is scored 1 to 5. This is what the levels mean, dimension by dimension. The graded report diagnoses where your own answers land and what to do about it.

Misfires get detected

  1. 1Customers tell us
  2. 2Manual spot checks
  3. 3Basic alerting
  4. 4Output monitoring
  5. 5Live drift detection

At the low end: If customers are your monitoring, the damage is done before you know. Put one automated check on the output of every AI system that touches a customer, even a crude one. What good looks like: Live drift detection is where you want to be. Wire the alerts to the named owner from question 3, so detection actually triggers action.

A playbook exists

  1. 1No plan
  2. 2Improvise each time
  3. 3Generic IT runbook
  4. 4AI-specific playbook
  5. 5Playbook, tested

At the low end: Improvised incidents are slow incidents. Write one page this month: how to pause the system, who to call, what to preserve. A rough playbook beats a perfect instinct at 2am. What good looks like: A tested playbook is rare. Keep it current as systems change, and rehearse it (question 6) so it is muscle memory, not a document nobody has opened.

Roles are named

  1. 1Nobody decides
  2. 2Escalates upward slowly
  3. 3Decision by committee
  4. 4Named owner
  5. 5Named, with authority

At the low end: When nobody owns the pause decision, everybody waits. Name one person per system who can pull it offline without asking permission, and tell them. What good looks like: A named owner with real authority is the difference between minutes and hours of exposure. Make sure they are reachable out of hours, not just on paper.

Communication is ready

  1. 1No idea
  2. 2Legal, after the fact
  3. 3Ad hoc drafting
  4. 4Templates ready
  5. 5Templates + obligations mapped

At the low end: Silence during an incident is its own scandal. Draft the three messages now: to customers, to staff, to your regulator. You will not write them well under pressure. What good looks like: Ready templates plus mapped obligations is mature. Review them when the law or your systems change, and make sure someone owns pressing send.

It has been rehearsed

  1. 1Never
  2. 2Discussed once
  3. 3Tabletop, once
  4. 4Annual drill
  5. 5Regular, cross-team drills

At the low end: A plan never rehearsed is a plan that fails on contact. Run one tabletop this quarter: pick a plausible misfire and walk it through with the people who would respond. What good looks like: Regular cross-team drills are what turn a playbook into readiness. Track what each drill exposes and feed the fixes back into the plan.