Home/Blog/AI Governance & Compliance
AI Governance & Compliance · 12 min read

AI incident response: how to detect, contain and learn from AI failures

The short answer

To respond to an AI incident, acknowledge the signal, set severity, take the smallest action that stops new harm (revoke an agent's access rather than asking it to stop), keep evidence, tell your data protection lead in time for any reporting clock, explain honestly to the people affected and run a blameless review. Most of the work is preparation: a register of your AI systems with owners and tested stop buttons, a simple reporting route, and practice.

Want the tools, not just the method? The The AI Incident Response Playbook kit turns this guide into ready-to-use templates, spreadsheets and printables, with an instant download.

Key takeaways
  • Near-misses count: log them, they are the cheapest lessons.
  • Stop agents by revoking access, never by instruction.
  • One AI incident can start several reporting clocks; escalate early so the right people decide in time.
  • Own the error in every customer message; write to everyone affected.
  • Blameless reviews and quarterly tabletop exercises keep the playbook alive.

What is an AI incident?

An AI incident is an event where an AI system you use caused, or nearly caused, harm to people, their rights, property or your organisation, or did something its owner did not intend. Near-misses count: an odd answer caught before a customer saw it is the cheapest lesson you will ever get.

The OECD's February 2025 paper on a common reporting framework for AI incidents separates incidents, where harm occurred, from hazards, where harm could plausibly follow, and proposes 29 criteria for describing them. A small organisation does not need all 29. It needs enough fields to see patterns: the system, the failure type, severity, the people affected, whether personal data was involved, and the times.

What should an AI system register contain?

You cannot respond to an incident in a system you do not know you run. Start with a register of every AI system you build, buy or switch on, including AI features that arrive inside tools you already pay for through upgrades. In the worked example, IT listed four systems and the register ended with twelve, because transcription, email drafting and photo checking had arrived as features nobody chose.

For each system record a named owner who can decide to narrow or pause it, the vendor and its incident contact, and five yes-or-no questions: is it customer-facing, does it act on its own, does it use personal data, does it make or support decisions about people, and does it affect people in the EU. Those answers set a risk tier. Add the kill-switch status (tested, written, none) and one monitoring check that would catch a problem before customers complain, such as a weekly sample of chatbot answers checked against your policies or a daily count of an agent's actions.

How do AI systems fail?

Most incidents in organisations that use AI fall into six types, adapted here from the OWASP Top 10 for LLM Applications 2025 into plain words.

  • Wrong information: confident and false, as in Moffatt v. Air Canada (2024) and New York City's MyCity chatbot (2024).
  • Data exposure: personal or confidential data in answers or summaries, or pasted into unapproved tools.
  • Manipulation: prompt injection, first on OWASP's list, where hidden instructions in text make the AI misbehave.
  • Unfair outcome: one group ranked or scored worse, often without complaints.
  • Agent beyond its permissions: OWASP's excessive agency, as when an AI coding agent deleted a live database in July 2025.
  • Runaway cost or outage: loops, usage spikes and failures.

Who should own AI incident response?

Name one incident lead and a deputy, usually from risk, operations or compliance rather than IT alone, because AI incidents often start in places security teams do not watch: a complaints inbox, a lettings desk, a recruitment shortlist. Each AI system has its own owner who chooses the containment rung. The data protection lead decides on personal data notifications with legal advice, and an executive sponsor approves the most serious decisions and any external reporting.

Make reporting easy and safe. Publish one route, such as a team chat channel and a phone number, with the words 'you do not need to be sure', and a five-minute form that asks for the exact words the AI used, the time, who may be affected and whether personal data, money or decisions about people could be involved. Thank every reporter and tell them what changed: the way the last reporter was treated predicts whether there will be a next one. Test the route by asking a new starter to report a made-up problem using only what they can find; if it takes more than five minutes, simplify it.

What should you do in the first hour?

Speed matters more than perfection in the first hour. Use a simple timetable and the same five facts every time you escalate.

  • 0-10 minutes: acknowledge the report; ask whether it is still happening and who could be hurt.
  • 10-20: set severity S1-S4; for S1 or S2 call the system owner and sponsor.
  • 20-35: contain with the lowest rung that stops new harm.
  • 35-45: keep evidence and ask the vendor in writing to preserve logs.
  • 45-55: start the clock tracker and tell the data protection lead if personal data may be involved.
  • 55-60: send a holding note to affected teams and set the next update time.

How do you contain an AI incident without switching everything off?

Use a four-rung ladder and pick the lowest rung that stops new harm. Warn and watch suits isolated, low-harm signals. Narrowing blocks the affected topic or function and routes those cases to people, which stops harm while keeping the value. Pausing switches the feature off and uses a manual fallback. Stopping an agent means revoking its permissions or using a tested kill-switch.

Never rely on telling an agent to stop. Instructions are followed most of the time; permissions work every time. For every agent, record each permission, whether it is needed, how to revoke it in minutes, who can, the fallback and the date it was last tested. In our worked example the first test took 40 minutes and the second took six.

Every containment needs a matching restart decision: restart only when the cause is understood, the fix passes a test against the case that failed, monitoring that would catch a repeat is running, and the owner has signed the decision log.

Measure where you are

Reading helps; measuring tells you what to work on. These AI-graded assessments on AssessAll pair with this topic:

Which reporting clocks can an AI incident start?

The clocks belong to the type of harm, not to the fact that AI was involved, and one incident can start several at once. The position below is correct at November 2026; check the official source and your adviser.

  • EU AI Act Article 73: providers of high-risk AI report serious incidents no later than 15 days after awareness, 2 days for widespread infringement or serious critical-infrastructure disruption, 10 days where a person dies. Deployers inform the provider. High-risk duties apply from 2 December 2027 (Annex III).
  • UK GDPR and EU GDPR: personal data breaches likely to risk people's rights are reported without undue delay and, where feasible, within 72 hours.
  • Canada: breaches with a real risk of significant harm are reported to the OPC; records of all breaches are kept for two years.
  • Australia: the Notifiable Data Breaches scheme requires notifying the OAIC and people when serious harm is likely.
  • New Zealand: notifiable privacy breaches go to the Privacy Commissioner as soon as practicable; the guidance says ideally within 72 hours.
  • India: DPDP Rules require a detailed report to the Data Protection Board within 72 hours (from May 2027); CERT-In requires cyber incident reports within 6 hours.
  • US: no federal AI incident law; state breach laws and the FTC Act apply, and Colorado SB 26-189 sets notice duties for consequential AI decisions from 1 January 2027.

How do you explain an AI incident to customers?

An honest explanation has four parts: what happened in plain words; what it means for this person and the remedy; that you own it; and what you are changing, with a named contact. Never write 'the AI did it'. In Moffatt v. Air Canada the British Columbia Civil Resolution Tribunal rejected the airline's argument that its chatbot was responsible for its own actions.

Write to everyone affected, not only those who complained. In the worked example, 37 tenants complained but the review found 96 who had been given the wrong timescale. None of the 59 who had not complained escalated after receiving an honest letter.

How do you learn from AI incidents?

Hold a blameless review within ten working days of containment: a timeline built from evidence, scope including silent affected people, five whys until the answer is about testing, monitoring, ownership or design, and fixes with owners and dates. Blame stops reporting; if people expect to be blamed, the next incident is found by a journalist.

Practise before you need it. Run a 60-minute tabletop exercise each quarter with timed injects, such as a chatbot that promised too much, an agent that kept going, or a summary sent to the wrong people, and score it on acknowledgement, containment, escalation and learning.

What does a twelve-week programme look like?

The kit's plan, followed by a fictional 240-person letting agency in Manchester, moves from a list of four AI systems to a register of twelve, from a self-assessment score of 12 of 40 to 32, and from 52 hours to escalate the first incident to 2 hours by the eighth.

  • Weeks 1-3: decision memo, register, self-assessment and gap plan.
  • Weeks 4-6: reporting route, severity levels, kill-switch tests.
  • Weeks 7-9: clocks with the data protection lead, message library, staff briefing.
  • Weeks 10-12: two tabletop exercises, re-score, report to leadership.
Bodhih Pro Kit

Run the whole AI incident job from one kit

The AI Incident Response Playbook gives you the workbook, forms, message library, tabletop exercises and a step-by-step e-book. Not affiliated with or endorsed by OWASP, NIST, ISO, IEC, the OECD, the European Commission or any regulator named.

Common questions

Questions people ask next

What is an AI incident?

An event where an AI system you use caused, or nearly caused, harm to people, their rights, property or your organisation, or did something its owner would not approve. Near-misses count. Chapter 1 gives a three-part definition and six failure types.

Do I need an AI incident response plan if we only buy AI tools?

Yes. Most incidents in organisations that buy AI start with a vendor's system, but the harm lands on your customers and staff, and you remain responsible for what your tools tell people. The kit is written for organisations that use AI rather than build it.

Do I have to report AI incidents under the EU AI Act?

Providers of high-risk AI systems report serious incidents; deployers who identify one must inform the provider. These duties apply with the high-risk rules, from 2 December 2027 for Annex III systems after the Digital Omnibus on AI. Personal data breaches have separate clocks under GDPR. Your adviser decides what applies to you.

Is there an AI incident reporting law in the UK or US?

At November 2026 there is no general AI incident law in the UK, and the US federal AI Incident Reporting Act is a bill. Data protection, breach-notice and consumer laws still apply. The Rules by market sheet has the dated position and official sources.

How do I stop an AI agent that is misbehaving?

Revoke its permissions or use a tested kill-switch. Do not rely on telling it to stop. The Agent kill-switch and permission check gives you one page per agent and a time-to-stop to improve.

How long does it take to set up?

The kit's plan runs over twelve weeks at about 140 hours of the incident lead's time, plus a 45-minute briefing for staff. Your first hour gets the register started and the reporting route published.