When an AI system goes wrong, stop it first, then work out how far the damage spread, tell the people affected, fix the cause and change something so it can't happen the same way again. That's the core of AI incident response, and it only works quickly if you prepared before anything broke: a kill switch that anyone on duty can use, logs that show what the AI did, and a named person who decides.

AI incidents look different from ordinary IT outages. The system is often still "up" and confidently doing the wrong thing. This post covers the types of AI incidents businesses actually face, a step-by-step response playbook, how to communicate, how to find the root cause, and the preparation that makes all of it faster.

What does an AI incident actually look like?

Most AI incidents fall into a handful of patterns:

  • Wrong answers at scale. A customer-facing assistant gives incorrect prices, policies or instructions, and because it's automated, it says the same wrong thing to everyone who asks. Often caused by an outdated document in its knowledge base or a confident hallucination.
  • Data exposure. An AI tool reveals information to someone who shouldn't see it: one customer's details to another, internal notes in a customer reply, or confidential files pasted into an unapproved tool by staff.
  • Runaway agent. An agent with tools gets stuck in a loop, sends hundreds of emails, creates duplicate records, or runs up a large model bill overnight.
  • Manipulation. Someone uses prompt injection to make your AI do things it shouldn't, from embarrassing outputs to unauthorized actions.
  • Harmful or biased outputs. Offensive replies, unfair treatment of certain customers or applicants, or advice the AI should never have given.
  • Silent degradation. A model update, prompt change or data change makes the system quietly worse. Nobody notices for weeks.
  • Vendor incidents. Your AI provider has an outage, a security breach or changes its terms.

Not every glitch is an incident. A single wrong answer caught by a human reviewer is the system working as designed. An incident is when something got past your checks and caused, or could cause, real harm.

How serious is it? A simple severity scale

Decide severity fast, because it determines who gets woken up. A three-level scale is enough for most SMBs:

SeverityExamplesResponse
HighPersonal or confidential data exposed; money moved or promised wrongly; harmful content sent to customers; agent taking unauthorized actionsStop the system immediately; owner and leadership involved within the hour; consider legal and regulatory obligations
MediumWrong information given to a limited number of customers; agent errors that are contained and reversible; noticeable quality dropPause or restrict the affected feature; fix within a day; contact affected customers
LowIsolated wrong answers caught internally; minor cost spikes; cosmetic issuesLog, fix in the normal course of work, review at the next check-in

When in doubt, treat it as one level higher until you know more. Downgrading later is easy.

What should you do in the first hour?

The first hour is about containment and facts, not blame or root cause.

  1. Stop the harm. Use the kill switch: disable the assistant, pause the agent, revoke its credentials or switch it to a "human only" mode. A chatbot can fall back to a contact form; an agent can drop its tasks into a queue for people.
  2. Preserve the evidence. Don't delete logs, conversations or outputs, even embarrassing ones. You'll need them to understand scope and possibly for regulators or insurers.
  3. Name an incident lead. One person makes decisions and keeps a timeline. In a small company, that's usually the AI system's owner or the founder.
  4. Establish the basic facts. What happened, when it started, which system, which customers or records, and whether it's still happening.
  5. Reverse what you can. Cancel queued emails, roll back records, recall messages where possible, stop payments.

If you don't have a way to stop the system within minutes, that's the first thing to fix after this incident. Our guide to AI agent guardrails covers kill switches and action limits in more detail.

How do you work out the scope of the damage?

This is where good logging pays for itself. You need to answer:

  • How many people or records were affected? Search conversation and action logs for the pattern, such as every conversation that mentioned refunds since the knowledge base was updated.
  • What exactly did they see or receive? Pull the actual outputs, not a summary.
  • What actions did the AI take? Emails sent, records changed, bookings made, money moved.
  • Was personal or confidential data involved? And whose?
  • When did it start? Often earlier than the first complaint.

If your logs can't answer these questions, scope becomes guesswork, and you'll likely have to treat everyone as potentially affected. Our post on AI agent monitoring and observability explains what to log so this step takes minutes rather than days.

Who should you tell, and what should you say?

Communicate early and plainly. Customers forgive mistakes far more readily than cover-ups.

Internally: tell staff who deal with customers what happened and what to say, so they're not surprised by calls. Keep leadership updated on a set schedule.

Affected customers: tell them what happened in plain language, what it means for them, what you're doing about it, and what they should do (for example, "the refund window our assistant quoted was wrong; we'll honor what it told you"). Honoring what the AI promised, where reasonable, often costs less than the damage to trust from refusing.

Regulators and partners: if personal data was exposed, data protection laws such as GDPR or India's DPDP Act may require notifying authorities and individuals within set timeframes. Sector rules and contracts may add obligations. This varies by jurisdiction and isn't legal advice; involve counsel quickly for high-severity incidents.

Your vendor: if a third-party AI tool was involved, report it to them and ask for their findings.

Don't blame "the AI" in external messages. Customers hear that as the business disowning its own system. Our post on AI disclosure and customer trust covers the tone that works.

How do you find the root cause?

Once things are stable, run a blameless review within a week. Ask "why" until you reach something you can change.

Typical root causes, and their usual fixes:

  • Stale or wrong source content behind wrong answers: an owner and review date for every document in the knowledge base.
  • Missing check on high-stakes outputs: add a human-in-the-loop approval step, or a validation rule, for that category.
  • Too much permission: narrow the agent's access, add rate limits and spending caps.
  • Untested change: a prompt, model or integration was changed without running the evaluation set. Make testing a release requirement.
  • No alert: the problem ran for days before anyone noticed. Add monitoring for the pattern.
  • Process gap around people: staff used an unapproved tool with sensitive data because the approved route was too slow.

Write up the incident in a page: timeline, impact, root cause, what changed. File it with your AI tools inventory. Over time, these write-ups become the most useful governance document you have.

What should you prepare before anything goes wrong?

A short preparation checklist, sized for a small business:

  1. An inventory of the AI systems you run and the AI tools staff use, each with a named owner.
  2. A kill switch per system, documented, tested, and usable by whoever is on duty.
  3. Logs of inputs, outputs, retrieved sources and actions, kept long enough to investigate.
  4. Alerts for unusual volume, cost spikes, error rates and repeated complaints.
  5. A one-page playbook: severity levels, who to call, customer message templates, where the logs are.
  6. A tabletop exercise once or twice a year: walk through a scenario ("our assistant quoted the wrong price to 200 customers") and see what breaks.

Fifteen minutes spent rehearsing usually exposes the missing piece, such as nobody knowing how to disable the chatbot.

Frequently Asked Questions

What counts as an AI incident?

Any event where an AI system caused or could cause real harm: wrong information reaching customers at scale, data exposed to the wrong people, unauthorized actions, harmful outputs, or a significant unexplained drop in quality. A single wrong answer caught by your normal review isn't an incident.

Do we need a separate incident plan for AI?

You can extend your existing IT or security incident process rather than writing a new one. Add AI-specific pieces: how to stop each AI system, what to search in AI logs, how to judge wrong-answer scope, and how to handle promises the AI made to customers.

Should we honor a price or promise our AI chatbot made by mistake?

Often it's the right call commercially, and in some jurisdictions you may be held to what your automated systems tell customers. Get advice for significant amounts, but plan for the possibility and design your assistant so it can't make binding offers in the first place.

Who should be responsible for AI incidents in a small company?

Each AI system should have a named owner who leads the response for that system, with the founder or managing director involved in high-severity incidents. The owner doesn't need to be technical, but they need access to whoever can stop and inspect the system.

How do we prevent the same incident happening again?

Find the root cause, not just the trigger, and change something concrete: a permission, a check, a test, an alert or a process. Then add the incident scenario to your test set so future changes are checked against it.

Want AI systems you can stop, inspect and fix quickly?

We design agents and assistants with kill switches, audit logs, alerts and approval steps built in, so an incident is a contained bad afternoon rather than a crisis. That's standard in our agentic AI development work. Talk to us about your AI setup and we'll help you find the gaps before they find you.