An AI incident playbook for a company of 50 to 500 people fits on one page. It names four things that can go wrong, four people, three severity calls, and what happens in the first hour. Anything longer was written for an organization that trains its own models and pays someone whose whole job is security.
The plans that rank for this tell you to roll back a model version, quarantine a poisoned retrieval corpus, and re-run an evaluation suite as a release gate. A company with ChatGPT seats, a Microsoft Copilot rollout, and Claude in two teams can do none of that, and none of it is the exposure. Its incidents are behavioral. A person with a deadline. A number nobody checked. A subscription on a personal card.
The four things that go wrong at this size
Every framework here offers the same six-part taxonomy: model failure, adversarial attack, data poisoning, privacy violation, physical safety, misuse. Good taxonomy, wrong company. A business that buys its AI has four.
Data went somewhere it should not have. A client contract, a patient list, or a pricing model pasted into a tool your agreements do not cover, usually from a personal account under deadline pressure. Verizon’s 2026 Data Breach Investigations Report put regular AI use on corporate devices at 45% of employees, up from 15% a year earlier, and 67% of it ran through non-corporate accounts.
Unverified output left the building. A number, a citation, or a clause reached a client, a regulator, or a board unchecked. The KPMG and University of Melbourne study of more than 48,000 people across 47 countries found 66% rely on AI output without evaluating its accuracy. No alert fires. You learn about it when the client does.
A tool nobody approved is in use. A free tier that solved a real problem, or an AI feature that switched itself on inside software you already pay for. IBM and Ponemon’s 2026 breach report, drawn from 602 breached organizations, found employee shadow AI involved in 43% of them, against 20% a year earlier.
A vendor changed what it does with your data. Retention, training use, subprocessors, or the model behind a product, revised in a release note nobody read. Your controls were written against last quarter’s behavior and quietly stopped describing it.
Add model extraction and rogue agents to that list and you will rehearse the 2% and skip the 98%.
Why a borrowed plan fails here
The right spine exists and is free. NIST finalized SP 800-61 Revision 3 on 3 April 2025, building incident response on the six functions of the Cybersecurity Framework. It is sound, and it assumes a security function is reading it.
Three things break when a COO inherits it as a second job. It assumes you know what you own, and the AI inventory is the harder document here: what a console will and will not show is set out in the options for logging and auditing company AI use. It assumes reproducibility, which sampling systems do not offer. And it assumes containment is technical, when yours is a phone call, a credential rotation, a deletion request, and a call on whether a deliverable gets recalled.
Four roles, named before you need them
Write four names on the page. Not departments, not a committee. The failure mode here is vacancy: operations assumes IT owns it, IT assumes legal does, and the first forty minutes go to discovering nobody was asked.
The receiver. Whoever hears it first, which means everyone. One channel, one name, no form.
The owner. One person who classifies severity and runs the hour, usually the COO, VP Operations, or chief of staff. Technical depth is not the requirement. Standing to tell a partner a document is not going out today is.
The counsel call. An outside lawyer’s name and mobile number on the page, agreed in advance. The notification question is cheap in hour one and expensive in week three.
The fixer. Whoever changes the workflow that produced it. Deliberately not the owner, because containment and prevention compete and prevention loses.
A plan whose owner left in March is not a plan. Review the four names when anyone leaves.
Your only sensor is a person
This is where a playbook for a small company stops resembling anything else, and it is the part the field skips.
You have no monitoring worth the name. With most AI use on personal accounts, the console sees a fraction of it, output quality triggers nothing, and a free-tier subscription appears in no system you query. Detection is one thing: somebody decides to tell you. Otherwise a customer asks a question you cannot answer.
So the reporting route is a control, and your answer to the first report is the most consequential thing you will do all year. Give the first self-reporter a disciplinary conversation and you have taught the next hundred to stay quiet, trading a contained hour for a discovery by a client. Make prompt self-reporting a stated mitigating factor in the policy, and say so in training.
Two thresholds, because without them nobody knows what to report:
- Report a near miss. “I almost pasted the contract and caught myself” names the workflow to fix before it breaks.
- Report uncertainty. “I do not know whether this counts” is a valid report. The owner classifies. Staff do not.
Pushing detection onto a monitoring product instead mostly fails here, for the reasons set out in enforcing an AI policy without surveillance. It costs more, misses the personal laptop, and erodes the trust the reporting route runs on.
Three severity bands instead of a grid
A five-by-five matrix gives twenty-five cells of invented precision. Use three bands, each tied to an action the owner can take alone.
| Band | What it covers | What happens | Clock |
|---|---|---|---|
| Note it | Near miss, non-sensitive content, approved tool, nothing sent | Log one line. Name the workflow to fix. | None |
| Contain and assess | Confidential or client material in an unapproved tool, or unverified output sent internally | Owner runs the first hour. Counsel call if a data category or contract is in play. | Same day |
| Notify | Regulated or personal data exposed, or client material already outside your agreements | Counsel engaged. Contract and statute reviewed. Notification decided on the earliest deadline. | Hours, not days |
If a report will not sort into one of the three, the description is too vague to act on. Ask two more questions rather than adding a band.
The first sixty minutes
Order matters here, and the usual instinct gets it backwards.
- Capture before you delete. Export or screenshot the conversation. Deleting the chat destroys the only record of what was exposed and proves nothing about whether it was logged or read.
- Rotate any credential that appeared. A key, password, or token in a prompt stays live until rotated, and deleting the chat leaves it working.
- Establish scope in writing. What was submitted, which product, which account, whether it keeps history, and whether anything was shared onward. Four sentences from the reporter, not an interview.
- Request vendor deletion and lock the account. A business tier gives you a deletion path and a support contact. A consumer login gives you neither.
- Classify and call. The owner assigns a band. If it is notify, counsel is on the phone inside the hour, since the earliest clock is often a contract term rather than a statute.
- Write six lines and stop. What happened, what data, which tool, what was done, what is open, who owns the fix. Then go work the cause.
The notification question, answered honestly
Most pages on this subject either wave at “consult your counsel” or imply an AI-specific clock is bearing down on you.
The EU AI Act does set a hard clock. Article 73 requires a serious incident to be reported immediately once a causal link is established or reasonably likely, and no later than 15 days after awareness, dropping to two days for a widespread infringement or a serious and irreversible disruption of critical infrastructure, and ten days where a person has died.
That duty lands on providers of high-risk AI systems, as Article 73 read with Article 16 and the Article 3(2) definition makes clear. A US company of 150 people using general-purpose assistants is not one. What does reach it is less exotic.
| What was exposed | What probably binds you | Who starts the clock |
|---|---|---|
| Name plus Social Security number, license number, or account number, and in many states biometric, health, or username-and-password pairs | A state breach-notification statute. All 50 states, DC, Guam, Puerto Rico, and the Virgin Islands have one, with differing definitions | State law, keyed to the data category |
| Protected health information, where you are a covered entity or a business associate | HIPAA breach notification | Federal rule, on discovery |
| Client material under a confidentiality clause, NDA, or master services agreement | The contract, often the shortest deadline on the page | The counterparty’s terms |
| A credential, key, or token | Nothing external. Rotate and move on | You |
The trigger is the data category and the contract, not the involvement of AI. Your NDAs predate anyone’s thinking about chat assistants and still apply, the argument in what your NDAs already say about pasting client material into AI. Read your top five client agreements for notice terms in advance, not at speed.
The review that changes a workflow
The standard response is to add a sentence to the policy and send an email. It feels like action and changes nothing, because the policy was not what failed. The paralegal knew the rule and had a filing deadline.
Three questions, twenty minutes.
- What was the person trying to get done? There was a real goal and a tool that reached it faster. Name both. If the sanctioned path is slower than the unsanctioned one, your policy is a speed bump.
- What judgment did we expect, and had anyone taught it? Recognizing that this document is covered. Noticing that an output asserts a number never present in the source. Knowing that this product sends content to a vendor you have no agreement with. Those are skills rather than rules, which is why a policy nobody was trained on changes nothing, the gap examined in the safe AI use playbook.
- Did detection work, or did we get lucky? If someone chose to tell you, find what made that feel safe and protect it. If a client told you, your reporting route is decorative.
Then make one change: a faster sanctioned path, an admin setting, or ninety minutes of applied practice for the team that owns the workflow. Add the incident to the AI risk register with its detection route filled in, and stop.
A rehearsal a non-security team will run
Tabletop exercises have a bad name here because the versions on offer run half a day and assume a security team. Yours takes 25 minutes a quarter on a meeting that already happens.
Read one scenario aloud. “A client emails to say the deliverable we sent on Tuesday cites a regulation that does not exist.” Or: “An account manager mentions she has been running client calls through a free transcription tool since August.” Then three questions. Who gets the call. What happens in the first hour. Who phones counsel, and do we have the number?
In the applied sessions we run with teams this size, the gap is never the sequence. It is recognition. People know the rule and do not see that it covers the document in front of them, or they assume someone in IT is watching. Talk one scenario through aloud and reports start arriving earlier, which is the only improvement this exercise is for.
What this costs
The document is an internal half-day. Reading your top five client agreements for notice terms runs one to three thousand dollars of counsel time at market rates, the cheapest money here. Business AI accounts, which supply the deletion path step four depends on, run roughly twenty to thirty dollars per seat per month. Applied AI enablement sessions sit in a two to fifteen thousand dollar range. A governance platform is low to mid five figures a year and is the wrong first purchase.
The ordering is the argument. Of the organizations in IBM’s 2026 report that had an AI-related incident, 92% lacked proper AI access controls and 68% lacked AI governance policies. What was missing was accounts, permissions, and a written position, not a dashboard. Monitoring bought before you issue business accounts and teach five data categories instruments a problem you have not addressed, which is the broader case for treating capability rather than license distribution as the work.
Frequently asked questions
Do we need a separate incident response plan just for AI?
No. Add one page to the plan you have: what counts, who owns it, the first-hour sequence, the notification triggers. A standalone plan duplicates escalation paths and goes stale separately.
Someone pasted a client contract into their personal AI account. Is that a breach?
Possibly, and the answer comes from the contract rather than from AI. Check that client’s confidentiality clause for a notice obligation, and whether the content held personal data a state statute covers. Capture the chat before deleting it, then call counsel.
What counts as an AI incident if nothing was hacked?
Most of them. Unverified output sent to a client is an incident with no attacker and no alert, as is an unapproved tool holding your data. Scope the definition to security events and your likeliest exposures never get reported.
Who should own AI incident response when we have no security team?
One named person with authority to stop a deliverable, usually the COO, VP Operations, or chief of staff, plus an outside counsel contact agreed in advance. Not a committee. The common failure is a role never given to a human.
Does the EU AI Act require us to report AI incidents?
For most US mid-market companies, no. The Article 73 duty, with its 15-day default and two-day and ten-day exceptions, applies to providers of high-risk AI systems. Buying licenses for general-purpose assistants does not make you one.
How do we find out about incidents without monitoring everything?
Design the reporting route and protect it. One channel, self-reporting as a stated mitigating factor, near misses welcome, “I am not sure if this counts” accepted. Surveillance tooling costs more and suppresses what you depend on.
Should we delete the chat right away?
Not first. Capture it, rotate any exposed credential, then request deletion through the vendor. Deleting early removes your evidence and leaves a leaked key working.
What should we log before we need it?
Enough to answer three questions: which products and accounts are in use, who has access to each, and what any admin console retains and for how long. Capturing every prompt is unavailable on most consumer tiers anyway.
Key takeaways
- Four classes cover a buyer: data in the wrong place, unverified output sent out, an unapproved tool, and a vendor changing its terms.
- Name four people: receiver, owner, counsel contact, fixer. Vacancy, not incompetence, loses the first hour.
- Your only real sensor is a person choosing to speak up, so your answer to the first report sets the rate for the next hundred.
- Capture before you delete, and rotate credentials first. Deleting the chat destroys the evidence and leaves the key live.
- The notification clock comes from the data category and the client contract, not from AI. Article 73 binds providers of high-risk systems, not mid-market buyers.
- Close reviews by changing a workflow, not a policy clause. Of the organizations with AI incidents in IBM’s 2026 report, 92% lacked access controls, not analytics.
If you want a second opinion on your one page, or on which judgments a ninety-minute applied session should cover, book a free AI-readiness call. We will read your playbook, name the class with no detection route, and say what it would take to close it.
Dirk Jan van Veen, PhD