What should a manager do when an employee violates the company AI acceptable use policy?

By the Shadow AI Policy team

Contain first, then classify: stop any ongoing data exposure, sort the violation into minor, moderate or serious based on what data left the building and whether the employee knew the rule, and match the response to that tier — a coaching conversation, a documented written warning, or a formal investigation that can end in termination.

This guide gives you a tiered response framework you can apply the same day, the documentation HR and legal will ask for, the signals that mean IT or Security should run forensics, and the point at which an AI mistake turns into a data breach with legal notification clocks. It also covers the distinction that decides most cases: an honest mistake versus deliberate circumvention.

1. What is the first thing a manager should do after spotting an AI policy violation?

Stop the exposure before you deal with the person. If an employee is still pasting client data into a personal chatbot account, or an unapproved AI notetaker is still joining meetings, the priority is to end that today: ask the employee to stop, have IT revoke the tool's access or block it, and preserve what already happened rather than deleting it.

Don't ask the employee to "clean it up" by deleting chat history or uploaded files. Deletion destroys the evidence IT needs to work out what was shared, and it can look like a cover-up later. Ask them to leave everything as it is and to write down, while it's fresh, which tool they used, what they put into it, and roughly when.

Then hold off on any disciplinary decision until you know three facts:

• What data went in — public information, internal documents, or personal, financial, health or client-confidential data.
• Where it went — an approved enterprise tool, a personal consumer account, or an unknown or free tool you haven't vetted.
• What the employee knew — whether they'd been trained on the policy, signed it, or been warned before.

Those three answers decide the tier. If you don't yet have a written policy to measure the conduct against, fix that before you discipline anyone; our AI acceptable use policy template guide walks through what it needs to say.

2. How do you decide whether a violation is minor, moderate or serious?

Score it on data sensitivity and intent, then take the higher of the two. A well-meaning employee who pasted a patient record into a free chatbot is still a serious data event, even though the intent was innocent. A deliberate workaround involving only public data is still a conduct problem, even though no data was harmed.

Tier	Typical example	Response	Who gets involved
Minor	First-time use of an unapproved tool with public or non-sensitive internal content; forgot to disclose AI-assisted drafting	Coaching conversation, point to the approved alternative, short written note to file	Manager; HR informed
Moderate	Confidential internal data (pricing, roadmap, unreleased code) in a personal AI account; a repeat of a minor violation after coaching	Documented written warning, retraining, tool access reviewed	Manager and HR; IT to assess the exposure
Serious	Personal, health, financial or client data in an unvetted tool; deliberately bypassing a block; hiding AI use after being told to stop	Formal investigation; possible suspension of access; outcomes up to termination	HR, legal counsel, IT/Security, and the privacy lead

Write your tier definitions into the policy itself, before you need them. Making up the scale during the first incident invites the argument that you made it up to fit the person. If your current policy only says violations "may result in disciplinary action, up to and including termination," that's a start, but it gives managers nothing to calibrate against.

Repeat violations move up a tier. A second minor violation after a documented conversation is a moderate one, because the employee can no longer say they didn't know.

3. How do you tell a good-faith mistake from deliberate circumvention?

Look at what the employee did to get around the rule, not at what they say they meant. Good-faith mistakes usually happen in the open: someone used a popular tool like ChatGPT on a personal account because nobody had told them about the approved version, and they'll tell you exactly what they did when asked.

Deliberate circumvention leaves different fingerprints:

• Using a personal device or personal network after the tool was blocked on the work one.
• Renaming, splitting or stripping identifiers from files specifically to get them past a filter.
• Continuing to use a tool after a direct instruction to stop.
• Giving an account of what was shared that IT's logs later contradict.

The distinction matters because it changes the response, not just the tier. A good-faith mistake is mostly a training and tooling failure: the fix is a conversation, a better approved tool, and a clearer policy. Deliberate circumvention is a trust and conduct issue, and it can justify serious discipline even when the data involved turns out to be low-risk.

Treat honest disclosure as a mitigating factor, every time. If people learn that admitting a mistake gets them punished as hard as hiding one, the next leak won't be reported — it'll be discovered months later by someone else.

Most shadow AI use comes from people trying to get their work done faster, not from bad actors. Our explainer on what shadow AI is and why it happens is a useful read for managers who are about to have this conversation for the first time.

4. What should HR and legal have documented?

Document enough that someone who wasn't there could reach the same decision from the file. That protects the company if the employee disputes the outcome, and it protects the employee from a decision that looks arbitrary.

For every violation, including minor ones, record:

• The date you learned of it, how you learned of it, and who reported it.
• The tool, the account type (personal or company), and what data was entered, described by category rather than by copying the sensitive data into the HR file.
• The policy clause that was breached, and evidence the employee had received it, such as a signed acknowledgment or a training completion record.
• The employee's own account, in their words, and the date you took it.
• The tier you assigned and the reason, plus the action taken and any follow-up date.

Keep the HR record factual. Write "pasted the Q3 customer renewal list into a personal account on an unapproved chatbot," not "was careless with client data." Opinions in the file are what opposing counsel quotes back to you.

For serious cases, bring legal in before the investigation interview, not after. Counsel may want the fact-finding to run under their direction, and the first interview is the hardest part to redo. If the employee is in a union or works under a contract with a defined disciplinary procedure, follow that procedure exactly; a fair outcome reached the wrong way can still be overturned.

5. When should IT or Security run forensics?

Bring in IT or Security whenever you can't answer, with confidence, what data left and where it went. That covers every serious violation and most moderate ones.

The specific triggers to escalate are:

• Personal, health, financial or client-confidential data may have been entered into any tool.
• The tool isn't on the approved list, so nobody has checked whether it keeps, trains on, or shares what it receives.
• The tool was connected to company systems — a browser extension, an AI notetaker with calendar access, or a plugin linked to email or file storage.
• The employee's account of what happened is vague, or doesn't match what others say.
• There's any sign of deliberate circumvention, such as personal devices or bypassed blocks.

Security's job here is narrow: establish the scope of the exposure, revoke tokens and integrations, request deletion from the vendor where its terms allow, and preserve logs. Managers shouldn't pull logs or browse an employee's account themselves, because an informal search can compromise the evidence and create privacy problems of its own.

The forensic findings feed straight into the next question: whether this is now a reportable breach.

6. When does an AI policy violation trigger breach notification obligations?

It can trigger them when personal data was disclosed to an unauthorized third party, and an unvetted AI tool can be exactly that third party. Whether you actually have to notify depends on what data it was, where the affected people live, which laws apply to your business, and what the vendor did with the data. That determination belongs to legal counsel and your privacy lead, not the line manager.

The clocks are short, which is why escalation can't wait for the disciplinary process to finish:

• GDPR: if you're subject to it, a personal data breach must be reported to the supervisory authority within 72 hours of becoming aware of it, unless it's unlikely to result in a risk to people's rights and freedoms.
• HIPAA: if a covered entity's staff put protected health information into a tool with no business associate agreement, that disclosure is presumed to be a breach unless a documented risk assessment shows a low probability that the information was compromised. Individual notices are due without unreasonable delay and no later than 60 days after discovery. Our comparison of HIPAA status across AI tools shows which vendors offer a BAA.
• US state breach laws: every state has one, and their definitions of personal information and their deadlines differ. Counsel will check which ones apply based on where the affected people live.
• Client contracts: many service agreements require you to tell the client about any unauthorized disclosure of their data, often on a tighter timeline than the law sets.

Your job as a manager is to start the clock correctly: report up the same day you learn of a possible exposure, and record when you found out. "We didn't know it was a breach" stops being a defense once a manager knew and sat on it. If you'd rather have the escalation path, tier definitions and acknowledgment forms written for you, you can generate a tailored policy kit and adapt it.

About Shadow AI Policy: We build AI acceptable use policy tools for HR and operations teams at 50–500 person companies. We publish guides on shadow AI, acceptable use policies, and AI governance, updated as regulations and AI tools change.