How should a small business respond to an AI data leak?

What should I do in the first hour after an AI data leak?

Treat the paste as a security incident, not a disciplinary meeting. Preserve the chat details, identify the tool and account, rotate any exposed secrets, and stop further sharing. Then classify what left the business. A quick, calm report gives you more options than a hidden mistake discovered weeks later.

I would put one person in charge and ask the employee to write down the service, model, account type, time, exact files or text shared, and whether a public link or shared project was created. Save the relevant screen or export if the product allows it. Do not delete the conversation before recording those facts. If a password, API key, token, connection string, or private key was included, revoke and replace it immediately. Check access logs for the period before revocation.

That sequence follows the spirit of NIST SP 800-61 Rev. 3, published in April 2025, which helps organizations prepare for, detect, respond to, and recover from incidents within the Cybersecurity Framework 2.0. The NIST publication is here: https://csrc.nist.gov/pubs/sp/800/61/r3/final

Which data needs the fastest response?

Secrets come first. Treat an API key, database password, OAuth secret, session cookie, or private key as compromised even if the employee used a paid AI plan. Rotate it, invalidate active sessions where possible, update the application, and inspect logs. Deleting a chat is cleanup. It is not credential revocation.

Next comes personal data and health or payment information. Then review client material covered by an NDA or MSA. Internal drafts and generic code still deserve cleanup, but they usually call for coaching and a written record rather than an emergency key rotation. This four-level order is my operating checklist, not a legal classification. Your lawyer or privacy lead decides the legal result.

For code, scan the source repository as well as the AI account. GitHub Secret Scanning, Gitleaks, and TruffleHog are practical tools for looking for exposed secrets. If the same secret was copied from a .env file into a prompt, assume it exists elsewhere too.

How do I find out what the AI provider kept?

Start with the exact product and account tier. Consumer and business accounts can have different training, retention, deletion, and administrator controls, so the logo alone tells you very little. Record the provider policy version you checked, ask the provider's security or privacy team for relevant account records, and request deletion or privacy assistance when personal data is involved. Provider policies change, so verify the current terms during the incident.

OpenAI says regular and archived ChatGPT chats stay in the account until deleted. Deleted chats are scheduled for permanent deletion within 30 days, subject to legal or security exceptions. Temporary chats may be kept for up to 30 days. Files in Library, Projects, and custom GPTs can be stored separately, so deleting the chat may not delete the file. Source: https://help.openai.com/en/articles/8983778-chat-and-file-retention-in-chatgpt

Commercial controls still need checking. OpenAI's enterprise privacy page, updated January 8, 2026, says business data is not used for model training by default and that enterprise customers can control retention. Anthropic's commercial terms say it may not train models on Customer Content. Microsoft says its enterprise data protection terms do not use prompts, responses, or Microsoft Graph data to train foundation models. These are provider commitments, not a reason to paste everything without classification. Sources: https://openai.com/enterprise-privacy/ | https://www.anthropic.com/legal/commercial-terms | https://learn.microsoft.com/en-us/copilot/microsoft-365/enterprise-data-protection

When do I need legal or client notification?

Do not make the notification decision from a chat deletion screen. Ask counsel or your privacy lead to assess the data, the recipients, the account terms, and the people or clients affected. If personal data is involved, start that review the same day and document the decision even when the conclusion is that notification is unnecessary.

For GDPR-covered processing, Article 33 sets a 72-hour supervisory-authority deadline when the breach is likely to pose a risk to people's rights and freedoms. Article 34 covers communication to individuals where the risk is high. Article 33(5) requires documentation of the facts, effects, and remedial action. The European Data Protection Board publishes Guidelines 9/2022 and a summary here: https://www.edpb.europa.eu/our-work-tools/our-documents/guidelines/guidelines-92022-personal-data-breach-notification-under_en. The regulation text is available from EUR-Lex: https://eur-lex.europa.eu/eli/reg/2016/679/oj

Check client contracts as well. An MSA may require notice in 24 or 48 hours, even when a privacy statute does not. That deadline comes from your contract, not a generic AI rule, so read the actual agreement. Keep a short incident record: discovery time, tool and account, data categories, number of records if known, containment actions, provider contact, and the reason for the notification decision.

How should I fix the process after the incident?

Give employees a safe way to report accidental pastes and replace vague warnings with a one-page traffic-light policy. Green covers public information and generic drafting. Yellow covers internal material only in a company-managed workspace. Red covers passwords, API keys, unredacted personal data, health records, payment data, and client-confidential documents unless a specifically approved workflow permits them.

Run a 15-minute exercise with the person involved. Show how to replace names with [CLIENT_A], remove email addresses and account IDs, and replace secrets with YOUR_API_KEY_HERE before asking for help with code. Then provision an approved business workspace with SSO where it fits your budget. Employees will use AI. The useful question is whether the company knows where and under which terms.

I would review the incident after 30 days, not to punish the employee but to check whether the control changed anything: was the tool blocked or approved, did the reporting path work, and can the next person find the policy in under a minute? If the answer is no, the policy is decoration.

Add the policy to the employee handbook and onboarding checklist. Managers should know where to send a report, and the person who owns the approved-tools list should review it at a fixed interval. That is enough structure for a small company to start.

What should I say to the employee?

Say: "Tell me exactly what you pasted, where, and when. We will contain it first, then decide what must be reported." That sentence matters. People hide mistakes when the first response sounds like a hunt for someone to blame.

A small company does not need a 40-page AI governance manual to handle this well. It needs a named responder, a short incident form, a current list of approved tools, and the nerve to rotate a credential before arguing about whose fault the paste was. If you need help turning that into a working workflow, contact DeuceWorks.

Sources and further reading: NIST SP 800-61 Rev. 3; EDPB Guidelines 9/2022; OpenAI Enterprise Privacy; Anthropic Commercial Terms; Microsoft Enterprise Data Protection.