Contents
Security teams have frameworks for AI. Owners of mid-sized companies usually need something shorter: a list of decisions, in order, that keeps the company's most sensitive data where it belongs.
Why a checklist, and why now
Two numbers from IBM's 2025 breach study explain the urgency. One in five organisations reported a breach involving shadow AI, the tools staff use without approval, and those breaches cost $670,000 more on average where shadow AI was widespread. Of organisations that reported a breach of an AI model or application, 97% lacked proper AI access controls, and 63% of breached organisations had no AI governance policy or were still writing one1.
The controls that were missing are mostly not technical. They are decisions about data, tools and permissions that someone in the company has to make.
The checklist
Data
- List the data that never goes into any AI tool. For most companies: customer personal data, documents under non-disclosure agreements, prices, margins and bids, personnel files and credentials. In Cisco's 2024 study, 48% of respondents said they had entered non-public company information into generative AI tools2.
- List the data that may go into approved tools only. Internal drafts, policies, product information. Tell staff the difference in one page.
Tools and terms
- Approve one business-grade assistant and retire personal accounts for work. Consumer plans of ChatGPT and Claude can train on conversations unless the user opts out34; business plans exclude training by default56.
- Read the vendor's data terms for four points. Training, retention, location and who at the vendor can see the data. Microsoft, for example, covers Copilot under its Data Protection Addendum and acts as data processor7; Azure may keep flagged prompts for human abuse review8.
- Turn on single sign-on. Access then ends when employment ends. SSO is included on Claude Team and ChatGPT Business95.
Permissions
- Clean up sharing before switching on AI in the office suite. Copilot only surfaces data a user has at least view permission for, and Microsoft tells customers to use its permission models to keep the right content with the right people10. Gemini only retrieves content the user can access11. An "everyone" folder becomes an answer for everyone.
- Label the most sensitive documents. Copilot inherits sensitivity labels and honours the usage rights of encrypted content10.
Retention and audit
- Decide how long prompts are kept, and who can review them. Microsoft 365 administrators can set retention policies for Copilot interactions in Purview10; Claude Enterprise has custom retention controls and audit logs9; Amazon Bedrock can be set to zero retention across an organisation12.
- Tell staff what the feedback button does. Anthropic may use thumbs-up and thumbs-down feedback for training after de-linking it from user and customer IDs6.
When AI acts
- Give each AI agent the least access it needs. OWASP's list of AI application risks puts prompt injection first, where instructions hidden in a website or file change the model's behaviour, and recommends privilege control and least-privilege access13. Its second risk is sensitive information disclosure, with strict access controls and restricted data sources among the defences14.
- Write approval rights for AI actions. For sending, paying and changing records, decide whether AI may act always, only after asking, or never, and name who approves. OWASP recommends human approval for high-risk actions13. Some systems build this in: OpenLazarus sets permissions for data, files, messages and spending to Always, Ask or Never, with the company deciding who approves; a teammate can hold a company credit card with its own spending permissions, and every action and approval is logged15.
Review
- Review the list every quarter. Vendors change terms; Anthropic changed its consumer terms in 20254, and Microsoft changed Azure's abuse-monitoring programme in November 20268.
Who owns each check
| Checks | Usually owned by |
|---|---|
| 1–2 Data lists | Chief executive with finance and legal |
| 3–5 Tools and terms | Head of IT or operations, with finance for the contract |
| 6–7 Permissions | Head of IT, with each department head |
| 8–9 Retention and audit | Head of IT, with legal |
| 10–11 AI actions | Chief executive and the head of each department where AI acts |
| 12 Review | One named person, on the leadership calendar |
What done looks like
Every employee knows the one approved tool and the short list of data that never goes into AI. The company's folders are shared on purpose, not by accident. Prompts are kept for a set period and someone can review them. Every AI action that sends, pays or changes something falls under a written rule with a named approver. And a date in the next quarter is booked to check it all again. The main comparison shows which products cover which checks, and the guide to private AI shows what each level costs.
Common questions
What is AI data security?
The set of controls that keep company data safe when AI tools read it, store it or act on it. It covers which data may go into AI, which tools are approved and under what terms, who the AI can see data for, how long prompts are kept, and which AI actions need a person’s approval.
What are the main AI data security risks?
Staff using unapproved tools, data reaching people through AI who could not see it before, vendors keeping or training on prompts, and AI that acts on hidden instructions. OWASP lists prompt injection and sensitive information disclosure as the first two risks for AI applications.
Should a company ban ChatGPT to protect its data?
A ban is hard to enforce when most AI users bring their own tools, 78% in Microsoft’s 2024 survey. Most companies get further by approving one business-grade tool, naming the data that may never go into any AI, and making the approved tool easier to use than a personal account.
Sources
- IBM Report: 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controls. IBM Newsroom, 30 Jul 2025. Accessed 19 Nov 2026.
- More than 1 in 4 organizations banned use of GenAI over privacy and data security risks: new Cisco study. Cisco Newsroom, 25 Jan 2024. Accessed 19 Nov 2026.
- ChatGPT pricing. OpenAI. Accessed 19 Nov 2026.
- Updates to consumer terms and privacy policy. Anthropic, 28 Aug 2025. Accessed 19 Nov 2026.
- ChatGPT Business. OpenAI. Accessed 19 Nov 2026.
- Is my data used for model training?. Anthropic Privacy Center, 18 Aug 2026. Accessed 19 Nov 2026.
- Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat. Microsoft Learn, 29 May 2026. Accessed 19 Nov 2026.
- Data, privacy, and security for Foundry Models sold by Azure. Microsoft Learn, 18 May 2026. Accessed 19 Nov 2026.
- Plans and pricing. Anthropic. Accessed 19 Nov 2026.
- Data, privacy, and security for Microsoft Copilot. Microsoft Learn, 9 Jul 2026. Accessed 19 Nov 2026.
- Generative AI in Google Workspace Privacy Hub. Google. Accessed 19 Nov 2026.
- Data retention in Amazon Bedrock. AWS Documentation. Accessed 19 Nov 2026.
- LLM01:2025 Prompt Injection. OWASP Gen AI Security Project. Accessed 19 Nov 2026.
- LLM02:2025 Sensitive Information Disclosure. OWASP Gen AI Security Project. Accessed 19 Nov 2026.
- OpenLazarus. OpenLazarus. Accessed 19 Nov 2026.
Updated 19 Nov 2026: First published.
