How to Use AI on Confidential Documents as an Executive

What counts as confidential material?

"Confidential" is too broad a word to act on. I sort what an executive handles into four classes, because each class gets a different rule.

If your organization already has a data classification policy, use its words and treat the table as a starting point.

Does a business account change what happens to my data?

Often yes, and the vendors say so in writing. The table below summarizes what each vendor's own published page said when I checked on October 1, 2026. These terms change often, so check the current page before you rely on any row.

Two reading rules. First, "not used for training" is a narrower promise than "not stored, logged or visible to administrators." Microsoft's page says interactions are stored and administrators can search and retain them. Second, a paid personal plan is still a personal plan. In the OpenAI and Anthropic pages above, Plus, Pro and Max sit on the personal side.

What does a business plan still not protect?

• Storage and access. Your content sits with the vendor, and your administrators can usually see, export and delete it.
• Feedback buttons. Anthropic's page says that explicit feedback, such as a thumbs up or down, can let it use that chat for training, even on commercial products (privacy center).
• Connected apps and agents. Each carries its own terms. Microsoft tells administrators to check an agent's privacy statement before enabling it.
• Privilege. In United States v. Heppner (S.D.N.Y., February 17, 2026), a federal judge held that a defendant's conversations with a public AI platform were not protected by attorney-client privilege or work product, partly because the platform's privacy policy allowed data collection and disclosure to third parties (Harvard Law Review). It is one trial-court ruling, and the commentary calls it too categorical. I am not a lawyer. If privilege matters to you, ask counsel before legal material goes near an AI tool.

What should I never paste into an AI tool?

• Passwords, API keys, recovery codes and anything else that works as a credential.
• Full personal records of customers or employees, such as government ID numbers, health information or account numbers.
• Privileged communications with your lawyer, unless counsel has approved the tool.
• Material under an NDA that bars sharing with third-party processors.
• Material nonpublic information about a public company, and anything a regulator requires you to handle in a set way.

If you are unsure, the answer is no until the person who owns that risk says yes.

How do I redact before I paste?

1. Paste the passage you need, not the whole file.
2. Replace names with roles: "Counterparty A," "the CFO," "our lender."
3. Replace exact figures with ranges or placeholders when the numbers are not what you are asking about.
4. Strip metadata, comments and tracked changes. Hidden text lives there.
5. Read the final prompt once. If you would not email it to an outside adviser, do not paste it.
6. Keep the key (who "Counterparty A" is) in a local note and put the facts back yourself in the output.

Redaction is imperfect. Context can still identify a deal or a person, so it reduces risk and does not remove it.

What goes in a one-page personal AI policy?

Write it once, date it and put your name on it. This is a template to adapt. It is not legal advice, and your organization's own policy outranks it.

What stays risky even with good habits?

• Terms move. Anthropic's 2025 consumer update is an example of a vendor changing training and retention for personal plans. Recheck on a schedule.
• People. Staff using personal accounts for work is the most common gap, and no vendor setting closes it. A short policy and a sanctioned tool do.
• Documents can carry instructions. A file or email you ask an assistant to read can contain text aimed at the assistant. See the approval gate guide.
• Copies. Chat history, exports and screenshots are copies of your material. Treat them with the same class as the original.

For regulated material, such as health, financial, legal or government data, I am not the right person to ask. Bring your counsel or compliance officer the vendor's current terms and the specific use. In coaching I help executives set up the tools and the habits for their own work, which is policy design and tool setup, not legal advice. You can read how that works on the executive AI coaching page, and the wider governance picture in AI governance for non-technical CEOs. A company-wide version of this policy is governance work of the kind described on the fractional CAIO page.