What Business Information Should Never Be Entered Into Public AI Tools?

The categories to keep out of public tools

• Customer information — identities, orders, complaints, account details
• Employee matters — performance, grievances, health, salaries
• Passwords, API keys, recovery codes and security configurations
• Bank details and payment information
• Unannounced products, plans and pricing strategies
• Legal disputes, board papers and privileged correspondence
• Contracts and commercially sensitive supplier terms
• Bids and tenders where confidentiality is a condition of participation
• Health and safeguarding information about any person
• Intellectual property the business has not chosen to publish

Why these categories are different

Three obligations sit behind the list. Confidentiality obligations — many businesses hold information under non-disclosure agreements, client contracts or professional duties, and entering it into an outside service can itself breach those obligations, whatever the service later does with it. Contractual restrictions — supplier and customer contracts frequently limit who information may be shared with, and an AI provider is a third party like any other. Client expectations — a client who discovers their dispute, bid or health matter was pasted into a free chatbot rarely asks whether the training toggle was off.

Anonymisation is harder than deleting a name

Removing a person's name may not make the information anonymous when the surrounding details still identify them. “A senior manager at our Norwich branch who raised a grievance in March” identifies one person to anybody who knows the business. Genuine anonymisation removes every combination of details that could identify someone — role, dates, locations, distinctive events. Where that is not possible, rewrite the scenario in general terms before asking the question.

How to get the work done anyway

1. 1Use placeholders — “[Customer]”, “[Supplier]”, “[£amount]” — and reinsert the real details after the AI has done its work.
2. 2Redact documents before uploading: strip names, addresses, account numbers and metadata, not just the obvious fields.
3. 3Generalise the scenario: ask about “a services business handling a late-payment dispute” rather than pasting the actual letter.
4. 4Use need-to-know discipline: the person prompting should only be handling information they are entitled to see in the first place.
5. 5Move genuinely sensitive work to an approved secure environment — an enterprise deployment with contractual commitments, admin controls and agreed data handling — rather than to a public tool.

Public tools and business tools are not the same

This guide is about public, consumer-grade AI tools. Enterprise services under a proper contract, with agreed retention, access controls and security commitments, can be appropriate homes for some confidential work — that is exactly what the supplier-assessment guide in this hub is for. The mistake is treating a free consumer account as if it carried enterprise protections. Free AI tools are not private business vaults.