What Should You Avoid Putting into AI Tools? A Small Business Guide

Do not paste passwords, confidential business material or identifiable customer information into an AI tool simply because it can process them. First establish whether the tool, account settings and contractual arrangements are suitable for that use. When you are unsure, use fictional or properly anonymised material instead.

This guide offers a practical starting policy for a small business. It is not a substitute for an assessment of a particular service or advice on a specific data-protection issue.

Think about the information, not just the prompt

A short question can reveal sensitive details. “Rewrite this message about ending our contract with our biggest customer” may disclose a commercial plan even if the attachment contains no passwords.

Files can contain information you did not intend to send, including comments, hidden sheets, tracked changes and identifying details. Connected tools may also expose material beyond the document you are currently looking at. Check the actual scope before granting access.

Keep secrets and access credentials out

Do not put passwords, recovery codes, private keys or access tokens into ordinary prompts. A tool does not need a real credential to explain a configuration example. Replace it with an obvious placeholder and keep the actual value in the appropriate secure system.

If a credential is accidentally shared, follow your incident process and have it revoked or changed where necessary. Deleting a conversation alone is not a reliable way to make an exposed credential safe.

Be cautious with personal and confidential information

Customer records, staff matters, medical details, payroll, contracts and unpublished financial information require care. Ask whether the task genuinely needs those details and whether the proposed processing is appropriate.

The ICO’s AI guidance explains how data-protection principles apply to AI systems. Using a new tool does not remove your responsibilities for personal information.

A hypothetical business owner wanting help with a complaint could describe the general situation with invented details, rather than upload the customer’s full email history. The draft can then be adapted and checked privately by the person handling the complaint.

Do not assume a paid account settles the question

Different services, plans and settings can handle data differently. Check the current documentation for the account you actually use. Questions include:

• Is submitted material used to improve models, and what controls apply?
• How long are prompts, files and outputs retained?
• Who can access them, including administrators or service providers?
• What contractual and data-processing terms apply?
• What happens when a file or conversation is deleted?
• What can connected apps read or change?

“Not used for training” is one relevant condition, not a complete privacy or security assessment. A business still needs suitable access control, retention and authorised use.

Use safer substitutes where possible

Ask whether the task can be completed with a blank template, a short excerpt or synthetic data. For a spreadsheet question, invented rows often show the calculation problem perfectly well. For copy editing, remove customer names, account numbers and unnecessary background.

Removing obvious names is not always enough. A distinctive job title, location and event can still identify someone. If you cannot confidently anonymise the material, use a different workflow or get appropriate help.

Check output before it reaches anyone else

The NCSC advises care with sensitive information in public language-model services. Its guidance also highlights inaccurate outputs. A tool can introduce invented details or reveal information from the context it was given.

Review the final result for confidential content as well as factual accuracy. Do not automatically send AI-generated replies to customers or use them to make consequential employment, legal or financial decisions without a suitable process and accountable human review.

Create a short team rule

A useful starting policy can fit on one page. Name the approved tools and accounts, identify permitted tasks, list information that must not be entered and explain who answers uncertain cases. Include what to do after a mistake.

For example: “Use the approved account for drafting from public or fictional material. Do not upload customer files, staff records or credentials. Check every output before use. Ask the designated owner before connecting another service.” Adapt that wording to your actual assessment rather than treating it as a complete legal policy.

If something is shared by mistake

Stop further sharing, record what was involved and tell the person responsible for security or data protection. Review the provider’s available controls and assess the consequences promptly. Any reporting duties depend on the facts; do not assume either that every mistake is reportable or that deleting the prompt resolves it.

Begin with the low-risk examples in our practical AI guide. Expand only when you understand the information, permissions and review process involved.