Small businesses are being sold AI tools constantly, and most of the marketing skips the part where you hand over your data. Here is what to ask before you do.
The questions
Where does the data go? Not "the cloud." Which company, and what do they do with it? If a tool is a thin layer over someone else's model, your information is going to at least two organisations.
Is it used for training? Some providers train on customer data by default, some never do, and some let you opt out. This is the single most consequential setting and it is frequently buried.
How long is it kept? Retention periods vary from days to indefinitely.
Who at the vendor can see it? Support staff, engineers, nobody?
Can you get it back and can you delete it? Both matter, and "delete" should mean deleted, not hidden from your view.
What happens if they are acquired or shut down? Small AI companies come and go. Your data goes with them.
If a vendor cannot answer these clearly and in writing, that is your answer.
What not to put in
Regardless of vendor, some things should not go into a general-purpose AI tool.
- Payment card details, bank information, government identifiers.
- Anything covered by a regulation you are subject to — health information in particular has specific rules.
- Employee personal records.
- Anything covered by an NDA with a customer.
- Credentials and API keys. People paste these into chat tools constantly.
That is not paranoia. It is the same discipline you would apply to emailing something to a stranger.
The customer-data question
The most common real scenario is wanting an AI tool to work with your customer list, your job history, your inbox.
That is legitimate and often genuinely valuable. Three things make it safe enough:
Use a business tier, not a consumer one. Business and enterprise plans generally carry different data commitments, including no training on your data. The free consumer version usually does not.
Send the minimum. If the task needs job descriptions, do not send names and phone numbers with them. Most useful AI work does not need identifying information at all.
Know what you promised. Your own privacy policy told customers what you would do with their data. Feeding it to a third party may or may not be consistent with that. Worth reading what you wrote.
The realistic risk
Being honest about proportion: the likely harm for most small businesses is not a dramatic breach. It is quieter.
Your data sitting on a platform you have no agreement with. A vendor changing terms. A tool being discontinued and taking your working history with it. Information about a customer ending up somewhere you cannot account for if they ask.
Those are manageable risks and they are managed by asking the questions above once, at the start, rather than never.
Running it on your own terms
There is a version of this where the data stays much closer to home — an assistant wired into your own documents and systems, with a vendor agreement that says the data is not trained on and is deleted on request.
That is more work to set up and it is the right shape for anything involving customer records. For a lot of small businesses it is also cheaper than the per-seat subscriptions they were considering.
Write down what you decided
One page. Which tools are approved, what may and may not go into them, and who to ask about anything not covered.
Your team is already using AI tools whether you have a policy or not. A short, sensible policy that people will actually follow beats a ban that gets quietly ignored.
The practical summary
Use business-tier tools with real data commitments. Send the minimum. Keep regulated and credential data out entirely. Read what your own privacy policy promised. Write down the rules.
None of that stops you getting value from AI. It stops the version where you find out later that value came with a cost you did not price.