AI and your data: the privacy questions every SMB should answer before rolling out a tool

The fastest way for an SMB to get into trouble with AI isn't a hallucinated answer or a clumsy automation - it's quietly handing customer data to a tool nobody vetted. Most data incidents we see in small and mid-sized businesses don't involve hackers. They involve a well-meaning employee pasting a client list, a contract, or a spreadsheet of personal details into a free chatbot to 'just summarise this'. The data leaves the building, and there's no easy way to get it back.

Why this matters more in 2026

Two things have changed. First, the volume of AI use inside the average SMB has gone up sharply, which means the surface area for accidental disclosure is larger. Second, regulators in the UK and EU have started taking a closer look at how businesses handle personal data inside AI workflows. The fines are still rare, but the reputational cost of a public mistake is real - and the cost of explaining one to a key customer is even higher.

The eight questions to ask of every tool

You don't need a fifty-page assessment. You do need to answer eight questions before any AI tool gets used with anything sensitive. Write the answers down. Keep them in a simple register. Review the register quarterly.

• Where is our data processed? UK, EU, US, or somewhere else?
• Is our data used to train the vendor's models? If yes, can we turn that off?
• How long is our data retained, and can we delete it on request?
• Who at the vendor can access our data, and under what conditions?
• Does the vendor have a data processing agreement we can sign?
• What happens to our data if we cancel the subscription?
• Has the vendor had a security incident in the last twelve months, and how did they handle it?
• What does the vendor do if law enforcement asks for our data?

If the vendor can't answer those questions inside a single phone call or short email exchange, that itself is the answer. A reputable AI tool in 2026 will have these answers ready, often on a public trust page. A vendor that gets defensive when asked is telling you something important.

The free-account problem

The single biggest privacy risk in most SMBs is not the tools the business has bought. It's the personal free accounts employees are using on the side. Free accounts almost always have weaker privacy terms, often train models on whatever you paste in, and live entirely outside your visibility. A workspace plan with proper admin controls is usually only ten or twenty pounds a month per user more, and it solves most of the problem at a stroke. Pay for the seats, set the policy, and remove the temptation.

What a sensible policy actually says

A workable AI data policy at SMB scale is rarely more than a page. It needs to cover, in plain English, what categories of information must never be pasted into a public AI tool, which approved tools the team can use for sensitive work, what the rule is for client data specifically, and what to do if a mistake happens. The tone should be practical, not threatening. People who feel scared by the policy will hide their AI use; people who feel guided will check before they paste.

Talking to clients and customers

If you're using AI in a way that touches a client's data - drafting a response, summarising a call, generating a report - it's worth being upfront. Most clients, in our experience, are perfectly comfortable with AI being used to help draft something a human reviews. Almost no clients are comfortable with their data being processed by a tool they've never heard of, in a country they didn't expect, with terms nobody read. The difference is disclosure, not the technology itself.

Vendor due diligence: a quick scoring approach

When evaluating any new AI tool, score it across three quick dimensions: data handling (how clearly the vendor explains what happens to your information), contractual coverage (whether they offer the agreements you actually need), and incident track record (whether they've handled past issues like adults). A simple traffic-light against each is enough at SMB scale. Anything red on any of the three is a 'no' until the vendor closes the gap. Anything amber needs a clear plan to get to green within ninety days.

Practical safeguards that don't slow the team down

Three small habits remove most of the risk without making AI use painful. One: default everyone to paid workspace accounts on the approved tools, never personal free accounts. Two: turn off model training in the admin settings of any tool that allows it. Three: agree a short list of categories that always need a human-only workflow - usually personal data, financial information, and anything covered by a specific NDA. With those three habits in place, the day-to-day risk drops sharply and the team can move fast.

What to do if something goes wrong

Mistakes happen. The difference between a small problem and a big one is almost always how quickly the business responds. Have a single named person to whom incidents get reported, a rough script for the first hour (what was disclosed, to which tool, by whom, when), and a habit of telling affected clients before they hear about it some other way. Most incidents at SMB scale are recoverable if handled openly and quickly. Almost none are recoverable if handled defensively.

The bottom line

Privacy doesn't have to be the brake on your AI programme. Done well, it's the thing that lets you go faster - because the team isn't second-guessing every paste, the leadership isn't lying awake about a leak, and the next big customer doesn't pull their RFP because your data answers were vague. A short register, a one-page policy, paid workspace accounts, and eight standard questions for every vendor. That's most of what an SMB needs to do AI safely in 2026.