Five questions to ask before an AI tool touches your business data

Most businesses adopting an AI tool aren't evaluating a foundation model — they're evaluating a SaaS product with an AI feature bolted on, and connecting it to data that matters: member records, payment history, client files, sometimes board-election details. The security questions that protect you are less about the AI itself and more about ordinary data-handling discipline that's easy to skip when the demo looks impressive. All five of these fit in a single email to the vendor. The answers — and how quickly they arrive — tell you most of what you need to know.

One: where does our data actually go? Ask plainly whether your inputs — documents, customer records, form submissions — are used to train the vendor's models, and get the answer in writing, not implied on a friendly sales call. Plenty of tools default to yes unless you opt out in a settings page nobody reads. "In writing" matters because the person reassuring you today may not work there when it matters.

Two: who inside our own organisation can see what? If a tool is used by several staff members, check whether access is scoped per person or whether everyone effectively shares one login with full visibility into everything — including data they don't need for their role. Most real-world leaks aren't hackers; they're over-broad access meeting an ordinary mistake.

Three: does it actually need the access it's requesting? A booking assistant needs your calendar. It does not need your full customer database, your email archive, and admin rights to your website — even if the integration flow cheerfully requests all three by default. Grant the minimum, see if the tool still works (it usually does), and expand only when something genuinely breaks.

Four: what happens if the vendor is acquired or shuts down? Small AI vendors churn faster than established SaaS. Know whether you can export your data in a usable format, and whether a shutdown notice gives you thirty days or three. A tool you can't leave isn't a tool; it's a dependency with a logo.

Five: who is accountable when it's wrong? AI features fail politely — a confident summary with a wrong number, an auto-reply that promises something you don't offer. Ask the vendor how errors are surfaced and corrected, and decide internally who reviews the output before it reaches a customer or a board. "The AI said so" has never once survived contact with an unhappy client.

For organisations handling anything genuinely sensitive — board elections, creditor arrangements, financial records — the standard is stricter and simpler: nothing connects to production data until a named person has read the vendor's data-handling terms, not just the pricing page. It's the slower first step, and it's the one that prevents the expensive kind of mistake. If you'd like a second pair of eyes on a tool you're evaluating, that's a conversation we have with clients regularly — usually before the contract, which is the only time it's cheap.