Is our data safe with AI? The questions that actually settle it

Not a reassurance piece. The five questions to put to any AI supplier in writing, what good answers look like, and the red flags that end the meeting.

"Is our data safe?" is the right question asked too vaguely. Vague questions get reassuring answers. Here is how to ask it in a form that produces commitments instead.

First, understand where the data actually goes

When an AI agent reads one of your documents, the text is sent to an AI provider's service, processed, and the answer is returned. Two things about that matter. The document itself stays in your systems — the agent reads from where it already lives. And with the business-grade services (as opposed to free consumer chatbots), the providers commit contractually that your content is not used to train their models. That commitment is the line between a professional deployment and pasting client records into a free website, which is what some of your staff may be quietly doing today — often the biggest genuine risk in the building.

The five questions, in writing

One: is our data used to train any AI model? The only acceptable answer is an unqualified no, backed by the provider's own terms, which your supplier should hand you rather than paraphrase.

Two: where is the data processed, and can that be constrained? If your industry or your clients require data to stay in a particular jurisdiction, that must be a configuration, not a promise. If it truly cannot leave your environment, self-hosted models exist — they cost more and perform a little worse, and a good supplier will say both parts of that sentence.

Three: what exactly can the agent reach? An agent should have its own login with the minimum access the job needs — the same discipline you'd apply to a new junior employee. "It uses the office admin account" ends the meeting.

Four: show me everything it did last Tuesday. The answer should be a log, produced in minutes: every document read, every action taken, every decision and what it was based on. If that record doesn't exist, you cannot audit the system, and neither can anyone else.

Five: what happens when it's wrong? You are listening for three specific things: actions that matter wait for a human; everything is reversible; and mistakes are surfaced rather than buried. A supplier who claims it won't make mistakes has just told you they haven't run one in production.

Red flags that don't need a second opinion

Reluctance to put any of the above in writing. A demo that needs your live data on day one. Pricing that requires your data to be "shared for improvement". And any answer to the training question that contains the word "anonymised" doing suspiciously heavy lifting.

The honest summary

Properly deployed, an agent is more auditable than the humans doing the same work today — every action logged, every decision reconstructable, nothing taken home on a laptop. The risk isn't the technology; it's deployments done casually. Which is, conveniently, exactly what the five questions detect.