Before you let an AI tool into your business: a 5-minute safety check

AI tools are genuinely useful, and most of them are perfectly safe to use. But “plug it in and hope” is how the occasional horror story happens — a tool given the run of everything, quietly doing something it shouldn’t. None of this is a reason to avoid AI. It’s just the same sensible care you’d take before handing a new supplier or a new hire the keys. Here’s the five-minute version, before you connect anything to your business.

1. What can it actually see?

Before anything else: what are you giving it access to? A tool should see what a sensible colleague in that role would see — and no more. Don’t hand it your whole inbox, drive or customer database when it only needs one folder. Scope it tightly; you can always widen later.

2. Can it act, or only suggest?

There’s a big difference between a tool that drafts a reply and one that sends it. Start every AI on suggest-only — it proposes, a person approves — and keep a human hand on anything that sends, spends or deletes. Widen its rope as you come to trust it, not before. It’s the same line we draw with an agent working inside your Slack or Teams: read and draft first, act later.

3. Where does your data go — and does it train the model?

Your data is going somewhere when you use a cloud tool. The one thing worth checking in the terms: is your data used to train their AI? Many providers offer a “no-training” or zero-retention option — turn it on. And never paste passwords, card numbers or other secrets into a tool; treat it like a chat that could, in theory, be read.

4. Is there a human on the high-stakes moves?

For anything expensive or hard to undo, a person signs off — every time. Money leaving the business, something going out to a customer, a record being deleted. The one that catches people out: never act on a supplier’s “changed bank details” on an AI’s say-so — that’s a classic fraud, and it needs a human phone call, not an automated nod.

5. Does it show its work?

A trustworthy tool lets you see why it did what it did, or where an answer came from — so a person can catch a mistake in seconds rather than trust it blind. If a tool is a total black box on things that matter, be more cautious with it. This is the same reason AI can confidently get things wrong: you want to be able to check.

The honest bottom line

None of this is about fear — it’s five minutes of setup that saves you the “oh no” later. Scope what it sees, let it suggest before it acts, know where your data goes, keep a human on the high-stakes calls, and prefer tools that show their work. Do that, and you get all the upside of AI with none of the daft, avoidable risk. It’s exactly how we run our own operation — the agent does the legwork, and a person holds the keys to anything that matters.