Is Your Team Using AI Safely? Six Questions Every Small Business Should Answer

Short answer: if your team uses AI at work, you don't need a 30-page policy to start. You need clear answers to six questions: who owns AI here, which accounts people use, what information goes in, what the AI can touch, who checks the output, and what happens when it goes wrong. Most small and mid-size businesses can answer all six in a week.

"We're too small for AI governance"

That's the first thing I hear. And I get it. Governance sounds like a big-company word, with committees and binders.

Here's the thing. You're not too small to have a client's CV sitting in someone's personal AI account. In one of my training sessions, a business owner pasted candidate CVs into an AI tool to tidy up the wording. Halfway through the session he realised he'd shared names, home addresses and phone numbers. He wasn't careless on purpose. Nobody had ever told him where the line was.

That's what AI governance means for a business your size. Not paperwork. Just knowing where the lines are, and making sure your team knows too.

Why this matters more this month

October 2026 gave us a few clear signals:

• Banks are drawing lines. RBC, one of Canada's largest banks, blocked third-party personal AI agents from logging into its online banking. Their reason: as agents get more independent, "the stakes go up when something goes wrong."
• Regulators say it's still on you. The UK's data regulator, the ICO, put it plainly: "the fact AI agents act with autonomy is not an excuse for poor compliance."
• Even the builders get surprised. OpenAI reported that one of its own internal coding agents switched off security alerts to humans, then reported that no messages were sent.
• Big firms aren't learning fast. In an IBM survey of 3,221 executives at firms with 1,000+ staff, 80% said an AI incident reached their board or leadership in the past year. Only 40% of those changed their governance afterwards.

If large companies with whole risk teams are getting caught out, a 60-person firm with no one in charge of AI is not safer. It's just less likely to notice.

The six questions

1. Who owns AI here?

In another session, someone couldn't connect her calendar to an AI assistant. She wasn't the admin. Nobody in the room knew who was.

That's the most common gap I see. AI arrives through ten different doors, and nobody is holding the keys. Ownership doesn't need a new job title. It needs one named person who decides which tools are allowed, who approves new connections, and who people go to when they're unsure.

This week: write down one name. Tell the team.

2. Which AI accounts are people actually using?

Ask your team one question: which AI account are you logged into right now? You'll usually get a mix. Company accounts, personal free accounts, personal paid accounts, and AI built into other software.

That matters because company plans and personal plans handle your data differently. Personal accounts often have a setting along the lines of "help improve the model" switched on by default. Check the provider's own settings page for the current rules. Don't rely on what someone remembers.

This week: list the tools in use and which account type each person uses.

3. What goes into the AI?

Your team doesn't need a long policy to start. They need three colours.

• Green: public or anonymised information. Go ahead.
• Amber: client documents, financials, business plans. Check the contract and the tool's terms first.
• Red: sensitive personal information, health data, children's data. Stop and ask.

One more thing. Your client contract beats your AI vendor's settings. A business plan that doesn't train on your data does not override a client contract that bans AI, or says their data can't leave the country. I've met business owners whose clients will fire suppliers for using AI on their work. Check your contracts before you check your settings.

This week: put the three colours on one page and share it.

4. What can the AI touch?

Reading an inbox is one thing. Sending, deleting, booking or paying is another.

When I connect AI to an inbox, the first thing I do is switch off its ability to send or delete. The instructions I give it have hard rules: never send, book, buy, pay or delete. Never invent anything. Ask when unsure.

In one session, an AI tool couldn't finish a bookkeeping task, so it suggested connecting the company bank account directly. The person said no. She was right. Tools will often ask for more access than they need. Decide in advance what is never allowed, before someone is halfway through a task and tempted to click yes.

The people who build these tools are going the same way. Anthropic switched off live internet access for all its internal testing after models submitted real web forms during tests.

This week: for every connected tool, write down whether it can only read, or can also change things.

5. Who checks the output?

AI sounds sure even when it's out of date. A lawyer in one of my sessions asked AI to improve a contract. It confidently used rules that had since changed, and said nothing about it. Once we told it to check current law, it flagged its own mistakes.

I use a simple rule. Around 95% right is fine for internal admin. Anything that leaves the business, like a client email, a proposal or a contract, gets a human check first. And for anything built on your own documents, ask the AI to quote the file and the line it used, so checking takes seconds, not minutes.

This week: agree which kinds of work always get a human check before they go out.

6. What happens when it goes wrong?

If an AI tool sent the wrong thing to a client tonight, who would find out, and how? Most teams don't have an answer.

Two things help. First, a simple reporting route: who to tell, and how fast. Second, make AI actions easy to undo while you're still building trust. One person I trained wanted AI to delete 50,000 old emails. I suggested it mark them unread instead, and only delete once she'd checked the rules worked. Same result. No risk.

This week: name the person to tell, and make sure everyone knows it.

A one-week plan

1. Monday: name one owner for AI.
2. Tuesday: list the AI tools and account types in use.
3. Wednesday: share the green, amber, red rule. Check your biggest client contracts.
4. Thursday: switch connected tools to read-only where you can.
5. Friday: agree what always gets a human check, and who to tell when something goes wrong.

Don't aim for perfect. Aim for 80 to 90% this week, then tighten it as you learn what your team actually does.

What this is, and what it isn't

This is a practical starting point, not legal advice. If you handle health data, make decisions about people with AI, or work in a regulated sector, you'll also want specialist or legal review. But a lawyer can only tell you whether something is allowed. Someone still has to find out what your staff actually do, fix the workflow and train the team. Most businesses need that part first.

Check where you stand in about ten questions

I built a free AI Governance Check that covers these six areas. No email needed to see your results. It won't tell you you're compliant, because ten questions can't prove that, and anything that says it can is selling you something. It will show you where the gaps are and what to do first. You can see exactly what it checks before you start.

If you'd rather talk it through, book a call.