An AI risk register for small business: keep it short and useful

Small businesses don’t need enterprise paperwork to manage AI risk. What they do need is somewhere to write down the risks they’ve spotted, the calls they’ve made, and the controls staff are expected to follow.

That’s the whole job of an AI risk register. Keep it short, keep it specific, and review it often enough that it still matches how the business actually runs.

If “risk register” sounds like the sort of thing that comes with a consultant and a workshop, strip the ceremony off and look at what’s underneath: a list of things that could go wrong, with a decision written next to each one. Small businesses run informal versions of this constantly, the mental list of which customers are slow payers, which machine is due to fail, which staff member is a flight risk. The AI version earns a written form for one reason: the risks are new enough that nobody’s instincts cover them yet. Nobody has twenty years of gut feel about what happens when a staff member pastes a client file into a chatbot, so the gut needs a page to lean on.

What belongs in the register

Start with the risks that can actually hurt you. Sensitive data pasted into unapproved tools. Wrong outputs used without anyone checking. Staff leaning on AI for legal, financial or safety decisions. Vendor terms that clash with your own obligations. AI-generated content going out to customers unread. Internal documents indexed with the wrong permissions so the wrong people can search them.

Notice what’s not on that list: robots, sci-fi, anything abstract. Every entry is a specific bad Tuesday. That’s the standard for inclusion, if you can’t describe the incident it would cause in one sentence, it doesn’t go in. “A quote goes out with an AI-invented price and the customer holds us to it.” “A tender response includes hallucinated experience we don’t have.” “The assistant answers a junior’s question from the salary spreadsheet.” Write the incident, not the category, and the register stays real.

Give each risk an owner, a likelihood, an impact, a control and a review date. If that feels like too much, cut the scoring back. The point is to force a decision, not to feed a template. A high/medium/low gut rating is fine; a 5x5 colour matrix is procurement theatre at this scale. The owner column is the one that matters most, because a risk with no name against it is a risk everyone has silently agreed to ignore.

Make controls practical

A control only counts if people can actually follow it. “Do not use AI irresponsibly” isn’t a control, it’s a wish. “Do not paste customer personal information into public AI tools” is something a person can do. “All customer-facing AI drafts get reviewed before they go out” is better still.

The test for every control is whether it survives 4:55pm on a Friday. Rules that require judgement under pressure (“use appropriate caution”) dissolve exactly when they’re needed; rules that are mechanical (“client names never go in the free tools, the approved tool is this one”) hold. It helps to write the control as the behaviour you want, not the behaviour you’re banning: “quotes drafted with AI get checked against the price list before sending” gives a person a step to follow, where “don’t trust AI pricing” just gives them a mood.

For a small business the controls that work are the plain ones. Clear rules, approved tools, access limits, a review point, and a log of what happened. If that sounds like the one-page AI policy, it’s because the two documents are siblings: the policy tells staff what to do, the register records why, and which risk each rule is holding down. Keep them consistent and one page each, and you’ve got governance a ten-person business can actually run.

Include shadow AI

Your staff are probably already using AI to draft emails, summarise notes, write formulas or tidy up text. Pretending that isn’t happening doesn’t make the risk go away. It just means you can’t see it.

So put the likely unsanctioned use in the register, and add a plan to swap it for an approved option wherever the use is reasonable. That way you can back the useful behaviour instead of leaving data handling to chance.

The instinct to respond with a ban is worth resisting, and not for soft reasons. A ban doesn’t stop the use; it moves it to personal phones and private accounts, where you have no visibility, no controls and no logs, which is strictly worse than the situation you banned. The register approach inverts it: assume the use, find it with a friendly amnesty (“reply with what you’re already using, no trouble”), and upgrade it, the staff member drafting emails with a free chatbot gets the paid business tier with training switched off, and the rule about what data goes in. Same productivity, most of the risk gone, and the next new tool gets asked about instead of hidden.

Vendors need scrutiny

When a product bolts on AI features, ask the obvious questions. What data does it use, does it train models on your inputs, where does the processing happen, how long is data retained, and are there admin controls. Write the answers down.

A lot of small businesses assume vendor AI must be safe because it lives inside a tool they already trust. Same familiar login, brand new path your data travels down.

This one deserves its own row in the register because it’s the risk that arrives without anyone deciding anything. Your accounting package, your CRM, your document tool, all of them are shipping AI features by default now, often switched on in an update you didn’t read about. The data that was sitting safely in the tool is suddenly being processed somewhere new, under terms that changed in an email nobody opened. The control is boring and effective: when a tool announces AI features, someone (the register says who) spends twenty minutes with the settings and the terms before the feature gets used with real data. The sovereignty questions are the checklist for that twenty minutes.

The register helps you say yes

A register isn’t there to block things. It’s there to make adoption easier to approve. When leaders can see the risks laid out next to the controls, they can back a useful project without having to pretend the risks are too fuzzy to name.

That’s the part that surprises people: the businesses with a register adopt AI faster, not slower. Without one, every proposal triggers the same circular meeting, someone’s enthusiastic, someone’s nervous, nobody can say precisely what they’re nervous about, and the default is to defer. With one, the nervousness gets converted into rows, each row gets a control, and the decision becomes “are we happy with these controls?”, which is a question a Tuesday meeting can actually answer. The register is how a small business says yes carefully instead of saying maybe forever.

Set the review rhythm when you create it, or it becomes a snapshot instead of a register. Quarterly is right for the first year, twenty minutes, three questions: what are we using now that we weren’t last quarter, did anything nearly go wrong, and does each control still describe what people actually do. The near-misses are the valuable part. The draft that almost went out with an invented figure, the file that almost went into the wrong tool, each one is a free lesson about which control is thin, and a register that never absorbs them is just laminated optimism.

For most teams, a one-page register and a short AI use policy is plenty to start with. The governance can grow later, as the use cases start to matter more. If you’d like help stress-testing yours, or you’re adopting AI and the risk conversation keeps stalling the whole thing, tell us what you’re using and what worries you and we’ll help you turn the worry into rows with controls, which is usually all it takes to get moving.