The founder asks the question at the end of a board prep call. “If one of these AI tools goes wrong, what’s our actual exposure?” The delegate nods, says they’ll look into it, and walks away with no honest answer ready.
What they need is a working document, one page, the kind of thing you can open on a laptop when the question comes up again. Not a consultant’s framework. Not an enterprise compliance report. A simple list of the risks, with a rating next to each one, a control in place, and a name against it.
What is a one-page AI risk register?
An AI risk register is a document that records the AI-related risks your business has identified, rates each one by likelihood and impact, notes the control in place, names an owner, and sets a review date. At the scale of a 90-person owner-managed business, it fits on a single page or a shared spreadsheet. Writing it is where the value appears.
Filling it in puts a direct question to the business that many teams haven’t faced. Have the risks been named? Does anyone own them? Many owner-managed businesses have adopted AI tools quickly, found them useful, and moved on without going back to ask what could go wrong. The register is the check-in they skipped.
Why does your business need one right now?
Your exposure to AI-related risk exists whether or not you’ve documented it. The ICO has been clear that using AI to process personal data activates UK GDPR obligations from the moment you start, regardless of business size. A Data Protection Impact Assessment may be legally required before deploying AI that processes personal data at high risk. The risk register is how you show you’ve thought about this.
The research is consistent on where governance failures come from at this scale. The gap is rarely ignorance of the risk. It’s the absence of anyone named as responsible for it, no owner, no review date, no escalation plan. When something goes wrong, the business improvises. Improvisation under pressure tends to be slower and more expensive than the original incident.
The other prompt is the board question itself. If the founder, an investor, or a client asks for a summary of AI exposure, a completed register is the answer. The delegate who can open it during the meeting looks prepared. The delegate who promises to follow up has the same problem as before, just with less time to fix it.
Sector regulators are beginning to raise this directly. The FCA expects documented AI governance for regulated firms. The SRA expects law firms to have controls in place when using AI with client data. Having a register that names the risk, the owner, and the mitigation is evidence of a considered approach when that conversation arrives.
What should the register actually cover?
At the scale of an owner-managed business using commercial AI tools, risks fall into six to eight categories. Data leakage, hallucination, bias and discrimination, intellectual property, vendor lock-in, regulatory non-compliance, reputational damage, and security vulnerabilities cover the real exposure. Each maps to a class of failure that happens in practice, and each deserves its own row in the register.
Data leakage is usually the highest-impact category. It covers employees putting confidential material into free AI tools that train on user inputs (Samsung lost semiconductor design data and source code this way in 2023, using a free LLM tier with no data processing protections), vendors experiencing security breaches, and client files entering AI systems without a Data Processing Agreement in place. Hallucination covers AI producing incorrect or fabricated outputs that get relied on without a human check, a routine risk for any business using AI for drafts, research summaries, or document analysis.
Bias and discrimination covers AI systems making or influencing decisions about people in ways that discriminate unlawfully. Under UK GDPR Article 22, decisions that rely significantly on automated processing and carry legal or similarly significant effects on individuals require human oversight. Intellectual property covers both the risk that AI outputs infringe third-party copyright and the risk that proprietary information is absorbed into a vendor’s training data. Vendor lock-in is frequently underweighted until a migration becomes necessary. Regulatory non-compliance covers the obligations that apply by sector: SRA rules for law firms, FCA rules for financial advisers, and GMC and MHRA considerations for healthcare practices.
How do you make it usable day to day?
The structure that keeps a risk register live rather than decorative has six columns and a split cadence. Risk description, likelihood rating (low, medium, or high), impact rating on the same scale, the mitigation already in place, the named owner, and the review date. Monthly for the high-likelihood high-impact rows. Quarterly for everything else.
The likelihood and impact ratings do two things. They force a judgment: is this genuinely high-impact, or are you giving disproportionate weight to a low-probability scenario? They also create a natural triage. The rows marked high/high get monthly attention. Everything else gets a quarterly check.
Build escalation triggers into the register before anything happens. An example entry: if an AI tool generates three incidents in a quarter where confidential data was at risk, the tool is reviewed immediately and either retired or given stronger controls. Pre-writing this rule means the decision is made before the pressure arrives. Deciding what counts as enough incidents to act is a harder conversation after the third one than it is when nothing has gone wrong yet.
For a 90-person business, the cadence is achievable without dedicated governance staff. Fifteen minutes at a monthly team meeting covers the live rows. Thirty minutes with the operations lead every quarter covers the full register. An annual sweep handles regulatory changes and new tools. That’s roughly 30 hours a year from one person, which is a reasonable overhead for the exposure it addresses.
What sits alongside the risk register?
The register doesn’t stand alone in a well-maintained governance approach. It sits between the AI policy, which sets the rules, and the vendor due diligence process, which checks new tools before adoption. Together the three documents answer what regulators, boards, and investors are increasingly asking about. Each one is a short document at owner-managed business scale, and none of them needs to be more than a page.
The AI policy covers which tools are approved, what data can be input into them, and what human review is required before output reaches a client or third party. The vendor due diligence checklist captures whether a vendor has a Data Processing Agreement, whether it trains on user inputs, and where data is stored. The risk register is what you maintain once tools are running.
For FCA-regulated firms, SRA-regulated law firms, and healthcare practices, the register also feeds directly into sector compliance. Having it well-maintained means that when the regulator asks, the answer is already prepared rather than assembled under time pressure.
The register also tends to be where shadow AI surfaces. When employees contribute to a monthly review, they mention tools they’ve been using informally. That visibility is genuinely useful, as it shows the operations lead where the business is finding value in AI and where new exposure may be building. The register becomes a feedback loop, not just a record.
