Quick answer: An AI risk register lists each realistic AI failure mode in your business, scores it by likelihood × impact (a 5×5 scale works), names an owner, and tracks a mitigation. The risks that dominate SMB registers are: confidential data entered into AI tools, hallucinated content reaching clients, bias in people-affecting decisions, vendor model drift, and regulatory non-compliance (EU AI Act). Below are 12 worked examples with suggested scores you can adapt, plus the column structure and scoring method.
The Register Structure
| Column | What goes in it |
|---|---|
| ID | R-01, R-02… |
| Risk description | Specific event, not a theme ("employee pastes client PII into unapproved chatbot," not "data risk") |
| System(s) | Which inventoried AI tools it applies to |
| Likelihood (1–5) | 1 = rare, 5 = expected this quarter |
| Impact (1–5) | 1 = annoyance, 5 = existential/regulatory |
| Score | L × I; ≥15 red, 8–14 amber, ≤7 green |
| Owner + mitigation | One name, one action, one due date |
12 Worked Examples
Data and confidentiality
R-01 — Confidential data entered into a non-approved AI tool. An employee pastes a client contract into a personal free-tier chatbot. Likelihood 4 (it's happening somewhere in your company today), impact 4 (breach of client confidentiality, possible GDPR exposure). Score 16 — red. Mitigation: approved business-tier tools with training opt-out, acceptable use policy with red-tier data rules, short training.
R-02 — Vendor trains models on your data by default. A tool you approved changes its data terms, or training was never actually off for your tier. L3, I4 = 12, amber. Mitigation: verify opt-out at account level, contractual DPA, annual re-check via your vendor assessment questionnaire.
R-03 — AI meeting assistant records and stores sensitive conversations. Transcription bot joins a call where layoffs or deals are discussed; transcript sits in a third-party cloud indexed for search. L4, I3 = 12, amber. Mitigation: recording rules in policy, restricted bot permissions, retention settings.
Output quality
R-04 — Hallucinated facts in client deliverables. A report cites statistics that don't exist. L4, I4 = 16, red for any services firm. Mitigation: "you own the output" rule, mandatory source-checking for factual claims, second review on client-facing work.
R-05 — AI-generated code introduces a security vulnerability. Copiloted code with an injection flaw merges to production. L3, I4 = 12, amber. Mitigation: same review/test bar as human code, security linting in CI.
R-06 — Inaccurate AI chatbot promises to customers. Your support bot confidently states a refund policy you don't have. L3, I3 = 9, amber. Mitigation: constrain the bot to vetted knowledge, log and sample-review conversations, clear AI disclosure (required anyway under EU AI Act Article 50 from 2 August 2026).
People decisions and bias
R-07 — Bias in AI-assisted hiring screening. Résumé-ranking tool systematically downranks a protected group. L3, I5 = 15, red. This is also Annex III high-risk territory under the EU AI Act, with deployer obligations applying by 2 December 2027 (risk tiers explained). Mitigation: documented human review of every decision, vendor bias documentation, consider whether you need this tool at all.
R-08 — Employee monitoring feature crosses into emotion recognition. A productivity or call-analytics tool ships a "sentiment/engagement scoring" feature — prohibited in EU workplaces under Article 5, in force since 2 February 2025. L2, I5 = 10, amber (low likelihood, but the impact tier is the AI Act's highest fine band: up to €35m or 7% of turnover). Mitigation: feature-level audit of HR and analytics tools, disable on discovery.
Vendor and operational
R-09 — Silent model change degrades a validated workflow. Vendor swaps the underlying model; your prompt templates start producing worse output. L4, I2 = 8, amber. Mitigation: change-notice clause in contract, monthly spot-checks of critical outputs.
R-10 — Critical-process dependency on a single AI vendor. The tool your delivery process depends on doubles prices or shuts down. L2, I4 = 8, amber. Mitigation: data export tested, fallback process documented.
Regulatory and reputational
R-11 — Missing EU AI Act obligations on a live deadline. No Article 4 literacy records, no Article 50 disclosure plan for August 2026. L3, I4 = 12, amber. Mitigation: run the SME compliance checklist, file training attestations, diary the 2027/2028 high-risk deadlines.
R-12 — Undisclosed AI-generated content damages trust. A client discovers deliverables were largely AI-generated against contract terms. L2, I4 = 8, amber. Mitigation: contract review for AI clauses, internal disclosure defaults.
How to Score Without Overthinking
Calibrate likelihood against a one-year horizon and your actual headcount. Two scoring traps to avoid: averaging optimism (scoring everything 2–3 so nothing demands action) and theoretical maximalism (scoring everything 5 so the register stops discriminating). The register's whole job is to rank — protect the spread.
Skip building the spreadsheet: the AI Governance Toolkit Pro ($99) ships with a pre-structured risk register including the 5×5 scoring model and starter risks, plus the governance charter, vendor assessment, and training deck that feed it.
Running the Register (So It Doesn't Die)
- Populate from the inventory. Every system in your AI inventory gets at least one register pass.
- One owner per row. "Team" is not an owner.
- Quarterly review, 30 minutes. Rescore, close mitigations, add new tools' risks.
- Feed it from incidents. Every reported AI incident either matches an existing row or creates a new one.
FAQ
What is an AI risk register?
A living document listing the specific AI-related failure modes in your organization, each scored by likelihood and impact, with a named owner and mitigation. It's the core tracking tool of an AI governance framework.
What are the most common AI risks for small businesses?
Confidential data entered into AI tools, hallucinated content reaching clients, bias in hiring or other people-affecting decisions, vendor model drift, and missed regulatory obligations under the EU AI Act.
How do you score AI risks?
Multiply likelihood (1–5, over a one-year horizon) by impact (1–5, calibrated to client loss and regulatory exposure). Treat scores of 15+ as requiring immediate mitigation, 8–14 as planned mitigation, 7 and below as monitored.
How often should we update the register?
Quarterly as a routine, plus immediately after any incident, any new AI tool adoption, or a regulatory milestone — the next being Article 50 transparency obligations on 2 August 2026.
Get the register pre-built. The AI Governance Toolkit Pro ($99) includes the scored risk register template plus charter, vendor assessment, and training deck — or take the whole governance and compliance document set in the Complete AI Compliance Stack ($199).
This article is for general information only and is not legal advice. Consult qualified counsel for your specific situation.