An AI risk assessment for a small business is a one-page check run before staff use AI for a new job. It names the job, the data involved and who sees the output. Then it scores how likely a mistake is and how much it would hurt, from 1 to 3 each. Multiply the two scores, and the result tells the owner to approve, add conditions or say no.
What should an AI risk assessment cover in a small business?
An AI risk assessment in a small business should cover one use case at a time. The same tool can be safe for one job and risky for another. ChatGPT drafting a job advert and ChatGPT summarising a client file are two different assessments.
The template below fits on one page. It takes about 20 minutes per use case. The owner or a named deputy fills it in with the person who wants to use the tool.
- The job. One sentence on what the AI will do. “Draft first replies to routine client emails” is a job. “Use ChatGPT” is not.
- The tool and account. The product name and whether it runs on a business account with a contract, or a personal one.
- The data. What goes in. Use the green, amber and red levels from your AI acceptable use policy.
- The person check. Who reads the output before it reaches a client, a customer or a decision.
- The score and decision. Likelihood times impact, the decision, any conditions, and a review date.
Questions 3 and 4 do most of the work. A use case with green data and a human check on every output rarely scores above 2.
How do you score an AI use case?
You score an AI use case by rating likelihood and impact from 1 to 3, then multiplying them. A three-point scale keeps the meeting short, because each level has a plain test that settles it. A finer scale spends the meeting arguing over the middle.
Likelihood asks how likely a mistake is to get past your checks:
- 1, rare. A person reads every output, and the data is green.
- 2, possible. Spot checks only, or amber data, or the tool’s terms are unclear.
- 3, likely. Output goes straight to clients, or red data goes in without a business contract.
Impact asks how bad that mistake would be:
- 1, minor. Fixed within a day, and no client notices.
- 2, costly. A client sees an error, or you lose time or money putting it right.
- 3, serious. Personal data is exposed, someone is treated unfairly, or a regulator or lawyer gets involved.
Two gates come before the score. If either is true, the use case cannot be approved on the score alone:
- The AI makes or shapes a decision about a person. Hiring, pay, credit, tenancy and disciplinary decisions all count. Screen for a full data protection impact assessment first (see below).
- Red data would go into a tool with no business contract. The answer is no until the account changes. The settings that decide this are covered in is ChatGPT safe for business use.
What does the approval rule look like?
The approval rule turns the score into one of four decisions, so two people scoring the same use case reach the same answer.
| Score | Decision | What happens next |
|---|---|---|
| 1-2 | Approve | Add the tool and job to the approved list. Review in 12 months. |
| 3-4 | Approve with conditions | Write down each condition, such as “a partner reads every draft”. Review in 3 months. |
| 6 | Rework first | Change the data, the account or the check, then score again. |
| 9 | No | Record the reason, and offer the person a safer way to do the same job. |
The 3-month review on conditional approvals matters most. Conditions drift as people get comfortable, and a quick re-score catches that.
Bedrock AI’s view is that most assessments should end in a conditional yes. A small firm that says no to everything sends staff back to personal accounts, which is how shadow AI starts.
What does a scored AI risk assessment look like?
A scored AI risk assessment is easiest to understand on real jobs. Here are three use cases from a hypothetical 12-person accountancy practice, scored with the template above.
| Use case | Data | Person check | Likelihood | Impact | Score | Decision |
|---|---|---|---|---|---|---|
| Drafting replies to routine client emails in ChatGPT Business | Amber, no client names pasted | Accountant reads every draft | 1 | 2 | 2 | Approve |
| AI note taker on client calls, free personal account | Red: client finances and personal details | None until someone reads the notes | 2 | 3 | 6 | Rework first |
| AI ranking job applicants’ CVs | Red: applicants’ personal data | Shortlist goes straight to interview | 3 | 3 | 9 | No, and the person gate trips |
The email drafting case scores 2 because two things hold. No client names go in, and a qualified person reads every draft.
If the practice later wants client names in the prompt, the data moves from amber to red, and the job needs a new assessment.
The note taker case is the instructive one. The job is useful, but a personal account holding client financial details scores 6.
The rework is specific: move it to a business account covered by a contract and tell clients at the start of the call. Then name one person who checks the notes before they go into the client file.
After those changes, likelihood drops to 1 and impact stays at 3. The new score is 3, an approval with conditions and a review in 3 months.
The CV case fails twice. It scores 9, and it shapes a decision about people. That means a data protection impact assessment before anyone revisits it.
When does an AI use case need a full DPIA?
An AI use case needs a full data protection impact assessment (DPIA) when it is likely to result in a high risk to people’s rights. That is the test in Article 35 of UK GDPR.
The ICO’s guidance on when you need a DPIA lists the features that point to high risk.
Three of those features catch small firms using AI:
- Innovative technology. The ICO names AI, machine learning and deep learning as examples. On its own this is not enough. Combined with one other feature it usually is.
- Evaluation or scoring. Ranking CVs, rating staff performance or scoring customers all count.
- Decisions about access. Decisions on whether someone gets a product, service, job or opportunity, made to any extent by automated means.
The same ICO page says employees can count as vulnerable people, because of the power imbalance at work. Any AI that scores staff or applicants should be treated as DPIA territory.
If you decide a DPIA is not needed, the ICO says to write down why. The one-page template does that job. Keep each completed sheet, because it is your record that you thought about the risk before you started.
Why not use the government or ICO AI risk templates?
The official AI risk templates were built for organisations with compliance staff, and a 12-person firm will struggle to finish them.
The UK government tested this directly. In November 2024 it consulted on AI Management Essentials, a self-assessment tool aimed at small and medium firms.
The government response, updated February 2026, accepts that the tool’s scope was too broad. It also notes concerns about complexity for smaller organisations, and says the tool will not be published.
The ICO’s AI and data protection risk toolkit is a free spreadsheet.
It is thorough, and useful when you build or buy an AI system that handles personal data at scale. For deciding whether the office manager can use an AI note taker, it is more than the job needs.
The sensible split is simple. Use the one-page template for everyday use cases. Move to a DPIA, and the ICO toolkit, only when a use case trips the person gate.
Who should approve AI use cases in a small firm?
The owner, or one named deputy, should approve AI use cases in a small firm, and should promise an answer within five working days. A slow answer is a no in practice, and staff route around it.
Keep the completed sheets in one shared folder, with the approved list on top. That list is what your acceptable use policy points staff to.
A risk assessment answers whether a use case is safe enough. It does not answer whether the use case is worth doing.
Score the job’s value first, using a method like scoring your workflows, then assess the risk of the jobs that are worth it. Diagnosing value before risk stops a firm spending an afternoon approving tools nobody needed.
FAQ
Is an AI risk assessment a legal requirement for a small business? An AI risk assessment is not a named legal requirement in the UK. A DPIA is required under UK GDPR when AI processing of personal data is likely to be high risk, such as scoring job applicants. The one-page check is how a small firm spots those cases.
How long does an AI risk assessment take? An AI risk assessment using a one-page template takes about 20 minutes per use case for a small business. A full DPIA takes longer, and is only needed when the AI shapes decisions about people or handles personal data at high risk.
Do I need a new AI risk assessment for every tool? A small business needs a new AI risk assessment for every new job. That applies even when the tool is already on the approved list. The same tool can score 2 for drafting adverts and 9 for screening CVs. Re-score when the data, the account or the human check changes.
What is the most common reason an AI use case fails the assessment? The most common reason an AI use case fails is red data going into a personal account with no business contract. The fix is usually moving the job to a business account and adding a named person who checks the output.
Get ahead with AI by getting The Business Brain | Would rather talk it through? Book a free 15-minute call