AI Risk Assessment for Small Business: A One-Page Method

An AI risk assessment for a small business does not need a consultant, a framework certificate or a fifty-page report. It needs one honest page that answers four questions: where do we use AI, what could go wrong, how bad and how likely is each problem, and what are we doing about it. Done in an hour, it replaces vague worry with a short list of actions and an owner for each. Done never, it leaves you finding out about risk when a customer, a regulator or a leaked document does it for you.

This guide gives a one-page method you can finish in an afternoon, a scoring table, the risks that matter most for small teams and a worked example. It is general guidance, not legal advice, and it uses Ask Mio only as an example of a tool you might be assessing.

What an AI risk assessment is, and what it is not

A risk assessment is a structured way to look at a use of technology and decide whether the benefits justify the dangers, and what to do to reduce them. For AI, the dangers fall into a handful of groups: data leaving your control, wrong output being acted on, unfair or harmful results, legal and contractual breaches, dependence on a supplier, and people using tools you do not know about.

It is not a compliance certificate. Formal frameworks exist, for example the NIST AI Risk Management Framework, and some uses of AI face specific legal duties; our overview of the EU AI Act in plain language explains how that regulation sorts uses by risk. A small business usually needs something lighter first: a record that you looked, decided and assigned responsibility.

The one-page method

The AI risk assessment below has five steps. Each fits in a short section of a single page or sheet.

1. List the uses. Every place AI is used or planned, by whom and for what.
2. Note the data. What goes in: public, internal, personal data, confidential client material.
3. Name the risks. What could go wrong in each use.
4. Score and decide. Likelihood times impact, then accept, reduce or avoid.
5. Assign and review. One owner, one action, one review date per risk.

Step 1: list every use of AI

Start by writing down what is actually happening, not what the policy says. Ask each person and each team two questions: which AI tools do you use, and for what? Include chat assistants, features inside existing software, browser extensions, transcription tools and anything on personal accounts. The gap between the official list and the real one is often the biggest finding. Our article on shadow AI at work explains why people use tools quietly and how to respond without punishing them.

For each use, write one line: who, which tool, what task, how often. Typical entries for a small business are drafting customer emails, summarising meetings, writing product descriptions, checking code, translating documents and answering support questions.

Step 2: classify the data that goes in

The most useful single question is what data enters the tool. Use four simple classes.

Class	Examples	Default rule for AI tools
Public	Published website text, marketing copy, press releases	Fine to use
Internal	Draft plans, internal notes, non-sensitive process documents	Allowed in approved tools
Personal data	Names, emails, orders, CVs, anything about identifiable people	Only where the tool and purpose are approved; minimise or anonymise
Confidential or regulated	Client contracts, health or financial records, credentials, trade secrets	Do not paste unless a specific approval exists

Our guides to what not to share with AI and anonymising text before pasting it give practical rules for the third and fourth classes.

Step 3: name the risks

For each use, think through the risk groups below. You will not find all of them everywhere; the point is to ask.

Data and privacy

Could personal or confidential data be exposed, kept longer than intended or used for training? Read the supplier’s terms and privacy information. For Ask Mio, chats and files are the user’s, can be exported or deleted at any time, and are not used to train models. For any tool, check what the contract says rather than assuming. If personal data is involved at scale or in a sensitive way, a formal assessment may be required; see our walkthrough of a DPIA for AI tools, and the European Data Protection Board for official guidance.

Accuracy and hallucination

Could a wrong answer cause harm if nobody checks it? A wrong figure in a quote, a made-up legal reference or an invented product claim is a real cost. Our article on AI hallucinations explains the behaviour; the control is human review sized to the stakes.

Bias and fairness

Does the use affect people, for example in hiring, pricing or credit decisions, where biased output could disadvantage them? Uses that decide things about individuals deserve extra caution and often legal advice.

Legal, contract and intellectual property

Do contracts with clients restrict use of third-party tools? Could output infringe someone’s rights, or could you lack the rights you assumed over what you publish? See AI-generated text and copyright for the questions to ask.

Security

Could the tool be used to attack you, for example through prompt injection in a document you ask it to read, or phishing text that looks genuine? Our guides to prompt injection and AI phishing emails cover both.

Dependence and continuity

What happens if the supplier changes terms, raises prices or disappears? Can you export your work? Our guide to switching AI assistants without losing your work describes the habits that reduce lock-in.

Reputation and customer trust

Would customers be upset to learn AI wrote or handled this? Is disclosure expected or required? A message that reads as robotic or wrong can cost more than the time it saved.

Step 4: score and decide

Use a three-point scale for both likelihood and impact so scoring takes seconds, not meetings.

Score	Likelihood	Impact
1	Unlikely in a year	Minor, easily fixed, no one outside is affected
2	Could happen once or twice a year	Costly or embarrassing, affects customers or revenue
3	Likely or already happened	Serious: legal exposure, major loss, harm to people

Multiply the two. A score of 1 to 2 is low, 3 to 4 is medium and 6 to 9 is high. Then pick one response per risk.

• Accept. Low risks with a note that you looked.
• Reduce. Add a control: review step, anonymisation, access limit, approved tool list, training.
• Avoid. Stop the use, or do it without AI, when risk is high and controls are weak.
• Transfer. Move the risk by contract or insurance where that is sensible and legal.

Step 5: assign owners and set a review date

A risk without a name next to it is not managed. For every medium or high item, write one action, one owner and one date. Add a calendar reminder to review the whole page every six months, and whenever you adopt a new tool or a supplier changes its terms. That is what turns a document into a practice. Many teams combine it with a short AI use policy that states the rules for the four data classes.

Controls that work for small teams

You do not need expensive tools. These controls reduce most of the risk for little cost.

• An approved tool list. Fewer tools, chosen on purpose, with business accounts rather than personal ones.
• A data rule on one page. The four classes and what may be pasted, written in plain words.
• A human check proportional to stakes. Nothing leaves the building to a customer, a regulator or the public without someone reading it.
• Anonymise by default. Replace names and numbers with placeholders unless the task needs the real ones.
• Short training. Thirty minutes covering the rules and three examples of what went wrong elsewhere.
• An incident habit. A named person to tell when something sensitive was pasted. Our guide on what to do after an AI data leak gives the steps.

A worked example: a ten-person marketing agency

An agency lists five uses of AI and scores them.

Use	Data class	Main risk	Score	Response
Drafting social captions	Public	Off-brand or inaccurate claims	2	Accept with editor review
Summarising client calls	Confidential	Client details exposed; contract limits	6	Reduce: approved tool, client consent, no names
Translating campaign copy	Internal	Wrong meaning in market	4	Reduce: native speaker check
Writing code for the website	Internal	Insecure or untested code	4	Reduce: code review and tests
Screening applicants’ CVs	Personal	Bias, legal duties	9	Avoid for now; revisit with legal advice

Result: one use stopped, two with new controls, two accepted. Each has an owner and a review date in six months. Total effort: about two hours, and the agency can now answer a client’s question about AI use with a page instead of a shrug.

Evaluating a tool like Ask Mio inside the assessment

When the assessed use involves a specific assistant, collect the supplier facts you need and record the date. For Ask Mio, the relevant points are: the free plan needs no card; paid plans start at €5 a month; the Business plan at €29 a month has 5 seats and shared projects; chats and files can be exported or deleted; data is not used to train models; and Ask Mio runs on its own servers. Compare those facts with your data classes and contracts, and read the pricing page and the terms for the current wording. Our AI privacy checklist gives twelve questions to put to any vendor.

Frequently Asked Questions

Do small businesses need an AI risk assessment?

There is no single rule for all businesses, but it is a sensible habit and may support legal duties such as data protection. A one-page record shows that you listed your uses, considered data and accuracy, and assigned owners. If your uses affect people’s rights, jobs or finances, take legal advice as well.

How long does an AI risk assessment take?

For a team of ten with a handful of uses, about two hours to gather information and score it, plus time to carry out the actions. The first one is the slowest. Reviews every six months usually take under an hour, because you are updating a list rather than starting over.

What is the biggest AI risk for a small business?

Usually it is confidential or personal data being pasted into tools without any rules, followed by unchecked output reaching customers. Both are controlled cheaply with a one-page data rule and a human review step. Specific uses, such as hiring decisions, can carry greater legal risk and need separate care.

Should we ban AI tools to be safe?

Blanket bans often push use out of sight, which is worse. A better approach is an approved list, clear data rules and training. Ban or avoid specific uses where risk is high and controls are weak, and revisit them when circumstances change.

Is this the same as a data protection impact assessment?

No. A DPIA is a specific legal assessment for processing likely to create high risk to individuals’ rights. The method here is a lighter overview of all your AI uses. It can show you which uses need a full DPIA, and its notes can feed into one.

Who should own the assessment?

One named person, usually the owner or an operations lead, who coordinates input from each team and keeps the page current. Individual risks get their own owners. If nobody is responsible for the whole page, it is rarely updated and quickly stops matching reality.

The Bottom Line

A useful AI risk assessment is short, honest and owned. List real uses, classify the data, name what could go wrong, score it simply and assign an action and a date to each medium or high risk. Review it twice a year and whenever tools or terms change. That gives a small business most of the protection of heavier processes at a fraction of the effort. Start with your top five uses on a single page, then compare the Ask Mio plans if you want a business account with shared projects for your team.