AI risk assessment: a practical method for one use case at a time
A seven-step AI risk assessment for one use case at a time, built on NIST, the EU AI Act, and ISO/IEC 42001.
BY STEFAN MARITZ··7 MIN READ
TL;DR
- An AI risk assessment is a structured look at one AI use case: what could go wrong, how likely and how harmful that would be, and which controls bring it to a level the business accepts. It is done per use case, before launch, and repeated when the use case changes.
- Cover seven risk types: privacy, accuracy, fairness, security, over-reliance, vendor dependence, and legal exposure. Most incidents we would worry about sit in the first two.
- Borrow structure from existing frameworks instead of inventing one. The NIST AI Risk Management Framework, the EU AI Act's four risk levels, and ISO/IEC 42001 all give you usable scaffolding.
- Rate each risk on likelihood and impact, then decide the control. For most internal use cases the main control is a named human reviewing the output, which only works if their review is given time.
- Decisions about people sit in the highest tier almost everywhere, including under POPIA section 71 in South Africa and the EU AI Act's high-risk category for employment.
Most organisations approach AI risk in one of two ways. Some write a policy that bans everything and watch staff use personal accounts anyway. Others approve tools one at a time with a security questionnaire that asks nothing about whether the output is correct.
This post sets out a practical AI risk assessment you can run on a single use case in an afternoon. It covers what the assessment is, which risks to include, a seven-step method, how to rate what you find, and which established frameworks to borrow from.
What is an AI risk assessment?
An AI risk assessment examines one specific use of AI and asks what could go wrong, how likely and how harmful that would be, and what controls reduce it to an acceptable level. The unit is the use case, such as "drafting replies to supplier queries", rather than the tool, because the same tool carries very different risk in different jobs.
It sits inside a wider AI governance framework, which decides who approves what. The risk assessment is the evidence that approval rests on.
Which risks should an AI risk assessment cover?
Seven types cover nearly everything a mid-sized organisation will meet. The table gives an example of each and the question that surfaces it.
| Risk type | Example | Question to ask | Typical control |
|---|---|---|---|
| Privacy | Employee records pasted into a consumer chatbot | What personal data goes in, and where does it go? | Approved tools only, and clear data rules in the usage policy |
| Accuracy | A hallucination, meaning confident output that is false, reaches a client | What happens if the output is wrong and nobody notices? | Human review, answers grounded in your own documents |
| Fairness | A screening tool scores some groups lower | Does this shape a decision about a person? | Human decision-maker, outcome monitoring |
| Security | Prompt injection, where hidden instructions in a document make the AI act on them | Can the system take actions or read untrusted content? | Limit permissions, keep a human approval step |
| Over-reliance | Staff stop checking because the output is usually right | Is the review step real, or a click-through? | Spot checks, review time built into the workflow |
| Vendor dependence | A price rise or model change breaks the workflow | Could we switch provider in a month? | Exit plan, avoid hard-coding to one model |
| Legal and regulatory | An automated decision breaches data protection law | Which laws apply to this data and decision? | Legal review for anything in the high tier |
Accuracy controls deserve a note. Grounding a model in your own approved documents, the approach behind retrieval-augmented generation, narrows the room for made-up answers, though it does not remove the need for review.
How to run an AI risk assessment in seven steps
- Describe the use case in two sentences. Who uses it, for what task, with what inputs, and what happens to the output.
- Map the data. List every type of data that goes in, where it is processed, and whether the vendor may retain or train on it.
- Name the decision. Say whether the output informs a decision, makes one, or takes an action. An AI that drafts is lower risk than one that sends, and AI agents that act across systems need the most scrutiny.
- Walk the seven risk types. For each, write the most plausible bad outcome in one sentence. Skip the ones that do not apply and say why.
- Rate likelihood and impact. Use a three-point scale for each (low, medium, high). The combination sets the tier.
- Choose controls and an owner. Each medium or high risk needs a control and a named person responsible for it.
- Set a review trigger. Reassess when the model, the vendor, the data, or the scope changes, and at least once a year.
The adoption step is in step six. A control that adds work nobody has time for will be skipped within a month, so build the review time into the workflow and the role.
How should you rate and tier AI risk?
Multiply likelihood by impact and sort the result into tiers. I find four tiers practical, and they line up loosely with the EU AI Act's four levels of risk: unacceptable, high, transparency, and minimal.
- Do not proceed: uses you would not defend publicly, or that the law prohibits.
- High: decisions about people's jobs, credit, or access to services. Legal review, a human decision-maker, and monitoring of outcomes. The EU classes AI tools for employment and management of workers as high-risk, and in South Africa POPIA section 71 restricts decisions based solely on automated profiling.
- Medium: customer-facing output or important internal decisions. Human review before anything leaves the building, and people told when they are dealing with AI.
- Low: internal drafting and summarising with a human in the loop. An approved tool and a usage policy are enough.
Most first use cases land in low or medium. That is by design: it is where you learn how your controls work in practice.
Which AI risk frameworks should you borrow from?
You do not need to invent a framework. Three established ones give you structure, and each suits a different need.
The NIST AI Risk Management Framework is voluntary and free, and is currently being revised. Its generative AI profile organises the work into four functions, govern, map, measure, and manage, which is a good checklist for whether your process is complete.
The EU AI Act gives you the tiering logic and, if you operate in Europe, legal obligations. ISO/IEC 42001 is the international standard for AI management systems, worth considering if your clients already expect ISO certification from you.
A risk assessment works like a pre-start check on heavy equipment. It takes minutes when the machine is in good order, and those minutes are what stop a small fault becoming an expensive incident.
Frequently asked questions
What is an AI risk assessment?
An AI risk assessment examines one specific use of AI to identify what could go wrong, how likely and harmful that would be, and which controls bring it to an acceptable level. It is done per use case, before launch, and repeated when the use case changes.
What are the main risks of using AI in business?
The main risk types are privacy, accuracy, fairness, security, over-reliance, vendor dependence, and legal exposure. For most internal uses, privacy and accuracy are where controls matter most.
Who should carry out an AI risk assessment?
The business owner of the use case should lead it, with input from IT or security, data protection, and legal for anything in the high tier. Ownership by the business keeps the assessment tied to how the tool is used.
Is the NIST AI RMF mandatory?
No. The NIST AI Risk Management Framework is voluntary, and it is a practical structure for organisations anywhere to borrow from.
How often should AI risk assessments be updated?
Whenever the model, the vendor, the data, or the scope changes, and at least once a year. Models and vendor terms change more often than most software, so set the trigger explicitly.
If you want an independent view of your AI risk and governance position, our AI readiness and value unlock audit assesses governance and risk alongside strategy, data, people, and execution.
WORK WITH PRAXES
Ready to move from reading about AI to getting your organisation ready for it? Start with the diagnostic.
Take the AI Readiness Diagnostic