How Do You Run an AI Risk Assessment? A Practical Framework Guide

TL;DR

This guide covers everything you need to run an AI risk assessment inside an SME or mid-market business:

• What an AI risk assessment is and how it differs from general risk management
• What an AI risk assessment framework does, and how the NIST AI Risk Management Framework and the ICO’s guidance on AI and data protection fit together
• The top AI risk assessment frameworks, a plain-English explainer on the NIST AI RMF, and what the EU AI Act’s risk tiers mean for your assessment
• Why assessments matter now, from shadow AI usage to the UK’s tightening data protection rules
• The five components every credible framework covers: governance, scope, risk identification, controls, and monitoring.
• The extra risks that generative AI and large language models (LLMs) bring, and how to test for them
• A step-by-step process you can run this quarter, plus a worked recruitment-shortlisting example
• A practical checklist and a set of best practices for keeping the assessment alive after launch
• Where senior-led support, including RSVR’s approach, fits in if you want the assessment run for you rather than by you

What is an AI Risk Assessment?

An AI risk assessment is a structured way of asking one blunt question: what could go wrong if we use this AI system, and how bad would it be? It sits inside the wider practice of AI risk management, which is the ongoing job of keeping those risks at a level your business can live with.

To put it plainly, an AI risk assessment examines an AI tool or model, identifies the harm it could cause, and assesses how likely those risks are and how serious their impact could be. That judgment drives the controls you put in place. Done well, an AI risk assessment turns a vague sense of unease about a new tool into a clear, ranked list of things to fix.

This is not a new idea dressed up in fresh language. Traditional risk assessment has existed in health and safety, finance, and IT security for decades. What makes an AI risk assessment its own discipline is the shape of the risk. Models can behave unpredictably, learn from data you never fully audited, and produce confident answers that are quietly wrong. So a good AI risk assessment borrows the discipline of classic risk work but adds checks for bias, explainability, data provenance, and model drift.

If your team is still deciding whether it is ready to adopt AI tools, start with the basics. Many organisations conduct an AI readiness review before assessing individual AI systems in detail.

What is an AI Risk Assessment Framework?

A framework is what stops your risk work from being a one-off spreadsheet that somebody forgets about. An AI risk assessment framework is a repeatable structure, a set of categories, questions, scoring rules, and control options that you apply to every AI system the same way.

Without a framework, one team judges a chatbot as “low risk” while another calls a near-identical tool “high risk,” and nobody can explain the difference. With an AI risk assessment framework, the criteria are fixed in advance, the scoring is consistent, and the results can be compared across projects and defended to auditors, clients, or regulators. In short, the framework is what makes every AI risk assessment repeatable rather than improvised.

The two frameworks most UK and international teams lean on are:

The NIST AI Risk Management Framework (AI RMF) is a free guide from the US National Institute of Standards and Technology. NIST released version 1.0 in January 2023. It is voluntary, and it works for any sector or size of business. It does not tell you exactly what to do. Instead, it gives you a clear way to think about AI risk.

The UK ICO’s guidance on AI and data protection. The ICO expects organisations to run Data Protection Impact Assessments that consider whether using AI is more or less risky than alternatives, and to document why less risky options were or were not chosen.

These are not competing choices so much as complementary lenses. NIST gives you the operating model; the ICO tells you what UK data protection law expects. Most mature programmes borrow from both.

Top AI Risk Assessment Frameworks

No single framework fits every business. Most teams pick one as their backbone and borrow ideas from the others. These are the top AI risk assessment frameworks worth knowing.

How to choose. Keep it simple:

• Start with the NIST AI RMF for structure.
• Add the ICO guidance for UK data protection.
• Add the EU AI Act if you sell into the EU or your AI outputs are used there.
• Add ISO/IEC 42001 if clients or auditors ask for certifiable proof.
• Add the OWASP LLM list if you build or deploy generative AI tools.

You do not need all five on day one. A good AI risk assessment framework is one your team will actually use.

The NIST AI RMF Explained

The NIST AI Risk Management Framework (AI RMF) is a free guide from the US National Institute of Standards and Technology. NIST released version 1.0 in January 2023. It is voluntary, and it works for any sector or size of business. It does not tell you exactly what to do. Instead, it gives you a clear way to think about AI risk.

Its goal is trustworthy AI. NIST describes trustworthy systems as valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair, with harmful bias managed.

The framework has four functions. They run in a loop, not a straight line:

1. Govern: Set the culture, policies, roles, and accountability for AI risk. This sits across everything else.
2. Map: Understand the context. What does the system do, who does it affect, and what could go wrong?
3. Measure: Test and track the risks you found, using numbers where you can and judgment where you cannot.
4. Manage: Prioritise the risks, act on them, and keep monitoring.

These four functions line up neatly with the five components earlier in this guide:

In July 2024, NIST also published a Generative AI Profile (NIST AI 600-1). It lists 12 risks that are unique to, or made worse by, generative AI, and it suggests actions for each. We cover these in the generative AI section below.

What the AI RMF is not. It is not a law, and it is not a certification. It also does not come with a ready-made scoring sheet. You build that yourself, which is where a simple high/medium/low grid (see the step-by-step section) comes in.

EU AI Act Risk Tiers and What They Mean for Your Assessment

The EU AI Act is the first major law to regulate AI by risk level. It is binding, unlike the NIST AI RMF. UK organisations are not exempt. If you place an AI system on the EU market, or its outputs are used in the EU, the Act can apply to you.

The Act sorts AI systems into four risk tiers:

The Act also has separate rules for providers of general-purpose AI models. Key dates. The timeline moved in 2026, so check it before you plan:

• Since 2 February 2025: Bans on unacceptable practices and the AI literacy duty apply.
• Since 2 August 2025: Rules for general-purpose AI model providers apply.
• From 2 August 2026: Transparency duties for chatbots and AI-generated content apply. Generative systems already on the market have until 2 December 2026 to meet the content-marking rule.
• 2 December 2027: Rules for stand-alone high-risk systems, such as recruitment and credit scoring, now apply. The Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026) moved this date from August 2026.
• 2 August 2028: Rules for high-risk AI built into regulated products now apply.

What this means for your assessment. Add one step at the start: work out which tier each system falls into. Then match the depth of your assessment to the tier.

• Unacceptable: Stop. Do not build or buy it.
• High risk: Run the full assessment, keep detailed records, and plan for human oversight.
• Limited risk: Check that your disclosure and labelling are in place.
• Minimal risk: A light-touch assessment is usually enough.

AI systems used in areas such as employment can carry significant risks and may be subject to stricter requirements. For example, a tool that ranks job applicants can affect who gets considered for a role, making human oversight and clear documentation of the decision-making process especially important. We explain this recruitment example in more detail later in the guide. This is a general guide, not legal advice. Check the current text of the Act and the Omnibus, or ask a lawyer, before you decide where a system sits.

Why an AI Risk Assessment Matters Right Now

Three forces have turned AI risk management from a nice-to-have into a board-level concern.

First, adoption is outpacing governance. Staff are already using generative tools, whether or not leadership signed them off. This unmanaged usage, often called shadow AI, is one of the fastest-growing exposure points for mid-sized businesses, and it is exactly the sort of thing a proper shadow AI review uncovers. If you do not know where AI is entering the working day, you cannot assess its risk.

Second, the regulatory floor is rising. In the UK, the ICO’s guidance is not a statutory code, but it sets out clearly how the regulator interprets UK GDPR and the Data Protection Act 2018 in an AI context. The Data (Use and Access) Act 2025 received Royal Assent on 19 June 2025, and its provisions affecting data protection law are now in force in stages, which raises the stakes for documented, defensible decisions.

Third, the failures are expensive and public. A biased hiring model, a chatbot that leaks customer data, or an automated decision nobody can explain does not stay quiet. A single AI risk assessment done early is far cheaper than the clean-up after a public failure, and a documented AI risk assessment is often the first thing a regulator or client asks to see.

There is also a simpler, human reason. Doing this well is part of treating AI as a responsible operator would, which ties directly into the broader question of AI and business ethics that every leadership team now has to answer.

The Essential Components of an AI Risk Management Framework

If you strip any credible framework back to its bones, the same building blocks appear. These are the essential components of an AI risk management framework, and any AI risk assessment worth the name touches all of them. Skip one, and your AI risk assessment has a blind spot exactly where trouble tends to start.

1. Governance and accountability. Someone owns AI risk. There is a clear line of responsibility, a stated risk appetite, and a policy that says what is and is not allowed. NIST puts this first for a reason, because without governance, the other three functions have nowhere to report to.
2. Context and scope (mapping). Before you score anything, you map what the system does, who it affects, what data it touches, and where it sits in your operations. This is where you decide the assessment’s boundaries.
3. Risk identification and measurement. Here, you name the specific risks- bias, inaccuracy, security exposure, privacy breach, lack of explainability- and measure each one for likelihood and impact. This is the analytical heart of any AI risk assessment, sometimes called AI risk analysis. A thorough AI risk assessment does not stop at naming risks; it grades them so you know what to tackle first.
4. Controls and mitigation. For each significant risk, you decide what to do: reduce it, avoid it, transfer it, or knowingly accept it. This step, AI risk mitigation, is where human review, access limits, testing, and guardrails get assigned.
5. Monitoring and review. AI systems drift. Data shifts, behaviour changes, and a control that worked at launch may quietly stop working. Ongoing monitoring, with a way for people to report harms, keeps the assessment alive rather than frozen at launch day.

Think of these five as a control framework wrapped around the model, not a form you fill in once. A useful AI assurance framework loops back through all five on a schedule.

Generative AI and LLM-Specific Risks

Classic risk checks assume a system behaves the same way each time. Generative AI and large language models (LLMs) do not. Ask the same question twice, and you may get two different answers. They can also be steered by the text they read. So your AI risk assessment needs a few extra checks. Here are the main generative AI and LLM risks to include:

How to adapt your assessment for generative AI:

• Test with your own data. Vendor demos hide real-world failures.
• Try to break it. Run red-team tests, such as prompt injection and requests for private data.
• Log inputs and outputs. Keep them long enough to investigate problems, and respect privacy rules while you do.
• Re-test after every model update. A new version can change behaviour without warning.
• Keep a human in the loop. This matters most for decisions that affect people.

For a deeper checklist, use the NIST Generative AI Profile (12 risks) and the OWASP Top 10 for LLM Applications alongside this guide.

How to Do an AI Risk Assessment, Step by Step

Here is a practical sequence for running an AI risk assessment. It works for a scrappy first pass and scales up for a formal AI safety assessment.

1. Define the system and its purpose. Write down what the AI does, what decisions it influences, and who is affected. A vague scope produces vague risk.
2. Identify a lawful basis and the data involved. Map where data comes from, where it goes, and how it is used. If the data includes personal information, document the lawful basis for processing it. For AI training at scale, consent is often impractical, so organisations commonly rely on legitimate interests and document why that basis applies.
3. List the risks across every category. Use the five components above as prompts. Cover bias, accuracy, security, privacy, explainability, and over-reliance.
4. Score likelihood and impact. A simple high/medium/low grid on each axis is enough to start. The point is consistency, not false precision.
5. Assign controls to each material risk. Decide who does what by when. Meaningful human review of significant automated decisions is a control that the ICO specifically expects.
6. Document the decision. Record what you found, what you chose, and, crucially, why you rejected the less risky alternatives. This is your defensible audit trail.
7. Set a review date and monitoring metrics. Decide how you will know if the risk profile changes, and who gets alerted when it does.

You might also wonder whether you have to do all seven steps every time. For a low-stakes internal tool, a light version is fine. For anything touching customers, money, or protected characteristics, run the full AI risk assessment and keep the paperwork.

An AI Risk Assessment Example

A worked AI risk assessment example makes this concrete. A short AI risk assessment walkthrough is often the fastest way to see how the theory lands in practice. Imagine a UK recruitment firm wants to use an AI tool to shortlist CVs

• System and purpose: an AI model ranks applicants for a shortlist that a human recruiter then reviews.
• Data involved: names, addresses, employment history, education, and inferred characteristics. Clearly, personal data, some potentially special categories.
• Risks identified: bias against protected groups if the training data reflects past hiring patterns; inaccuracy in ranking; lack of explainability if a rejected candidate asks why; privacy exposure of applicant data.
• Scoring: bias is scored as high likelihood, high impact, because a discriminatory shortlist creates both harm and legal exposure under the Equality Act 2010.
• Controls: test the model’s outputs across different demographic groups; require a human recruiter to review and be able to override every shortlist; provide a plain-English explanation route for rejected candidates; restrict data access.
• Documented decision: the firm records that a fully automated shortlist was rejected as too risky, and that a human-in-the-loop version was chosen instead, with the reasoning written down.
• Monitoring: quarterly bias re-testing and a channel for candidates to raise concerns.

That single example touches all five framework components, which is exactly what a good AI risk assessment should do. Notice how the AI risk assessment did not just flag bias; it forced a concrete design change and a documented reason. Another common question is whether this differs from assessing an off-the-shelf tool versus one you build yourself. It does, mainly in how much you can see inside the model, which is one reason the custom versus off-the-shelf AI decision feeds directly into how deep your assessment needs to go.

An AI Risk Assessment Checklist You Can Use Today

Use this AI risk assessment checklist as a first-pass screen for any AI system before it goes live. Treat it as the minimum bar for an AI risk assessment: if you cannot tick a box, that gap is your risk.

• The system’s purpose and the decisions it influences are written down.
• Every data flow is mapped, and personal data has a documented lawful basis.
• Bias and fairness have been tested across affected groups.
• The model’s accuracy has been measured against a real benchmark.
• Security and access controls are in place and documented.
• There is a plain-language way to explain decisions to affected people.
• A human can review and override significant automated decisions.
• Less risky alternatives were considered, and the choice was recorded.
• Monitoring metrics and a review date are set.
• There is a channel for people to report harms.
• An owner is named and accountable for this system’s risk.

This checklist is not a substitute for a full framework on high-stakes systems, but it stops obvious problems from slipping through, and it forces the right conversations early.

Effective AI Risk Assessment Framework Best Practices

A few habits separate a framework that works from one that gathers dust. These are the AI risk assessment framework best practices worth building in from day one.

• Assess across the whole lifecycle, not just at launch. The ICO stresses that an AI lifecycle is not linear, so risk should be revisited from problem formulation through to decommissioning.
• Keep a human in the loop for anything that matters. Automated decisions with significant effects on people need meaningful human review, not a rubber stamp.
• Write down why you rejected the safer option. The single most valuable artefact from any AI risk assessment is the documented reasoning, because that is what turns a decision into a defensible one.
• Revisit the assessment on a schedule. A yearly or quarterly AI risk assessment catches the drift that a launch-day check never could.
• Score consistently, even if roughly. A shared large/medium/low scale beats precise-looking numbers that nobody trusts.
• Tie the framework to real controls. A risk register with no assigned owners or actions is theatre. Every material risk needs an action and a name against it.
• Measure the model, not just the paperwork. Track accuracy and speed in production, because an AI system’s performance in the real world is itself a risk factor once it starts drifting.
• Favour explainability where you can. Models you can interrogate are easier to govern, which is part of the broader case for explainable AI in any decision that affects people.

How RSVR Supports AI Risk Assessment

Most SME and mid-market teams do not lack the will to assess their AI risk; they lack the spare senior time to do it properly alongside everything else on their plate. Software can score a form, but it cannot interview your teams about where a model is actually being used, or judge whether your risk appetite matches what a regulator would expect to see.

This is where senior oversight becomes important. RSVR Tech runs structured AI risk assessments for growing businesses through its AI Data Safety service, which maps shadow AI usage, scores the risks against a framework rather than a gut feeling, and hands leadership a prioritised action plan rather than a static report. The work is built around WorkLex AI, so the assessment can be repeated regularly after the system goes live rather than becoming a one-time exercise. Each automated recommendation also has a designated person responsible for reviewing it and making the final decision.

If you would rather have this diagnosed and built for you than figure it out from a blog post, RSVR Tech works with COOs, General Counsel, and operations leaders across the UK, US, and ANZ to turn a vague sense of AI exposure into a governed, defensible programme.

Common Pitfalls to Avoid

• Treating it as a one-off. An assessment frozen at launch is worthless once the model drifts.
• Ignoring shadow AI. You cannot assess tools you do not know staff are using.
• Confusing a policy with a control. A written rule nobody enforces is not risk mitigation.
• Over-engineering low-stakes tools. Match the depth of assessment to the actual stakes.
• Skipping the documentation. If you did not write down why, you cannot defend the decision later.
• Treating the framework as separate from the work. The framework only earns its keep when a real AI risk assessment runs through it on a live system.

Conclusion

An AI risk assessment is not about predicting everything that could go wrong with an AI system. It is about understanding the risks you can identify, deciding which ones matter most, and putting clear controls around them.

The process does not need to be complicated. Define the system, understand the data and people involved, identify and score the risks, assign controls and owners, document the decisions, and review the assessment as the system changes.

The important part is to keep the assessment connected to how the AI is actually being used. A spreadsheet completed before launch will not tell you much if the model, data, vendor, or way people use the system changes six months later. Regular reviews, meaningful human oversight, and clear accountability keep the assessment useful.

For SMEs and mid-market businesses, the goal is not to build an enormous governance programme. It is to have a repeatable process that helps you make informed decisions about where AI is safe to use, where controls are needed, and when a system should not be used at all.