Method first, service second
How to run an AI risk assessment (and when to get help)
An AI risk assessment answers one question: of all the ways your organisation now depends on AI, which ones could hurt you, and how badly? This page sets out a five-step method you can run internally, grounded in the NIST AI RMF, ISO 42001 and the EU AI Act's risk categories, with a free template to run it on. And when you would rather have practitioners do it, the fixed-price service is at the bottom.
The territory
What an AI risk assessment actually examines
Not just the models. The risks that reach a UK business come from the whole estate: confidential data flowing into tools with unclear retention, fluent but wrong outputs feeding real decisions, AI features arriving inside existing software without anyone opting in, vendors whose failures become yours, and, for systems you build, attacks on the systems themselves.
If you searched for an AI risk management framework, this is where the recognised ones earn their keep: the NIST AI RMF supplies the assessment questions, ISO 42001 defines the management discipline around them, and the EU AI Act supplies the legal risk categories. The method below borrows from all three; the wider AI governance framework guide shows how they relate.
The deliverable is unglamorous and valuable: a scored register of AI risks with a named owner and a treatment decision against each, current enough to survive a client's due diligence questionnaire.
The method
A five-step AI risk assessment method
Runnable by an internal team with the template below. Each step feeds the next, so run them in order and resist starting at step three because the inventory feels like admin.
Identify every AI system in scope
List the tools you bought, the AI features switched on inside software you already own (Copilot across Microsoft 365 is the classic blind spot), the models and integrations you build, and the tools staff adopted on their own initiative. Expense records, browser logs and a short staff survey surface most of the last category. Nothing off this list gets assessed, so completeness here decides the value of everything after.
Classify each use case
For every system, record what data goes in, what decisions or outputs come out, and who is affected by them. The EU AI Act's risk categories (prohibited practices, high risk, limited risk with transparency duties, minimal risk) make a workable classification scheme even where the Act does not bind you, and using them means later applicability questions arrive pre-answered.
Assess against a framework
Structured questions beat instinct. The NIST AI RMF's Map and Measure functions supply them for organisational risk; ISO 42001 defines what a repeatable AI risk process must look like if you want the discipline auditable. For systems you build or expose to users, add the security threats: prompt injection, data leakage through outputs, and poisoning of anything the model learns from.
Score and prioritise
Rate likelihood and impact on a scale your organisation already uses for other risks, resist the temptation to mark everything amber, and rank the register. A short register with owners against the top items beats a long one that flatters its own thoroughness.
Treat and monitor
Each risk gets one of four decisions: accept it and record why, constrain the use case through policy, add a control such as mandatory human review, or retire the system. Then set the re-assessment triggers: any material change to the system or its data, and a calendar date so drift gets caught even when nothing announced itself.
The treatment step usually lands work on two other documents: the AI governance policy for use cases you decide to constrain, and, where you want the whole discipline auditable, the requirements of ISO 42001.
The working document
What the AI risk assessment template contains
Five parts matching the five steps: an inventory sheet with the prompts that surface staff-adopted tools, classification questions per use case, a risk register with a ready-made scoring scale, a treatment plan that forces a decision and an owner per risk, and a re-assessment log so the register stays a living document.
Each part carries a worked example, because a blank register is where most first attempts stall. Adapt the examples, do not inherit them: your risks live in your use cases, not ours.
Take the template
The full five-part document, free, with no form between you and it. If it does the job on its own, that outcome suits us fine.
Assessed for you
The fixed-price AI risk assessment service
When the estate is large, the stakes are contractual or the internal hours simply do not exist, our consultants run the full method for you: discovery workshops and technical checks to build the inventory, classification and framework-based assessment of every use case, and a scored register with a treatment roadmap and a board-ready summary at the end.
Pricing is banded by the number of AI use cases in scope, starting from £8,500 as a fixed fee set by scope, and every band is on the pricing page rather than behind a sales call. You know the number before we know your name.
The same engagement doubles as groundwork: the register and evidence it produces are the risk backbone of ISO 42001 readiness if certification is on your horizon.
Quick answers
AI risk assessment questions, answered
What is an AI risk assessment?
A structured review of where artificial intelligence is used across an organisation and what could go wrong: data leakage into tools, incorrect outputs being acted on, legal and regulatory exposure, security weaknesses in AI systems, and dependence on the vendors behind them. The output is a scored risk register with named owners and treatment decisions, refreshed on a defined cycle.
How often should AI systems be risk-assessed?
On two triggers. Event-driven: whenever a system, its data sources or its use case materially changes, and whenever a new system is adopted. Time-driven: at least annually across the estate, because AI products change under you between your own changes. Systems whose outputs affect people or client commitments deserve a shorter cycle than back-office conveniences.
What should an AI risk assessment template include?
Five parts: an inventory sheet for the systems, classification questions per use case, a risk register with a scoring scale, a treatment plan recording the decision and owner for each risk, and a re-assessment log. Our free template contains all five with worked examples, structured so the register can be maintained after the first pass rather than rebuilt.
Is an AI risk assessment the same as a DPIA?
No, though they overlap. A data protection impact assessment is a UK GDPR exercise focused on personal data and the rights of the people it describes. An AI risk assessment is broader: accuracy, security, legal exposure, vendor dependence and misuse, across systems that may touch no personal data at all. Where an AI use case processes personal data, the assessment will often show a DPIA is also needed, and the two share inputs.
Know your exposure
Get your AI estate assessed
Run the method yourself with the free template, or book a scoping call and have every AI system you rely on assessed, scored and roadmapped at a published fixed price.