AI risk management is how an organisation identifies, assesses and controls the risks that artificial intelligence introduces, from data leakage and bias to unreliable output and vendor failure. This guide from the Institute of AI sets out what a working framework looks like, where it commonly goes wrong, and how a UK board can put real controls in place.
A plain answer first, because most boards are researching this while the technology is already in the building.
AI risk management is the process of identifying, assessing and controlling the risks that artificial intelligence introduces across an organisation. It answers three questions a board cannot afford to leave open: where could AI cause us harm, how serious and how likely is that harm, and what are we doing to keep it within a level we are willing to accept. Done well, it is not a one-off audit. It is an operating discipline that runs for as long as the systems do.
The risks are specific and manageable once they are named. They include personal data leaking into a third-party model, biased or unfair automated decisions, confident but wrong output being acted on, security and prompt-injection attacks, over-reliance on a single vendor, and models that quietly drift as the world moves away from their training data. In the UK these sit alongside existing obligations, particularly data protection under UK GDPR, rather than replacing them. Good AI risk management makes those duties easier to meet, not harder.
Crucially, risk management assigns responsibility to people, not to the technology. A model cannot be held to account, so accountability rests with named individuals, from the board that sets the risk appetite to the employee who checks an output before acting on it. The Institute of AI, the UK's professional body for AI, exists to define that standard and help organisations meet it. The sections below break the framework into its parts and show where it most often breaks down.
A working framework is made of concrete, named parts. Miss one and the gap tends to be exactly where the next incident appears.
A living register of every place AI is used, built in-house, bought in, or embedded in a product you already run. Each entry names what the system does, what data it touches, and what could go wrong. Most uncontrolled risk hides in shadow AI that teams adopted quietly, so identification has to reach beyond the systems IT signed off.
A consistent scale that rates each use by likelihood and by the harm it could cause, from low-stakes drafting to decisions that affect a person's money, health or rights. Rating turns a long list into a short queue: the handful of high-impact uses that earn tighter controls, human oversight and, where the law requires it, a formal impact assessment.
The technical and human measures that reduce a risk to a level the board is willing to accept: access limits, human review of high-stakes output, testing for bias, guardrails on prompts, and a documented decision to accept, reduce, transfer or avoid each risk. A control you cannot evidence is a control you do not have.
Clear rules on what data may train, prompt or fine-tune a model, on what lawful basis, and for how long. This is where AI risk meets UK GDPR: minimise personal data, document the purpose, and never treat a customer record as prompt fuel. Data leakage into a third-party model is one of the most common and most damaging AI incidents.
Every material risk needs a named owner who is answerable for it, and a board that sets the appetite the whole organisation works to. A model cannot be held to account, so responsibility rests with people: the executive who funds the controls, the risk function that challenges, and the person at the keyboard who checks an output before acting on it.
Ongoing measurement of how deployed systems actually behave, with logging, drift detection, incident reporting and a clear route to escalate or switch a system off. Models degrade as the world moves away from their training data, so controls signed off at launch and never revisited quietly stop protecting you. Monitoring has to run for the life of the system.
Most AI incidents are not exotic. They are a small set of predictable failures, each with a control that would have prevented it.
Teams paste sensitive data into consumer chatbots and browser extensions that no one has reviewed. The organisation carries the risk without ever seeing it.
Control: maintain a live AI inventory and an acceptable-use policy, so every system is visible and every tool has an owner.
Customer records are used as prompts or training data with no lawful basis, and that data resurfaces in outputs or is retained by a vendor indefinitely.
Control: set clear data rules tied to UK GDPR, minimise personal data, and run a data protection impact assessment before any high-risk use goes live.
A system trained on skewed history quietly disadvantages a group in hiring, lending or triage, and no one notices until a complaint or a regulator arrives.
Control: test for disparate impact before launch, keep meaningful human review on decisions that affect people, and document the fairness checks.
A generative model fabricates a figure, a citation or a policy, and a member of staff acts on it because it read as authoritative.
Control: require human verification of AI output before it is relied on externally, and train staff to treat fluent output as a draft, not a fact.
A supplier updates or retires a model, and a workflow that depended on it silently degrades, with no fallback and no one accountable.
Control: run due diligence on suppliers, contract for notice and stability, and keep a documented fallback for every business-critical AI use.
A policy was written, signed and filed, but no one measures whether reality matches it, so the controls decay while the risk register says all is well.
Control: assign owners, review the register on a fixed cadence, and treat monitoring as a standing operating task rather than a one-off project.
A risk framework does not have to start from a blank page. The Institute of AI offers three practical routes, from a free public commitment to hands-on advisory.
A public commitment to getting your organisation ready for AI, set out as five practical pledges, including human oversight on decisions that affect people. Signing is free and you choose at least three, which anchors your risk management in a recognised standard people can see.
Sign the CharterIndependent assessment of your organisation's AI practice against the Institute of AI's standard. Accreditation moves you from a stated intention to an externally verified one, giving clients, regulators and partners assurance that your risk controls are real and not just declared.
Explore accreditationPractical, vendor-neutral guidance to design a risk framework that fits your size, sector and appetite, and to build the internal capability to run it. Advice from an independent professional body that resells no third-party software and takes no implementation kickbacks.
Talk to an adviserStart with a free, public commitment to getting your organisation ready for AI. Signing the UK AI Readiness Charter is a credible first step towards risk controls you can evidence to clients and regulators.
Five practical pledges on getting your organisation ready for AI, responsible use among them. Free to sign and a credible first step in managing your AI risk.
Read the CharterIndependent assessment against the Institute of AI standard, so your risk controls are verified rather than merely declared.
Explore accreditationVendor-neutral guidance to design a risk framework that fits your size, sector and appetite, and to run it well.
Talk to an adviser
