Short answer
AI risk management is how an organisation identifies, assesses and controls the risks that come from building, buying or using AI, such as wrong answers, bias, data leaks, security attacks and breaches of law. Most organisations base it on the NIST AI Risk Management Framework or ISO/IEC 42001, keep an inventory and risk register of their AI systems, and check their exposure under laws such as the EU AI Act.
AI has moved from pilot projects into everyday work: drafting reports, screening documents, answering clients, scoring risks. The risks have moved with it. An AI system can be confidently wrong, treat people unfairly, leak confidential data or be manipulated by a well crafted prompt. This guide explains what AI risk management means, the risks that matter most, the frameworks and laws that cover them, and how to put a working process in place.
Table of Contents
What Is AI Risk Management?
AI risk management applies the ordinary risk management cycle (identify, assess, treat, monitor, report) to AI systems. It covers AI you build yourself, AI built into the software you buy, and general AI tools your people use at work.
It has two jobs. The first is to stop AI causing harm: to the organisation, its clients, its employees and the public. The second is to let the organisation use AI with confidence, because the risks are known, owned and controlled rather than ignored or feared.
Why AI Risk Is Different
- It is probabilistic: the same question can produce different answers, and a fluent answer is not necessarily a correct one
- It is hard to explain: many models cannot show exactly why they produced an output
- It changes: models are updated by vendors, data drifts, and performance can shift without anyone touching your systems
- It sits with third parties: most organisations use models built by someone else, so vendor risk and AI risk overlap
- It spreads quickly: staff can start using a new AI tool in minutes, often before anyone has assessed it
The Main AI Risks
| Risk | What goes wrong | Typical controls |
|---|---|---|
| Accuracy | Wrong or invented answers presented as fact | Human review before use, source citations, testing against known answers |
| Bias and fairness | Some groups are treated worse, for example in hiring or lending | Bias testing before launch and in use, diverse test data, human decision makers |
| Privacy | Personal data used without a lawful basis, or exposed in outputs | Data minimisation, approved tools only, privacy impact assessments |
| Confidentiality | Client or company information pasted into tools that keep or train on it | Enterprise agreements that exclude training, clear rules on what may be shared |
| Security | Prompt injection, data poisoning, model theft, insecure integrations | Security testing, input and output filtering, least privilege access |
| Transparency | People do not know they are dealing with AI, or cannot get an explanation | Disclosure to users, documentation, explanations for significant decisions |
| Over reliance | People stop checking and accept outputs without thought | Training, review steps that require sign off, sampling and quality checks |
| Third party | A vendor changes or withdraws a model, or suffers an incident | Vendor due diligence, contract terms, fallback plans |
| Legal and regulatory | Breaching AI, data protection, consumer or employment law | Classifying each use case against the rules that apply, legal review |
| Intellectual property | Outputs that copy protected material, or unclear ownership of what AI produces | Approved tools with IP terms, review of published outputs |
AI Risk Management Frameworks
| Framework or law | Publisher | Type | Use it for |
|---|---|---|---|
| NIST AI RMF 1.0 (January 2023) | NIST (US) | Voluntary framework | A practical structure for AI risk: Govern, Map, Measure, Manage. A Generative AI Profile was added in July 2024 |
| ISO/IEC 42001:2023 | ISO / IEC | Certifiable management system | Showing clients and regulators that AI governance has been independently audited |
| ISO/IEC 23894:2023 | ISO / IEC | Guidance standard | Applying ISO 31000 style risk management to AI |
| ISO/IEC 42005:2025 | ISO / IEC | Guidance standard | Assessing the impact of an AI system on people and society |
| EU AI Act (Regulation (EU) 2024/1689) | European Union | Law | Legal obligations based on how risky an AI use is, for anyone placing AI on the EU market or whose AI output is used in the EU |
| OECD AI Principles (2019, updated May 2024) | OECD | Non binding principles | The shared values many national AI policies are built on |
| COSO, Achieving Effective Internal Control Over Generative AI (February 2026) | COSO | Guidance | Applying the COSO internal control framework to generative AI, with templates |
These fit together. A common setup is NIST AI RMF or ISO/IEC 23894 for the method, ISO/IEC 42001 if you need certification, and the EU AI Act and local law for the legal obligations. AI risk then sits in your main register alongside every other risk; see our risk management frameworks comparison for how the general frameworks fit.
The NIST AI RMF in Practice
| Function | What it means | In practice |
|---|---|---|
| Govern | Culture, policies, roles and accountability for AI | An AI policy, a named owner for each system, board oversight |
| Map | Understand the context and identify risks | An inventory of AI systems: what each does, who uses it, what data it touches, who could be affected |
| Measure | Analyse, assess and track the risks | Testing for accuracy, bias and security; metrics tracked over time |
| Manage | Prioritise and act on the risks | Controls, human oversight, incident response, decisions to retire a system |
The EU AI Act in Brief
The EU AI Act sorts AI uses by risk:
- Unacceptable risk: banned practices, such as social scoring and certain manipulative techniques, prohibited since 2 February 2025
- High risk: uses listed in Annex III, such as AI in recruitment, credit scoring, education and access to essential services, plus AI in regulated products. These carry the heaviest duties: risk management, data governance, documentation, human oversight, accuracy and security
- Transparency duties: for example, telling people when they are talking to a chatbot or seeing AI generated content
- Minimal risk: most other uses, with no specific obligations
Key dates: the Act entered into force on 1 August 2024; rules for general purpose AI models have applied since 2 August 2025. After the Digital Omnibus amendment (Regulation (EU) 2026/1744), obligations for high risk Annex III systems apply from 2 December 2027, and for AI in regulated products from 2 August 2028. Fines for prohibited practices reach EUR 35 million or 7% of worldwide annual turnover, whichever is higher.
It Can Apply Outside the EU
The Act covers providers placing AI on the EU market and, in some cases, providers and deployers outside the EU whose AI output is used in the EU. If you have EU clients or users, check your exposure even if you have no office there.
How to Build an AI Risk Management Process
- Write a short AI policy. Which tools are approved, what data may and may not be shared, and who approves new uses.
- Build an inventory. List every AI system and tool in use, including AI features inside software you already buy. Ask teams directly; usage is often wider than expected.
- Classify each use. Rate how much harm it could do (an internal drafting aid is not the same as a hiring filter) and which laws apply.
- Assess the risks. For each system, work through the risks in the table above and score them with your normal likelihood and impact scale.
- Apply controls. Human review, testing, access limits, vendor terms, disclosure. Record them in your controls register.
- Decide where humans stay in charge. Make clear which outputs must be checked and signed off by a person before they are used.
- Monitor. Track accuracy, complaints, incidents and vendor changes, and review high risk uses more often.
- Plan for incidents. Know how you will pause a system, investigate and tell the people affected.
- Train people. Everyone using AI should know its limits and your rules.
The AI Risk Register
An AI risk register is a normal risk register with a few extra fields: the AI system, its purpose, its risk class, the data it uses, the vendor and the human oversight in place. For example:
| AI system | Risk | Rating | Key control | Owner |
|---|---|---|---|---|
| Report drafting assistant | Inaccurate statements reach a client report | High | Reviewer signs off every report before release | Engagement lead |
| CV screening tool | Unfair treatment of some applicant groups | High | Bias testing each quarter; human makes every decision | Head of HR |
| Public chatbot | Prompt injection exposes internal data | Medium | No access to internal systems; output filtering; security test | IT security |
| General AI tools used by staff | Client data shared with unapproved tools | Medium | Approved tool list, enterprise agreements, training | Compliance |
Who Does What
- Board: sets appetite for AI risk and oversees the most significant uses
- System owners: own each AI system and its risks day to day
- Risk and compliance: set the method, classify uses against the law, challenge owners
- IT and security: assess vendors, test security, control access
- Internal audit: gives independent assurance that AI risks are managed
Common Mistakes
- No inventory: you cannot manage AI you do not know is in use
- Treating all AI the same: heavy controls on a spelling aid, light controls on a hiring tool
- A separate AI register nobody connects to the main one: AI risks belong in the same picture as every other risk
- Assuming the vendor has it covered: you stay accountable for how you use a third party's model
- Review on paper only: "a human checks it" means little unless the check is defined, recorded and sampled
AI Risk Management in 2027
- EU deadlines are close: high risk Annex III obligations apply from 2 December 2027, so inventories and classifications need to be done in 2027
- Internal control over generative AI: COSO's February 2026 guidance gives auditors and finance teams a structure and templates for AI controls
- Impact assessments are standardised: ISO/IEC 42005:2025 gives a common way to assess effects on people and society
- Certification is growing: more clients ask suppliers for evidence of AI governance, and ISO/IEC 42001 is the certificate available
Summary
- AI risk management applies the normal risk cycle to AI you build, buy or use
- The main risks are accuracy, bias, privacy, confidentiality, security, transparency, over reliance, third party and legal risk
- NIST AI RMF gives the method, ISO/IEC 42001 the certificate, ISO/IEC 23894 the ISO style guidance
- The EU AI Act sorts uses by risk; high risk Annex III obligations apply from 2 December 2027
- Start with a policy and an inventory, then classify, assess, control and monitor
- Keep AI risks in your main risk register, with a named owner and human oversight defined
Frequently Asked Questions
What is AI risk management?
It is the process of identifying, assessing and controlling the risks that come from building, buying or using AI, such as wrong answers, bias, data leaks, security attacks and breaches of law, so the organisation can use AI with confidence.
What is the NIST AI Risk Management Framework?
A free, voluntary framework from the US National Institute of Standards and Technology, released in January 2023. It organises AI risk management into four functions: Govern, Map, Measure and Manage. NIST added a Generative AI Profile in July 2024.
What is the difference between NIST AI RMF and ISO/IEC 42001?
The NIST AI RMF is voluntary guidance on what to think about. ISO/IEC 42001 is a management system standard with auditable requirements, so an organisation can be certified against it. Many organisations use the NIST framework to do the work and ISO/IEC 42001 to prove it.
Does the EU AI Act apply to companies outside the EU?
It can. It covers providers placing AI systems on the EU market wherever they are based, and in some cases providers and deployers outside the EU whose AI output is used in the EU. Organisations with EU clients or users should check their position.
What should an AI risk register include?
The usual risk register fields (risk, cause, rating, controls, owner, actions) plus the AI system, its purpose, its risk class under the rules that apply, the data it uses, the vendor and the human oversight in place.
How do I start with AI risk management in a small organisation?
Write a one page AI policy, list the AI tools people use, decide which are approved and what data may go into them, and add the main AI risks to your existing risk register with an owner. Then review it every quarter.
- The NIST AI Risk Management Framework, including the Generative AI Profile
- ISO/IEC 42001:2023, ISO/IEC 23894:2023 and ISO/IEC 42005:2025, from ISO
- The EU AI Act, Regulation (EU) 2024/1689, and its amendment, Regulation (EU) 2026/1744, on EUR-Lex
- The OECD AI Principles
- COSO's Achieving Effective Internal Control Over Generative AI
- The MIT AI Risk Repository, a public database of AI risks
For educational purposes only, not legal, audit or professional advice. Standards and laws change, so check the sources above before acting. Last reviewed: November 2026.