"Be responsible with AI" is good advice, but not a method. Organisations need a repeatable way to find AI risks, judge how serious they are, decide what to do and prove they did it. Two frameworks have become the global reference points: the NIST AI Risk Management Framework from the United States and the international standard ISO/IEC 42001. This chapter shows how both work and gives you the practical tools behind them.
By the end of this chapter you can
- Define AI risk and name the main sources of AI-specific risk.
- Explain the four functions of the NIST AI RMF and its characteristics of trustworthy AI.
- Describe what ISO/IEC 42001 requires and how it differs from NIST and from law.
- Use an impact assessment, a risk register, a heat map and model documentation.
What makes AI risk different
A risk is the combination of how likely a harm is and how severe it would be. AI brings familiar risks (security, privacy, outages) plus new ones, and its harms can fall on three different groups at once: people (an applicant wrongly rejected, a patient misdiagnosed), organisations (fines, lost trust, failed projects) and society (misinformation, discrimination at scale, environmental cost).
Biased or poor data
Unrepresentative, outdated, mislabelled or unlawfully collected data.
Model behaviour
Errors, hallucinations, opacity, and brittleness when inputs change.
Misuse & over-reliance
Using a model outside its intended purpose, or people trusting it blindly.
Attacks
Prompt injection, data poisoning, model theft and extraction of training data.
Drift
The world changes and performance quietly decays (Chapter 4).
Third parties
Vendor models, open-source components and datasets you do not control.
The NIST AI Risk Management Framework
Published in January 2023 by the US National Institute of Standards and Technology, the AI RMF 1.0 is voluntary, free and sector-neutral, which is why organisations worldwide, including in India, use it as a practical playbook. It starts by describing what "trustworthy AI" means through seven characteristics:
| Characteristic | In plain words |
|---|---|
| Valid and reliable | It does what it claims, consistently, including on new data. |
| Safe | It does not endanger life, health, property or the environment. |
| Secure and resilient | It withstands attacks and recovers from failures. |
| Accountable and transparent | People can see what it is, how it is used and who answers for it. |
| Explainable and interpretable | Its outputs and how it works can be understood at the right level for the audience. |
| Privacy-enhanced | It protects personal data and people's autonomy. |
| Fair, with harmful bias managed | It avoids unjust differences in outcomes between groups. |
The core of the framework is four functions. Govern sits at the centre because it enables the other three:
Swipe sideways to see the full diagram.
In July 2024 NIST added a Generative AI Profile (NIST AI 600-1), which applies the framework to risks that are new or amplified by generative AI, such as confabulation (confidently stated false content), harms to information integrity, data privacy, intellectual property, harmful bias and information security.
ISO/IEC 42001: an AI management system you can certify
Published in December 2023, ISO/IEC 42001 is the first international standard for an AI management system (AIMS). If you know ISO 27001 (information security) or ISO 9001 (quality), it works the same way: it does not judge a single model, it checks that the organisation has a disciplined, continuously improving system for managing all of its AI. Organisations can be independently audited and certified against it, which is why it increasingly appears in procurement requirements.
It follows the familiar Plan–Do–Check–Act cycle and requires, among other things:
- Understanding the organisation's context and its role (developer, provider or user of AI).
- Leadership commitment, an AI policy and defined roles.
- AI risk assessment and treatment, plus AI system impact assessments on individuals and society.
- Controls across the lifecycle; its Annex A lists 38 reference controls, from data quality and documentation to responsible use and third-party relationships.
- Internal audit, management review and continual improvement.
It sits within a family of related standards, including ISO/IEC 23894 (guidance on AI risk management) and ISO/IEC 42005 (guidance on AI system impact assessment).
Frameworks vs. law: how they fit together
| NIST AI RMF | ISO/IEC 42001 | EU AI Act | |
|---|---|---|---|
| Type | Voluntary framework | International management-system standard | Binding law |
| Focus | Practical risk-management activities | Organisation-wide AI management system | Obligations based on a system's risk category |
| Certifiable? | No | Yes, by accredited bodies | High-risk systems need conformity assessment |
| Best used for | "How do we actually do this?" | "Can we prove we do it systematically?" | "What must we legally do?" |
They are complementary: many organisations use NIST for day-to-day practice, build an ISO/IEC 42001 management system to prove it, and map both to the legal obligations of each market they serve.
Practical tool 1: the AI impact assessment
Before building or buying a significant AI system, teams complete an impact assessment. It forces the right questions early, while changes are cheap:
- Purpose: what decision does it support, and is AI the right tool?
- People: who is affected, including people who never use the system directly?
- Harms: what could go wrong for them: wrong decisions, exclusion, discrimination, privacy loss, safety?
- Data: is it lawful, representative and minimised?
- Oversight: where can a human review, override or stop it, and how can people appeal?
- Decision: proceed, proceed with conditions, or stop.
Practical tool 2: the risk register and heat map
Each identified risk goes into a risk register with a likelihood score, an impact score, the controls applied and an owner. Plotting them on a heat map shows at a glance where attention is needed.
Swipe sideways to see the full diagram.
For each risk you choose a treatment: avoid it (do not build that feature), reduce it (add controls), transfer it (contracts, insurance) or accept it (if it is within the organisation's risk appetite). Controls can be preventive (stop the problem: data checks, access limits), detective (spot it: monitoring, audits) or corrective (fix it: rollback, appeal process).
Practical tool 3: documentation that travels with the model
Residual risk and risk appetite: the conversation leaders must haveGo deeper
No control removes risk completely. What remains after controls is residual risk. Leadership must decide, in advance and in writing, how much residual risk is acceptable for different kinds of systems: their risk appetite. A hospital may accept almost no residual risk of missed diagnoses but tolerate some administrative errors; a marketing team may accept more. Without an explicit appetite, every launch decision becomes a debate, or nobody makes it at all.
Key takeaways
- AI risk = likelihood × severity of harm to people, organisations and society, from data, models, use, attacks, drift and third parties.
- NIST AI RMF: seven trustworthy characteristics and four functions (Govern, Map, Measure, Manage); the GenAI Profile extends it.
- ISO/IEC 42001 is a certifiable AI management system using Plan–Do–Check–Act with 38 Annex A controls.
- Use impact assessments, a risk register with heat map, clear treatments and model cards to make governance real and auditable.
Sources and further reading: NIST AI Risk Management Framework 1.0 (NIST AI 100-1, January 2023) and Generative AI Profile (NIST AI 600-1, July 2024); ISO/IEC 42001:2023, ISO/IEC 23894:2023 and ISO/IEC 42005:2025; Mitchell et al., "Model Cards for Model Reporting" (2019); Gebru et al., "Datasheets for Datasets" (2021).