AI Risk Assessment: A Step-by-Step Guide

How does an AI risk assessment work?#

An AI risk assessment works by pairing two questions for every risk: how likely is it, and how bad would it be? The NIST AI Risk Management Framework defines risk as a combination of an event's probability and the size of its consequences. You rate each risk this way, decide what is acceptable, add controls and test them.

Three published approaches fit together well:

NIST says its framework is intended for voluntary use and that, once governance is in place, most users start with the Map function before moving to Measure or Manage. It does not prescribe risk tolerance: deciding how much risk is acceptable is up to each organisation, and NIST notes it can be shaped by legal and regulatory requirements.

Step by step: how do you assess AI risk?#

Work through nine steps in order: describe the use, identify relevant risks, rate them, set thresholds, choose controls, test, record and assign owners, then decide and review. The steps follow IMDA's Identify, Test and Assess approach inside NIST's Govern, Map, Measure and Manage cycle, and suit a small team as well as a large one.

1. Describe the use. Write down what the system does, who uses it, who it affects and what data it touches. This is NIST's Map function: establishing the context. If you have no list of AI uses yet, start one.
2. Identify the relevant risks. Start with IMDA's five common risks for apps built on a large language model (LLM): hallucination and inaccuracy, bias in decision making, undesirable content, data leakage and vulnerability to adversarial prompts. Add privacy risks involving personal data and any risks specific to your sector.
3. Rate severity and likelihood. Use the factors in the next section. IMDA suggests treating a risk as material if problematic output has reputational or legal implications, or affects safety or life opportunities.
4. Set thresholds before testing. Decide the pass mark for each material risk. IMDA's Starter Kit says setting it before testing prevents moving goalposts, and suggests "hard-gates" for outputs linked to severe or irreversible harm.
5. Choose controls. Match each risk to a safeguard, such as human review, limited access, input rules or a narrower scope. For an AI agent, IMDA suggests limiting access to tools and external systems.
6. Test. Run tests against your thresholds. Project Moonshot, from the AI Verify Foundation, implements the benchmarks recommended in the Starter Kit. See also evaluation for the basics.
7. Record and assign owners. Enter each risk in the register with a named owner and review date.
8. Decide. Go ahead, go ahead with extra controls, or stop. Record who decided and why.
9. Monitor and review. IMDA notes that testing is not a one-off exercise. Revisit the assessment when the system, data or use changes.

How do you rate severity and likelihood?#

Rate severity by asking how much harm a failure could cause and to whom; rate likelihood by asking how often a failure could slip through. IMDA's Starter Kit lists practical factors for both, and its framework for agentic AI adds factors for systems that take actions. Use high, medium and low ratings unless you need a finer scale.

The NIST framework lists characteristics of trustworthy AI that make a useful cross-check: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed.

Risk register template#

A risk register is the written record of your assessment: one row per risk, with its rating, controls, test, owner and review date. Copy the template below into a spreadsheet and add a row for every material risk. The two filled rows are illustrative examples only, showing the level of detail to aim for.

Agree what "high", "medium" and "low" mean in your context before you start, so different people rate risks the same way.

What might this look like in a Singapore workplace?#

Picture a small recruitment agency in Paya Lebar that wants to use a generative AI tool to summarise candidate CVs for consultants. The tool would save reading time, but it touches personal data and could shape decisions about people's jobs. That combination makes a risk assessment essential before anyone uses it with real CVs.

This is an illustrative scenario, not a real case. A sensible approach would be to:

• describe the use: the tool drafts summaries and a consultant reads every CV before shortlisting
• identify the material risks: inaccurate summaries, bias against certain groups and leakage of candidate data
• rate bias as high severity, because IMDA's guidance treats effects on life opportunities as material
• set thresholds before testing, using sample CVs made up for the purpose rather than real ones
• add controls: no automated ranking, input rules in the AI use policy and access limited to consultants
• record each risk with an owner and review the assessment if the tool adds features that act on its own

PDPC's advisory guidelines on generative AI describe organisations that deploy these systems as bearing primary responsibility for meeting their PDPA obligations, so the agency's data rules belong in the register too.

Common mistakes#

Most weak AI risk assessments share the same faults. They list generic risks without linking them to the actual use, set no pass mark before testing, name no owner, or are filed once and never revisited. A useful assessment is specific to one use, honest about likelihood and severity, and updated when anything changes.

• Generic risk lists. "AI may be biased" helps nobody. Say where, how and who would be affected.
• Moving the goalposts. Setting the pass mark after seeing results undermines the test.
• Forgetting people who are not users. IMDA's guidance notes that impacted parties are not always the app's users.
• No owner. A risk without a named owner is rarely managed.
• Assessing once. NIST says risk management should be continuous across the AI lifecycle.
• Skipping the record. If the decision and its reasons are not written down, nobody can review them later.

What to learn next#

A risk assessment tells you which uses need the strongest controls. The next steps are to turn those controls into staff rules, document your practice against a recognised framework and build hands-on confidence in testing. The guides below cover each one, and the course lets you practise on a simple AI app.

• Turn your controls into rules with how to write an AI use policy, which includes a template.
• Document your practice with AI Verify, Singapore's testing framework.
• Explore more governance guides in this category or return to the AI learning pathway.
• Practise building and testing a simple AI app on the GenAI Builder Sprint course.

Sources#

1. Artificial Intelligence Risk Management Framework (AI RMF 1.0), National Institute of Standards and Technology (NIST), accessed 8 November 2026
2. AI Risk Management Framework, National Institute of Standards and Technology (NIST), accessed 8 November 2026
3. Starter Kit for Testing LLM-Based Applications for Safety and Reliability, Infocomm Media Development Authority (IMDA), accessed 8 November 2026
4. Model AI Governance Framework for Agentic AI, Infocomm Media Development Authority (IMDA), accessed 8 November 2026
5. What is AI Verify Testing Framework, AI Verify Foundation, accessed 8 November 2026
6. Project Moonshot, AI Verify Foundation, accessed 8 November 2026
7. Advisory Guidelines on Use of Personal Data in Generative AI, Personal Data Protection Commission (PDPC), accessed 8 November 2026

Frequently asked questions#

What is an AI risk assessment?+

A structured check of what could go wrong when you use an AI system, how serious and how likely each problem is, and which controls will keep it within limits you accept. It ends with a written record of risks, controls, owners and review dates.

How does NIST define risk?+

In the NIST AI Risk Management Framework, risk is a composite measure of an event's probability of occurring and the magnitude or degree of its consequences. NIST's framework is intended for voluntary use and does not prescribe how much risk an organisation should tolerate.

What are the five key risks in IMDA's Starter Kit?+

Hallucination and inaccuracy, bias in decision making, undesirable content, data leakage, and vulnerability to adversarial prompts. IMDA describes these as the five most common risks enterprises face from LLM apps.

How do you rate the severity of an AI risk?+

IMDA's Starter Kit suggests considering whether the app is in a high-stakes domain, who is affected (especially vulnerable groups), whether it is internal or external facing, and whether it is connected to other systems that could be affected by a failure.

What is a threshold in AI testing?+

IMDA's Starter Kit describes the threshold as the passing mark for baseline safety and reliability. Each organisation sets its own threshold for each app before testing, which prevents moving goalposts after the results come in.

What should a risk register include?+

For each risk: the AI use, what could go wrong, who could be harmed, severity, likelihood, an overall rating, the controls, the threshold or test, a named owner, a review date and the current status.

How often should an AI risk assessment be repeated?+

Whenever the system, its data or its use changes, and on a regular review date. NIST says risk management should be continuous and run throughout the AI system lifecycle, and IMDA notes that testing is not a one-off exercise.