AI Risk Assessment: How to Do It Step by Step, Per Use Case, With a Worked Example
November 12, 2026 · 8 min read · Tim Etzal Group

An AI risk assessment is done per use case, not once for the whole company: list every place AI is used, rate how likely and how severe the harm would be if it went wrong, set controls and a level of human oversight that match the rating, test before go-live, and monitor afterwards. Singapore's Model AI Governance Framework and its 2026 framework for agentic AI rate risk this way, and AI Verify's published crosswalks let the same worksheet be read against the NIST AI Risk Management Framework and ISO/IEC 42001.
Updated: 12 November 2026
What an AI risk assessment is
An AI risk assessment is a documented judgement, made for each AI use case, of what could go wrong, how likely and how harmful that would be, and which controls and level of human oversight bring the risk to a level the organisation accepts. AI risk management is the ongoing cycle around it: assess, control, test, monitor and re-assess. Singapore's agentic AI framework puts it simply: risk is a function of likelihood and impact (section 2.1.1).
When an AI risk assessment should be conducted
- Before you build or buy: some use cases fail the assessment and are better served by a deterministic, rule-based workflow.
- Before go-live: with the controls and test results in hand.
- After a significant change: a new model, new data, wider tool access or more autonomy.
- On a fixed review cycle: the Model AI Governance Framework points to a "periodically reviewed risk impact assessment" (paragraph 3.13).
How to do an AI risk assessment in six steps
Step 1: Build the use-case inventory
One row per use case, not per tool, because the same chatbot can be harmless in one workflow and risky in another. Record the business purpose, the owner, the people affected, the data used (and whether it is personal data), the vendor or model, what the AI can read and write, and whether it only advises or also acts.
Step 2: Tier the risk
The Model AI Governance Framework rates each use case on two axes, the probability of harm and the severity of harm, and notes that what counts as harm depends on context (paragraphs 3.15 and 3.16). For AI that takes actions, the agentic framework adds factors that raise impact or likelihood:
| Factor | Lower risk | Higher risk |
|---|---|---|
| Tolerance for error in the domain | Summarising internal meetings | Executing financial transactions |
| Access to sensitive data | Public information only | Personal or confidential data |
| Scope of actions | Read only, a few fixed tools | Can write to systems, many tools |
| Reversibility | Rescheduling a meeting | Sending external emails, entering a contract |
| Autonomy | Follows a set procedure | Decides every step itself |
| Exposure to external systems | Internal knowledge base | Open web with untrusted data |
| Who operates the agent | Built in-house, full visibility | Third-party agent, limited visibility |
A three-tier scale (low, medium, high) is usually enough. Write down the reason for each tier; the reason is what an auditor or a new manager will read.
Step 3: Choose controls and build the control matrix
For each risk, name the control, its owner and the evidence that it works. Typical controls: data minimisation and pseudonymisation, least-privilege access to tools and systems, input and output filters, logging of every AI action, and limits on what an agent may spend, send or change. For security controls, the Cyber Security Agency of Singapore published Guidelines and a Companion Guide on Securing AI Systems on 15 October 2024.
Step 4: Set the level of human oversight
The Model AI Governance Framework defines three approaches (paragraph 3.14):
- Human-in-the-loop: the AI only recommends; nothing happens without a person's affirmative action.
- Human-over-the-loop: a person monitors and can take over when the AI meets unexpected events.
- Human-out-of-the-loop: the AI acts with no human override.
The higher the tier, the closer to human-in-the-loop. For agents, the 2026 framework asks for significant checkpoints that require human approval, such as high-stakes or irreversible actions, and for regular audits that oversight still works, because people who approve many correct outputs tend to stop checking them (automation bias).
Step 5: Test before go-live
Test against real cases from your own operation, not vendor demos. For traditional machine learning models, the AI Verify testing framework pairs 11 principles with process checks, and its toolkit runs fairness, explainability and robustness tests. For applications built on large language models, IMDA's Starter Kit for Testing LLM-based Applications covers five risks: hallucination and inaccuracy, bias in decision-making, undesirable content, data leakage, and vulnerability to adversarial prompts. An Agentic Testing Starter Kit adds tests of tool use, handoffs and human approvals. Both are listed on IMDA's artificial intelligence page.
Step 6: Monitor and re-assess
Keep logs, an incident record and a review trigger list. Track a few signals per use case: error rate on a sample, how often people override the AI, complaints, and data access outside the expected pattern. Feed what you learn back into Step 2.
A worked example
For example, take a distributor with about 200 staff and two AI use cases. The numbers below are illustrative.
| Item | Email reply drafts | Credit limit agent |
|---|---|---|
| What the AI does | Drafts replies to order status emails | Proposes credit limits for new trade customers |
| Data | Customer names, order numbers | Financial details, some of sole proprietors |
| Impact if wrong | Low: a person reads every draft | High: bad debt, or a fair applicant refused |
| Likelihood factors | Read-only order data, simple task | Mixed data, judgement, write access requested |
| Tier | Low | High |
| Human oversight | Human-in-the-loop by design | Human-in-the-loop for every limit |
| Key controls | No payment data in prompts, logging | No write permission, written reasons, PDPA notice |
| Test before go-live | 100 past emails reviewed | 200 past applications compared with actual decisions |
| Monitor | Monthly sample of sent replies | Override rate and bad debt by tier, quarterly review |
The assessment changed the second design: the agent lost the write access it asked for and became a recommender. A useful AI risk assessment changes designs, not only paperwork. The PDPA side of the credit case is covered in our guide to PDPA and AI in Singapore.
NIST AI RMF vs ISO/IEC 42001 vs Singapore's frameworks
| Framework | What it is | Status | Use it for |
|---|---|---|---|
| NIST AI RMF 1.0 | US framework released 26 January 2023, four functions: Govern, Map, Measure, Manage; Generative AI Profile 26 July 2024 | Voluntary; NIST says it is being revised | A detailed risk vocabulary and playbook |
| ISO/IEC 42001:2023 | International AI management system standard; Annex A controls include AI system impact assessment | Voluntary standard with audit schemes | A management system that customers or auditors can check |
| Singapore Model Framework 2020 and agentic framework 2026 | Probability and severity matrix, three levels of human involvement, agentic risk factors | Voluntary guidance | Practical tiering and oversight decisions |
| AI Verify testing framework | 11 principles with process checks, evidence and test toolkits | Voluntary self-assessment | Evidence mapped to NIST AI RMF and ISO/IEC 42001 |
The NIST AI RMF is described by NIST as "intended for voluntary use". AI Verify publishes crosswalks to both the NIST AI RMF and ISO/IEC 42001, including an ISO/IEC 42001 crosswalk that maps the standard's controls to AI Verify process checks. A mid-size company in Singapore can run the assessment on the Singapore frameworks and use the crosswalks when a customer asks about NIST or ISO. Our overview of the AI governance framework Singapore publishes explains how the documents fit together.
Key takeaways
- Assess per use case, rating probability and severity of harm, plus access, reversibility and autonomy when the AI acts.
- Match human oversight to the tier: human-in-the-loop for high-risk decisions and irreversible actions.
- Write a control matrix: risk, control, owner, evidence.
- Test on your own real cases before go-live, then monitor overrides, errors and complaints.
- Singapore's frameworks are enough to run the assessment; AI Verify crosswalks connect the result to NIST AI RMF and ISO/IEC 42001.
AI risk assessment with Etzal Group
Per-use-case risk tiering, a control matrix and the level of human oversight are the core of the written output of Etzal Group's AI Consulting package. The worksheet aligns to Singapore's Model Framework and can be read against the NIST AI RMF. It is not a certification and not legal advice. If you are earlier in the journey, start with what an AI readiness assessment should deliver or build vs buy AI agents, or book a consultation.
Frequently asked questions
How do you conduct an AI risk assessment?
Work per use case: build an inventory, rate the probability and severity of harm, choose controls, set the level of human oversight, test on real cases before go-live, and monitor and re-assess afterwards. Write down the reason for each rating.
What is an AI risk assessment?
It is a documented judgement, made for each AI use case, of what could go wrong, how likely and harmful that would be, and which controls and human oversight bring the risk to an acceptable level.
When should an AI risk assessment be conducted?
Before you build or buy, again before go-live, after any significant change such as a new model or more autonomy, and on a fixed review cycle. Singapore's Model AI Governance Framework points to a periodically reviewed risk impact assessment.
What is the NIST AI Risk Management Framework?
It is a voluntary framework that NIST released on 26 January 2023, organised into four functions: Govern, Map, Measure and Manage. A companion Playbook and a Generative AI Profile support it, and NIST says AI RMF 1.0 is being revised.
What is AI model risk management?
It is the part of AI risk management that deals with the model itself: validating it before use, monitoring its performance and drift, and controlling changes. It sits inside the wider assessment, which also covers data, people and process.
Is ISO/IEC 42001 the same as an AI risk assessment?
No. ISO/IEC 42001:2023 is an AI management system standard; an AI risk and impact assessment is one of the processes it expects an organisation to run, alongside policies, roles and continual improvement.
Related service
WhatsApp AI Chatbot
We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.
