AIGP AI Risk Matrix: Likelihood, Severity and Oversight

Quick answer

An AI risk matrix rates each identified risk by likelihood and severity (impact), often on 1–5 scales, and multiplies or plots them to set a risk level that drives treatment. Key points for the AIGP: assess inherent risk (before controls) and residual risk (after controls); severity weighs harm to individuals and society, not just the organisation; and the risk level should guide human oversight (in, on or out of the loop) and approval levels. The exam tests the reasoning, not arithmetic.

Risk matrices appear throughout AI governance: impact assessments, NIST AI RMF practices, ISO/IEC 42001 risk treatment and Singapore's human-involvement model. AIGP questions use them to test whether you can reason about risk. Here's how they work.

A basic AI risk matrix

Likelihood ↓ / Severity →	1 Minimal	2 Minor	3 Moderate	4 Major	5 Severe
5 Almost certain	5	10	15	20	25
4 Likely	4	8	12	16	20
3 Possible	3	6	9	12	15
2 Unlikely	2	4	6	8	10
1 Rare	1	2	3	4	5

Typical bands: 1–4 low, 5–9 medium, 10–16 high, 20–25 critical. Organisations set their own scales and bands.

What's different about AI risk

Severity includes harms to people and society: discrimination, loss of access to services, safety, and fundamental rights. Not just financial or reputational impact on the organisation.
Scale matters: an AI error repeated across thousands of decisions can turn a minor individual harm into a major aggregate one.
Reversibility: harms that can't easily be undone, such as a wrongful denial of benefits, rate higher on severity.
Likelihood changes over time: drift, new users and model updates mean you need to reassess.
Opacity: if failures are hard to detect, the effective likelihood of harm goes up.

Inherent vs residual risk

Inherent risk is the level before controls. Residual risk is the level after controls such as human review, testing, guardrails and monitoring. Governance decisions, such as whether to deploy and who approves, should be based on residual risk compared against your risk appetite. Document both.

From risk level to human oversight

Singapore's Model AI Governance Framework links the probability and severity of harm to the degree of human involvement:

Risk profile	Oversight model	Example
High severity, high probability	Human-in-the-loop (a human makes the decision)	Medical diagnosis support
Severe if it happens, low probability; or moderate	Human-over-the-loop (a human monitors and can intervene)	Fraud alerts with analyst review
Low severity, low probability	Human-out-of-the-loop	Product recommendations

See Singapore's framework vs NIST.

Worked example

A bank plans an AI model that recommends whether to approve small-business loans.

Risk: the model systematically under-approves applicants from certain postcodes, which act as a proxy for ethnicity.
Inherent rating: likelihood 3 (possible, because historical data reflects past lending bias) × severity 5 (discrimination; denial of access to credit; regulatory exposure) = 15, high.
Controls: remove or test proxy features, run fairness testing before release, require human review of declines, monitor approval rates by group monthly, and give applicants a route to explanation and appeal.
Residual rating: likelihood 2 × severity 4 = 8, medium.
Decision: approve deployment at the committee level, with quarterly reassessment. Also note that AI credit scoring is high-risk under the EU AI Act, with a FRIA required from 2 December 2027.

Exam traps: (1) rating severity only by cost to the organisation; (2) treating a high inherent score as an automatic ban instead of a trigger for controls; (3) forgetting to reassess after changes; (4) assuming human oversight lowers risk when reviewers just rubber-stamp outputs (automation bias).

Factors that raise AI likelihood and severity

Raises likelihood	Raises severity
Training data unrepresentative of the people the system serves	Decisions affecting access to jobs, credit, housing, healthcare or benefits
Little testing on real-world conditions or subgroups	Harm to vulnerable groups (children, patients, people in financial difficulty)
High autonomy and no human review	Harms that are hard to reverse or detect
Frequent model or data changes	Large numbers of people affected
Exposure to untrusted inputs (prompt injection, poisoned data)	Safety, legal or fundamental rights consequences
Third-party models with limited transparency	Reputational or regulatory exposure for the organisation

Linking risk scores to governance actions

Residual risk	Approval	Required controls	Review cycle
Low	Use-case owner, with a checklist	Basic documentation, standard monitoring	Annual
Medium	AI governance team	Impact assessment, testing evidence, user transparency	Six-monthly
High	AI governance committee	Full assessment, bias and robustness testing, human oversight design, incident plan	Quarterly or on change
Unacceptable	Rejected, or executive decision with strong justification	Redesign or stop	–

The AIGP rewards answers that tie the level of oversight and documentation to the level of risk. A one-size-fits-all process is usually the wrong answer.

Legal categories override scores

A matrix is an internal tool. It doesn't change legal classifications:

A prohibited practice under Article 5 of the EU AI Act is prohibited however low your score.
An Annex III system is high-risk under the AI Act unless the Article 6(3) filter applies, whatever your internal rating.
A DPIA is required where GDPR Article 35 says so, even if your AI score is medium.

Good practice is to apply legal classifications first, then use the matrix to set controls within each category.

Common mistakes with AI risk matrices

Scoring the model in the abstract rather than a specific use in a specific context.
Ignoring who is affected and how vulnerable they are.
Treating "we have a human in the loop" as an automatic low score without checking the oversight is meaningful.
Never re-scoring after model updates, new data or new uses.
Adding up scores across very different harms, which hides the worst one.

Sample AIGP-style question (original Archuz item)

An AI system has a low likelihood of error, but an error would cause severe, irreversible harm to individuals. Which oversight approach is most appropriate?

Human-out-of-the-loop, because errors are unlikely
Human-in-the-loop or human-over-the-loop with strong intervention capability, because severity is high
No oversight, but extensive logging
Annual audits only

Answer: B. Severity, especially irreversibility, justifies meaningful human involvement even when likelihood is low.

Risk reasoning, practised

The AIGP Complete Pack includes risk-assessment and oversight scenarios across its 300 BoK v2.1 questions.

Get the Complete Pack →

Frequently asked questions

How do you calculate an AI risk score?

Commonly, by multiplying likelihood by severity on defined scales (for example 1–5 each), then mapping the result to risk bands that drive treatment.

What is residual risk in AI governance?

The risk that remains after controls are applied. Deployment decisions should be based on residual risk against your risk appetite.

Will the AIGP ask me to calculate risk scores?

Questions focus on reasoning, such as what drives severity, how controls change residual risk, and which oversight fits, rather than arithmetic.

Does an internal AI risk score override the EU AI Act?

No. Legal classifications such as prohibited practices and Annex III high-risk uses apply regardless of internal scores. Use the matrix to set controls within those legal categories.

How often should AI risk scores be reviewed?

On a schedule matched to the risk, for example annually for low risk and quarterly for high risk, and whenever the model, data, use or affected population changes.

Sources

Last reviewed: October 2026.