Risk Matrix

A risk matrix is a decision-support tool that compares the likelihood of a harmful event with the severity of its consequences. A common implementation uses a three-by-three or five-by-five grid, with each cell mapped to a risk category and required action. For example, an unlikely disclosure of highly sensitive health information may require stronger controls than a frequent formatting error in an internal document assistant. The matrix makes these judgments explicit, reviewable, and consistent across teams.

Start with a specific risk scenario rather than a broad label such as “AI risk.” Document the system, intended use, affected people, triggering conditions, potential harm, and assessment period. For a cloud-hosted hiring assistant, a scenario might be that biased ranking recommendations systematically disadvantage a protected group. Assess the likelihood using relevant evaluation results, deployment conditions, and exposure, then assess consequences across legal, financial, operational, privacy, safety, and individual-rights dimensions. Define how multiple consequence dimensions determine the final impact rating so that severe harms are not hidden by averaging.

Organizations should define the matrix scales before assessing individual projects. “Likely” should have a shared interpretation, such as an expected frequency within a stated period, rather than depending on each assessor's intuition. Impact thresholds should reflect organizational risk appetite and the people affected. Some organizations multiply numerical likelihood and impact ratings, but these numbers often represent ordinal categories rather than measured quantities. A score of twelve is therefore not necessarily twice as risky as a score of six, and equal scores can conceal very different risk scenarios.

Apply the matrix throughout the AI lifecycle. At intake, classify the use case and identify legal restrictions before committing to development. During data preparation and model evaluation, update assessments using evidence about data provenance, privacy, bias, security, and output reliability. Before deployment, require controls and approval appropriate to the risk category. In production, reassess after incidents, model changes, new data sources, increased usage, or changes in the population affected. Retirement also requires decisions about retained data, access removal, and evidence preservation.

Distinguish inherent risk, assessed before the proposed controls, from residual risk remaining after controls are implemented and their effectiveness is evaluated. For a customer-support assistant, retrieval restrictions, sensitive-data filtering, adversarial testing, and human escalation may reduce particular risks. Record the evidence supporting any reduction rather than assuming that enabling a control guarantees effectiveness. Assign each risk an owner, treatment deadline, review trigger, and an authorized decision-maker for accepting residual exposure.

A matrix supports risk treatment through avoidance, mitigation, transfer or sharing, and acceptance. Higher-risk AI applications may require independent validation, legal review, human oversight, and formal release gates. Lower-risk applications can follow a lighter process. Cloud infrastructure controls, model evaluations, application telemetry, and audit logs provide evidence, but no single service determines the appropriate governance classification.

A common misconception is that an internal matrix replaces a regulatory AI classification framework. Legal classifications can depend on intended use, jurisdiction, and prohibited practices rather than a likelihood-times-impact score. A prohibited use cannot become acceptable merely because its estimated likelihood of harm is low. Frameworks such as the NIST AI Risk Management Framework can structure risk management, but do not prescribe a universal matrix or themselves establish regulatory compliance. Use explicit escalation rules for severe harms, document uncertainty, and supplement the matrix with specialist assessments where simple categories would obscure important consequences.