How to conduct a risk assessment with AI

Set scope and decision context

Identify the system, process, information, people, objective, assessment period, accountable owner, and decision the assessment supports. Define impact dimensions and rating language before reviewing individual risks. Record excluded systems and dependencies.

NIST SP 800-30 Rev. 1 provides guidance for conducting information-system risk assessments and covers threats, vulnerabilities, likelihood, impact, and controls. NIST’s AI Risk Management Framework provides a broader voluntary framework for managing AI risk. Apply only the guidance relevant to the organization’s context and authority.

Build risk scenarios, not labels

Write each risk as: “Because [threat] could exploit [vulnerability or condition], [event] may occur, affecting [asset or objective] and causing [consequence].” Avoid single nouns such as “privacy” or “cybersecurity.”

Worked synthetic assessment

Suppose a team is considering an external AI tool for drafting summaries from synthetic assessment materials.

Scenario: Because workspace users may paste material outside the approved synthetic dataset into a tool whose processing route has not been verified, information could be handled outside the approved environment, causing unauthorized disclosure and loss of evidence control.

| Element | Current evidence | Confidence | | --- | --- | --- | | Asset | Assessment materials and workflow records | High | | Threat/event | Unapproved transfer to external service | Medium | | Vulnerability | Tool access available without enforced data boundary | Medium | | Existing control | Written acceptable-use instruction | High for design, unknown for operation | | Impact | Confidentiality and operational response | Context dependent | | Likelihood | Insufficient event and control data | Low confidence |

This table does not assign a numeric probability. It shows what is and is not known.

Separate three risk views

Inherent risk asks about the scenario before considering controls. Control assessment examines whether controls are appropriately designed, implemented, and operating with evidence. Residual risk is the remaining exposure after evidenced controls, including uncertainty.

A policy document supports control design, not operation. Access logs, configuration tests, sampled behavior, and incident exercises may support other claims, depending on quality and scope.

Use calibrated ratings

Define likelihood and impact bands with observable anchors. If evidence cannot distinguish “possible” from “likely,” retain a range or mark confidence low. Do not multiply ordinal labels and present the result as an exact quantity.

Ask AI to find missing assets, alternate threat paths, common-cause failures, controls with no operating evidence, and assumptions that dominate the rating. A human owner verifies relevance and chooses treatment.

Design treatments and verification

For each proposed action record the risk element it changes, owner, due date, expected artifact, test, dependencies, and residual decision. Possible synthetic actions include restricting tool access, enforcing an approved data route, training users with realistic examples, monitoring transfers, and rehearsing response. Their effectiveness remains unverified until tested.

Produce the risk record

Include scope, scenario, evidence table, inherent assessment, control assessment, residual range, uncertainty, treatment options, acceptance authority, review trigger, and source links. Preserve dissent and overdue evidence.

Verification checklist

Confirm assets and objectives; make threat and vulnerability explicit; verify control evidence; distinguish design from operation; calibrate ratings; retain uncertainty; test common-cause failure; assign treatment owners; identify acceptance authority; and schedule reassessment after material change.

Sources and claim map

The tool, scenario, controls, and ratings are synthetic. This educational method does not establish legal conformity, security adequacy, or an assured risk reduction.

• NIST, SP 800-30 Rev. 1: Guide for Conducting Risk Assessments: supports the threat, vulnerability, likelihood, impact, control, and risk-assessment structure for its intended information-system context.
• NIST, AI Risk Management Framework: provides the official voluntary framework for managing risks associated with AI systems.

Sources: [1] [2]