AI Audit: A Practical Guide for Business and Tech Leaders

An AI audit that examines only the algorithm can miss important risks. Data, integrations, human oversight, and day-to-day operations all shape how a system behaves. This guide explains how to assess a specific AI solution, gather evidence, and turn findings into action before or after deployment.

What an AI audit is and when to conduct one

An AI audit is a structured assessment of a specific solution. It examines whether the system serves its intended purpose, which data and components influence its outputs, and what controls are in place to address failures. The scope depends on the use case. A generative tool that drafts text calls for different checks than a model that supports credit decisions or critical operations.

Before deployment, an audit can uncover limitations while there is still room to address them. It is also worth conducting one when data, the model, prompts, integrations, or the intended use change. Incidents, complaints, and changes in the system’s impact on people are further reasons to reassess it. An AI risk assessment should consider both the likelihood of failure and the consequences in that particular context.

An audit is not the same as organization-wide governance, which sets principles and responsibilities across multiple solutions. Nor is it a certification or an automatic guarantee of legal compliance. For related organizational practices, read our article on AI acceptable use policies for businesses.

Define the scope and map the whole system

Together, these elements define how to audit an AI system beyond its technical configuration. Document its purpose, users, context, supported decisions, and approved limits. Also specify what falls outside the intended use. For example, a tool that supports analysis may not be authorized to make decisions automatically or to be used by another department.

Next, trace the path from data to output. Include sources, preparation, the model, prompts, interfaces, APIs, and external services. Record dependencies that cannot be inspected and explain how they limit your conclusions. Identify who maintains each component, receives alerts, and can stop use when a critical risk arises. Once the scope is agreed, you can assess the solution as a whole. If your company plans to build a custom solution, custom AI development can incorporate integrations and controls from implementation onward.

Assess data, machine learning models, and algorithmic bias

• ✓Source and permissions: confirm where the data comes from and whether its use is authorized for this purpose.
• ✓Quality and freshness: look for incomplete, inconsistent, or outdated data that could distort results.
• ✓Representativeness: check which groups, situations, and exceptions are or are not represented in the data.
• ✓Performance and limitations: test relevant cases, known errors, and situations outside expected conditions.
• ✓Differences across groups: compare results in relevant scenarios and connect disparities to the impact of the decision.

This AI audit checklist helps guide your questions, but it does not replace criteria tailored to each solution. A machine learning model audit needs to connect tests to the system’s purpose. An overall metric can hide failures on specific inputs or a decline in quality in a particular context. Record sources, transformations, data and model versions, and known limitations. That record helps you interpret results and repeat tests after changes.

An algorithm audit should also examine how errors are distributed across relevant groups or scenarios. To investigate algorithmic bias, select appropriate samples, compare metrics suited to the problem, and consider false positives and false negatives. Interpretation must account for the real effect of the decision, not just a statistical difference. Document the method, the criteria, and the reasoning behind each choice. What counts as acceptable depends on the context, the people affected, and the alternatives available.

When a change to the data or model affects important results, repeat tests under comparable conditions and compare findings with a baseline. If the assessment involves personal data, consider purpose, access, and protection. Our article on AI and data privacy compliance explores safeguards for using that data.

Review integrations, security, human oversight, and production use

The model’s observed behavior does not reveal every risk. Integrations can expand access to data, people may treat recommendations as final decisions, and production changes can alter results. Use the table to connect checks to these points of interaction and include an AI security assessment in your scope.

For generative AI, test instructions that try to bypass rules or obtain information that should not be exposed. Also check what happens when an external service is unavailable or returns unexpected content. Record the permissions for each integration and whether third parties can store or reuse submitted data. These tests can reveal risks that a model-only assessment would miss.

Human oversight works only when the responsible person has context, understands the system’s limits, and can challenge an output. That person also needs the authority to stop or escalate its use. Define how challenged decisions are recorded and who reviews those cases. In production, monitor quality, changes, and incidents. Decide who receives alerts, when tests are repeated, and which conditions require suspension or investigation. The right frequency depends on the solution’s impact and how it operates.

Follow audit steps and record verifiable evidence

1. 1 Set the scope Describe the purpose, components, users, and impacts that the tests need to cover.
2. 2 Gather evidence Collect relevant documentation, configurations, samples, access logs, and vendor information.
3. 3 Run the tests Connect each test to the component, use case, and expected evidence.
4. 4 Analyze the results Compare the evidence with criteria defined for the use case, then describe risks and limitations.
5. 5 Document findings Record evidence, criteria, owners, and proposed actions for each finding.

This sequence makes an AI audit verifiable and helps you understand how each conclusion was reached. For every test, record the question, conditions, component assessed, interpretation criteria, and observed result. Preserve the relevant versions of data, the model, and configurations. Without that record, a repeat test may compare different conditions.

A sample of outputs may point to an error pattern, but the conclusion depends on the system’s purpose and the effect of that error. Record where the sample came from, how it was selected, and the conditions under which it was produced. Connect every piece of evidence to the criteria used to interpret it. Incomplete evidence may call for further investigation rather than a definitive conclusion.

Also report unavailable data, untested dependencies, and limits on what could be observed. If you could not observe the system under real-world conditions, explain how that gap affects the conclusions. The report should not imply more confidence than the evidence supports. It should help accountable people decide what to fix, monitor, or investigate.

Prioritize findings, assign owners, and answer common questions

• Impact: consider how serious the consequences could be for people, operations, and the system’s objectives.
• Likelihood and exposure: assess the chance of occurrence and how many situations or people could be affected.
• Reversibility: determine whether an unsuitable output can be corrected or may cause effects that are difficult to undo.
• Action and follow-up: assign each measure an owner, deadline, completion evidence, and a way to monitor residual risk.

Use these criteria to distinguish blockers that need to be resolved before use from improvements that can go into a monitored plan. Priority depends on context, not just a score. For each action, specify who will make the change, who approves resuming use, and how the fix will be verified. Track residual risk as well, since a measure can reduce exposure without eliminating the possibility of failure. No single checklist works for every solution.

Move forward with a safer, more controlled AI solution

Agence conducts audits, monitoring, and governance for AI solutions through its Safe AI service. The work connects evidence to controls applied before deployment and to follow-up afterward. This gives you clear criteria, assigned owners, and measures that reflect the system’s impact.

The work focuses on the components, integrations, and uses of the solution you plan to deploy or already use. Corrective actions may include access restrictions, usage limits, human validation, and conditions for stopping use. Monitoring tracks relevant changes and signals when new tests are needed. Talk with Agence to define the right work for your solution.

Talk to an AI specialist