How to audit AI: a complete guide to reviewing AI systems

An AI audit is a structured review of an AI system to confirm it behaves as intended, within accepted boundaries and in line with relevant rules. What is an AI audit? It is the process of examining a model's data, logic and outputs so that decision makers can trust what the system produces. Whether you call it an AI audit or ask what is an ai audit, the goal is the same: verified confidence in your AI. This guide walks through every stage.

What an AI audit actually is and what it covers

An AI audit is a structured review of an AI system's behaviour, data, logic and outputs. The goal is to confirm the system works as intended and stays within accepted boundaries. Unlike a standard IT audit, it looks at probabilistic outputs rather than deterministic code. That distinction matters because an AI model can behave differently each time it runs. A proper audit covers the training data used to build the model, the logic the model applies, the outputs it produces and the controls around each of those layers. It also examines whether the system's behaviour matches what was documented at the design stage. Auditors compare stated intent with observed results, note any gaps and recommend corrective action. The scope varies depending on the risk level of the application, but every audit shares these four layers: data, model, output and governance.

How to audit an AI system step by step

Auditing an AI system follows a clear sequence that any team can apply. Starting without a defined scope leads to wasted effort and incomplete findings, so begin there.

1. Define the scope by identifying which model, version and use case is under review and what risks matter most.
2. Gather evidence by collecting training data documentation, model cards, design specifications and any previous audit reports.
3. Test the model by running it against controlled inputs and edge cases to observe how it behaves under varied conditions.
4. Review governance by checking that accountability structures, approval records and policy documents are in place and current.
5. Document findings by recording every gap, risk and observation with enough detail for a decision maker to act on it.
6. Report and seek sign-off by presenting findings to the relevant owner and agreeing remediation steps with clear owners and timelines.

How to build an AI audit checklist that works

An effective AI audit checklist is built around risk areas rather than a flat list of questions. Group items by the layer they address: data, model behaviour, outputs, governance and monitoring. Within each group, write checks that produce a clear yes or no answer so that reviewers cannot skip past ambiguous items. Include a column for evidence references so every finding links back to a document or log. Revisit the checklist whenever the model is updated, the regulatory context changes or a new use case is added. A checklist that was accurate at launch becomes unreliable once the model is retrained on fresh data. Build in a version control field so auditors know which checklist revision they used and can compare results across audit cycles. The checklist is a living document, not a template to be filed once and forgotten.

AI governance and compliance checks to run first

Before examining the model itself, auditors should verify that the right governance structures exist. Without them, technical findings have nowhere to land.

1. Confirm that a named owner exists for the AI system, with documented accountability for its outputs and risks.
2. Check that an approved AI use policy is in place, current and accessible to everyone who interacts with the system.
3. Review any data processing agreements and confirm they cover the personal data the model uses.
4. Verify that a risk assessment was completed before the model was deployed and that its findings were acted on.
5. Confirm that there is a clear escalation path when the model produces an unexpected or harmful output.
6. Check that third-party components, including pre-trained models or external APIs, have been reviewed under the same governance standards as internal ones.

How to assess bias fairness and transparency in an AI model

Bias and fairness checks examine whether a model treats comparable inputs differently based on protected characteristics such as age, gender or ethnicity. Begin by reviewing the training data for imbalances that could introduce skewed patterns. Then test the model by running matched input pairs that differ only in a sensitive attribute and compare the outputs. Document every case where the outputs diverge without a justified reason. Transparency checks look at whether the model can explain its decisions in terms a non-specialist can understand. Where a model cannot produce an explanation, auditors should note that gap and assess its impact given the application. In high-stakes contexts such as credit, hiring or healthcare, a lack of explainability is itself a risk finding. Record the methods used, the results obtained and any remediation steps agreed, so the fairness review is reproducible in the next audit cycle.

Security and adversarial risk in an AI audit

Security testing in an AI audit goes beyond standard vulnerability scanning. Auditors probe for risks that are specific to machine learning systems.

1. Test for data poisoning by reviewing how training data is sourced, validated and protected from manipulation before it enters the pipeline.
2. Check for prompt injection vulnerabilities by submitting carefully constructed inputs designed to override the model's intended behaviour.
3. Assess adversarial robustness by running inputs that have been subtly altered to see whether the model's output changes in unexpected ways.
4. Review access controls to confirm that only authorised parties can query the model, update its weights or access its training data.
5. Verify that inference outputs are logged so that a malicious query can be identified and traced after the fact.
6. Confirm that a process exists to respond to a security incident involving the model, including steps to isolate it and notify affected parties.

Post deployment monitoring as part of ongoing AI auditing

A single audit at deployment captures the system as it was at one point in time. Once the model is live, the data it encounters changes, its outputs drift and new risks emerge. Continuous monitoring is the mechanism that catches those shifts between formal audit cycles. Set up automated checks that flag when output distributions move outside the range observed during the initial audit. Define thresholds that trigger a human review rather than waiting for a complaint or an incident. Log every significant change to the model, its data sources or its operating environment as a formal event that requires a reassessment. Schedule periodic reviews at intervals proportionate to the risk level of the application. A low-risk internal tool may need a light review annually, while a model making consequential decisions in a regulated sector warrants more frequent scrutiny. Monitoring is not optional: it is the mechanism that keeps audit findings valid over time.

How AI auditing maps onto compliance frameworks

Audit activities become more useful when they map onto recognised frameworks. Findings then feed into compliance evidence without duplication.

1. Review ISO 42001, the international standard for AI management systems. Align your audit checks to each control it requires.
2. Identify any sector rules that apply. Financial services and medical devices each carry their own obligations. Confirm the audit covers them.
3. Map each finding to a specific control. This lets you prioritise remediation by compliance impact.
4. Use the framework's own terms in your reports. Compliance teams can then absorb findings without translating them.
5. Record which framework version each audit cycle used. Frameworks are updated regularly, and your evidence must reflect the version in force at the time. Where a framework is revised mid-cycle, note the transition point clearly so auditors in subsequent cycles understand which controls applied during each period.

The audit trail and documentation an AI system must produce

An AI system that cannot be reconstructed from its records cannot be audited properly. Documentation requirements fall into several categories. Training data records should capture the source, version, processing steps and any exclusions applied before training began. Model version histories should record every retrain, fine-tune or configuration change alongside the reason it was made. Decision logs should store enough information about each output that an auditor can trace the input, the model version and the result. Access logs should show who queried the model and when, so that unusual usage patterns can be investigated. Incident records should document every case where the model produced an output that was reviewed, escalated or corrected. All records should be stored in a format that cannot be altered after the fact. Without this documentation, an auditor cannot confirm that the system operating today is the one that was approved.

The Agency builds AI systems and AI agents for businesses.

Read about our What does a marketing agency do.

Take our free AI growth audit: use it and see it before you buy.

How to audit AI across different industries and use cases

Audit priorities shift depending on the sector and the consequences of a wrong output. Understanding which risks matter most in a given industry helps auditors focus their effort where it counts.

In financial services, the audit focuses heavily on explainability and fairness. Regulators in many jurisdictions expect that credit or fraud decisions can be explained to those affected, and protected characteristics should not influence lending or insurance outcomes in unjustified ways.

In healthcare, the audit centres on clinical safety. Does the model align with validated clinical guidance, and what safeguards exist when it does not?

In recruitment, bias testing across demographic groups takes priority. A biased screening model can expose an employer to discrimination claims.

In customer service, the audit checks whether the model escalates correctly, since some queries should always involve a human reviewer.

What changes across sectors is which risk carries the greatest weight and how strictly explainability is expected to be demonstrated.

Tools and techniques used to carry out an AI audit

Auditors draw on several types of tooling to carry out a thorough review.

Model testing tools let auditors run structured test suites. They measure output consistency and flag edge cases where the model behaves oddly.

Bias detection libraries scan outputs for gaps across groups. They produce reports that can go straight into audit records.

Explainability tools generate plain-language summaries of model decisions. Auditors use these to check whether the model's reasoning is clear and traceable.

Document review tools help auditors scan large sets of policy files and version histories. They surface gaps quickly.

Log analysis tools examine inference and access logs. They detect unusual patterns and help reconstruct events around a specific output.

The right mix depends on the model type, risk level and regulatory context. Always choose tools that produce exportable evidence. Summaries that cannot be verified are not enough.

Common mistakes that undermine an AI audit

Certain errors appear repeatedly in AI audits. Each one weakens the credibility of the findings.

1. Scoping too narrowly by focusing only on the model. The data pipeline, governance controls and monitoring all matter too.
2. Treating the checklist as a rubber stamp. Each item needs real evidence, not just a tick.
3. Testing only expected inputs. Edge cases and rare scenarios are where models most often fail.
4. Writing findings without enough detail. Anyone not present must be able to understand what was found and why it matters.
5. Running a one-off audit and assuming findings stay valid. Models, data and environments all change over time.
6. Leaving out third-party components. Pre-trained models must be in scope even if your team did not build them.

What qualifications and skills an AI auditor needs

A credible AI auditor brings together technical skill, sector knowledge and an understanding of relevant rules.

On the technical side, you need to understand how models are trained, tested and deployed. Without this, you cannot judge whether a model's behaviour is acceptable.

Statistical literacy matters too. Many audit tests produce numerical outputs that must be read correctly.

Sector knowledge is equally important. An auditor reviewing a financial services model must understand the rules that apply there.

Certifications in AI governance, information security and data protection provide a recognised baseline. They also signal credibility to the organisations you audit.

Experience with structured audit methods transfers well. Financial audit, IT audit and risk management all build the same core skill: gathering sound evidence.

Clear written communication matters just as much. If findings are not understood by decision makers, they will not lead to action.

How to choose the right AI auditing approach for your situation

The right audit type depends on the risk level of the system and where it operates.

A model making low-stakes internal suggestions needs only a light review. Focus on basic governance and output consistency.

A model influencing important decisions in a regulated sector needs a full audit. Cover data, model, outputs, governance, security and fairness.

Decide whether to audit internally, use an external reviewer or combine both. External audits carry more weight with regulators. They are not affected by internal conflicts.

Match the depth of testing to how often the model is updated. A system that retrains often needs a more flexible audit structure.

Whatever approach you choose, write down why you chose it. Stakeholders need to see that it suited the risk level. They also need to judge whether a more thorough approach is needed as the system grows.