Illustrative cover photograph: Olena Kholina / Unsplash. Not a client engagement.
Decide what the review must answer
An AI audit should tell you something specific about a system already in use: whether its controls work, where the evidence is weak, and what needs to change. Before asking for a quote, write the decision the report will support.
“Is our AI safe?” is too broad to commission sensibly. “Can our support assistant retrieve another customer’s records, and do our controls prevent it?” gives a reviewer something to investigate.
Buy a defined examination, not a reassuring label. The UK government’s introduction to AI assurance describes a range of assurance techniques, including impact assessment, performance testing and compliance audit. Different techniques answer different questions. An audit of an AI system is also different from using AI to audit financial accounts.
The working method below is for commissioning a review of an operating system. For setting internal responsibilities, use the AI governance framework.
Choose the review before choosing the reviewer
An operational review examines how the system behaves in its actual workflow: inputs, outputs, exceptions, human decisions and incidents. A security assessment examines attack paths and technical controls. A data-protection review needs the relevant privacy expertise. Certification concerns a specified scheme and its requirements, not a general declaration that a product is trustworthy.
Ask a prospective reviewer which of these they can actually provide and which require another specialist. If certification is required by a customer or procurement process, identify the scheme, scope and required recognition before commissioning work. A consultancy report is not automatically an accredited certificate.
For a UK or UAE organisation, record where the system operates, who it affects and which contractual or regulatory requirements need specialist interpretation. This worksheet does not establish compliance in either jurisdiction. My consultancy support can help define the review and operational questions; it is not an offer of accredited certification or specialist legal assurance.
Put the operating boundary in writing
Name the application, model or service version, connected systems, business process and review period. Record where the reviewer can inspect configuration directly and where they must rely on supplier evidence. A hosted model’s underlying training process may be outside your access; that limitation belongs in the brief, not in a footnote discovered after payment.
Include the people who use and supervise the system. A technically correct answer delivered to the wrong customer is still a failure. So is a warning that nobody has time to review.
Define excluded work explicitly. For example: review the internal support-drafting assistant and its retrieval permissions; exclude autonomous customer messaging, which is disabled, and exclude an opinion on the foundation model’s training-data legality. Verify that excluded functions really are disabled.
Agree permitted testing, accounts, windows and escalation contacts. Destructive testing or access to real customer records must not be an informal surprise. Use a representative controlled environment where possible, then document any differences from production.

Build an evidence request the team can fulfil
Request configuration snapshots, permission rules, model and prompt versions, test cases, redacted incident records, change history and a small sample of real operational outcomes that can lawfully be shared. Name an owner and a delivery date for each item. Do not put credentials or personal records in the worksheet itself.
A policy says what should happen. A test shows what happened under particular conditions. A log may show what happened in operation. These are different kinds of evidence and should not be presented as interchangeable.
Use a simple evidence register:
| Review question | Evidence requested | Limitation to record |
|---|---|---|
| Can users retrieve only permitted records? | Access rules and cross-account negative tests | Test accounts may not cover every production role |
| Are unsupported answers caught? | Reviewed outputs and escalation records | Sample covers named topics and dates only |
| Can an incident be reconstructed? | Redacted event logs and version history | Missing logs restrict conclusions |
| Are fixes maintained? | Change approval and regression evidence | Later releases need fresh evaluation |
“Not evidenced” is not the same as “passed”. Make that distinction part of the report specification.
Check who is marking whose work
Ask the proposed reviewer to disclose whether they designed, sold, configured or operate the system, and whether they earn money from recommended remediation products. Ask who selects the test sample and whether the reviewer can report an adverse finding without commercial approval from the implementation supplier.
Independence is not achieved by putting a different logo on the report. If the implementer performs the assessment, describe it as an internal or supplier review and commission separate scrutiny where the risk warrants it. A specialist who helped design the system may still provide useful evidence; that relationship should be visible.
Request a redacted example report and the named team’s relevant experience. Assess whether findings contain reproducible evidence, consequences, practical remedies and clear limitations. A long checklist with green ticks is not enough.
A support-assistant review, worked through
Consider a fictional business whose assistant drafts replies from approved product documents and customer tickets. A person sends every reply. The commissioning question is whether account separation and review controls remain effective after a new retrieval connector was added.
A proposed initial sample might contain 12 ordinary requests, 8 attempts to retrieve a different account’s material, 6 questions with no approved answer and 4 escalation cases: 30 cases in total. These are illustrative planning numbers, not an assurance standard or a statistically representative sample. The reviewer should adjust them for the system’s actual risk and diversity.
Require a case-level result and evidence reference. If one cross-account test exposes restricted data, stop the affected workflow, investigate and retest the boundary. Do not average that failure into a 97% success headline. If the reviewer cannot inspect the connector’s access rules, the conclusion must say so.
Specify the final deliverables: scope and versions tested, methods and sampling rationale, findings with severity and evidence, limitations, management responses and a retest plan. Each accepted remediation needs an owner, date and closure evidence. A changed model, connector or permission scheme should trigger a decision about renewed review.
Use the commissioning worksheet
The download contains separate rows for scope, reviewer conflicts, evidence access, sample design, reporting and remediation, with a worked support-assistant example and blank fields for your own system. Replace the examples before circulating it. It is an editable CSV, with no formulas or external connections.
Send the same completed scope to each shortlisted reviewer. Compare their proposed methods and exclusions before comparing fees. If one quote covers only documentation and another includes technical tests, you have two different purchases.
Download the editable worksheet (CSV) ↓
Frequently asked questions
Can the implementation supplier carry out the review?
They can provide a useful supplier review, but disclose that they built or operate the system. Decide whether the decision requires a separate independent reviewer. Do not describe self-review as independent assurance.
What if the reviewer cannot access the underlying model?
Commission the system-level checks they can perform and record inaccessible components. Require the report to distinguish direct tests from supplier statements and to limit its conclusion accordingly.
Does a passing audit cover later changes?
Treat the report as evidence about the scope, versions and period examined. Agree triggers for renewed review, such as a new connector, model release, permission change or material incident.
Sources checked 8 November 2026. Worksheets and worked examples are suggested working methods, not official standards or reported client results.

