EU AI Act transparency duties apply now; high-risk duties from December 2027. Check your exposure
Insights About us Careers
Contact us
AI Governance, Security and Compliance

AI Audit and Third-Party Assurance

An independent assessment of whether your AI systems and governance do what you say they do, tested against evidence, for readers who will not accept a self-assessment.

6 to 12 weeks
Typical engagement
Fixed fee
Commercial model
Independent
Not a certificate

The demand for this is arriving from three directions at once: boards asking whether the controls described to them exist, enterprise customers no longer accepting a vendor's own questionnaire responses, and regulators expecting evidence rather than assertion. All three are asking the same question, which is whether the documentation matches reality.

In one paragraph

An AI audit is an independent assessment of AI systems and the governance around them against a stated standard — a regulation, a framework, a customer's requirements or your own published policies — in which claims are tested against evidence rather than accepted as described.

Being clear about what this is

Worth knowing

We are not a certification body, and independence has limits

We cannot issue a certificate — that requires an accredited body. Nor will we audit a governance framework or management system that we designed and built, because a firm auditing its own work is not providing assurance. Where we have done the implementation, we will say so and recommend a different assessor for the audit.

What an audit produces is a report with an opinion, the evidence behind it and the findings, addressed to a reader who needs more than your word. That is different from certification and, for many purposes, sufficient.

Auditing against a stated standard

StandardThe question answeredTypical reader
Your own published policyDo you do what you say you do?Board, audit committee
Your governance frameworkIs it operating, or only documented?Executive, internal audit
NIST AI RMFIs the risk process real and evidenced?US enterprise customers
ISO/IEC 42001Are you ready for certification audit?Certification body, executive
EU AI Act readinessIs classification and documentation defensible?Regulator, counsel, customers
A customer's requirementsDo you meet what the contract demands?The customer's risk function

The standard has to be named before the audit starts. An audit against unstated criteria produces observations rather than findings, and observations are not something anyone can act on or rely on.

How we test rather than accept

Sample the systems, not the summary

We select systems from the inventory ourselves, including ones that were not offered, and trace them through the controls end to end. An audit that examines only the examples provided is testing the selection rather than the estate.

Re-perform where feasible

Where a control produces a measurable result — a bias test, an evaluation, a monitoring alert — we re-run or independently reproduce it rather than reading the reported figure. Reported results and reproducible results diverge more often than organisations expect.

Interview the operators, not only the owners

The person who runs a review daily knows whether the process described actually happens. The gap between the documented process and the practised one is the single most productive area of any audit.

Trace an incident and a decision

Pick something that went wrong and follow what actually happened, and pick a deployment decision and follow the approval trail. Both reveal in an afternoon whether governance is operating or performing.

Test the inventory's completeness

Not only whether the listed systems are governed, but whether unlisted systems exist. We look in procurement records, expense data and engineering repositories, because completeness is the finding that most often changes an executive's understanding of their exposure.

Rate findings by consequence

With a remediation plan attached, owners named and dates agreed. An audit report with thirty findings and no priority ordering produces paralysis rather than improvement.

What the report contains

  1. Scope and the standard audited against, stated precisely, including what was excluded and why.
  2. An opinion, expressed plainly, on whether the controls operate as described.
  3. Findings rated by consequence, each with the evidence, the affected systems and the recommended remediation.
  4. Evidence appendix, so a third party can see what was tested and how the conclusion was reached.
  5. Remediation plan with owners and dates, agreed with you rather than imposed.
  6. A customer-facing summary where the audit exists to answer enterprise buyers, written to be shared.
Process

How the engagement runs

Systems are sampled by us, including ones that were not offered.

Week 1

Scope and criteria

The standard named precisely, systems in scope agreed, exclusions documented.

Weeks 2 to 3

Evidence gathering

Documentation reviewed, inventory completeness tested against procurement and engineering records.

Weeks 4 to 7

Testing

Controls traced end to end on sampled systems, results re-performed where feasible, operators interviewed.

Weeks 8 to 9

Findings

Rated by consequence, with evidence attached and affected systems named.

Weeks 10 to 12

Report and remediation

Opinion issued, remediation plan agreed with owners and dates, customer-facing summary where needed.

Deliverables

What you receive

An independent opinion backed by tested evidence, and a remediation plan someone owns.

01

Scope and criteria statement

The standard audited against, systems in scope, and documented exclusions.

02

Audit opinion

Whether controls operate as described, stated plainly.

03

Findings

Rated by consequence, with evidence, affected systems and recommended remediation.

04

Evidence appendix

What was tested and how conclusions were reached, inspectable by a third party.

05

Inventory completeness assessment

Whether unlisted AI systems exist, tested against procurement and engineering records.

06

Remediation plan

Owners and dates, agreed rather than imposed; plus a shareable summary where the audience is customers.

Fit check

Is this the right engagement?

Worth being direct. AI Audit and Third-Party Assurance is the wrong spend in some situations, and those are listed rather than buried.

Good fit if

  • A board or audit committee has asked whether the described controls exist.
  • Enterprise customers no longer accept self-assessment questionnaires.
  • Certification is planned and readiness needs independent testing first.
  • An incident has raised questions about whether governance operates.
  • A regulator has asked for evidence rather than description.

Choose something else if

  • You need a certificate, which requires an accredited certification body.
  • We designed and built the framework in question, which disqualifies us from auditing it.
  • Nothing is in place yet; build first. See framework design.
  • The intention is a favourable report rather than an accurate one.
Questions

Frequently asked questions

Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.

Can you certify our AI systems?

No. Certification requires an accredited certification body, and we are not one. What we provide is an independent audit report with an opinion and tested evidence, which is what boards, enterprise customers and regulators are usually asking for — and we support certification separately where a certificate is the actual requirement.

Can you audit governance you helped us build?

No, and we will say so rather than take the work. A firm auditing its own implementation is not providing assurance to anyone, whatever the report says. Where we did the build, we recommend a different assessor and will support you through their process.

What do you actually test?

Systems we sample ourselves, including ones not offered to us. We trace controls end to end, re-perform measurable results rather than reading reported figures, interview the people who operate processes daily rather than only those who own them, and follow an incident and a deployment decision through the trail they left.

What if you find serious problems?

They go in the report, rated by consequence, with evidence and recommended remediation. That is the purpose of commissioning an audit. Findings are reported to you under NDA; where the audit exists to answer customers, you decide what is shared, and we will not write a summary that misrepresents the findings.

How is this different from a certification audit?

A certification audit assesses conformity to a specific standard and results in a certificate from an accredited body. Ours assesses against whatever standard you name — your own policy, a framework, a customer's requirements — and produces an opinion with evidence. It is more flexible, faster, and carries no certificate.

Is this the right engagement?

Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.