AI assurance rules are emerging. Boards still need evidence of who chose the evaluator, what they could examine and what findings customers will receive.

Issue 29

AI Has Auditors. Boards Still Need to Test Their Independence.

THE TAKE

Three September announcements put AI assurance in front of boards. On September 18, Accenture and Anthropic announced an embedded evaluation partnership, led by Faculty, Accenture’s specialist AI business. The work includes red-teaming, alignment assessments and testing safeguards. Each company expects to invest at least $1 billion in evaluation capacity over five years. The same day, California’s Executive Order N-9-26 accelerated implementation of its new AI-assurance laws and requested recommendations by November 16 on measures including onsite verification at frontier labs. A joint international statement, published by the Netherlands on September 22, called for mandatory pre-deployment testing and qualified independent evaluators with sufficient access.

The assurance market is taking shape. Boards still need to examine what independence means in each engagement.

There are already rules worth reading. California’s AB 1405 establishes auditor independence, competence and reporting requirements; SB 813 creates a framework for designating independent verification organizations. Implementation is still ahead. Illinois’s SB 315 requires annual independent compliance audits for covered large frontier developers starting in 2028, and specifies auditor competence, conflicts, access and public reporting. These are scoped requirements with different timelines, not a single assurance regime covering every AI vendor.

The commercial arrangements matter too. Anthropic says it will fund Accenture’s work directly. Its CEO has outlined intended publication protections, including publication of key findings without Anthropic’s editorial control, subject to specified redactions. The announcements do not publish the Accenture-specific contract. Buyers should ask how those protections operate, what access is excluded and how disagreements are resolved.

You already know the questions from external assurance: who appoints the assurer, who pays them, what scope they can refuse, what they must report and to whom, and what happens when management dislikes the answer. Bring those questions to AI procurement. A label on a vendor’s website cannot answer them for you.

THREE THINGS BOARDS SHOULD KNOW

Treat “independently evaluated” as a claim to examine. Ask which standards and legal requirements apply, what was tested, and which independence safeguards govern the engagement.
Statutory incident reporting does not guarantee notice to your company. The enacted frontier-AI laws in California, New York and Illinois cover specified incidents beyond completed mass harm. Their deadlines and recipients differ; New York’s and Illinois’s duties begin January 1, 2027. Check your contract’s notification terms.
Audit committees are prioritizing AI. In Deloitte and the Center for Audit Quality’s 2026 survey, 75% of the 72 respondents whose committees oversaw AI ranked it a top-three priority beyond financial reporting and internal controls.

THE QUESTION I’M HEARING

“Every AI vendor in our stack says its models are independently evaluated. How do we tell whether that claim means anything?”

Ask five questions, in writing. Who selected the evaluator and who pays them? What scope or access was excluded? Can the evaluator publish without the vendor’s consent, and what redactions are permitted? What did the evaluation find in the last twelve months? What obligation does the vendor have to notify you when it finds a problem? An unanswered question belongs on the diligence list.

Run it live at the next audit committee meeting. Pick the two AI vendors with the deepest access to your data, put that grid on one page and let management fill it in, with “unknown” permitted as an answer. The unknowns are the finding.

The Deloitte–CAQ survey drew 248 responses overall. Among the 126 respondents whose committees oversaw cybersecurity, 82% were confident or very confident in that oversight. Among the 72 overseeing AI, the figure was 56%. Those are different groups, but the AI confidence gap is reason enough to ask for evidence at the next meeting.

My recommendation: the audit committee chair asks the general counsel and the CIO to produce that one-page profile for the top five AI vendors before the next quarterly meeting, and to flag every contract carrying no incident notification obligation. Put those on the renewal list. The international statement calls for exploring a new institution; it does not itself create binding duties or a timetable for them. Your procurement work can start now.

THE BOTTOM LINE

“Independently evaluated” should prompt a second question: who chose the evaluator, what could they examine, and what will you be told? Put the answers in the vendor file before you need to rely on them.