Audit owns the evidence standard. Delivery operationalizes it.

Engineering and the CIO function are measured on velocity, cost, and feature parity. Their incentives push toward shipping AI capability and away from slowing it down. That is not a character flaw. That is the role.

An Audit Chair, a Risk Committee, a General Counsel, and a Chief Compliance Officer are measured on defensibility. Their job is to be unsurprised. The first economic sponsor for a certification effort is usually the owner accountable for evidence, risk acceptance, and external review. The CIO becomes an ally once the gate is documented. The conversation changes from "you are slowing me down" to "you are giving me a release decision I can sign."

The pressure on Audit and Risk Committees is already here.

Each number below has already landed in an Audit Committee inbox somewhere. None of them is a vendor talking point. They are the floor of the conversation, not the ceiling.

• 95% of GenAI pilots produce zero measurable P&L impact.^[1] Capital is flowing on reputational pressure rather than operational conviction. The Audit Chair is one quarter of board questions away from explaining why.
• EUR 35 million, or 7% of global turnover, whichever is higher.^[2] The EU AI Act, Article 99, specifies the maximum administrative fine for prohibited AI uses. High-risk uses carry their own penalty tier. The penalty schedule applies regardless of where the model was developed if the output reaches the EU.
• SEC AI-washing enforcement.^[3] In March 2024 the SEC charged two investment advisers for allegedly misleading statements about their use of AI. The agency has signaled continued scrutiny. Disclosures must match what the firm actually does.
• Self-reported board AI knowledge gap.^[4] NACD director surveys report persistent director concern about board readiness for AI oversight. The Audit Committee is usually the first to act because the question always lands on the audit side first.

Three of these four signals can be cited verbatim in a board pre-read. The fourth, the NACD self-assessment, explains why the others go unanswered. Boards know they are exposed and are openly asking for help.

Four artifacts decide whether an AI program is auditable.

The certification model is intentionally narrow. It does not require a new platform, a new framework, or a new committee. It requires four artifacts the Audit Committee can read in a sitting and external reviewers can sample on their own schedule. Each is aligned to NIST AI RMF^[5] and ISO/IEC 42001^[6] by design.

A note on language: in this brief, "certification" means a documented evidence-readiness process and an Audit Committee-facing evidence pack. It is not accredited ISO/IEC 42001 certification, legal advice, or an independent assurance opinion.

1. Model and data register.

Every AI use case, model, dataset, vendor, owner, decision impact, and exit criterion in one signed register. The register names a human owner, not a team. The register lists the decision the model influences in plain English. The register names the date the model can be retired and what would have to be true to retire it earlier. The register is reviewed quarterly and the Audit Committee receives a one-page diff. Aligned to the NIST AI RMF GOVERN function and ISO/IEC 42001 planning and operational-control requirements.

2. Release-of-funds gates.

Funding, deployment, and exception approval are tied to documented readiness thresholds, not opinions. A release does not move because a vendor demo went well. It moves because a defined set of evidence is in the file. Evidence categories include security review, data lineage, fairness testing where applicable, human review points, rollback plan, monitoring plan, and named owner for each. Aligned to the NIST AI RMF MANAGE function.

3. Decision and exception log.

Every approval, override, retrain, and incident is captured as a record a regulator can read in plain English. Each exception has an owner, a scope, and an expiration date. The log is the single source of truth when the question lands. A program with no exception log effectively says "we approved everything by default" when read by an external reviewer.

4. Board-ready brief.

A short document, on the order of 10 to 12 pages, that the Audit Chair can hand to a regulator, plaintiff, or external auditor without translation. The brief is not a tour of the model. It is the four artifacts above, with names, dates, owners, and a short narrative of how they fit together. If the Audit Chair cannot read it cold and follow it, the brief is not finished.

An uncertified AI program leaks evidence at five places.

When the four artifacts are absent, the gaps cluster in predictable places. The categories below are aligned to NIST AI RMF^[5], ISO/IEC 42001^[6], and ISACA's Auditing Artificial Intelligence guidance^[7]. They are the categories an internal auditor will sample first and the ones an external reviewer will ask about by name.

Gaps compound. A single uncertified model produces a paper trail. Forty produce noise no committee can defend. The engagement is designed to close the five gaps where possible, bound the remainder, or convert them into owned remediation plans within 90 days.

What changes when the program is certified.

The right column below is what the engagement leaves behind. Each row is something an Audit Committee, Risk Committee, or external reviewer can ask about by name and receive a documented answer for without reconstructing the decision trail from scratch.

From exposure to evidence pack, in three phases.

The certification motion is 90 calendar days. The clock starts at the kickoff with the Audit Committee or its delegate. A senior cross-functional team enters the program, exposes what is missing, remediates the gaps, and leaves the committee with a record designed for external review.

Eight artifacts. One audit-grade record.

Each item below is concrete, owned, and dated. None is a slide deck about a slide deck. Naming conventions and templates are aligned to NIST AI RMF and ISO/IEC 42001 so internal audit can sample any of them on its own schedule.

• Signed model and data register. Every AI use case, model, dataset, vendor, owner, decision impact, and exit criterion in one document. Reviewed quarterly.
• Release-of-funds gate definitions. Documented thresholds for funding, deployment, and exception approval. Each gate has named evidence categories and an owner.
• Vendor diligence evidence pack. AI-specific procurement questions, expected evidence, exit clauses, and a scoring rubric. Used before the next contract signature.
• Decision and exception log. Every approval, override, retrain, and incident captured as a plain English record. Each exception has an expiration date.
• Policy-as-code rule set. Written governance translated into executable rules, dashboards, and exception workflows that engineering can run inside delivery.
• Incident response ritual and runbook. A short, repeatable response sequence that separates facts from blame and produces an external lesson release plus an internal controls package.
• Staff training artifacts and attestation. Role-specific training for engineering, legal, procurement, audit, and executive owners. Attestation captured in the register.
• Audit Committee brief. A 10 to 12 page document the Audit Chair can hand to a regulator, plaintiff, or external auditor without translation.

Certify before the next examination, not after the next incident.

The right moment to start is a quarter before the question lands. The wrong moment is the morning the regulator, plaintiff, or external auditor arrives. If any of the four signals below are visible in the company today, the engagement reduces evidence-retrieval friction before the committee is asked to approve or defend an AI release.

The question is not whether the program will be examined. The question is whether the examination will arrive before the signed record is ready.

Sources.

Every figure used in the hero strip and Section 02 is sourced. Hyperlinks resolve to the canonical source. The slide deck reproduces a condensed version of this list on its own Sources slide.

Nobody on the committee wants to slow innovation. Nobody wants approval to outpace evidence. Certifying the program is the path between those two sentences.