Subscribe
Analysis
7 min read

The AI Audit Checkbox: What ISO 42001 Certification Actually Proves

Three years after publication, only a few hundred organizations hold accredited ISO 42001 certificates, yet procurement teams behave as if the checkbox settles vendor trust. What the audit proves, what it costs, and what to read on the certificate.

Retro illustration of two figures inspecting a long wall of identical framed certificates with blank seals
Retro illustration of two figures inspecting a long wall of identical framed certificates with blank seals

Three years after ISO published the world's first certifiable standard for managing artificial intelligence, a few hundred organizations hold a certificate. Enterprise procurement teams now behave as if the answer to "are you ISO 42001 certified?" decides whether a vendor is trustworthy. The honest answer is more complicated, and the paperwork is more complicated still.

What ISO 42001 actually is

ISO/IEC 42001:2023 was published on December 18, 2023. It is a management system standard, built on the same architecture as ISO 27001: an organization documents how it governs the thing in question (here, AI systems it develops, provides, or uses), runs risk and impact assessments, assigns roles, monitors, audits itself, and gets reviewed by management. The standard's clauses 4 through 10 set the requirements, and Annex A carries 38 controls organized into nine control areas, as the Cloud Security Alliance summarizes it.

Certification is voluntary and follows the familiar three-year cycle: a two-stage initial audit, annual surveillance audits, and a recertification audit at year three. The first accredited certificate in the world went to fintech infrastructure firm ORO Labs on July 16, 2024, issued by certification body Mastermind under IAS accreditation.

What the certificate does not do is certify any model. It says nothing about whether a system is accurate, fair, or safe, and it is not a product endorsement. As Reco puts it in its explainer, certification "provides independent assurance that an organization's AI management system conforms to the standard's requirements," which is a statement about process, not output. A vendor can hold a valid certificate and ship a bad model on the same afternoon.

The audit itself is regulated at one remove: certification bodies need accreditation (ANAB in the US, UKAS in the UK), and a companion standard, ISO/IEC 42006:2025, sets the competency requirements for auditors and the rules for computing audit days from an organization's AI-related headcount and complexity. Audit time, in other words, scales with how much AI you touch.

A rubber stamp pressing an official seal over a network of nodes

Who holds certificates, and how fast that is growing

The certified population is small enough that it can be listed. The ISO Survey 2025, the first edition to count the standard, recorded 271 valid accredited ISO/IEC 42001 certificates covering 572 sites in 40 countries as of December 31, 2025, with the United States holding the most at 100, followed by India with 25 and South Korea with 24. Because companies can hold multiple certificates, that is fewer than 271 organizations. Atoro, which compiles announcements because no official global register exists, estimates roughly 350 certified organizations worldwide by spring 2026.

Growth is real but off a tiny base. The ISO42k directory, which tracks publicly announced certifications, counts 229 announcements through late October 2026: 19 in 2024, 89 in 2025, and 120 so far in 2026, spread across 30 countries and 28 observed certifying bodies. BSI, one of the larger certification bodies, issued 11 certificates in all of 2025 and says its volume more than tripled in 2026, awarding its 50th certificate on October 21, 2026.

The marquee names are certified. AWS announced in November 2024 that it was the first major cloud provider accredited against the standard. Google announced certification on December 19, 2024 covering ten products including Google Workspace and Gemini, and notably publishes what sits outside scope, naming Antigravity in Gemini Enterprise as one exclusion. Anthropic followed on January 13, 2025, certified by Schellman under ANAB accreditation. Microsoft's certificate covers Azure AI Foundry Models, which includes Azure OpenAI models, plus GitHub Copilot, Microsoft 365 Copilot, Copilot Studio, Security Copilot and Dragon Copilot, with the service list published on Microsoft Learn.

Coverage is uneven below the first tier. The AI Provider Trust Registry, a third-party tracker that grades evidence rather than endorsing vendors, lists public evidence of certification for the OpenAI API, Anthropic's API, Bedrock and Vertex as platforms, and Azure OpenAI, while marking Cohere as partial and recording "no public evidence" for first-party certificates from Mistral, xAI and DeepSeek. Absence of evidence in a spreadsheet is not the same as absence of a certificate, but it is what a procurement analyst can verify.

What it costs

ISO does not publish prices, and certification bodies quote per scope, so every number circulating is a vendor or consultancy estimate. The audit fee is the small part. Estimates cluster at $5,000 to $25,000 for a small organization's combined Stage 1 and Stage 2 audit, roughly three to five audit days, with annual surveillance running 20 to 40 percent of the initial fee and recertification about 60 to 70 percent, per Scytale's breakdown.

The larger bucket is the work before the audit: gap assessment, policies, risk registers, control implementation, internal audit. Konfirmity puts a realistic first-year band at $20,000 to $60,000 for most teams. Fruggr's planning figures run $85,000 to $150,000 all-in for a mid-sized organization once consulting, training and certification fees stack up, and 4 to 12 months end to end, or 3 to 4 months if an ISO 27001 system already exists. There is also a personal certification market: PECB sells Foundation at about $299 and Lead Implementer or Lead Auditor credentials at $599 to $659, and APMG accredits training organizations to deliver them.

Why anyone buys it, given no law requires it

The regulatory pitch has weakened even as sales grew. The EU AI Act's general-purpose model obligations applied on August 2, 2025, and the high-risk rules were scheduled for August 2, 2026. Then the AI Omnibus arrived: proposed by the Commission on November 19, 2025, finally adopted by the Council on June 29, 2026, and in force since July 27, 2026, it moved the standalone Annex III high-risk obligations to December 2, 2027 and the embedded Annex I product rules to August 2, 2028. Article 50 transparency obligations, including machine-readable marking of AI-generated content, did still apply from August 2, 2026. Penalties for prohibited practices reach 35 million euros or 7 percent of global turnover.

The confusion is its own market signal. A great many compliance explainers still dated August 2, 2026 as the operative cliff months after the omnibus was adopted, and the Cloud Security Alliance's research note through its drafting period advised treating the original date as binding because the extension "has not been enacted into law." That note is now stale, which is precisely the point: the goalposts moved, many sellers of guidance did not update, and buyers cannot tell which guides are current.

Meanwhile, the standards that would formally connect certification to the law are not finished. ISO 42001 has been adopted in Europe as EN ISO/IEC 42001:2026, but European adoption does not make it a harmonized standard that confers presumption of conformity. CEN-CENELEC made EN 18286:2026 available in July 2026 as the first European standard written specifically for the quality management system that Article 17 of the AI Act demands of high-risk providers, and its citation in the Official Journal is still pending. In the US, NIST's AI Risk Management Framework remains voluntary with no certification mechanism at all.

So the demand is not legal. It is procurement. As one consultancy's framing of the market puts it, customers and procurement teams asking for the certificate contractually is "currently the most common driver by a distance." For a vendor selling AI features into regulated buyers, one accredited certificate answers a governance questionnaire that would otherwise be rebuilt deal by deal. For the buyer, asking for a certificate is considerably cheaper than reading forty pages of policy.

What to do with the certificate someone hands you

If you are buying, treat an ISO 42001 certificate the way security teams learned to treat SOC 2 reports: useful, narrow, and due for scrutiny. Four checks matter. Read the scope statement, because certification is per legal entity and per product, not per company, and Google's willingness to name exclusions is the disclosure standard to expect. Verify the accreditation mark behind the certification body, since an unaccredited "certificate" can be a training diploma in disguise. Check the dates, because a three-year certificate with a missed surveillance audit is an expired control system. And ask the vendor how they monitor the controls between audits, because the certificate attests to an assessment on a past date, not to present behavior.

If you are selling, the calculus is mostly arithmetic: a six-figure first-year program that shortens enterprise sales cycles. Build the management system to operate, and the audit passes as a consequence. Build it only to pass, and, as practitioners in the audit community note, you will be rebuilding it within eighteen months.

The industry label for what is happening is "trust." The mechanical reality is more prosaic: AI governance became a checkbox, and checkboxes are tradeable. That is neither a scandal nor a solution. It is what a procurement filter looks like in its first year of existence, before it becomes table stakes and stops meaning anything on its own.

Sources

AnalysisGovernance