Compliance as Proof: When Governance Becomes a Certificate

In July 2026, Figma announced it had achieved ISO/IEC 42001:2023 certification: 38 controls across nine control objectives, audited by Schellman, an ANAB-accredited body, covering nine of its products. The company summarized the shift in one line: “Saying you use AI responsibly is easy, but proving it to an accredited auditor is harder.”

That sentence marks a transition. For two years, “responsible AI” lived on slides. It was a claim, a value statement, a paragraph on a trust page. In 2026 it started becoming a certificate, a document produced by an independent auditor who examined your controls and signed their name to the result. The distance between saying and proving is now the competitive terrain.

Two forcing functions arrived together

The move from assertion to evidence is being driven by two mechanisms that landed within weeks of each other, one voluntary and one regulatory.

The voluntary one is ISO/IEC 42001, the first international management-system standard for artificial intelligence. It works the way ISO 27001 works for security: you build a management system, an accredited third party audits it, and the certificate attests that the system exists and functions. Figma’s certification covers nine control objectives that read like a governance table of contents: AI impact assessment, governance and accountability, AI risk management, lifecycle management, data governance, third-party AI risk, monitoring, human oversight, and responsible use. None of those are novel ideas. What is new is that an external auditor now verifies each one against evidence rather than intention.

The regulatory one is the EU Cloud and AI Development Act, or CADA, proposed in June 2026. Per the proposal, CADA introduces a four-tier sovereignty framework for public-sector cloud, with direct consequences for national rail operators, EU banks, and telecom providers. Where ISO 42001 asks you to prove your AI is governed, CADA asks you to prove where your data and models physically live and who can compel access to them. Sovereignty becomes an auditable property, not a marketing adjective.

Why “we use AI responsibly” stopped being enough

A claim has no cost of entry. Any vendor can write it, and by 2026 almost every vendor had. When every homepage says the same thing, the sentence carries no information. Buyers learned to discount it.

An audited certificate carries cost. Someone had to build the management system, gather the evidence, submit to examination, and pass. That cost is exactly what makes the certificate a signal. It cannot be copied from a competitor’s marketing page. Figma’s framing is precise: the easy part is the claim, the hard part is the proof, and the hard part is now the one that closes deals in regulated procurement.

This is the compounding logic of verification. As explored in Model Cards Are Governance as Product, governance artifacts gain value when they become externally legible. A model card documents a system. A certificate does something stronger: it puts a third party’s accreditation behind the documentation. The internal artifact says “here is what we did.” The external audit says “an accredited body confirmed they did it.”

What an auditor actually checks

Each of the nine ISO 42001 objectives carries a specific demand for evidence. A written policy is where you start. The auditor wants proof it runs.

AI impact assessment asks for documented analysis dated before deployment. Analysis produced after an incident does not count. Governance and accountability asks who owns each decision, by name and role. Risk management asks for a living register with a recent revision history. Lifecycle management asks how models are versioned, retired, and re-evaluated. Data governance asks for lineage. Third-party AI risk asks what you know about the models you did not build. Monitoring asks for the telemetry. Human oversight asks where a person can intervene and whether that intervention is logged. Responsible use asks for the boundaries and the enforcement behind them.

An accredited auditor treats each objective the way a financial auditor treats a ledger. Assertions without evidence fail. This is why the certificate means something: it is adversarial verification, performed by someone whose accreditation depends on being rigorous.

Sovereignty is becoming platform-enforced, not documented

CADA’s four-tier framework changes how sovereignty gets demonstrated. A PDF describing your data-residency policy is documentation. A regulator asking a national bank to prove residency wants enforcement, something the platform guarantees by construction.

Per the CNCF analysis published in July 2026, teams building for sovereignty are converging on a specific pattern: Kubernetes plus OpenStack plus GitOps plus policy-as-code with OPA or Kyverno, rather than buying a hyperscaler’s “sovereignty” SKU. The reason is that policy-as-code turns a residency rule into an executable constraint. A Kyverno policy that blocks any workload from scheduling outside an approved region enforces the rule at runtime. It admits the workload or rejects it, and the rejection is logged. When the auditor arrives, the evidence is the policy engine’s decision history, not a signed attestation from a sales team.

This mirrors the argument in Agent Specs Are Governance Artifacts: the durable governance asset is the machine-readable, enforced constraint, not the human-readable description of intent. Sovereignty follows the same trajectory. The tier you claim is only as real as the policy that enforces it.

The asymmetry this creates

Two vendors bid on the same regulated contract. One says it uses AI responsibly and keeps data in-region. The other hands over an ISO 42001 certificate from an accredited auditor and a policy-engine audit trail showing every workload’s placement decision for the last twelve months.

The second vendor ends the argument. The buyer’s compliance team can forward the certificate to their own auditor and check a box that the first vendor leaves open. In regulated procurement, an open box is a stalled deal.

This asymmetry compounds. The organization that builds the management system once can certify repeatedly, extend coverage to new products, and answer every future audit from a standing position. The organization that treats governance as a slide has to reconstruct evidence under deadline pressure every time a regulator or a large customer asks. One has a standing audit posture. The other has a fire drill.

Do this now

Pick one AI system that touches a regulated customer or a regulated market, and run it against the nine ISO 42001 objectives as a gap assessment. Not to certify tomorrow, but to find out honestly which objectives you could already evidence and which are still only claims. For each objective that is only a claim, identify whether the evidence should come from a document, a log, or an enforced policy. Document answers satisfy the weakest objectives. Enforced-policy answers satisfy the strongest, and they are the ones a sovereignty regime like CADA will eventually demand.

The organizations that start this assessment now, while certification is a differentiator, will hold a signed advantage when it becomes a requirement. The ones that wait will build the same management system later, under audit pressure, at higher cost, and without the head start.

The claim was always easy. The proof is the product now.

---

This analysis synthesizes Trust You Can Verify: Figma is now ISO 42001 Certified (Figma, July 2026), and How data sovereignty is changing cloud native infrastructure design (CNCF, July 2026).

Victorino Group helps organizations turn responsible-AI claims into audit-ready evidence that survives third-party scrutiny. Let’s talk.