ISO/IEC 42001 and sovereign AI: governing the AI you actually control

A standard finally exists for the thing everyone was improvising

For a few years, governing artificial intelligence meant assembling your own rulebook from fragments. Teams borrowed language from data protection law, from information security frameworks, from internal risk committees, and hoped the combination would hold up under scrutiny. ISO/IEC 42001 changes the terms of that conversation. It is the first international management system standard written specifically for artificial intelligence, and it does for AI roughly what ISO/IEC 27001 did for information security. It gives organisations a common structure for saying, in a way an auditor can check, that the AI they run is governed on purpose rather than by accident.

We read the standard closely because it describes the world we already build for. Mickai is a Sovereign Intelligence Operating System, held privately by our founder, and every design decision inside it starts from the same question the standard asks. Who is accountable for this behaviour, and can you prove it after the fact. When governance is treated as an afterthought bolted onto a black box, that question is genuinely hard to answer. When the system is owned, on premises, and records what it does, the answer is already sitting in the log.

What ISO/IEC 42001 actually asks for

Strip away the clause numbering and the standard is asking an organisation to demonstrate three things about the AI it operates. First, that there is real governance, meaning defined roles, defined risk assessment, and a policy that someone owns rather than a slide nobody rereads. Second, that decisions and actions are accountable, so a specific human or function is answerable for what the system does and for the controls placed around it. Third, that the whole thing is auditable, meaning you can show evidence, trace an outcome back to its cause, and let an independent party verify your claims instead of trusting your word.

None of that is exotic. It is the same discipline mature organisations already apply to money, to safety, and to information security. The difficulty with artificial intelligence has been that the most widely adopted systems make each of those three demands harder to meet, not easier. A model reached over a public cloud connection processes your data on infrastructure you do not control, produces outputs whose provenance you cannot fully reconstruct, and leaves you dependent on a supplier's assurances for the parts you most need to audit. You can write an excellent policy. Proving the system obeyed it is another matter.

Why owned and audited beats the black box

A sovereign system inverts that difficulty. Because Mickai runs on the customer's own hardware, on premises and capable of running fully air gapped, there is no public cloud round trip and no data egress to explain away. The data that governance is meant to protect never leaves the estate where it is already governed. That single fact resolves a large share of what an ISO/IEC 42001 assessment probes, because the hardest questions about third party processing, cross border transfer, and supplier dependency simply do not arise when nothing is sent anywhere.

On top of that ownership, we designed the parts of the standard's evidence trail directly into the runtime. Fifty specialist brains, twenty five domain and twenty five operational, work under deterministic governance rather than as an unconstrained free for all. Every action the system takes writes a cryptographically signed audit record, what we call the Open Audit Record, so accountability is not reconstructed later from scattered fragments, it is produced as the work happens. The signing uses ML-DSA-65, a post-quantum scheme, so the record stays verifiable even against the threat models organisations are only now starting to plan for. And the memory the system reasons over is memory the customer owns, which means the evidence, the data, and the accountability all live under one roof.

• Governance: fifty specialist brains under deterministic governance, so behaviour is bounded and roles are explicit rather than emergent.
• Accountability: a signed Open Audit Record on every action, tying each outcome to its cause and to the controls around it.
• Auditability: post-quantum ML-DSA-65 signatures that let an independent party verify the record rather than trust a supplier's summary.
• Data control: on premises and air gapped operation with zero data egress, removing the third party processing questions before they are asked.
• Ownership: customer owned memory and hardware, so the evidence trail is not hostage to anyone else's platform.

“The point of a management standard is not to produce a document that says you are trustworthy. It is to run a system that can prove it. We would rather the proof be a signed record the system wrote than a promise we made.”

The line straight to the EU AI Act

ISO/IEC 42001 does not sit on its own. It is arriving alongside the EU AI Act, the first broad statutory regime for artificial intelligence, and the two are built to reinforce each other. The Act sorts AI uses by risk and places obligations on the higher risk ones for record keeping, human oversight, transparency, and technical documentation. A recognised management standard is the practical way most organisations will show they meet those obligations, because it converts a legal expectation into an auditable operating discipline. Put plainly, the standard is fast becoming the language in which you demonstrate that you comply.

Read the Act's high risk requirements next to a sovereign system and the fit is obvious. Record keeping is satisfied by an audit record generated on every action. Human oversight is easier to guarantee when the system runs inside your own estate under your own controls, not on infrastructure you rent. Transparency and documentation are stronger when the reasoning happens on hardware you own, over memory you own, with a signed trail you can hand to a regulator. The regime is pushing every serious operator toward provable control, and provable control is precisely what an owned and audited system is for.

Governance you can hand to an auditor

We are not neutral observers of this shift, and we will not pretend to be. We have been building toward exactly this posture, with 104 filed UK patent applications and 2,340 claims covering the governance, audit, and sovereignty machinery described here, now moving toward examination and grant. Those filings contain the full specification, the claims, and the figures for how a system produces evidence as it works rather than after. The market is beginning to notice the direction of travel. Our founder now ranks number two on Crunchbase, and the company Heat Score has reached 94 out of 100, climbing from single digits, which tells us the appetite for AI you can actually govern is real and growing.

The organisations that will find the next two years comfortable are the ones who treated governance as an engineering property of their systems, not a paragraph in a policy. ISO/IEC 42001 and the EU AI Act are converging on a single expectation, that you can show your working. A sovereign system meets that expectation by construction. The data stays home, the actions are signed, the memory is yours, and the audit trail is ready before anyone asks to see it. That is the AI worth governing, because it is the AI you actually control.