A Douglas firm made headlines this month for landing ISO 42001, the international standard for AI management systems, alongside ISO 27001, in a combined audit, with zero non-conformances on the first attempt. The press release calls it a first for the island. It probably is; nobody has published a register to check against.
What the release does not say, because press releases never do, is what the certificate actually attests to. So let's be precise about that, and then let's do something more useful than repeating the marketing: build the thing the certificate is meant to prove, for free, in an afternoon.
Published December 2023, ISO/IEC 42001 is barely two and a half years old. The accreditation infrastructure behind it is younger still; ANAB only opened its accreditation programme for certification bodies in January 2024, and even a year into that programme only a handful of bodies had cleared the process. One organisation was called out publicly for claiming to be first certified against the standard before the final draft was even published; an unaccredited claim that misrepresented what "certified" is supposed to mean. The standard is credible. The ecosystem around it is still finding its feet.
Structurally, 42001 is not a new species. It is built on the same Annex SL scaffold as ISO 27001 (information security) and ISO 14001 (environmental management): clauses 4 through 10 covering context, leadership, planning, support, operation, evaluation, and improvement, plus a domain-specific set of Annex A controls, 38 of them in this case, covering risk assessment, AI lifecycle management, and data handling. Any firm already holding 27001 can bolt 42001 on in a combined audit, because the audit machinery already exists: the PDCA loop, the internal review cycle, the Stage 1/Stage 2 process. That is precisely how the Douglas firm did it.
Here is the part that matters most and gets said least: ISO 42001 certifies the management system, not the AI. A certified firm can still ship a broken, biased, or dangerous model and remain fully compliant, because compliance is about the governance wrapper, not the payload. It is the same category as SOC 2 Type II being waved around as "we're secure"; what it actually says is "we have controls and we followed them for the audit period," and the controls chosen could be thin.
Where the scope gets gamed
Annex A has 38 controls; an organisation chooses which apply to its declared scope. Draw the AI management system boundary narrowly, say around documentation for internal tool use rather than around bias testing and human-oversight boundaries for a live product, and you can clear an audit cleanly while never being tested on the parts that are genuinely hard to evidence. Anyone who has been through a real combined 27001/42001 audit knows what "zero non-conformances on the first attempt" usually means: either the scope was microscopic, or the relationship with the certification body was remarkably cosy. A real, messy, living AI system always has friction points. A clean sheet on the first pass is not proof of discipline; it is a claim that needs the scope statement put next to it before anyone believes it.
It is not proof the AI works. It is not EU AI Act conformity; that is a product-level obligation for high-risk systems, a different exercise entirely, though a mature AIMS makes it easier. It is not, on its own, evidence of anything a client can act on beyond "this organisation has a documented, audited process." Whether that process is any good depends entirely on what got scoped in and what got scoped out; information the certificate itself does not carry.
We run The Haunted Lighthouse under Cyber Essentials, ICO registration, and a documented AI usage policy (THL-POL-AI-001) that has already been through several revisions. What we did not have, until a conversation started by this very news story, was a risk register mapping each AI use case to a documented control.
So we built one. Seven rows: infrastructure work on the primary server, the client demonstration platform, a two-way research-and-verification loop with a research assistant, the article production pipeline, live codebase work, an agentic task-delegation tool, and general use. Each row gets a likelihood, an impact, the control that already operates, and, critically, the gap that does not yet. Two rows came back exposed rather than clean: the verification loop, because the override decisions inside it were not being logged, and the agentic tool, because it had no written scope boundary for what it is and is not permitted to touch.
Those gaps stayed in the document. That is the tell. A register built to pass an audit gets scoped until the hard rows disappear. A register built to actually manage risk leaves them in, dated, waiting to be closed.
Total cost: an afternoon. Total certificate spend: £0. Total accredited third-party validation: none, which is the honest trade-off, and the one worth being clear-eyed about. What we have is not a stamp for someone else's procurement questionnaire. It is a working document that does the actual job the standard describes: demonstrating, on paper, that AI use across the business is identified, assessed, and controlled, and updated when it stops being true.
The obvious objection: this register only got written because a news story forced the accounting. Left to itself, without a looming auditor, does the document just gather dust? Fair challenge, and it deserves a straight answer rather than a concession. Something does have to force the accounting. £60,000 and an auditor arriving on a Tuesday morning is not the only thing capable of doing that. A deadline, a client's question, a public commitment to keep the register current, or simply the discipline of writing the article you are now reading, all did the job here for free. The forcing function matters. The invoice attached to it does not.
For most small firms, no, not yet. ISO 42001 earns its cost when a specific client or tender demands the accredited stamp; procurement teams in financial services, healthcare, and government are increasingly asking for exactly that. Certification has moved, in those sectors, from differentiator to gate, and it is a brutal, unthinking one. Nobody in that process reads your risk register. A procurement checklist does not care that your honest handling of a verification loop is better risk management than a competitor's narrowly-scoped certificate; it looks for a certificate number and moves on. In that world, the fee is not spent on risk management at all. It is protection money paid to the gods of bureaucracy, for the right to enter the room. If none of your pipeline sits in a room that demands it, £15,000 to £60,000 and three to six months of audit preparation buys you a document that says what you could have written yourself for nothing: that a governance loop exists and operates.
Write the register first. Certify only when someone with a purchase order asks you to.
The Haunted Lighthouse's AI use risk register, built in the course of writing this piece, sits alongside THL-POL-AI-001 in our compliance documentation and is available on request.
The Sovereign Auditor covers digital sovereignty, cybersecurity governance, and data protection policy—with particular focus on Isle of Man jurisdiction and Crown Dependency issues.
Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.
