ISO 42001 and AI Governance in Public Sector Tenders
ISO/IEC 42001:2023 is the international standard for an artificial intelligence management system, and it is the AI equivalent of what ISO 27001 is for information security. It is not currently a blanket mandatory requirement for UK public sector bidding. What is happening is that buyers increasingly ask AI suppliers to evidence how they govern models, handle data, keep humans in the loop and monitor for drift, and certification is the most straightforward way to answer that. If you hold it, say so early. If you do not, you can still score well by documenting the governance you already practise.
Key facts at a glance
- What ISO/IEC 42001 is
- The international standard for an artificial intelligence management system, published in 2023. Certifiable, and structured like ISO 27001 so the two align closely
- Is it mandatory to bid
- Not as a blanket UK requirement. It appears as a scored differentiator or a desirable rather than a pass or fail condition in most tenders today
- Who should seriously consider it
- Suppliers already holding ISO 27001, those selling AI into regulated sectors, and those competing in crowded technology lots where differentiation is thin
- What buyers ask instead
- How models are approved and tested, where humans stay in the loop, how drift and failure are monitored, and whether client data trains your models
- The question that trips suppliers
- Whether customer data is used to train or improve models. A vague answer here costs marks and a wrong one creates a contractual problem later
- Realistic timescale
- Months rather than weeks for certification, so it is a pipeline decision rather than something to start when a tender lands
The honest answer to "do we need it"
Usually not yet, and a consultancy telling every AI supplier to go and get certified before they can bid is selling urgency rather than advice.
What is genuinely happening is that buyers have moved from not asking about AI governance at all, to asking about it in general terms, to asking specific questions that a supplier without a written management system struggles to answer well. Certification is one answer to those questions. It is not the only one.
Where it does change the calculation: if you already hold ISO 27001, much of the management system structure, the risk methodology, the internal audit cycle and the management review, is already in place, and the incremental work is smaller than a first certification. If you are selling into a crowded lot where every bidder claims the same capability, a certificate is a differentiator an evaluator can verify.
Where it does not: if your pipeline is a handful of contracts with buyers who are not asking, spending months and a fee to answer a question nobody has posed is the wrong sequencing. Document what you do, bid, and let the tenders tell you when the market has moved.
What AI suppliers actually get asked in tenders
How do you decide a model is fit to deploy. Buyers want a process with a named decision maker, not a statement that models are tested thoroughly.
Where does a human stay in the loop. Which decisions the system makes alone, which it recommends, and who can override it. For anything touching individuals, this is usually the highest weighted part of the answer.
How do you monitor after deployment. Drift, degradation, changing input distributions, and what triggers retraining or rollback. Saying you monitor performance is not an answer; naming the metric, the threshold and the response is.
What happens when it gets something wrong. Incident handling, notification, and remediation. Buyers assume failures will occur and are scoring how you handle them rather than whether you claim they will not.
Is our data used to train your models. Asked almost universally. The answer must be unambiguous, must match your contract terms and privacy notice, and must survive a procurement lawyer reading all three.
What third party models sit underneath. If your product calls an external provider, buyers increasingly want to know which, where processing happens, and what happens if that provider changes terms or withdraws the model.
Bias, fairness and accessibility. What you test for, how, and what you found. An honest account of a limitation you have mitigated scores better than a claim of no bias, which no evaluator believes.
The wider assurance picture
Central government has been moving towards published transparency for algorithmic tools used in public decision making, through the Algorithmic Transparency Recording Standard. If your product supports a decision affecting members of the public, expect the buyer to need a record they can publish, and expect them to ask you for the content of it.
Government has also published guidance for departments on buying and using AI. The practical effect for a supplier is that buyers arrive with a checklist rather than curiosity, and vague answers are more visible than they were two years ago.
In healthcare the stack is denser again. Digital technology assessment criteria, clinical safety standards, and, where a tool has a medical purpose, regulation as software as a medical device. An AI supplier entering NHS procurement without understanding which of these apply will fail on compliance rather than on capability.
If you sell into the EU as well as the UK, the EU AI Act imposes obligations that scale with risk classification, and buyers on both sides increasingly ask suppliers to state where they sit under it.
None of this is a reason to avoid the market. It is a reason to treat governance evidence as a bid asset built once, rather than as something assembled under deadline for each submission.
What we do for AI and machine learning suppliers
Write the governance and assurance answers. These recur across every AI tender with minor variation, so we build them once from your actual practice, with named owners, document references and dates, and they become reusable.
Translate engineering reality into evaluator language. The gap for most AI suppliers is not that governance is absent; it is that it lives in engineering process and nobody has written it in the form a procurement evaluator scores.
Handle the data protection and transparency sections properly, including the training data question, so the answer is consistent across the bid, the contract and the privacy notice.
Tell you when certification is worth pursuing and when it is not, based on what the buyers in your actual pipeline are asking for rather than on what would generate work.
Run the rest of the submission too: framework applications, the compliance matrix, social value, and portal submission, because AI suppliers lose on those as often as on anything technical.
A note on accuracy
This area is moving quickly. Standards are revised, guidance is republished and requirements that were desirable become mandatory. Anything on this page describing what buyers currently ask reflects the position as we see it across live tenders, and we would rather you verified a specific requirement against the tender in front of you than took a general statement from any website, including this one.
If a tender you are looking at names a standard, a framework or an assurance requirement we have not covered here, send it over. Reading the actual requirement is a short job and it is more useful than a general answer.
Frequently asked questions
Do we need ISO 42001 to bid for UK public sector AI contracts?
Not as a blanket requirement. In most tenders today AI governance appears as a scored question or a desirable rather than a pass or fail condition. Certification is one way to answer it well, and documenting the governance you already practise is another. Check the specific tender rather than assuming either way.
What is ISO 42001?
ISO/IEC 42001:2023, the international standard for an artificial intelligence management system. It sets out how an organisation governs the AI it develops or uses: risk assessment, controls, oversight, monitoring and continual improvement. It is structured like ISO 27001, so organisations holding that standard find the overlap substantial.
We already have ISO 27001. Does that help?
Considerably. The management system structure, risk methodology, internal audit cycle and management review carry across, so the incremental work for ISO 42001 is smaller than a first certification. It also means you can answer information security questions in AI tenders from an existing certified position.
How do we answer governance questions without certification?
Document what you actually do, with specifics. Who approves a model for deployment, what testing precedes it, where humans remain in the loop, what you monitor after release and at what threshold you act, how incidents are handled, and what your data protection position is. Named owners, real process and honest limitations score; general assurance does not.
Buyers keep asking whether their data trains our models. What is the right answer?
Whatever is true, stated unambiguously, and consistent with your contract terms and privacy notice. Evaluators and procurement lawyers read all three. A hedged answer reads as a yes you are reluctant to give, and an answer that contradicts your own terms creates a problem after award rather than before it.
How long does ISO 42001 certification take?
Months rather than weeks, covering gap analysis, building or adapting the management system, an operating period producing evidence, then a two stage external audit. It is a decision to take against your pipeline, not something to start when a tender appears.
Does this apply to us if we only resell AI features in someone else’s product?
Often yes. Buyers increasingly ask suppliers to declare AI in the delivery chain, including models embedded in tools you resell or integrate. Being unable to say which third party models sit underneath your service is itself a weak answer, and an undeclared one discovered later becomes a contract management issue.
Is AI procurement really growing, or is it hype?
The volume of tenders explicitly naming AI is genuinely growing, and the governance questions inside otherwise ordinary technology tenders are growing faster. The practical implication for a supplier is less about chasing AI-labelled opportunities and more about being able to answer AI questions inside contracts you were bidding for anyway.
Send us the AI questions from your tender
If a tender is asking how you govern models, handle data or keep humans in the loop, send us those questions and what you actually do today. We will tell you whether you can answer them now and what is missing. Call 020 3668 5488.