Opens in a new tab

Six valuable lessons from implementing ISO/IEC 42001

Six valuable lessons from implementing ISO/IEC 42001
Summary

ISO/IEC 42001 is moving from “nice to have” to a condition for doing business, and the gap between reading the standard and passing the audit is bigger than most teams expect. In summer 2026, Deeploy achieved ISO/IEC 42001 certification with BSI. We ran our own AI management system (AIMS) on our own platform, and BSI’s auditors assessed it independently. These are the lessons we’d have wanted before we started.

In this post:

  • six lessons from our own certification, from setting the scope to keeping up with change
  • what BSI auditors look for at Stage 1, Stage 2 and surveillance audits
  • a checklist to use before you book your Stage 1 audit

Why BSI’s perspective matters

BSI perspective

Trust isn’t claimed. It’s proven. Anyone can say they govern AI responsibly. An independent certification body tests that claim against ISO/IEC 42001. Its auditors follow real AI systems through their life cycle and ask to see the evidence.

BSI is the UK’s National Standards Body. Its information security standard, BS 7799, evolved into ISO/IEC 27001. BSI was also involved in developing ISO/IEC 42001. Today BSI works with more than 100 clients worldwide on ISO/IEC 42001 certification. Its global AI assessment team brings experience across sectors and geographies.

Accreditation checks the certification body in turn. In March 2026, BSI became the first certification body globally to achieve triple accreditation for ISO/IEC 42001. Together, the three accreditation bodies recognise BSI’s competence, impartiality and consistent approach to certifying AI management systems.

Accredited for ISO/IEC 42001 by
ANAB (US), UKAS (UK) and RvA (Netherlands)
ISO/IEC 42001 clients
More than 100 worldwide
Ali Moshayedi, Senior Digital QMS Specialist at BSI

After each lesson below, Ali Moshayedi, Senior Digital QMS Specialist, AI Services Delivery at BSI, explains what BSI auditors look for and why. You’ll find his perspective in a blue box like the one above.

Why ISO/IEC 42001 matters now

The pressure is coming from the supply chain first. Microsoft for example now asks suppliers that provide AI systems for independent proof that they manage AI responsibly, and accepts an ISO/IEC 42001 certificate as that proof (Microsoft SSPA Program Guide).

When a buyer that size makes a standard part of its buying rules, others will follow. The EU AI Act adds to the pressure, because much of what it requires overlaps with ISO/IEC 42001. So the question is no longer “should we?” but “how do we get there without drowning in paperwork?”

Our starting point

We went from kick-off to certification as a team that already held ISO/IEC 27001. Deeploy builds an AI governance platform, so customers rightly ask whether we practise what we sell. The certification was our answer.

We scoped the AIMS to how we use third-party AI internally: general-purpose assistants, coding agents and AI-powered process automation. Every AI use case was registered in Deeploy with an owner, a risk and impact assessment, and the controls it relies on. Policies, objectives and the Statement of Applicability (SoA) lived alongside our existing ISMS documentation.

The rough sequence was:

  1. Form the AIMS team and agree the scope.
  2. Write the AI policy, objectives, risk method and SoA.
  3. Register use cases and collect evidence of operation.
  4. Run an internal audit and a management review.
  5. Stage 1 audit (documentation), close findings, then Stage 2 audit (operation).

We were certified with a handful of minor findings to close in the first surveillance cycle. Those findings, and the ones we caught ourselves, are where most of the lessons below come from.

Three themes get the most space below: merging with ISO 27001 without losing the AI detail (Lesson 2), grounding risk in real incidents (Lesson 3), and keeping the system current as AI changes (Lesson 6). Lesson 5 shows how we used Deeploy to run the controls day to day.

Lesson 1: Scope is a strategic decision, not a formality

ISO/IEC 42001 asks a question ISO 27001 doesn’t: what is your role towards each AI system? Do you build it, provide it to others, or use someone else’s?

Your answer decides which controls apply:

  • If you build or provide AI, you must show how you manage its whole life cycle and inform its users.
  • If you only use AI from suppliers, the focus shifts to managing those suppliers.

Your scope sets your role, and your role sets the work.

Start with the scope you can prove works today

And write down when you will widen it. We deliberately started with our own internal use of AI, on use cases we fully control. That let us build a working system rather than a paper system across everything. The scope statement records when we will reassess it, which auditors expect to see.

A narrow scope is not a loophole. Check it against what your customers actually need: a buyer asking about the AI inside your product will not be satisfied by a certificate that only covers your internal tooling.

BSI perspective

How do auditors judge whether a scope is credible?

A credible scope is one that can be audited. At Stage 1, BSI auditors ask what each AI system does and what it’s for. They also ask what goes in, what comes out and who it serves.

Auditors then check that the documented scope matches what they see: the systems included, the boundaries by site and activity, and any exclusions. The scope statement also names the role for each system: producer, provider or user. The certificate records that role, the AI techniques used and the industry they’re applied in.

Common scoping mistakes BSI auditors see at Stage 1:

  • Scoping the future. A product planned for 2028 can’t be audited today.
  • Open-ended statements, such as “all public sectors”. A scope needs clear boundaries.
  • A scope that doesn’t match reality. The walk-through shows more than the agreed scope, such as an extra AI system used for sensitive purposes. That can change the time needed for the Stage 2 audit.
  • No defined role, even though ISO/IEC 42001 requires organisations to determine one.

An independent auditor tests the scope against what’s running, so the certificate reflects real AI use.

Why independent assessment matters

Lesson 2: Merge with ISO 27001: share the machinery, keep the AI detail

Already certified for ISO 27001? Then you have a head start. Processes like document control, internal audit and management review work for both standards. What you can’t copy is the AI content: your AI stakeholders, risks, impacts and controls. Knowing which is which is most of the work.

Our pitfall was sharing too much. A combined stakeholder analysis, for example, easily misses which requirements the AI management system itself addresses. Incident processes often have the same gap: they work, but don’t have the detail AI incidents need.

This is the split we ended up with:

ElementReuse from your ISMSMake AI-specific
Document control, internal audit, management review, corrective actionsThe process, tooling and backlogAI inputs on the management review agenda
Context and interested partiesThe analysis formatAI stakeholders, their requirements and which ones the AIMS addresses
Risk assessmentScoring method and registerImpact types for AI, plus impact on people and society
Supplier managementThe onboarding processModel terms: training on your data, model changes, AI service levels
Incident managementThe processExplicit scope for harmful output, misuse and model failure
ControlsOverlapping controls, mapped onceAnnex A controls for AI life cycle, data and impact

One practical habit that helped: a single table of context topics, each mapped to both its ISO 27001 and ISO/IEC 42001 controls. It shows the overlap, and the AI gaps, at a glance.

BSI perspective

What does a well-integrated ISO/IEC 27001 and ISO/IEC 42001 system look like?

ISO/IEC 27001 and ISO/IEC 42001 share the Harmonised Structure used across ISO management system standards. That lets organisations share processes such as document control, internal audit, management review and corrective action. They can also share the format of the interested parties analysis.

In an integrated system, one internal audit programme covers both standards. One management review covers the performance, findings and corrective actions of both the AI management system and the information security management system, with top management present for both.

The AI-specific processes can’t be shared, and that’s where integration can create risk. ISO/IEC 42001 asks for an AI risk assessment. A register built for business or information security risk isn’t enough. The AI system impact assessment has no ISO/IEC 27001 equivalent. The life cycle and data controls are specific to AI, so they can’t simply be mapped across from information security.

An independent audit looks at what’s shared and what’s AI-specific, so integration doesn’t hide gaps.

Why independent assessment matters

Lesson 3: Ground your risk assessment in real incidents

A shared risk taxonomy and real incident data moved our AI risk assessment from guesswork to real-world insight. Without them, most teams either start from a blank page and end up with the same three risks (privacy, security and hallucinations), or rely on static benchmarks that say little about how AI behaves in real use.

Start from a taxonomy, not a brainstorm

We use the AI Risk Repository from MIT FutureTech. It brings together more than 1,700 AI risks from 65 frameworks and sorts them into seven domains:

DomainExample risks
Discrimination and toxicityUnfair treatment of groups, exposure to toxic content, unequal performance across groups
Privacy and securityLeaking sensitive information, attacks on the AI system itself
MisinformationFalse or misleading output, pollution of the information ecosystem
Malicious actors and misuseFraud and scams, disinformation at scale, cyberattacks
Human-computer interactionOver-reliance and unsafe use, loss of human agency
Socioeconomic and environmentalConcentration of power, effects on jobs, environmental harm
AI system safety, failures and limitationsLack of robustness, lack of transparency, risks between interacting agents

Why it is built this way. The repository is a meta-review. Instead of adding yet another framework, MIT FutureTech collected the risks named in existing frameworks and sorted them into one shared structure, which gives risk owners, auditors and regulators a common language.

Next to the domains, a causal taxonomy classifies each risk by three questions: who causes it (a person or the AI), whether it is intentional, and whether it arises before or after deployment. Those answers point to the kind of control you need. An unintentional failure after deployment calls for monitoring; deliberate misuse by a person calls for access control and usage rules.

We tag every risk in our register with one of these domains. An empty domain then stands out as a question to answer, not a gap nobody noticed.

Then check what actually happens

MIT’s AI Incident Tracker sorts around 1,600 real incidents by the same domains and by severity of harm. Looking up incidents per risk grounds your likelihood and impact scores in reality, and shows which controls would have helped.

This is how the theory became concrete for us. Incidents of professionals relying on made-up AI output, such as court filings citing cases that don’t exist, made over-reliance a real risk for our own use of AI. It led to two treatments: training on over-reliance, and a guardrail that flags legal and regulatory questions.

Close the loop

Log your own AI incidents and near misses with the same domains, so internal and external lessons sit side by side. Then trace every context issue to a risk, and every risk above your threshold to a treatment with an owner and a date.

This is where most gaps appear: an issue is named in the context analysis but never becomes a risk, or a treatment is agreed but never evidenced.

BSI perspective

Which AI risks and impacts do organisations overlook?

  • AI-specific risk. A register that only covers business risk, or risk from other management systems, falls short of ISO/IEC 42001. The standard asks for AI risks to be identified and treated through life cycle controls. These include verification, validation and data quality checks.
  • Risk that spans life cycle stages. A risk can start in one stage and affect others: incorrect training data, for example, affects every output after release. The standard expects treatment to cover the whole life cycle, not just go-live.
  • Impacts beyond the direct user. ISO/IEC 42001 asks for impact assessments to cover individuals, groups and society.
  • Suppliers. A supplier’s patch or model update can change how an AI system behaves. Auditors look for evidence that those risks are identified and that responsibility for them is clearly shared.

An independent auditor brings a fresh view of risks that teams close to the system can miss.

Why independent assessment matters

Lesson 4: Read Annex A as a system, not a checklist

The Statement of Applicability is where auditors test whether you understood the standard. Treat it as the bridge between your risks and your controls, not as a list of ticks.

Work from risk to control

For every Annex A control, you should be able to say which risk or requirement makes it necessary, or why none does. A default control framework gives you a fast start, but the decisions per control are yours.

Watch the umbrella controls

Several Annex A groups open with a control that sets the overall process or objective, followed by detailed controls that sit underneath it. If you mark a detailed control as applicable, the umbrella above it almost always applies too. Excluding the umbrella while keeping its children is an easy inconsistency for an auditor to spot.

Umbrella controlWhat it setsDetailed controls that depend on it
A.4.2 Resource documentationWhich resources your AI systems rely onA.4.3 to A.4.6: data, tooling, computing and human resources
A.5.2 AI system impact assessment processHow you assess impactA.5.3 to A.5.5: documenting impact, impact on individuals and on society
A.6.1.2 and A.6.1.3 Responsible development objectives and processesHow you develop AI systems responsiblyA.6.2.2 to A.6.2.8: requirements, design, validation, deployment, operation, documentation and logs
A.7.2 Data for development and enhancementHow you manage data for AI systemsA.7.3 to A.7.6: data acquisition, quality, provenance and preparation
A.9.2 and A.9.3 Responsible use processes and objectivesHow you use AI systems responsiblyA.9.4: intended use
A.10.2 Allocating responsibilitiesWho is responsible across the AI supply chainA.10.3 and A.10.4: suppliers and customers

Before you finalise the SoA, read each group from the top down and check that its umbrella is consistent with what sits below it.

BSI perspective

What separates a strong Statement of Applicability from a weak one?

A strong Statement of Applicability (SoA) reads as the output of the risk assessment process. For each Annex A control, auditors expect to see whether it applies, why, and which risks it treats. For example, a risk register may list a bias risk treated by testing across different groups. The life cycle control covering that testing is then marked as applicable, with the link visible.

One weakness auditors see is excluding a control without a reason that fits the scope and role. An organisation that only uses AI and trains no models can justify excluding some life cycle controls. But exclusion is never automatic: if data is selected, prepared or fed into a system, data controls often still apply.

An independent auditor reads the Statement of Applicability against the risk assessment, so inconsistencies between them come to light.

Why independent assessment matters

Lesson 5: Run the controls where the AI work happens

The audit is won or lost on evidence that your AI systems are managed day to day. That evidence is far easier to produce when the controls sit next to the AI use cases themselves, not in a separate document.

We split the AIMS into two layers. The management system itself (policy, objectives, context, clauses 4 to 10) lives in our documentation. The Annex A controls are applied per use case in Deeploy, our own AI governance platform. That meant the operational side did not have to be built from scratch:

What the standard asks forWhat Deeploy provides
Annex A controls applied consistentlyA default ISO/IEC 42001 control framework on every use case, with evidence attached per control
Documented assessmentsTemplates for use case intake, AI impact assessment, verification and validation, and vendor assessment, plus an EU AI Act risk classification
Proof that controls operateAutomated checks, for example that monitoring alert rules are configured, and approval rules before a deployment or change goes live
Ongoing reviewScheduled use case reviews per workspace, with reminders to owners
Evidence for auditors and customersAll evidence per use case in one place

Where this is heading: evidence that collects itself

Most evidence is still assembled by hand: someone finds it, writes it up and attaches it to a control. Yet the runtime traces of an AI system already show much of what an auditor asks for, such as whether guardrails fire, whether the system is used as intended and how it performs.

We are building Deeploy to check controls automatically against those traces, so evidence comes straight from production instead of being put together for the audit. Anyone working with AI agents will also be able to declare what they use and propose evidence straight from their agent, through the Deeploy MCP server. People stay in charge of approval: automation proposes, a person accepts.

Make one complete record the condition for going live

Whatever tooling you use, require one complete record per use case before it goes live:

  • a named owner
  • the risk and impact assessment, kept as evidence
  • how the system is monitored and supported
  • validation before go-live
  • approved changes

It turns documentation into a habit and gives the auditor one place to sample from.

The gap most teams miss is AI they buy. It is tempting to rely on supplier due diligence for monitoring, updates and support, but the standard (control A.6.2.6) expects you to document how you operate every AI system in scope, bought or built.

BSI perspective

What evidence of operation do auditors look for at Stage 2?

Stage 1 asks whether processes are in place. Stage 2 asks whether they run effectively. Auditors choose in-scope AI systems using risk-based sampling and follow them through. The evidence can include, but isn’t limited to:

  • completed risk and impact assessments
  • the validation results that the procedure calls for
  • evidence that data quality and preparation checks were done
  • monitoring dashboards and event logs
  • the tickets and release records behind recent changes

A template or procedure on its own shows design, not operation. Auditors also speak to people, including top management and a sample of the engineers or users who work with the AI systems.

How much history is enough for a new management system? By Stage 1, auditors expect to see at least one completed internal audit and one management review. By Stage 2, auditors expect records for at least one in-scope AI system. It will have completed the life cycle stages that match the role in the scope. For example:

  • Producer: at least one build.
  • Provider: at least one system made available to others.
  • User: real operational use.

Independent assessment asks for evidence that processes run, not only that they’re documented. That’s what gives buyers confidence in the certificate.

Why independent assessment matters

Lesson 6: Build for change, not for the audit date

AI changes every month; a standard can’t. That is why ISO/IEC 42001 uses deliberately broad definitions. What counts as an AI system, what counts as development, what counts as a significant change: the standard leaves those lines for you to draw.

Draw them well and your AIMS keeps pace with the technology. Draw them once and forget them, and it is out of date before you realise it.

This is where we drew those lines. Use them as a starting point for your own:

Open term in the standardWe draw the line when…Then we…
AI systemother people use it, or it automates a recurring taskregister it as a use case. One-off personal use, like drafting an email, does not need a separate use case.
Developmenta change can alter what the AI outputs or who sees ittreat it as development, even if no model is trained.
Significant changethe model, intended use, level of autonomy, users or configuration changesreopen the risk assessment, the impact assessment and the controls.
Regular reviewa quarter has passed, or an incident has happenedcheck the use case on value, impact and compliance.

Development shows how fast these lines move. Many organisations mark the life cycle controls in Annex A.6 as not applicable because they never train a model. Yet building prompts, agents, tools and integrations on top of a model is AI system development, and agentic AI has made that the norm. The definition didn’t change; what falls under it did.

Make it run by itself

  • Put reviews on reminders, and let approval rules stop a changed use case from going live before it is reviewed. Deeploy has both built in.
  • Bring the outside world into your management review: supplier release notes, new AI features from existing vendors, regulation, and new risk categories such as multi-agent risks.
  • Revisit your scope as your use of AI grows.
BSI perspective

How do auditors judge whether an AI management system keeps up with change?

Keeping up with change. BSI auditors test this on real changes. They select a recent one, such as a model swap, a new feature, a dataset change or a new customer. Then they trace it through three questions:

  • Was the change planned and reviewed?
  • Did it go through the release and change process, with records?
  • Did it trigger a new risk and impact assessment, where the organisation’s own definition of “significant change” says it should?

A change that bypassed the process can be raised as a finding, even if the change itself worked.

What surveillance audits look for. Stage 2 shows the system works; surveillance shows it keeps working. Surveillance audits take place at least once a year between initial certification and recertification. At each one, auditors expect to see:

  • corrective action on Stage 2 nonconformities, with evidence that it was effective
  • documentation updated to reflect changes since Stage 2
  • a further internal audit and management review, at the planned intervals
  • risk and impact assessments revisited for changed or new AI systems
  • AI objectives tracked, with a record of results

Using versus developing AI. ISO/IEC 42001 requires each organisation to determine its role for each AI system. BSI auditors then check that the stated role matches what they see in practice. An organisation can hold different roles for different AI systems.

  • Producer: designs or trains models or builds them into an AI system.
  • Provider: makes an AI system available to others, including inside its own organisation.
  • User: uses an AI system without significantly changing its intended use or its life cycle.

Building agents, tools or integrations on a third-party model counts as AI system development once the organisation shapes what the system does. That development can fall within the production or provision of AI systems, or both. Annex A.6 controls can still apply when no model is trained, because they cover the requirements for every stage of the AI system life cycle, not just training. Which controls apply depends on the organisation’s operational risks, and these differ from one organisation to another.

Organisations set their own roles and their own threshold for significant change. An independent auditor checks both against what’s actually running at every surveillance audit.

Why independent assessment matters

When to book your Stage 1 audit

If you can tick these, you are ready for the auditor.

  • Scope names your role (user, developer, provider) and when you will review it.
  • AI system development is defined, and the SoA reflects it.
  • Shared ISMS processes are reused, while AI-specific context, interested parties and risks are kept separate.
  • Every SoA control, included or excluded, has a justification, and umbrella controls are consistent.
  • Every AI use case, bought or built, has an owner and one record holding its controls, assessments and evidence.
  • Risks are tagged to a taxonomy and checked against real incidents, and each one above threshold has an implemented treatment.
  • Significant-change triggers and review dates are set for every use case.
  • Internal audit and management review are done, with findings fed into corrective actions.

We can help you get
Stage 1‑ready

We went through it ourselves, with BSI.
Let us help you do the same.

Getting ISO/IEC 42001
certification through Deeploy
Tim Kleinloog
About the author

Tim Kleinloog

Co-founder · CTO

Tim Kleinloog is co-founder and CTO of Deeploy.

Want more? Download
the whitepaper

Co-developed with 12+ partners including Deloitte and BearingPoint. The question isn't whether you'll need robust AI governance, it's whether you'll build it in time.