Throughline.
    Insights
    Compliance Planning

    ISO 42001 for AI-Native Startups: Where to Start Your AI Management System

    Build an AI management system around your real use cases: an agent register, risk and impact assessments, shared responsibilities, governance policy and Statement of Applicability.

    Paul Wenham
    2026-11-04

    Picture a small startup that gives its customers AI agents to work alongside their employees. It was founded on an AI opportunity, so AI isn't only what it delivers to users. It's how the company runs its operations, support and almost everything else.

    For a company like this, ISO 42001 covers a lot of ground. This guide sets out where to start building the AI management system, the core pieces to put in place and the order to build them, and what else the standard expects.

    If you're still deciding whether certification is worth pursuing, start with ISO 42001 for AI companies: what it covers and when it makes sense. This guide is the practical next step once you want to build the management system.

    What is ISO 42001, and why does it matter for an AI-native startup?

    ISO/IEC 42001 is the international standard for an AI management system: the policies, roles, processes and controls an organisation uses to govern how it develops, provides and uses AI. It's certifiable, and it shares the same management system structure as ISO 27001, so the two can run as one integrated system.

    For an AI-native startup, the standard reaches further than it would for most companies. An employee agent platform plays several roles at once:

    • AI provider to its customers, who deploy its agents in their businesses.
    • AI user internally, across operations, support and development.
    • AI customer of the model providers its product depends on.

    Each role brings different responsibilities, so it pays to be clear on them from the start.

    Where should you start?

    For a company already using AI, start with its real use cases, then assess them and build the policy around what you find. Many people will tell you to start top-down with the AI policy, because it defines why you use AI. For a company already using AI in practice, it works better the other way round:

    1. A register of AI use cases, agents and models, so you know exactly what you're governing. It also becomes the home for the next two steps.
    2. A risk and impact assessment of each entry, done together.
    3. Responsibilities between the platform, its customers and its model providers, mapped for each entry.
    4. An AI governance policy that rolls all of this up into how the company governs AI.
    5. A Statement of Applicability, last, setting out which controls apply and how.

    An AI-native company has already worked through the policy questions intuitively; they just aren't written down. Doing the register and risk assessment first gives you the context for what really matters in the policy. This is a practical building sequence, not a requirement to follow the standard's clauses in this order.

    What goes in an AI use case and agent register?

    With AI used so broadly, cataloguing every use case is the natural starting point, and it supports everything that follows. The register should identify each use case individually, and ideally each specific agent and what it has access to.

    FieldWhat it captures
    Use case or agentA unique name and ID for each AI application
    OwnerThe person ultimately responsible for its governance and oversight
    PurposeWhat it's for, and the outcome it's meant to deliver
    Scope and boundariesWhat it may and may not do, and the systems it can reach
    Authorised usersWho may use it
    Authorised changersWho may change its configuration, prompts, tools or model
    Data in and outWhat data it takes in, what it produces, and where that data goes
    Model and providerThe underlying model and the third party behind it

    The register is also the natural home for each entry's risk assessment, impact assessment and responsibilities, covered below.

    How far you go is a judgement call. Much casual, general-purpose AI use is hard to define one case at a time, and can be grouped. But every material application deserves its own entry.

    For an employee agent platform, the first entries might look like this:

    Use caseOwnerPurpose and boundaries
    Customer-deployed agentsHead of ProductAgents customers deploy in their own businesses, limited to the tools and data each customer approves
    Support agentHead of SupportAnswers customer questions from documentation; can't change customer accounts
    Coding agentsCTOWrites code for human review; can't merge or deploy without approval
    General AI assistantsOperations leadStaff use for drafting and research; no customer data

    For how agent ownership, access and changes translate into audited controls, see SOC 2 controls and criteria for AI agents in production. For specific risk-to-evidence chains, read three worked AI control examples: coding assistants, access-review agents and agents changing customer data.

    How do you assess the risks and impacts of each AI use case?

    Work through the register one entry at a time. For each use case, ask where it could go wrong, where that could cause harm, who would be affected, and what matters to them.

    For customers, that includes your terms of service, the risks to their business, their insurance and their personal data. Then come the risks specific to what the agent is meant to do. For an employee agent platform:

    RiskWhat it looks likeWays to mitigate it
    MisconfigurationAn agent deployed with the wrong scope, permissions or instructionsGuided onboarding; configuration guardrails built into the product
    UnavailabilityThe agent goes down when a customer depends on itClear service commitments; monitoring; contingency guidance for customers
    Erroneous changesThe agent changes data or takes actions incorrectlyGuardrails that hold the system's boundaries; logging; clear customer responsibilities
    Missing contextThe agent lacks the context to make the right callProducts that require context before an agent goes live; onboarding and training
    Unexpected judgementThe agent processes or decides something in a way the customer didn't expectClear terms of service; data flow diagrams; agreed responsibilities with each customer

    How is an impact assessment different from a risk assessment?

    Risk assessments consider what could go wrong from using AI. Impact assessments consider the effects on individuals and society. ISO 42001 asks for both, and they look at different things:

    Risk assessmentImpact assessment
    FocusWhat could go wrong, and its implications, from using AIThe effects of an AI system on individuals and society
    Typical concernsFinancial losses, compliance failures, reputational damage, AI straying from its intended purposeEthical, legal, social and privacy effects on the people the AI touches
    For an agent platformAn agent's error costs a customer money, and costs you the accountAn agent's decision treats a customer's employee unfairly, or exposes their personal data

    Run the two together, and house both in the register. Each flows logically from the purpose, scope and objectives the register defines for every use case, model and agent. That context is what tells you what the risks and impacts actually are.

    Keeping everything in one place also supports governance and oversight, and gives everyone involved a shared understanding of each AI use.

    How should responsibilities be split between the platform, its customers and model providers?

    Map responsibilities in the register for each use case, model and agent, then put them where each party can understand and act on them. The split differs from entry to entry. An internal use case puts no responsibilities on customers. Every entry carries some responsibility for the model provider, but the use case shifts how far it extends and how much you depend on it.

    Each party's responsibilities then need to land somewhere they'll be understood and acted on:

    PartyWhere their responsibilities land
    CustomersYour terms of service, covering the use cases customers rely on, whether that's one or many
    The platformYour acceptable use policy, security policies and AI governance policy, so the right people across the company understand and perform their roles
    Model providersThird-party governance: due diligence, a risk assessment of each provider, and mitigations such as model settings, access restrictions and limits on how far certain models are used

    This also helps you answer enterprise security reviews, where buyers need to understand what you control and what remains their responsibility.

    What should the AI governance policy cover?

    With the register and risk assessment done, the policy almost writes itself. Roll the use cases up and you can see what AI means for the company, and the governance needed to keep it under control. At a high level, the policy should set out:

    • Why the company uses AI, drawn from the register as a whole.
    • Ownership. Who owns each model and agent, and how that ownership is maintained.
    • Oversight. How the company keeps appropriate oversight of every material use case.
    • Risk assessment cadence. How often assessments are refreshed to stay current.
    • Customer support. Clear terms of service, onboarding, training, monitoring and guardrails for customers deploying agents.
    • Product direction. Guardrails built into the product, such as not letting customers deploy an agent without the context its purpose needs.
    • Responsibilities. Who is ultimately accountable, and who helps manage the system.
    • Audit. The internal and external audit requirements for the management system.

    The product direction point deserves emphasis. You wouldn't deploy an employee without training and confidence that they understand their role, and you shouldn't deploy an agent that way either. An agent platform can help its customers get this right.

    How do you build the Statement of Applicability?

    Build it last in this practical sequence. Once the register, assessments, responsibilities and policy are in place, they give you the context for which controls apply, and how.

    The Statement of Applicability works through the predefined controls in ISO 42001's Annex A. For each one, it records whether it applies, with a justification for any exclusion, how it's implemented, where the evidence lives and who is responsible. Together, that becomes the map of controls supporting the whole AI management system.

    ControlApplies?How it's implementedEvidenceOwner
    A.6.2.8 Recording of event logsYesEvery agent action is logged with its inputs, outputs and approvalsLogging platform and retention settingsCTO

    For an AI-native agent platform, expect many Annex A controls to apply; what varies is how each one looks in practice. Applicability and any exclusions depend on your scope, risk assessment and AI roles. Areas to examine include:

    • Data for development (A.7). A company that doesn't train or fine-tune models may scope down the controls on training data. Data quality and provenance still matter if agents act on customer data or retrieved context.
    • The development life cycle (A.6). A company that only uses third-party AI tools, without building or configuring its own AI systems, may exclude some development controls.
    • Customer relationships (A.10). An AI management system that covers internal use only may exclude controls relating to customers.

    An AI-native agent platform builds agents and serves customers, so expect broad control coverage. Auditors look closely at each exclusion, so justify every one against the register.

    What else does an ISO 42001 AI management system need?

    The pieces above form the core. The rest of the standard builds on them:

    ComponentWhere in ISO 42001What it means for an agent platform
    Context, interested parties and scopeClause 4Define what the system covers, who has a stake (customers, their employees, model providers, regulators) and your AI roles
    AI objectivesClause 6.2Set measurable goals for responsible AI, such as agent accuracy or incident rates
    Competence and awarenessClause 7Make sure staff understand AI risks and their responsibilities
    AI system life cycleAnnex A.6Requirements, verification and validation, deployment, monitoring and event logs for each agent
    Data for AI systemsAnnex A.7Quality, provenance and preparation of the data agents learn from or act on
    Information for interested partiesAnnex A.8Documentation for users, and how you tell customers about incidents
    Responsible useAnnex A.9Rules for how the company itself uses AI, matched to each intended use
    Internal audit, management review and corrective actionClauses 9 and 10Regular checks that the system works, leadership review, and fixing what doesn't

    Certification follows the same pattern as ISO 27001: a two-stage audit by an accredited certification body, then annual surveillance audits over a three-year cycle. Because the management system structures match, ISO 27001 and ISO 42001 can share governance processes. A company with SOC 2 can also reuse relevant evidence where the scope and requirements align.

    For choosing and combining frameworks, read SOC 2 vs ISO 27001. For certification support, see ISO 42001 with Throughline.

    Related reading

    Paul Wenham
    Written by
    Paul Wenham

    Former PwC manager, founder of AssuranceLab and former Sensiba partner and Chief Innovation Officer. Now founder of Throughline, focused on compliance and assurance for AI-native companies.

    More from Paul
    Compliance Planning · Audit & Assurance
    SOC 2 Cost and Timeline Benchmarks by Company Size

    Typical SOC 2 costs and timelines for startups, mid-sized companies and enterprises, by report type. Compare planning, readiness, observation, fieldwork and reporting, plus related costs in USD.

    Audit & Assurance
    How Australian Companies Should Choose a SOC 2 Audit Firm

    Australian companies need a SOC 2 report issued under the AICPA standards by a US-licensed CPA firm, delivered by a strong local team in Australian time zones. Here's how the provider options compare, and what to check.

    Compliance Planning · Audit & Assurance
    How Much Does SOC 2 Cost? The True All-In Cost, Category by Category

    A category-by-category SOC 2 budget in USD, covering audit fees, readiness, platforms, penetration testing, consultants, internal assessments and operational costs. Learn what is required, what is optional, and what a quote can leave out.

    Audit & Assurance
    SOC 2 Isn't Pass/Fail: What It's Really For

    SOC 2 is a risk report, not a certificate. Understand how enterprise customers use it, why having a report is only base camp, and how to turn your audit into a trust asset for growing revenue.

    Compliance Planning
    SOC 2 Readiness Assessments: Your Questions Answered

    What readiness should cover, when to involve your auditor, which gaps matter most, and how readiness, Type 1 and Type 2 can work as one continuous engagement. A practical guide to preparing for a report your customers can rely on.

    SOC 2
    What makes a SOC 2 report quality?

    Enterprise buyers use SOC 2 as an initial qualifier, then look deeper. A quality report is what carries you through that assessment and turns your compliance investment into real trust.

    SOC 2
    What makes a SOC 2 audit high quality

    Audit quality isn't service quality, and it isn't only professional standards. The real test is whether a reader of your report thinks: this tells me what I need to know.

    Audit & Assurance
    Getting Through Enterprise Security Reviews

    Enterprise security teams review you in the background and rarely give feedback. What they look at beyond SOC 2, what counts as a genuine pen test, and how a quality report does the heavy lifting.

    Agentic Compliance
    Agentic Compliance: AI-Native, Human-Led

    Compliance from first principles, the 95% problem, the compliance metadata AI needs, rethinking screenshots, spreadsheets and policies, and why audits should meet clients where they are.

    Last updated