What You Need to Know
ISO/IEC 42001:2023 sets the requirements for an AI management system: the policies, roles, risk and impact assessments, controls and reviews that govern how you build or use AI. On a mid-market data platform, implementing it means building inventory, access control, lineage, logging and incident handling into the lakehouse, then running audits and reviews over that evidence.
- Conformance vs certification: you can follow the standard without a certificate; certification adds an external audit.
- Time: six to twelve months from gap assessment to certificate is the commonly reported range.
- Cost: the data controls largely ship with the platform; the real spend is owner time plus audit fees.
- Fit: it earns its keep when enterprise customers start asking for it in procurement.
ISO/IEC 42001 was published in December 2023 as a 51-page first edition from the joint ISO and IEC committee on artificial intelligence. ISO describes it as a standard that “specifies requirements for establishing, implementing, maintaining, and continually improving an Artificial Intelligence Management System (AIMS),” for organizations of any size that develop, provide or use AI.
The certification path is not quick. CASRAI’s certification guide reports that certification bodies and consultancies commonly quote six to twelve months from the first gap assessment to certificate issuance, and that organizations already running ISO/IEC 27001 tend to move faster because the management scaffolding already exists.
Most mid-market teams that ask us about it share the same starting position. Their AI touches data in several systems, nobody can produce one list of models and agents, and the logs an auditor wants sit in four tools with four retention settings. The standard does not fix that. It asks you to prove you have.
What Is the Difference Between Conformance and Certification?
Conformance means you run your AI management system the way the standard requires. Certification means an external body audits it and issues a certificate. You can do the first without the second, and many companies should, at least for a year.
ISO itself does not certify anyone. Its certification page states plainly that “ISO does not perform certification or issue certificates,” and that certification is performed by external certification bodies. It advises checking that a certification body is accredited before you hire one. Any vendor that implies ISO granted it a certificate has misread the process.
The audit runs in two stages. Schellman, an accredited certification body, describes Stage 1 as typically one to two days, with four to twelve weeks before Stage 2, which runs three to nine days or more. The certificate is valid for three years, kept alive by annual surveillance audits. Advisera reports implementation itself typically takes three to 12 months depending on company size.
To be clear about our own position: Analytics AIML is not a certification body and does not issue or imply certificates. We build the data controls and evidence an auditor examines.
What Does an AI Management System Contain?
ISO describes 42001 as a management system standard built around Plan-Do-Check-Act. Like other ISO management systems, it expects a defined scope, leadership commitment and an AI policy, risk and impact assessment, support resources, operational controls, internal audit, management review and continual improvement.
The controls sit in Annex A. CASRAI reports 38 controls in nine areas: policies, internal organization, resources, impact assessment, the AI system lifecycle, data management, information for interested parties, system use, and third-party relationships. A Statement of Applicability records which controls apply to your actual AI activities and why any are excluded. That document is where scope discipline lives or dies.
Notice how many of those areas are data problems. Resources, lifecycle, data management, system use and third-party relationships all ask questions a governed lakehouse answers with a query. That is the mid-market advantage: you do not need a large compliance function if the platform records the evidence as a byproduct of running.
In What Order Should You Build the AIMS on a Lakehouse?
Build it in this order. Each step produces the input the next one reads, and the Annex A areas are as reported by CASRAI, mapped by us.
| Build step | Annex A area it supports | Lakehouse control | Evidence the auditor sees |
|---|---|---|---|
| 1. Scope and inventory | Resources; AI system lifecycle | Models in Unity Catalog plus an agent inventory Delta table with owner, purpose and status | One queryable list of in-scope AI systems, matching the Statement of Applicability |
| 2. Classification and impact assessment | Impact assessment; data management | Governed tags and automated data classification on the tables each system reads | Impact assessment per system, tied to tagged sensitive columns |
| 3. Access control | System use; internal organization | ABAC row filters and column masks, owners set to groups, one service principal per agent | Exported grants and policies, reviewed on a schedule |
| 4. Lineage and data provenance | Data management; AI system lifecycle | Automatic column-level lineage from source through feature tables to model versions | Source-to-model trail for any system in scope |
| 5. Third-party AI routing | Third-party relationships | External models, agents and MCP servers reached only through Unity AI Gateway | Approved provider list, usage and guardrail logs |
| 6. Logging and retention | System use; information for interested parties | system.access.audit copied to a retained Delta table, decision records written by scoring jobs | Who accessed, changed or decided what, and when |
| 7. Incident handling | AI system lifecycle; system use | Incident table, gateway endpoint disable and principal revoke runbook, timed drills | Incident log with containment times and corrective actions |
| 8. Internal audit and management review | Policies; internal organization | Control metrics on one gold-layer dashboard, sampled monthly | Audit findings, review minutes and the trend behind them |
A few platform facts hold the table up. Unity Catalog captures lineage automatically to the column level for queries run on Databricks, though not through UDFs or path-referenced tables. ABAC policies, governed tags and data classification reached general availability on 13 May 2026, per Databricks’ announcement. Unity AI Gateway became generally available on 4 August 2026 with identity-aware access, rate limits, spend caps, guardrails and tracing across models, agents, MCP servers and tools. The audit log system table is Public Preview with a 365-day default, so copy what you need; our audit log retention guide covers how long to keep it.
Steps one through four are the same foundations we describe in the AI governance framework. What 42001 adds is the management layer on top: the Statement of Applicability, the documented impact assessments, the internal audit and the management review that prove the controls run.
What Goes Wrong on the Way to Certification?
- Scope too wide on the first pass. Teams put every model, report and spreadsheet macro in scope, then stall writing impact assessments for systems nobody uses. Scope the AI systems that touch customer data or a consequential decision, and grow from there.
- Policy written before the process is mapped. An AI policy that describes a workflow nobody follows fails the first internal audit. As a Lean Six Sigma Master Black Belt, I map the real process first, remove the waste, then write the policy that describes what now happens.
- Evidence assembled by hand. If every audit means screenshots and exported spreadsheets, surveillance audits become an annual fire drill. Evidence should be a query against tables the platform already writes.
- Data outside the catalog. A SaaS assistant wired to the CRM with a personal API key sits outside lineage, access policies and logs. Auditors find these by asking staff what tools they use. Our data governance consulting notes cover how to bring those flows in.
- Retention assumed, not configured. A 365-day default on a preview audit table does not cover a three-year certificate cycle. Decide retention per record type and schedule the copy job.
- Management review held once, for the auditor. The standard expects leadership engagement as a cycle. A review with no metrics behind it reads as theatre.
Should You Build Now, Buy a Platform, or Hold Off?
Build the AIMS on your lakehouse now when your AI workloads and the data behind them already run on Databricks, or will within two quarters. The deciding signal: an enterprise customer or prospect has asked for ISO/IEC 42001 or a roadmap to it in a security questionnaire. A conforming system with evidence in hand answers that question, and certification becomes an audit rather than a project.
Buy a GRC or compliance-automation platform when your AI runs across many systems you do not control and the work is mostly policy, attestations and vendor assessments. The deciding signal: auditors ask for sign-off trails more than for data evidence. Feed that platform lineage and logs from the lakehouse rather than duplicating them.
Hold off on certification, but keep conforming, when no customer has asked and your AI footprint is small. The deciding signal: your inventory lists only a handful of systems, none touching a consequential decision. Run the controls, hold one internal audit, and revisit in six months.
If the build path fits, our AI governance and compliance work starts with a gap assessment of your inventory, access, lineage and logs against the standard. This article is general information, not legal or certification advice; an accredited certification body makes the conformity call.
Frequently Asked Questions (FAQs)
What is ISO 42001?
ISO/IEC 42001:2023 is the international standard for an AI management system. It specifies requirements for establishing, implementing, maintaining and continually improving one, for any organization that develops, provides or uses AI.
Is ISO 42001 certification mandatory?
Not as a general rule. It is a voluntary standard, and demand usually comes from customers and procurement teams that want third-party assurance about how you govern AI.
How long does ISO 42001 certification take?
CASRAI reports a commonly quoted range of six to twelve months from gap assessment to certificate. Companies with an existing ISO/IEC 27001 system tend to move faster.
Who issues ISO 42001 certificates?
External certification bodies, not ISO. ISO states it does not perform certification or issue certificates, and recommends choosing an accredited certification body.
How many controls are in ISO 42001 Annex A?
CASRAI reports 38 controls in nine areas, from AI policies and impact assessment to data management and third-party relationships. Your Statement of Applicability records which apply to you.
Can a lakehouse produce the evidence ISO 42001 auditors ask for?
Much of it. Unity Catalog inventory, access policies, lineage and audit tables cover the data side; impact assessments, internal audit and management review remain documented human processes that read that evidence.