ISO 42001: when AI stops being an experiment and starts requiring management

ISO/IEC 42001 doesn't judge whether your AI model is any good. It checks whether your company knows where it uses AI, who is accountable for each use, and what happens when something goes wrong. For most mid-sized companies that already have a chatbot live, a forecasting model running, and half the team using generative assistants on their own, that second question is harder to answer than the first.

ISO and IEC published the standard in December 2023, and it is the first certifiable international standard for an AI management system (ISO, 2023). Our position, after using it in governance work: for a mid-sized company, certification is rarely the first step. Using the standard as an organizing roadmap almost always is.

What the standard actually asks for

42001 follows the same structure as other management system standards many companies already know, such as ISO 9001 (quality) and ISO/IEC 27001 (information security). The main body covers organizational context, leadership, planning, support, operation, performance evaluation, and improvement. If your company has been through a 27001 audit, the skeleton will look familiar.

The AI-specific part sits in Annex A: 38 controls grouped under nine objectives, A.2 through A.10, covering AI policy, internal organization, resources, impact assessment, system life cycle, data, information for interested parties, responsible use, and third-party relationships (Vanta). You are not required to apply all of them. You assess your risks, select the controls that apply, justify the ones you exclude, and record those choices in a Statement of Applicability.

Day to day, that becomes very concrete:

• An inventory of AI systems: what exists, who uses it, with which data, and for which decision
• A named owner for each system, with the authority to pause it
• An impact assessment before anything goes to production, especially when the system affects people
• Documented criteria for selecting and monitoring model and API vendors
• Post-launch monitoring, with a log of incidents and what was done about them

A family of standards has grown around it. ISO/IEC 42005:2025 guides AI system impact assessments, and ISO/IEC 42006:2025 sets requirements for the bodies that audit and certify against 42001. That matters to buyers: the certification market is maturing, and a certificate from an accredited body now carries more weight than it did in the standard's first year.

Who has certified, and why it reaches you

The first certificates came from vendors. AWS announced 42001 certification in November 2024 for Amazon Bedrock, Amazon Q Business, Amazon Textract, and Amazon Transcribe (AWS, 2024). Anthropic, which builds Claude, announced its own in January 2025, valid through January 2028 with annual surveillance audits (Anthropic, 2025).

The effect on your company runs both ways. You gain an objective criterion for vendor selection: asking whether a model provider holds 42001, and what the exact scope of the certificate is, is a legitimate procurement question. And if you sell AI-enabled software or services to large companies or regulated industries, the same question will come back to you in a due diligence questionnaire.

The Brazilian context

In Brazil the standard is voluntary. It does, however, map directly onto three fronts that already affect companies operating there.

The first is the LGPD, Brazil's data protection law. Article 20 of Law 13.709/2018 gives individuals the right to request review of decisions made solely through automated processing of personal data, including consumer and credit profiling, and requires the controller to provide clear information about the criteria used. If one of your models denies credit, prioritizes service, or segments customers, you already need to be able to explain how it decides. The 42001 inventory and impact assessment are the organized way to have that answer ready before someone asks.

The second is the ANPD, the national data protection authority. It runs a regulatory sandbox on AI and data protection whose stated focus is algorithmic transparency (ANPD). That is a clear signal of where enforcement will look.

The third is the legal framework. Bill PL 2338/2023, passed by the Senate in December 2024 and now in the Chamber of Deputies, follows the logic of the EU AI Act: it classifies systems by risk level and places heavier obligations, such as algorithmic impact assessments, on high-risk ones. The final text may still change, but the direction is known. If you also operate in Europe, the AI Act has been in force since August 2024, with obligations phasing in over time.

None of these fronts requires ISO 42001. All of them get cheaper to meet when the company already has the basics the standard asks for.

Certify, or just adopt

Certification has a real cost: an internal team or consultants to build the management system, an audit by an accredited body, annual surveillance audits, and periodic recertification. There is also the risk of compliance theater, with a polished policy on the intranet and no control actually working in operations.

The rule we use with clients:

• Pursue certification if you sell AI as a product or service to large enterprises, operate in a regulated industry (financial services, healthcare, insurance), or have already seen it required in a tender
• Adopt without certifying if AI is an internal tool, the use cases are still few, and your biggest risk today is not knowing what is running
• Do neither before at least one AI system is in real use; at that stage the job is to validate a use case first

For transparency: Cortex IQ is not ISO 42001 certified. We use the standard as a working reference in governance engagements and say so on our trust page.

Example: a distributor with three AI use cases

This example is illustrative, built from situations common in mid-sized companies. A distributor runs a WhatsApp assistant that reports order status, a model that suggests purchase quantities to the procurement team, and sales reps who use generative tools to draft proposals.

Applying 42001 logic here requires neither certification nor a six-month project. It requires answering five questions for each use:

1. Who owns it? The WhatsApp assistant belongs to the customer service manager, not IT. The purchasing model belongs to the procurement manager.
2. What data goes in? The assistant looks up name, tax ID, and address, so it processes personal data under the LGPD. Proposals carry negotiated prices and terms, confidential information that should not pass through personal accounts on free tools.
3. What happens when it's wrong? A wrong delivery estimate creates a complaint. A wrong purchase suggestion creates a stockout or dead inventory. The impact differs, and so do the controls.
4. Which controls apply? For the assistant: a notice that the customer is talking to an AI, a clear path to a human agent, and conversation logs. For the purchasing model: a weekly comparison of suggested versus actual orders, with a threshold above which the buyer must justify the decision. For generative tools: a one-page usage policy and company accounts.
5. How is it reviewed? A 30-minute monthly meeting with the three owners, looking at incidents and metrics.

That covers, in lean form, much of what the Annex A objectives ask for. And it leaves the company ready to answer a large customer, an auditor, or a regulator without improvising.

A 90-day plan

1. Weeks 1 and 2: inventory. Ask every department which AI tools and models it uses, including those nobody formally approved.
2. Week 3: an owner for each system and one person accountable for AI governance overall, with time allocated to it.
3. Weeks 4 to 6: impact classification, starting with uses that affect customers, credit, pricing, or decisions about people.
4. Weeks 7 to 10: minimum controls on the highest-impact uses and an acceptable use policy that fits on one page.
5. Weeks 11 and 12: first formal review and a decision, with data on the table, on whether to pursue certification.

If you want outside support to run that cadence, that is the scope of Governance As a Service: weekly sessions on real decisions, documented criteria, and controls that work in operations. Our methodology page shows how it fits with our other engagement models.