ISO 42001 generative AI work is the inventory entry for a feature that takes a prompt, may pull retrieval data, often calls a vendor model, and then says something a customer will act on. The management system has to name an owner, a purpose, and a record of how that output was tested. A policy PDF is not that record.

What ISO 42001 generative AI has to name

A feature that writes is four surfaces, not one model card. Someone can edit the prompt. Retrieval data can change the answer. A vendor can host the weights. The model may say something a person will act on. ISO 42001 generative AI work fails when the prompt library has no owner.

The AI management system, the AIMS, is still clauses 4 to 10. Annex A still has 38 controls, from A.2 through A.10. A generative feature does not get a private standard. It gets a row in the inventory, an AI risk assessment, and an AI impact assessment. Seven Step Consulting is a consultancy in Gurugram. It does not issue ISO certificates. A certification body does, after Stage 1 and Stage 2.

The prompt

Treat the system prompt as a controlled record under ISO 42001 generative AI, the same way you treat a configuration that changes behavior. Name who may edit it, where the current text lives, and what the last change was for. A file called Final_v7 on a laptop is not that record.

User prompts matter when they carry personal data, a client document, or an instruction that overrides the system text. You do not need to archive every keystroke on day one. You do need a rule for what is kept, for how long, and who can read it. Version the prompt. ISO 42001 generative AI will ask which version produced the answer you are defending.

If a product manager can change the prompt in a console with no second person, write that down as the real control, or change the console.

Retrieval data

Retrieval data belongs in the ISO 42001 generative AI inventory because it changes what the model says. Name the source, the purpose, the refresh, and who may add a document. A shared drive poured into an index is a data set with no provenance. Annex A asks about data resources, quality, and provenance for that reason.

Customer contracts, old tickets, and a wiki page that was never meant to be spoken aloud all become part of the output. If the index can see it, the model can repeat it. Log the prompt, the retrieved passage, and the output if ISO 42001 generative AI is going to be more than a policy.

When the index is rebuilt, ISO 42001 generative AI expects the impact and the risk to be looked at again. A new corpus is a change to the system, not a quiet job.

What the model may say

ISO 42001 generative AI asks what the model may say to a customer, not only what the demo said on a good day. Intended use is operational. Who the user is. What decision the words can influence. What the system must refuse. “Helpful assistant” is not a purpose line.

An evaluation set is part of ISO 42001 generative AI evidence when you claim the refusal actually works. A refusal that works in a test and fails after a prompt edit is a failed control.

ISO 42001 generative AI monitoring means someone reads failures, not that a dashboard exists. Name that person. Give a customer, an employee, or an affected person a way to report a bad answer. Keep one real report, or an honest note that none arrived, and keep the channel visible.

A red-team note, even a short one, is better ISO 42001 generative AI evidence than a paragraph of intentions.

How ISO 42001 generative AI treats a vendor model

A customer-facing feature needs an ISO 42001 generative AI entry even when the weights sit with a vendor. Your use is in scope. The vendor’s full training recipe may not be. The Statement of Applicability should mention supplier controls when ISO 42001 generative AI scope includes a rented model. Say what the contract lets you ask, and what you will do when the model version changes under you.

ISO 42001 generative AI does not let you mark a pilot out of scope once staff rely on it. A proof of concept that answers clients is already in production. ISO 42001 generative AI scope includes the tool a product manager pasted into a workflow if customers feel the result. A bought copilot belongs on the same list as a model you fine-tuned.

Personal data in a prompt is an ISO 42001 generative AI issue and a DPDP Act issue at the same time. As of 31 October 2026, Rules 1, 2 and 17 to 21 of the DPDP Rules, 2025 are in force. Rule 4 on Consent Managers commences on 13 November 2026. Rules 3, 5 to 16, 22 and 23, covering notice, security, breach intimation, rights and the main fiduciary duties, commence on 13 May 2027. Do not treat the full rulebook as already in force. Do not paste a client file into a vendor tool and call the index internal.

ISO 42001 generative AI controls for prompts, retrieval data and vendor models
Surface What fails a sample Record that holds
Prompt No owner and no version Current text, editor, last change
Retrieval data An index nobody reviewed Source, purpose, refresh, access
Vendor model A version change with no note Supplier, version, the contract question
What the model may say A demo and no refusal test Intended use, a dated test, a report path

Check that ISO/IEC 42001:2023 sits in the certification body’s accredited scope on the NABCB or IAF directory before you book dates. NABCB, under QCI, accredits management-system certification bodies in India. This article names no logos. Stage 1 can look tidy on the policy. Stage 2 opens the console. A surveillance audit later asks what changed in the prompt, the index, and the vendor.

Harm, not only downtime

GenAI risk management scores a wrong answer, a leaked prompt, and a retrieval hit that should not have been retrieved. GenAI risk management is not a renamed row from the ISO 27001 register. Confidentiality still matters. So does a confident sentence that sends a customer the wrong way.

GenAI risk management gives each risk an owner and a treatment you can show. GenAI risk management covers the prompt, the retrieval data, the vendor, and the words the model may say. GenAI risk management links to the AI risk assessment the AIMS already requires. ISO/IEC 23894:2023 guides that risk work. ISO/IEC 42005 guides the AI impact assessment when a person is affected: a borrower, a patient, a job applicant, a customer who will act on the sentence.

GenAI risk management expects you to retest when the model version or the index changes. GenAI risk management includes a path for a customer to report a bad answer. GenAI risk management stays in the register a sample can open, not in a chat channel. If the harm can reach a person, the impact note names that person in ordinary words.

LLM governance starts with the feature a customer can already touch, not with a model card you downloaded. LLM governance names who may change a system prompt, and who may not. LLM governance includes the retrieval index, because the answer is more than the weights. LLM governance records what the model is allowed to say, and what it must refuse.

LLM governance puts the vendor on the inventory when you do not host the weights. LLM governance asks for a test of the output, not a screenshot of a playground. LLM governance fails when a pilot in one team is invisible to the person who owns the product. LLM governance is the operating rule for that system inside the AIMS, not a second policy that contradicts the first.

The EU AI Act has applied most Annex III high-risk obligations since 2 August 2026. ISO/IEC 42001 does not by itself create a presumption of conformity. It gives an Indian exporter an inventory and a test record when a European buyer asks how the feature behaves. MeitY’s India AI Governance Guidelines, released 5 November 2025, are non-binding guidance, not a statute.

In 2026 India market ranges, not a Seven Step rate card, consulting for a single-product firm often sits between ₹4 lakh and ₹8 lakh. Several writing features often push the same work into the ₹8 lakh to ₹15 lakh band. A focused firm with a small inventory can be audit-ready for Stage 1 in about 90 days if leadership shows up. That is readiness, not a certificate. A full build often takes three to six months.

GenAI governance assessment

GenAI governance assessment. Bring one live prompt, the index source, and the vendor terms. We will say which of the four surfaces a sample would open first. Use https://www.sevenstepconsulting.com/contact-us/.