Lancashire Digital • Kevin Lancashire, October 2026
Where Governance Ends and Reality Begins
ISO/IEC 42001 helps you govern AI. It does not control what public AI models tell your customers.
The Framework vs. The Wild
ISO/IEC 42001 requires monitoring, measurement and evaluation (Clause 9.1), and its controls on verification, validation and operation (Annex A.6.2.4, A.6.2.6) apply to the AI systems you operate. Public third-party models are not among them.
Where it fits in your risk assessment
Whether what they say about your products belongs in your AI risk assessment at all (Clause 6.1.2) is your organization's call. If you decide to monitor it, the evidence has to come from outside. The standard gives you the framework. It does not do the looking.
ISO/IEC 42001 vs. AI Answer Auditing
What it covers
ISO/IEC 42001: Documents policies, risk controls and workflows for the AI you operate.
AI Answer Auditing: Checks how public AI assistants answer the questions patients, doctors and customers ask.
Inside-Out vs. Outside-In
ISO/IEC 42001: Works inside-out: it governs how you develop, provide and use AI.
AI Answer Auditing: Works outside-in: it tests what ChatGPT, Claude, Gemini and Perplexity answer about your products, automatically via their APIs and, where it matters, captured directly from the apps.
Frameworks vs. Factual Accuracy
ISO/IEC 42001: Looks at frameworks, controls and the AI systems within your organization.
AI Answer Auditing: Looks at the factual accuracy of public answers against locked, approved source texts.
Assessment & Evaluation
ISO/IEC 42001: Assessed by the organization that operates the system, and a certification body assesses the management system.
AI Answer Auditing: Adds independent evaluation against approved source texts, with expert human review.
In the Wild: What Actually Happens
In a recent independent analysis of an OTC medicine, I tested how a public AI model (Gemini via its API, with and without web search) answers ten typical questions from patients and healthcare professionals, and cross-checked every answer against the approved product information. In Switzerland, that is the text published in Swissmedic’s AIPS; for centrally authorised medicines in the EU, the equivalent reference is the EMA’s SmPC (Summary of Product Characteristics).
Audit Evaluation Parameters
The audit checks every answer for three kinds of gaps: safety information that is missing or softened, factual statements that contradict the approved text, and claims that cannot be traced to the approved source at all. It also records whether the product is named in the first place.
Safety Information Softened
An age restriction in the approved text was missing from the answers, and on pregnancy and breastfeeding the answers said “generally yes” where the approved text calls for consulting a doctor or pharmacist first.
Factual Contradiction
The answers stated a compatibility that the approved text explicitly warns against.
Visibility Gaps
In a general category question, the model named the product in one phrasing and omitted it in another.
The Takeaway
In this analysis, the dosing information was correct in every answer. The gaps were elsewhere. None of these would show up in a process audit. All of them showed up in answers to everyday patient questions.
Two Layers, Not One
A common objection: “Our internal tech team has already built monitoring tools.” Those tools are valuable, and they are the first layer, not the whole structure.
For what public AI models say about your products, there is no mandated self-check and no external review. An internal script is a good first layer. It cannot substitute for independent, source-locked evidence when a regulator, board risk committee or compliance lead asks for proof of due diligence.
Independent verification requires no access to your internal IT, but it does require:
1. A locked ground truth: Every response is checked strictly against the approved Swissmedic or EMA text, not against what seems plausible.
2. Two evaluation layers: An automated judge scores every answer against the locked source text; in the full audit, a human expert verifies every finding against that text before it is reported.
3. A versioned response log: Model, date and verbatim answers documented every 14 to 28 days, so that drift becomes visible.
The Bottom Line
ISO/IEC 42001 is necessary. It is not sufficient. You cannot inspect from the inside what public AI says on the outside.
Claim Your Complimentary AI Answer Snapshot
Want an independent external look at what public AI says about your products? I am offering a complimentary AI Answer Snapshot for qualifying Pharma, OTC and MedTech companies: [5 / 10] questions, scored against your official Swissmedic AIPS or EMA product information. It is the first evaluation layer, enough to see whether the gap exists for your product.
Contact me with one product name to claim your complimentary AI Answer Snapshot.
Sources
- ISO/IEC 42001:2023 – Clause 6.1.2 (AI risk assessment); Clause 9.1 (Monitoring, measurement, analysis and evaluation); Annex A.6.2.4 (AI system verification and validation); Annex A.6.2.6 (AI system operation and monitoring)
- EU GMP Guidelines, Chapter 9 (Self Inspection); Commission Directive (EU) 2017/1572, Article 14
- EMA Guideline on good pharmacovigilance practices (GVP)
