The most consequential chart in Stanford’s AI Index 2026 for anyone selling software into a large enterprise is not about model performance. It is a survey question about which regulations influence responsible-AI decisions (Ch3, Fig 3.3.11).
| Framework cited as an influence | 2025 | Change |
|---|---|---|
| EU GDPR | 60% | −5pp |
| EU AI Act | 43% | +2pp |
| ISO/IEC 42001 | 36% | new entry |
| NIST AI Risk Management Framework | 33% | new entry |
| US Presidential Executive Order on AI | 21% | +2pp |
| OECD AI Principles | 16% | −5pp |
| No regulatory influence at all | 12% | −5pp |
Two frameworks appeared from nothing and went straight past the OECD Principles and the US executive order. In the same period, the share of organizations with no responsible-AI policy at all fell from 24% to 11%, and AI-specific governance roles grew 17% (Ch3).
That is what a compliance requirement looks like in the year it becomes normal. AI governance has moved out of the ethics committee and into the vendor questionnaire.
What are these three things, actually?
They get named together and they are not the same kind of object.
The EU AI Act is binding law. It tiers obligations by use case, not by technology, and it reaches you if your users are in the EU regardless of where your company sits. Its first measures, the prohibited practices, took effect on 2 February 2025, and obligations for providers of general-purpose AI models, covering risk assessment, transparency disclosure, and systemic-risk mitigation, took effect on 2 August 2025 (Ch8). It determines whether a feature can legally ship.
ISO/IEC 42001 is a certifiable AI management system standard. That word, certifiable, is why it climbed the chart so fast. A procurement team can ask for a certificate instead of trusting a questionnaire, and if you already hold ISO 27001 the scaffolding is familiar. It answers “do you manage your AI systems deliberately?”
The NIST AI Risk Management Framework is voluntary, American, and structural: Govern, Map, Measure, Manage, with a generative-AI profile added in 2024. There is no certificate. Its value is as the backbone of an internal AI risk program, and increasingly as a named requirement in US enterprise contracts.
One is a law, one is a certificate, one is a vocabulary. Buyers ask about all three, often in the same document, and frequently without distinguishing between them.
What buyers are actually asking
Strip away the framework names and enterprise AI due diligence converges on six questions. They are worth rehearsing whether you are answering them or asking them.
1. What decisions does this system influence, and about whom? The risk tier follows the use case. A model that summarises internal documentation and a model that scores insurance claims are the same technology and completely different regulatory objects. Insurance risk assessment and pricing for life and health cover, creditworthiness assessment, and employment screening sit in the high-risk tier, which matters directly if you build for insurers, lenders, or HR platforms.
2. Can a human overrule it, and does that actually happen? Nominal oversight that rubber-stamps model output satisfies nobody’s auditor. The question behind the question is whether overrides are recorded as overrides, because an oversight mechanism no one has ever used is indistinguishable from an absent one.
3. Can you reconstruct a decision from six months ago? This is where most engagements find their real gap. Reconstruction requires the model version, the prompt version, and the retrieval corpus version logged together with the output. Teams that log only the output discover this at the worst possible time.
4. Where does the data come from, and where does it go? Lineage, retention, deletion, and whether anything leaves the jurisdiction. GDPR is still the most-cited influence at 60% for a reason: the data questions did not go away, they acquired a second regulator.
5. What did your model vendor tell you, and can you verify it? Increasingly you cannot. The Foundation Model Transparency Index fell from 58 to 40 in 2025, with the largest gaps in training data, compute, and post-deployment impact (Ch3). The defensible answer is not “our provider assured us”; it is a design that does not depend on assurances you cannot check.
6. How do you know it still works? Accuracy thresholds agreed in advance, evaluated continuously, with a defined response when the system drifts below them.
The engineering behind the questionnaire
Governance obligations sound abstract until you translate them into the artifacts that satisfy them. Then they turn out to be an ordinary, buildable backlog.
| What the framework asks for | What you build |
|---|---|
| Traceability and record-keeping | immutable decision logs; model, prompt, and corpus versioned per inference |
| Human oversight | review queues, confidence surfacing, override paths recorded as overrides |
| Transparency to affected people | disclosure that AI is in use, plus a plain-language explanation surface |
| Accuracy and robustness | eval suites in CI, drift monitors, regression gates on model upgrades |
| Data governance | data contracts, lineage, retention and deletion paths |
| Risk management | documented failure modes, incident runbook, kill switch |
| Vendor risk | a model abstraction layer that makes provider substitution real |
None of that is exotic. Most of it is what a well-built system has anyway, which is the point worth making to a nervous board: the compliance backlog and the reliability backlog are largely the same backlog, and the work is reusable across all three frameworks because they are asking for the same evidence in different dialects.
The honest caveat, and it belongs in every conversation: the frameworks trade off against each other, and so do the properties they ask for. The AI Index reports empirical studies where training techniques that improved one responsible-AI dimension (safety, say) consistently degraded another, such as accuracy (Ch3). There is no configuration that maximises safety, fairness, privacy, and accuracy at once. Governance is the discipline of choosing the tradeoff deliberately and being able to show your work.
Why we think this favours European delivery
One more number, from the public opinion chapter. Across 25 countries surveyed by Pew, a median of 53% trust the EU to regulate AI effectively, against 37% for the United States and 27% for China (Ch9). The US reported the lowest trust in its own government to regulate AI of any country surveyed, at 31%.
For a team building regulated software from inside the EU, that is not a talking point so much as an operating advantage. The frameworks that are becoming global procurement defaults are the ones our clients’ regulators wrote, and the engineering habits they demand (traceability, oversight, data lineage) are the ones EU-domiciled teams have been building under GDPR for eight years.
If you are preparing for this kind of diligence, compliance-ready MVP architecture covers the same ground for earlier-stage products, and pilot to production in regulated enterprises covers the crossing itself. Our AI for regulated industries practice exists to make these answers demonstrable rather than aspirational.
All figures cited from the Stanford HAI Artificial Intelligence Index Report 2026 (9th edition); regulatory-influence data originates in McKinsey’s 2025 survey and is self-reported. This article is a technical and commercial reference, not legal advice. Regulatory phasing has changed before and may change again; verify current obligations against the primary texts before relying on a date.
