ISO 42001, NIST, the EU AI Act and MAS agree on the capabilities but not the operating model. NKK's tiering matrix, six pillars and evidence chain turn them into decisions, owners and proof.
Written for: Chief risk officers, heads of AI and data, general counsel, internal audit leaders, board risk committees
12 chapters, 57 cited sources, 30 models and charts. The first 2 chapters are open; the rest unlock with the form.
Values in score (1 to 4).
Most surveyed businesses now report a responsible-AI policy: the share with none fell from 24% to 11% [1]. The surveys we reviewed do not measure whether businesses can show how an AI system was approved, who watches it, and what happens when it fails. Our reading is that the gap is missing decisions, owners and evidence, not principles.
Our thesis is that the major frameworks converge on the same capabilities, but none supplies the operating model that joins them. None of the nine instruments we compare gives an enterprise a ready-made method for turning its own use cases into control tiers. The EU AI Act supplies legal risk classes; Singapore's regulator (MAS) and US bank guidance SR 26-2 expect a materiality assessment but leave thresholds to the firm [2][3]; NIST "does not prescribe risk tolerance" [4]; ISO/IEC 42001 "avoids specific guidance on management processes" [5], and the preview we read has no tiering method (Annexes A and B unread). How much governance each use case gets is left to you.
This paper offers NKK's operating model: Six Pillars, a Tiering Matrix, a Policy-to-Evidence Chain, an Incident Escalation Ladder and a scorecard. The 2026 changes raise the bar. The EU moved Annex III high-risk duties to 2 December 2027 [6], US bank model-risk guidance now excludes generative and agentic AI [2], and MAS finalised AI risk guidelines on 7 October 2026 [3]. No source we reviewed shows that any governance model reduces incidents or losses; outcome evidence is self-reported.
The AI Index 2026 puts global average responsible-AI maturity at 2.3 on a four-point scale, where level 2 is "integrating" and level 3 means all necessary practices are in place. Its own reading is that "most organizations are still integrating RAI practices rather than having them fully operational" [1]. Maturity rose in every region between 2024 and 2025, but regional scores sit in a narrow band and no region reaches level 3 [1].
Values in score (1 to 4).
The obstacles respondents name point to know-how rather than principles. Knowledge gaps top the list at 59% (51% in 2024), followed by budget constraints at 48% and regulatory uncertainty at 41% [1]. Ownership is also unsettled. Information security is the most common primary owner of AI governance at 21%, ahead of risk and compliance at 19% and AI-specific governance roles at 17%; 5% of organisations have no designated owner [1]. A separate IAPP survey, fielded in spring 2024, found a different split (privacy 22%, legal and compliance 22%, IT 17%) [7]. The two surveys use different respondents, years and categories, so they cannot be ranked against each other. Our reading is narrower: no single function has become the default owner.
A separate EY survey tests control knowledge directly. It surveyed 975 C-suite leaders at companies with revenue above USD 1 billion in 21 countries and reported that only 12% (11% of chief risk officers) "answered correctly" on a question asking them to identify the appropriate controls for five AI risks [8]. The release does not define the scoring, so we cannot tell whether 12% means all five were right or an average per risk. It is one consultancy's survey of large companies, and we do not know how representative it is, so we use it as a direction-of-travel signal only.
of C-suite respondents answered correctly on a question about appropriate controls for five AI risks (11% of chief risk officers; EY does not define the scoring)
The vocabulary is part of the problem. A 2026 meta-review in Patterns synthesised 74 AI risk frameworks and found 1,725 distinct risks, and it attributes 38% of them to human decisions and 42% to the AI systems themselves [9]. The study classifies risks listed in frameworks by cause; it does not measure failures or incidents. Even so, a large share of the risks catalogued in these frameworks are attributed to human decisions, so governance is not a layer bolted on to the technology; it covers the decisions people make about it.
A policy states intent. A framework lists capabilities. An operating model specifies who decides what, on which criteria, using which workflow, producing which evidence, reviewed at which cadence. The test is practical: could a new risk analyst follow it on Monday and reach the same decision as a colleague?
The AI Index reports that, compared with 2024, 7 percentage points more organisations say their responsible-AI policies improved business outcomes and 8 points more report a drop in the number of AI incidents [1]. EY reports that respondents with real-time monitoring were "34% more likely to see improvements in revenue growth" and "65% more likely to see improved cost savings", and says the survey "suggests greater adherence to RAI principles is correlated with positive business performance" [8]. Both are self-reported and neither establishes cause. We found no study showing that a particular operating model reduces incidents, losses or regulatory penalties. Everything below is therefore a design argument built on the requirements regulators and standards bodies have written down, not a performance claim.
The rest of this paper is organised around one frame. Every instrument discussed below describes some subset of the same capabilities in different words. The NKK Six Pillars arrange them in a single chassis, and the later sections follow their order. The foundation is the board-approved AI policy and risk appetite; the cap is a legal and standards crosswalk kept current. Ordering matters: each pillar produces an input the next one consumes.
Every figure and regulatory statement in this piece traces to a numbered source below. Links open the original publisher in a new tab.
The EU delayed its high-risk rules, not the board's duty to show it was watching. What the evidence says, what it does not, and an NKK framework of five lenses and a one-page dashboard.
Dated, source-checked status of AI regulation across the EU, US, UK, India, China and Asia-Pacific as of 13 October 2026, with three NKK models for managing regulatory change.
Prompt injection remains a residual risk, so agent security is a privilege-and-containment problem. Evidence, controls and three NKK models: a Threat-to-Control Map, Autonomy Tiers and an Agent Identity Stack.
