Written by: Jason Veen
Published:
Eighteen months into the framework mapping, someone asks what the system is allowed to say. Not whether it is safe. Not whether it is biased. What it is allowed to say to a customer, in the organization's own terms.
The mapping does not answer that question. It was never built to.
This is not a failure of effort. The program is real work. Risk is mapped, ownership is assigned, trustworthiness is measured, decisions are documented. An auditor would find the file in order.
Then a customer receives an answer no instrument in the program ever evaluated. Nothing in the program defines what a good answer would have been.
The gap is not measurement. Measurement is required by name, in more than one place. The gap is the criterion the measurement runs against.
Key Takeaways
The EU regime in force in October 2026 requires that people be told they are interacting with an AI system. It also prohibits a defined set of practices. It evaluates no answer against a particular organization's own standard.
The high-risk regime is deferred to December 2, 2027 and August 2, 2028. Reading that as a durable condition rather than a short transition is this article's interpretation, not a Commission statement.
NIST AI RMF 1.0, which NIST states is under revision, requires output to be measured, interpreted in context and documented. It prescribes no content criterion.
ISO/IEC 42001:2023 specifies requirements for an AI management system. ISO states the standard does not look at the details of specific AI applications.
The standard of correctness is an input each of these instruments requires and none of them supplies. It belongs to the organization whose name is on the answer.
What an AI Governance Framework Covers
An AI governance framework is the set of policies, controls, roles and records governing how an AI system is built, deployed, measured and overseen. It establishes who is accountable, what risks are in scope, and what evidence survives the decision.
That is the whole of what the published frameworks do, and they do it well.
Every instrument an organization can map to governs the system. None of the three examined here governs the individual sentence. A complete framework therefore has three parts, not two.
The system, the output, and the standard the output is measured against. The published frameworks supply the first two. They require the third as an input they do not write.
Why This Matters Now, and Until the High-Risk Regime Applies
The dates below are from the European Commission's regulatory framework for AI, retrieved October 5, 2026. They apply in the European Union.
From August 2, 2026, the AI Office and the authorities of the Member States implement, supervise and enforce the AI Act. Rules for systems used in certain high-risk areas apply from December 2, 2027. For systems integrated into products such as lifts or toys, the rules apply from August 2, 2028.
The AI Omnibus was proposed by the Commission on November 19, 2025 and agreed on May 7, 2026. The European Parliament and the Council adopted it as Regulation (EU) 2026/1744 of July 8, 2026, and it entered into force on July 27, 2026.
The transparency duty was not deferred. Regulation (EU) 2024/1689 was adopted by the European Parliament and the Council on June 13, 2024. Under its Article 50, people must be made aware when they are interacting with an AI system.
Two facts have to be set against that, because the easy version of this argument is wrong. The Act's first prohibitions on AI practices, including harmful manipulation, have applied since February 2, 2025, along with AI literacy obligations. Obligations for general-purpose AI models followed on August 2, 2025.
So EU law does reach the content of what an AI system says. It sets a floor. Do not manipulate, do not exploit, do not deceive a person about the nature of the AI.
What it does not set is a standard of correctness for one organization's customer-facing answer. On the dates currently in force, the regime prohibits defined harms and requires disclosure. It evaluates no answer against any particular organization's standard of what that answer should have said.
Reading the deferral as a durable condition rather than a passing one is an interpretation drawn from the dates. The Commission has not stated it.
Where the Gap Shows Up in the Workflow
What follows is a composite workflow, assembled as a representative pattern. It is not a named organization's process.
An enterprise deploys a customer-facing AI system. The program is mapped to one or more of NIST AI RMF, ISO/IEC 42001 and the AI Act. The output is generated at interaction time, as a new statement that has never existed before.
The program reviews the system, periodically, against trustworthiness characteristics. The individual statement is not reviewed. Review runs on the thing that produces answers, never on an answer.
Deployment approval covers the system. The individual statement passes through without an approval step of its own. What is logged is the measurement and documentation the framework requires, at system level. The answer then reaches the customer, inside an interaction correctly disclosed as AI.
Trace that sequence and the release point is clear. The answer is correctly measured for safety, bias and privacy. It is correctly disclosed under Article 50. It says something no instrument in the program has defined as acceptable or unacceptable.
Give that sequence one statement. A customer asks when a feature they need will arrive. The assistant answers that it is coming next quarter.
The company has never announced a date. The answer is not unsafe, not biased, and correctly disclosed as AI. It is a commitment the organization did not make. Nothing in the program was positioned to catch it.
The exposure the program was commissioned to reduce is still open. It is found after an incident, not before one.
That shape is not unique to the unregulated enterprise. The companion analysis of model risk management sets out where the same gap appears under supervisory guidance.
What NIST and ISO Require of the Output, and Where They Stop
What NIST AI RMF MEASURE Requires
AI RMF 1.0 was released on January 26, 2023, and NIST describes it as intended for voluntary use. NIST states it is being revised as part of the White House AI Action Plan. The text quoted here therefore has a known expiry and no published date for it.
The MEASURE function, in NIST's own words, "employs quantitative, qualitative, or mixed-method tools, techniques, and methodologies to analyze, assess, benchmark, and monitor AI risk and related impacts." It further requires that "The AI model is explained, validated, and documented, and AI system output is interpreted within its context, as identified in the map function, to inform responsible use and governance."
Read that closely. It requires measurement, benchmarking, documentation, and interpretation of output in context. It requires nothing about what the output says.
The benchmark is named. What the benchmark contains is not. Treating that absence as a deliberate scoping choice rather than an oversight is an interpretation, and NIST does not characterize it either way.
What ISO/IEC 42001 Covers
ISO/IEC 42001:2023 was published in December 2023 as Edition 1. By its own published scope it "specifies requirements for establishing, implementing, maintaining, and continually improving an Artificial Intelligence Management System (AIMS) within organizations." ISO describes the standard as working "rather than looking at the details of specific AI applications."
ISO does not say why. This article reads the boundary as scope, not omission. A standard written for every AI system in every sector cannot also define what one brand's answer should say.
The Strongest Objection, Stated Fairly
The best argument against everything above runs like this. MEASURE reaches output evaluation. ISO/IEC 42001's Annex A controls reach it too. A release gate is a control inside the framework, not a layer beneath it.
A competent management system implementer was always supposed to define the organization's own acceptance criteria. On that reading there is no gap, only an implementation nobody finished.
That objection is correct.
The gate does belong inside the framework. That is this article's position too. The framework tells the organization to build it.
It does not tell the organization what to put in it. What goes in it is the one artifact a standards body is not scoped to write for any particular organization.
The implementer can write that artifact. Writing it and enforcing it are different jobs. A producer grading its own output has no independent reading to report. The judgment has to come from outside the system that generated the sentence.
A criterion for customer-facing language is brand-specific and jurisdiction-neutral. A framework is jurisdiction-specific and brand-neutral. The two cannot be the same document. That is why the frameworks correctly left it out.
The deferral of the high-risk regime is a sequencing decision, not an absence of law.
These frameworks produce real governance: documented risk mapping, measured trustworthiness, accountable ownership, auditable decisions. The unfinished work is a single, nameable thing, and it sits with the organization.
The Standard Belongs to the Brand
The standard does not belong to the model, the platform or LuxeDetect™. It belongs to the brand. LuxeDetect™ turns that standard into something enforceable.
Until it is written down, versioned and enforced at the release point, no measurement has anything to run against.
Release is also the only point where it can bind. The statement did not exist before the customer asked for it, so no earlier review could have approved it. A standard enforced any later is a standard enforced after the customer has read the answer.
LuxeDetect™ is AI Brand Integrity Infrastructure: the independent, deterministic brand-integrity release gate across the enterprise AI stack. It evaluates AI-generated or AI-assisted language against a benchmark the brand owns.
It enforces tiered actions at the release gate, with reason codes and an audit trail. It never writes the content it judges, and that independence is the point.
The enterprise stack still lacks a consistent, independent control layer that evaluates a generated statement against the organization's own standard before release.
AI scales language. We scale the standard.
What Each Instrument Requires of the Output
Each instrument gets the same question: what does it require about the output, and what does it leave to the organization?
EU AI Act, in force October 2026
Requires about the output: Disclose that the user is interacting with an AI system. Do not engage in the Act's prohibited practices.
Leaves to the organization: Whether the answer is accurate, on-standard, or something the brand would have approved.
EU AI Act, high-risk regime
Requires about the output: Nothing yet. The regime applies from December 2, 2027, and from August 2, 2028 for systems embedded in products.
Leaves to the organization: Whether the answer is accurate, on-standard, or something the brand would have approved, for at least another year.
NIST AI RMF 1.0, voluntary
Requires about the output: Measure, benchmark and monitor risk. Explain, validate and document the model, and interpret output in context. NIST states AI RMF 1.0 is under revision.
Leaves to the organization: The criterion the measurement runs against.
ISO/IEC 42001:2023
Requires about the output: Establish and continually improve a management system for AI.
Leaves to the organization: The details of specific applications, which ISO states the standard does not look at.
Frequently Asked Questions
What Should an AI Governance Framework Cover?
Three parts: the system, the output, and the standard the output is measured against. The published frameworks cover the first two thoroughly. The third is the organization's own definition of an acceptable customer-facing answer. None of the published frameworks examined here supplies it.
Does the EU AI Act Require AI to Be Accurate?
No, and the reason needs care. The Act prohibits a defined set of practices, the first of them in application since February 2, 2025, and has regulated general-purpose AI models since August 2, 2025. It reaches content, as a floor. It does not evaluate whether one organization's answer was correct against that organization's own standard.
What Does NIST AI RMF MEASURE Actually Require?
Measurement, benchmarking, monitoring and documentation. In NIST's words it analyzes, assesses, benchmarks and monitors AI risk and related impacts, and requires that output be explained, validated, documented and interpreted within its context. It does not prescribe what the output should say, and NIST states AI RMF 1.0 is under revision, so verify this text before relying on it.
Who Sets the Standard an AI Output Is Measured Against?
The organization. Not the model, not the platform, not the regulator and not the standards body. Each of those either requires the standard as an input, or declines by scope to write it.
What to Do with This
Take these four instruments to whoever owns the AI governance program. Read the "Leaves to the organization" line for each and ask one question: which instrument here defines what an acceptable customer-facing answer says?
If the honest answer is none of them, that is not a finding against the program. It is the work this window leaves to the organization. LuxeDetect™ is being built to carry it at the release point.
Request early access to LuxeDetect™ and enforce the answer standard your organization writes.
AI content risk is not solved after publication; it has to be controlled before release.