It is a Monday in the second week after the new policy shipped, and the head of learning is in a glass conference room with the chief compliance officer, who has printed one page and circled one sentence. The sentence reads: "AI-generated content should be reviewed for accuracy where appropriate." The compliance officer taps the words "where appropriate" with a pen and asks the only question that matters: "Who decides when it is appropriate, and where is that recorded?" Nobody in the room can answer, because the policy was written to sound responsible rather than to be operated. Across the enterprise, 340 learning professionals in six business units are already using AI to build courses, and not one of them knows what this policy actually requires of them on a Tuesday. A policy that cannot answer "who, when, and where is it recorded" is not a policy. It is a press release that happens to live in a governance folder.

Why Most Learning-AI Policies Fail on Contact

By the time a learning organization reaches the enterprise tier, it usually has a learning-AI policy already. The problem is rarely that one does not exist. The problem is that it does not survive contact with a Tuesday. Most learning-AI policies fail for the same three reasons, and an enterprise leader has to recognize all three before writing the next one.

The first failure is the aspirational verb. A policy built on words like "should," "encourage," "strive to," and "where appropriate" reads beautifully and binds nobody. An enforceable policy, by contrast, is a document whose every clause names an actor, a trigger, an action, and a record. Why you care: when a regulator, an accessibility auditor, or a CFO reopens a course eighteen months from now, "we encouraged review" produces silence in the room, while "the named SME signed this claim on this date against this source, and here is the log entry" produces a closed finding. The difference between those two outcomes is grammar. Aspirational verbs describe a culture you hope for; enforceable verbs describe a control you can show.

The second failure is the scope gap. A policy that governs content production but says nothing about assessment, delivery, or analytics leaves three of the four places AI actually lives in a learning function completely ungoverned. The AI that scores a quiz, the AI tutor that answers a learner at 11pm, and the AI that infers an employee's skills from their click data are each capable of a quiet, expensive failure, and a production-only policy never touches them. Enterprise learning AI is not one activity. It is at least four, and a policy that covers one and calls itself comprehensive is a policy with three blind spots.

The third failure is the orphaned policy. A learning-AI policy that does not connect to the organization's broader AI governance, its information-security controls, its data-privacy regime, and the EU AI Act literacy duty (the legal obligation, in application across the EU since 2 February 2025, to ensure staff have a sufficient level of AI literacy, with enforcement by national authorities beginning 2 Aug 2026) becomes an island. It contradicts the security policy on data handling, ignores the privacy team's consent rules, and uses different language than the enterprise AI management system. An auditor reading three contradictory policies does not see governance. They see a function that has not decided what it actually believes.

A policy you cannot operate on a Tuesday is not a policy. It is a wish written in the imperative mood.

The One Policy That Holds Across Four Domains

The enterprise leader's job is to author a single learning-AI policy that holds across the four domains where AI touches learning, with one consistent spine of rules and four domain-specific applications. The four domains are production, assessment, delivery, and analytics. They share a spine and differ in their specifics, and the architecture of the policy should make both the shared spine and the differences visible on the page.

The shared spine is the program's iron rule, written as policy: AI assists, a named human verifies, that human owns the decision, and "the AI generated it" is never a defense to a compliance officer, an accessibility auditor, or a CFO. Everything else in the policy is that sentence applied to a specific domain, with a specific actor, trigger, action, and record. Hold that spine constant and the policy reads as one coherent instrument rather than four stapled memos.

Consider each domain in turn, and notice that the spine never bends, only the specifics change.

Production: The Verified Claim

In production, AI drafts modules, scripts, scenarios, job aids, and media. The load-bearing rule is provenance: every regulated, safety, or policy claim in AI-drafted content must trace to a human-approved source of truth, and a named subject-matter expert must sign that the claim matches the source before it ships. The trigger is "content contains a regulated, safety, or policy claim." The actor is the named SME. The action is verification against the source. The record is an entry in the sign-off log with the claim, the source, the approver, and the date. This is not a suggestion to be careful. It is a gate with a logged crossing.

Assessment: The Human-Owned Pass/Fail

In assessment, AI drafts items, distractors, rubrics, and feedback, and sometimes scores responses. The load-bearing rule is the bright line that AI does not certify a learner as competent. AI may draft the item and the feedback; a human validates that the item measures the objective (this is constructive alignment, the requirement that the objective, the assessment, and the content all point at the same performance), and a human owns the pass/fail and credential decision. The trigger is "an assessment outcome affects a credential, a certification, or a record of competence." The actor is the assessment owner. The record is a validation note tying each scored item to its objective. A well-formed question that measures nothing still certifies the wrong people, and the policy has to make the validity check mandatory, not optional.

Delivery: The Grounded Answer

In delivery, AI tutors, copilots, and adaptive engines answer learners directly, in real time, often without a human in the loop at the moment of the answer. The load-bearing rule is grounding: a learning assistant that answers questions about policy, safety, or procedure must answer from approved content (this is grounded generation, also called RAG, retrieval-augmented generation, forcing the model to answer from your source rather than its training data) and must cite or refuse. The trigger is "an AI system answers a learner's substantive question without human review at the moment of answering." The action is grounding plus a logged content source plus a monitored escalation path for questions the system cannot ground. Because no human reviews each answer live, the control moves upstream: into what the system is allowed to draw from and what it must do when it cannot ground an answer.

Analytics: The Inference You Can Defend

In analytics, AI infers skills, recommends paths, flags at-risk learners, and summarizes learning data for leadership. The load-bearing rule is that an inference about a person that affects an opportunity (a promotion path, a development plan, a flag to a manager) is a decision a human must be able to inspect and contest. The trigger is "an AI inference about an individual feeds a decision about that individual." The action is a documented, inspectable basis for the inference and a route for the learner to see and challenge it. The failure mode here is the quietest in the whole function: a skills-inference model can route a capable person away from an opportunity, and nobody sees a wrong sentence, only a path not offered.

The Structure of an Enforceable Policy

An enforceable learning-AI policy has a predictable skeleton, and using a predictable skeleton is itself a control: it forces every clause to declare who, when, what, and where it is recorded. The skeleton below turns the four-domain spine into a document an auditor can read top to bottom and a designer can operate without a meeting.

The single most important column is the third one. If a clause cannot pass its test, it is aspirational, and aspirational clauses are where audits go to die. Run every sentence in the draft against its test before the policy ships. The sentence that fails the test is the sentence the compliance officer will circle.

Getting It Followed Across the Enterprise

A correct policy that nobody follows is worth exactly as much as no policy, and at the enterprise tier the gap between "published" and "followed" is the whole game. Six business units, hundreds of designers, dozens of contractors, and a steady churn of new tools mean that a policy lives or dies on whether it is operable, discoverable, and enforced, not on whether it is well written.

Operable means the rule shows up at the moment of work, not in a PDF nobody reopens. The provenance rule belongs inside the SME sign-off log the designer already uses; the validity rule belongs inside the item-bank workflow; the grounding rule belongs inside the configuration of the learning assistant; the inference rule belongs inside the analytics review. A policy embedded in the tools and templates people already touch gets followed by default. A policy that requires someone to remember it gets followed by the conscientious and ignored by everyone under deadline.

Discoverable means a designer can answer "what does the policy require of me for this task" in under a minute. That is a job aid, a one-page decision guide, and a "cite or refuse" instruction baked into the shared prompt library, not a forty-page document. The enterprise leader who ships the forty-page document and skips the one-page job aid has confused completeness with compliance.

Enforced means there is a consequence and a check. Not a punitive culture, but a real audit: a sample of shipped courses pulled each quarter and tested against the policy, with findings that go somewhere. The first time a unit learns that "shipped without a logged SME sign-off" is a finding that reaches their director, the behavior changes. A policy with no audit is a policy the busy will route around, and at enterprise scale the busy are the majority.

The policy that gets followed is the one embedded in the tools people already use, not the one filed in the folder people never reopen.

A Worked Example: Before and After

Return to the glass conference room and the circled sentence, and watch the same enterprise rewrite one clause two ways.

Before (the aspirational clause). The policy reads: "AI-generated learning content should be reviewed for accuracy where appropriate, and teams are encouraged to maintain records of their review." A designer in the manufacturing unit reads this, builds a lockout/tagout refresh with an AI assistant, decides that "where appropriate" does not obviously apply to a routine annual refresh, ships it to 1,400 technicians, and keeps no record. Three months later an incident review asks who verified the de-energization sequence on screen 22. There is no SME on record, no source the step traces to, and no log entry. The clause did not fail because people were careless. It failed because it never told anyone who, when, or where to record. The grammar was the vulnerability.

After (the enforceable clause). The rewritten policy reads: "Any AI-drafted content containing a safety, regulatory, or policy claim must be verified by a named subject-matter expert against an approved source before publication. The SME records the claim, the source, their name, and the date in the course sign-off log. Content with an unverified safety claim must not be published." Now the same designer reads an unambiguous trigger (a safety claim), a named actor (a SME), a required action (verify against source), and a required record (the log). The lockout/tagout refresh does not ship until the SME signs the de-energization sequence against the plant SOP, and the log entry exists before launch. When the incident review asks who verified screen 22, the answer is a name, a source, and a date, produced in one breath. Same course, same tool, same speed. The grammar changed, and the grammar was the control.

The lesson is not that the first policy was unethical and the second is virtuous. Both intended the same thing. The first expressed the intent in verbs that bind nobody; the second expressed it in verbs that name an actor, a trigger, an action, and a record. At enterprise scale, across hundreds of designers and four domains, that difference is the difference between a policy and a liability waiting for an incident.

Key Takeaways

• Most learning-AI policies fail on three fronts: aspirational verbs that bind nobody, a scope gap that covers production but ignores assessment, delivery, and analytics, and orphaning from the organization's broader AI, security, and privacy governance.
• An enforceable policy clause names four things: an actor, a trigger, a required action, and a required record. If a clause cannot name all four, it is a wish, not a control.
• One policy must hold across all four domains where AI touches learning, sharing a single spine (AI assists, a named human verifies, the human owns the decision) and differing only in the domain-specific application.
• The four domain rules: production requires provenance to a verified source; assessment keeps the pass/fail human-owned; delivery requires grounding and "cite or refuse"; analytics makes any inference about a person inspectable and contestable.
• The most important test for any clause is whether it can be operated on a Tuesday by a designer under deadline without calling a meeting.
• A policy gets followed when it is operable (embedded in the tools and templates people already use), discoverable (a one-minute job aid, not a forty-page PDF), and enforced (a real quarterly audit with findings that reach someone).
• The EU AI Act literacy duty (in application since 2 February 2025, enforced from 2 Aug 2026), the ISO/IEC 42001 AI management system, and the privacy and security policies are the context the learning-AI policy must align with, not contradict.
• The grammar is the control: "should review where appropriate" produces silence in an audit, while "a named SME verifies against an approved source and logs it" produces a closed finding.