A head of learning at a 40,000-person company stands in front of a slide titled "Our AI Pilots." There are nineteen of them. Two designers using a writing assistant, a video team trialing avatars, an enablement squad running a grounded chatbot, a compliance group quietly testing item generation. Each pilot works. None of them adds up to anything. The CFO asks one question: "So what is the operating model?" The room has no answer, because nineteen pilots are not a transformation. They are nineteen experiments waiting for someone to decide what the learning function will actually become. This lesson is about being the person who decides, and the staged path you walk to get there.

From a Pile of Pilots to an Operating Model

Most enterprise learning functions in 2026 are not short on AI activity. They are short on an operating model, which is the durable arrangement of how the function produces value: who does what work, against which standards, on which tools, measured by which numbers, governed by which controls. A pilot answers "can this tool do something useful." An operating model answers "how does this entire function now build, verify, ship, measure, and govern learning, every day, at scale, in a way a regulator and a CFO both accept." Those are different questions, and the gap between them is where most learning transformations quietly stall.

The stall has a signature. Adoption looks wide and is genuinely shallow. LinkedIn's 2025 Workplace Learning Report found that roughly 71 percent of L&D professionals are already exploring, experimenting with, or integrating AI, while only about 25 percent factor it into their work routinely. Read those two numbers together and you have the exact disease the transformation leader has to cure: almost everyone has touched the tool, almost no one has wired it into a repeatable, defensible workflow. Treat those as numbers to verify against your own function, not slogans to repeat, and you will find your own version of the same split: lots of pilots, no spine.

The reason this matters at the enterprise level, and not just the team level, is the asymmetry the whole program is built on. AI collapsed the cost of producing a course. The Josh Bersin Company, in February 2026, framed AI as disrupting a roughly 400 billion dollar corporate-learning market and reported that 74 percent of companies say they are not keeping up with skill demand. When a first-draft module is minutes of work, producing more learning faster is no longer the constraint. The constraint, and therefore the value, moved to verifying that the learning is correct, proving it changed behavior, keeping it accessible, and governing how AI is used across the whole catalog. A transformation that scales production without scaling verification does not save the enterprise money. It manufactures liability faster.

Nineteen pilots that each work are not a transformation. A transformation is one operating model that every pilot was secretly auditioning for, and the leader's job is to choose it and make it the standard.

The Iron Rule at Enterprise Scale

Before any staged path, the leader has to internalize what does and does not change as the function scales. The iron rule of this entire program is simple: AI assists, the human verifies, the human owns the decision, and "the AI wrote it" is never a defense to a compliance officer, an accessibility auditor, or a CFO. At the individual-designer level, that rule lives in one person's verification habit. At the enterprise level, it has to live in the operating model itself, because you can no longer watch every build.

That is the whole reason an operating model exists. When two designers use AI, you can trust their judgment directly. When 200 people across eleven business units use AI to build regulated training in nine languages, you cannot personally check anything. The iron rule then has to be encoded into how work flows: grounding on an approved source of truth is the default, not a choice; a named human signs off on every regulated claim and the sign-off is logged; accessibility is a gate that blocks shipping, not a polish step; and an AI-generated scenario about people is bias-checked before a learner ever sees it. The transformation is precisely the act of turning four individual disciplines into four structural guarantees that hold whether or not the leader is in the room.

This is why the program insists the L5 verbs stay operational. The transformation leader does not "champion an AI vision" or "publish a point of view." The leader runs an engine, owns a governance standard, sets the source-of-truth architecture, and signs the number that goes to the board. The difference between a strategist and a transformation leader is the difference between recommending the operating model and being accountable for running it.

The Five-Stage Path a Learning Function Walks

An enterprise learning function does not jump from scattered pilots to a mature operating model. It walks a staged path, and naming the stages lets the leader say honestly where the function is, what it is allowed to do at this stage, and what gate it has to clear to advance. The five stages below are a maturity ladder, not a calendar. A function can sit at a stage for a year, and a function that skips a stage almost always ships an incident that drags it back down.

The discipline the table enforces is that you do not advance a stage by buying a tool. You advance by clearing a gate, and every gate is a verification or governance artifact, not a feature. A function can have the most advanced AI-native LMS on the market and still be stuck at Stage 1, because it never wrote the standard. Conversely a function with modest tools can reach Stage 3 because it built the spine. Tools change monthly; the gates do not. That is what makes this a playbook rather than a procurement list.

Why the Order Is Not Negotiable

Leaders under pressure want to start at Stage 4, because dynamic enablement is the exciting part and the part the CEO read about. Starting there is the most expensive mistake in the playbook. Dynamic enablement means putting AI-grounded capability directly in front of the workforce in the flow of their job, often without a human in the loop on each interaction. If you do that before you have a Stage 1 standard and a Stage 2 pipeline, you have built a system that ships unverified, potentially inaccessible, possibly biased content to thousands of employees at machine speed, with no provenance and no gate. That is not transformation. It is the incident the rest of the curriculum exists to prevent, scaled to the enterprise.

The order is evidence-and-accessibility-first for the same reason a building inspector signs structure before paint. The standard (Stage 1) and the pipeline (Stage 2) are the load-bearing walls. Measurement (Stage 3) is what tells you the walls are holding as you add floors. Only then is it safe to put capability in the flow of work (Stage 4) and build the reskilling engine on top (Stage 5). A leader who can explain to the board why the function is deliberately at Stage 2 and refusing to skip to Stage 4 is demonstrating exactly the judgment the role exists to provide.

What the Leader Actually Does at Each Stage

The playbook becomes real when the leader's own work is named per stage, in operational verbs. The strategist built the roadmap and the standard at L4. The transformation leader runs them.

At Stage 1, the leader stops the bleeding and writes the standard. Scattered pilots are not a problem to be celebrated or banned; they are uncontrolled exposure. The first act is to inventory them honestly (the nineteen-pilot slide is the artifact), surface which ones touch regulated, safety, or people content, and convene L&D, compliance, legal, IT, and accessibility to ratify one written standard every AI-built course must clear. The standard is the function's constitution. Without it, every later stage is built on sand.

At Stage 2, the leader makes the pipeline real on a flagship build. Pick one high-stakes, high-volume program, often a quarterly compliance refresh or a safety curriculum, and run the full source-to-certified-course pipeline on it: grounding on the approved source of truth, aligned objectives and a validated item bank, accessible media verified to WCAG 2.2 AA (the W3C accessibility conformance target), and a tamper-evident SME sign-off log. The goal is not to digitize the catalog. It is to prove the spine works on the hardest case, so the rest of the function has a template to copy and a success to point at.

At Stage 3, the leader institutionalizes the model and turns on enterprise measurement. The pipeline becomes the default, governance becomes a standing practice with a real cadence, and roles are redrawn so humans sit on evidence, judgment, and governance while AI sits on production and personalization. Critically, the leader stands up measurement that reports speed, scale, behavior, results, and risk posture together, so the board never sees a production-speed win without the liability ledger next to it.

At Stage 4, the leader shifts the center of gravity from the catalog to the flow of work. This is the move toward dynamic enablement, which means grounded, on-demand capability delivered where work happens rather than a library of courses learners visit. The same standard and the same grounding discipline apply, now to assistants and in-the-flow support, which is harder precisely because there is no editor between the model and the learner. The leader extends the gate, not relaxes it.

At Stage 5, the leader runs the engine. The function now operates a skills architecture and a reskilling engine at workforce scale, feeding the WEF-scale reskilling demand the enterprise faces. The leader's job is steady-state operation and continuous re-verification: of content, of accessibility, of bias posture, and of the regulatory state itself, which moves. That last point is not optional. The EU AI Act's Article 4 AI-literacy duty has been in application since 2 February 2025 with enforcement beginning 2 August 2026, and the Digital Omnibus amendment was in flight through 2026 and not yet in the Official Journal. A transformed function re-checks the live legal state rather than operating on last year's wording.

A Worked Example: The Bank That Skipped a Stage

Watch two versions of the same transformation at a large regulated bank with a 40,000-person workforce and a head of learning under pressure to "show the board an AI story by Q3."

Before (skip to the exciting stage). The head of learning, energized by the CEO's enthusiasm, greenlights a dynamic-enablement push: an AI assistant embedded in the teller and advisor workflow that answers policy and procedure questions on demand, drafted and shipped in a quarter. It demos beautifully. Six weeks later, the assistant confidently tells branch staff a wrong threshold for a suspicious-activity report, a number it generated from training data rather than retrieving from the bank's approved anti-money-laundering procedure. The wrong guidance reached an estimated number of advisors before anyone noticed, and it surfaced not in a learning dashboard but in a compliance review. The regulator's examiner asked the question every transformation leader fears: "Who verified what this system tells your staff, and where is the record?" There was no standard, no grounding, no sign-off log, because the function had skipped Stages 1 through 3 to reach the part the CEO wanted. The transformation was paused, the assistant was pulled, and the head of learning spent the next two quarters rebuilding trust instead of capability.

After (walk the stages). A different head of learning, same pressure, refuses to lead with the assistant. She spends the first quarter at Stage 1: an honest pilot inventory and one ratified standard, signed by compliance, legal, and accessibility. The second quarter she runs Stage 2 on the bank's mandatory anti-money-laundering refresh, the highest-stakes program in the catalog, with full grounding on the approved procedure, a validated item bank, a conformance report, and an SME sign-off log naming the compliance officer who approved every threshold. When the board asks for the AI story in Q3, she does not show a flashy assistant. She shows a rebuilt regulated curriculum that cut build time substantially while producing, for the first time, a tamper-evident record proving every threshold traces to an approved source. The regulator's examiner, in the next review, asks the same question, and this time the answer is a document. Only after that foundation holds does she move to Stage 4 and put a grounded assistant in the flow of work, now gated by the same standard the rest of the function clears. The exciting part shipped later and survived, because the load-bearing walls went in first.

The lesson is not that dynamic enablement is dangerous. It is that an operating model is a structure, and you cannot hang the roof before the walls. The leader who can hold the line against pressure to skip stages is doing the single most valuable thing the role requires.

Key Takeaways

• A pile of pilots that each work is not a transformation; an operating model is, and the leader's job is to choose one model and make every team's pilot resolve into it.
• Wide-but-shallow adoption (roughly 71 percent exploring, about 25 percent routine in the 2025 LinkedIn data, both numbers to verify against your own function) is the exact disease: almost everyone has touched AI, almost no one has wired it into a defensible workflow.
• Because AI collapsed production cost, scaling production without scaling verification does not save money, it manufactures liability faster, so the transformation must be evidence-and-accessibility-first.
• The five stages are scattered pilots, standardized practice, operating model, dynamic enablement, and transformed function, and each gate between them is a verification or governance artifact, never a tool purchase.
• The iron rule has to move from one person's habit to four structural guarantees: grounding by default, logged human sign-off on regulated claims, accessibility as a blocking gate, and bias-checking of scenarios about people.
• The order is not negotiable: starting at dynamic enablement before the standard and the pipeline exist ships unverified content to the workforce at machine speed, which is the enterprise-scale version of the incident the whole program prevents.
• The leader's verbs stay operational: write the standard, run the pipeline on the hardest case, institutionalize the model, turn on enterprise measurement, shift to the flow of work, and run the reskilling engine, never "publish a vision."
• A transformed function re-verifies the regulatory state as it moves, because Article 4 enforcement began 2 August 2026 and the Digital Omnibus wording was still in flight, so operating on last year's text is itself a risk.