A newly hired head of learning at a 12,000-person insurer inherits a function that already uses AI everywhere, and it is quietly falling apart. Every designer drafts modules with an AI assistant. Build time is down 70 percent. And yet in her first month, a compliance officer pulls a claim from an AI-drafted anti-money-laundering course that misstated a reporting threshold, an accessibility reviewer fails three AI-narrated videos in a row, and the CFO asks for the behavior-change number on the sales-onboarding program and gets a slide of completion rates instead. The tools work. The speed is real. The function is undefended. She realizes the problem is not that they adopted AI badly. It is that they bolted AI onto a structure designed for a world where producing content was the hard part, and never redesigned the function around the fact that producing content is now the easy part. This lesson is that redesign.

The Inversion That Forces the Redesign

The previous lesson drew the new org chart: three roles, humans on judgment and accountability, AI on production and personalization. This lesson is about the harder thing, which is redesigning the whole function so that split is true by construction rather than by hope. Naming three roles on a chart changes nothing if the work still flows the old way underneath them. Redesign means changing how the work moves, where the gates sit, what gets measured, and what the function is actually organized to produce, so that the human owns evidence, judgment, and governance not because someone remembered to, but because the structure makes it impossible to skip.

Start from the inversion, because everything follows from it. For thirty years the scarce, expensive, valuable thing in L&D was producing content: writing the module, building the SCORM package, recording the narration, assembling the fifty-item question bank. The function was organized around that scarcity. Designers were hired to produce, timelines were built around production, and quality was something that happened invisibly inside the act of building, because a human who was slowly, carefully building a module was also, in the same motion, reading the source, checking the claim, and aligning the item. Verification rode along for free inside slow production. Then AI collapsed production cost, and the free ride ended. When a module is drafted in minutes, the reading, checking, and aligning do not happen automatically anymore, because the human is no longer slowly building the thing. The Josh Bersin Company's February 2026 framing of AI disrupting a roughly 400 billion dollar corporate-learning market is, underneath the headline number, exactly this: the value of the function inverted. What was cheap (judgment about whether the content is right) is now scarce, and what was scarce (production) is now cheap.

Here is a term to anchor the redesign. A function's design is the deliberate arrangement of its work, roles, gates, and measures so that the outcomes it needs happen by default rather than by heroics. Why you care: if your function's design still assumes verification rides along inside slow production, then every time production speeds up, verification silently drops out, and you get exactly the insurer's situation, faster and less defensible at the same time. Redesigning L&D around AI means deliberately rebuilding the function so that the expensive judgment, which used to be a free byproduct of slow work, becomes an explicit, owned, gated, measured part of the flow.

Verification used to ride for free inside slow production. AI ended the free ride. If you do not redesign the function to pay for verification on purpose, you stop paying for it at all.

The Four Things the Human Must Own by Design

Redesign is not vague. It comes down to making four specific things the human's explicit, structural property: evidence, judgment, governance, and the sign-off record that ties them together. Each one was previously implicit inside production. Each one now has to be a named part of the function's design, or it disappears.

Evidence: The Function Owns Whether It Worked

The first thing the human owns is evidence: the proof that the learning changed behavior and produced results, not just that it was completed and liked. In the old function this was often skipped entirely, because production ate the whole budget and there was no time left to measure. AI hands that time back. When a module takes two days instead of three weeks, the freed capacity should not go into producing ten times more content; it should go into the measurement that finally proves the content mattered. Redesigning around AI means the evaluation plan is designed in at the start, using Kirkpatrick's four levels (reaction, learning, behavior, and results) with a real path to Level 3 (did behavior change on the job) and Level 4 (did results move), captured through xAPI statements (the standard for recording learning-related behavior beyond the LMS, so you can see whether the trained action actually happens). Why you care: the CFO's question, "did any of this change what people do," is now answerable only if evidence is a designed-in property of the function, not an afterthought nobody had time for. The redesign spends the AI dividend on evidence.

Judgment: The Function Owns Alignment and Validity

The second thing the human owns is judgment: the decisions AI cannot make because they require weighing a claim against a source and an item against an objective. This is verification of every regulated or safety claim against an approved source of truth, and validation that every assessment item actually measures the objective it claims to (assessment validity, the difference between a question that is well-written and a question that tests the right thing). In the redesigned function these are not steps a busy designer might do; they are gates the work cannot pass without. A gate means the build physically cannot advance to the LMS until a human has verified the regulated claims against the source and validated the item bank against the objectives, and the passing of that gate is recorded. Why you care: AI does not certify a learner as competent and does not author a regulated claim that ships unverified, so the function has to be built so those two decisions are structurally impossible to skip, because when they are optional, speed guarantees they get skipped.

Governance: The Function Owns How AI Is Used

The third thing the human owns is governance: the standard that says how AI may be used across the whole function, applied consistently rather than reinvented by each designer. This is one written standard covering grounding (AI drafts from approved sources, never from its training data on regulated content), accessibility (every AI-generated experience clears WCAG 2.2 AA, the accessibility conformance target, before it ships), bias-checking (every AI-generated scenario about people is checked before it reaches a DEI, hiring, or harassment module), synthetic-media disclosure, and learner-data handling (whether the workforce's data trains a vendor's model). Redesigning around AI means the governance standard is a real control the function operates inside, plugged into the organization's AI management system under ISO/IEC 42001 (the AI management systems standard, December 2023). Why you care: governance is the difference between eleven designers each making a private, undocumented decision about whether AI can see learner data and one function-level answer an auditor can read.

The Record: The Function Owns Provenance

The fourth thing binds the other three: the sign-off record, a tamper-evident log of who verified what, against which source, on what date. Evidence, judgment, and governance are only defensible if you can prove they happened. In a slow function the record was implicit in the fact that a known human built the thing over weeks. In a fast function where AI drafted it in minutes, the record has to be an explicit artifact, or "a human verified this" is just a claim with nothing behind it. Why you care: when the compliance officer asks who verified the threshold, "we usually check those" is not an answer, and "here is the SME who signed it, against this policy, on this date" is. The redesign makes provenance a produced artifact of every build.

Redesigning the Flow, Not Just the Titles

The redesign that matters is the redesign of how work moves, because titles without a new flow are theater. Compare the two flows directly. In the old flow, work moved linearly through a single owner: a designer received a request, built the module over weeks, and shipped it, with verification, accessibility, and measurement all happening (or not) invisibly inside that one long act of building. In the redesigned flow, work moves through a pipeline with explicit human gates, where AI does the production between gates and humans own the gates themselves.

Read the right column as a single sentence: production is fast and continuous, and the human-owned gates are the load-bearing structure the fast production flows through. The redesign does not slow production down to make it safe. It lets production run at machine speed and inserts the human exactly and only where judgment, evidence, and governance live. That is the whole art. Put the human everywhere and you have thrown away the AI dividend. Put the human nowhere and you have the insurer's undefended function. Put the human on the gates, and you get speed and defensibility at once.

Redesigning L&D around AI does not mean slowing AI down to make it safe. It means letting production run fast and building the human into the gates, so the function is fast everywhere and defended exactly where it must be.

A Worked Example: The Insurer, Redesigned

Return to the insurer and watch the head of learning redesign the function over two quarters.

Before (AI bolted onto the old design). Eleven designers, each independently using an AI assistant, each deciding privately how to use it. There is no shared grounding, so some draft from the real policy library and some draft from the open model. There is no verification gate, so whether a regulated claim gets checked depends on how rushed each designer feels. Accessibility is a personal habit some have and some do not. Measurement is completion rates, because nobody has an evaluation plan. The function is producing four times the old volume and cannot answer a single one of the three questions that hit her desk in month one. The design still assumes verification rides along inside slow production, but production is not slow anymore, so verification fell out, invisibly, everywhere at once. The speed made the old design's hidden weakness fatal.

After (the function redesigned around the inversion). She does not add headcount. She redesigns the flow. First, governance: one written standard that says AI drafts only from the approved source library, every regulated claim is verified against its source before ship, every experience clears WCAG 2.2 AA, every people-scenario is bias-checked, and learner data does not train any vendor model. Second, the pipeline: one designer becomes the learning engineer and wires the approved policy and product libraries into every build so grounding is automatic and traceable, and builds the reusable prompt library that locks the audience, reading level, and cite-or-refuse rule. Third, the gates: two senior designers become learning architects who own the verification gate and the assessment-validity gate, and no build reaches the LMS without passing both, with the SME sign-off logged against the source. Fourth, evidence: because production is now fast, the freed time goes into designing evaluation plans in from the start, with xAPI capturing whether the trained behavior actually happens on the job. Same eleven people. Same tools. Same speed. But now the compliance officer's threshold question has a logged answer, the accessibility failures stop because conformance is a gate not a habit, and the CFO gets a Level 3 behavior number instead of a completion rate. The function did not get slower. It got a structure.

The redesign paid for itself the first time an auditor asked a question and the function had an artifact instead of a shrug. That is the point of redesigning around AI rather than merely adopting it: adoption makes you fast, and only redesign makes you fast and defensible. The insurer's function did not fail because AI was dangerous. It failed because a structure built for slow production cannot survive fast production without being rebuilt around the new location of value.

The Traps in the Redesign

Three traps recur when a function redesigns around AI, and each one quietly reverses the whole point of the exercise.

The efficiency trap: spending the AI dividend on more content instead of on evidence and verification. The most seductive move when production gets cheap is to produce ten times more, because volume is visible and impressive on a dashboard. But the workforce did not need ten times more courses; it needed the courses it had to be correct, valid, accessible, and proven to work. A redesign that pours the entire AI dividend back into volume has kept the old function's obsession with production and simply industrialized it. The disciplined redesign spends the dividend on the four things the human now owns: evidence, judgment, governance, and the record. Volume is the reward you take last, not first.

The abdication trap: letting the tools define the operating model. When a function has not deliberately redesigned, the vendors redesign it by default. The AI-native LMS decides how content flows, the authoring assistant decides what a build looks like, and the function's actual operating model becomes an accidental emergent property of whatever tools got purchased. A redesigned function decides its own flow and gates first, as a matter of design, and then selects tools that fit the flow, rather than letting the tool catalog dictate where the human sits. The obligation never transfers to the platform, so the operating model must never be outsourced to it either.

The verification-drift trap: letting the gates soften under deadline pressure. A gate that can be waived when the quarter is tight is not a gate; it is a suggestion. The first time a big launch slips and someone asks to skip the verification gate "just this once, we will check it after," the redesign begins to unwind, because the whole reason the gate exists is that verification does not survive being optional. A real redesign makes the gates structural, so that skipping one is a visible, logged, accountable decision by a named person, not a quiet convenience. The gate that bends under pressure is exactly the gate the pressure was going to break anyway.

Key Takeaways

• AI inverted L&D's value: producing content, once the scarce and expensive part, is now cheap, and judgment about whether content is correct, once a free byproduct of slow production, is now the scarce and expensive part.
• Verification used to ride for free inside slow production; when AI made production fast, that free ride ended, so a function must be redesigned to pay for verification on purpose or it stops paying for it at all.
• Redesigning around AI means making four things the human's explicit, structural property: evidence that it worked, judgment on alignment and validity, governance of how AI is used, and a tamper-evident sign-off record that proves all three happened.
• Titles without a new flow are theater; the real redesign changes how work moves, letting AI produce fast between explicit human gates rather than hiding verification inside one long act of building.
• The design principle is to put the human on the gates, not everywhere and not nowhere: production runs at machine speed, and humans own exactly the points where judgment, evidence, and governance live.
• Verification, assessment validity, and accessibility become gates a build cannot pass without, recorded when cleared, because a decision that can be skipped under deadline pressure will be skipped when production is fast.
• The efficiency trap is spending the AI dividend on ten times more content instead of on evidence and verification; the disciplined redesign takes volume last, after correctness, validity, accessibility, and proof.
• The iron rule rendered as a function: AI assists production and personalization, the human owns evidence, judgment, and governance by design, and "the AI wrote it" is never a defense to a compliance officer, an accessibility auditor, or a CFO.