The professional learning industry has spent a decade promising "learning in the flow of work." The promise was simple: instead of pulling people out of their jobs to sit through training, deliver knowledge where they already are — in their tools, their workflows, their daily routines.
It was a beautiful idea. And every implementation of it has been a lie.
Here is what "learning in the flow of work" actually looks like today: your company buys a platform. You get an email. You click a link. You leave your workflow to log into an LMS. You watch a video. You take a quiz. You earn a badge no one will ever check. You return to your work having lost thirty minutes and gained a completion record.
The flow of work was never entered. It was interrupted.
This is not a marginal failure of execution. It is a failure of architecture. The industry spent a decade optimizing delivery — better interfaces, mobile-first design, AI-powered recommendations, microlearning formats — while leaving the underlying system untouched. Every innovation assumed the same thing: that a catalog of pre-produced content, if surfaced at the right moment, would constitute learning in the flow of work.
It did not. It could not. The architecture was wrong.
COVID accelerated a structural break that was already underway. The assumption that people would physically go somewhere to learn — a campus, a conference room, a training center — collapsed and has not recovered. What replaced formal institutions was not better institutions. It was YouTube tutorials, peer networks on Discord, and large language models that answer questions at two in the morning without requiring enrollment. A generation learned to learn informally. They can pull a video and follow instructions to complete a task. They are resourceful, self-directed, and entirely comfortable outside institutional walls.
That is the easy part. The hard part — the part no informal channel solves — is how you take someone from "I can follow instructions" to "I can exercise judgment." How you move from pattern execution to the kind of professional capability that matters when the playbook runs out and the situation demands a decision no template anticipated.
That is the problem worth solving. And solving it requires an architecture the professional learning industry has never built.
The Collapse
The $400 billion professional learning market is experiencing the simultaneous collapse of three equilibria that held the old model together.
The first is the credential signal. Only 18% of job postings now require degrees, down from over 50% a decade ago. Google, Apple, IBM, and sixteen U.S. state governments dropped degree requirements. The logic is straightforward: when employers can observe ability directly — through portfolios, project work, AI-assisted screening — the expensive signal of a credential becomes redundant. The signal still exists. It is just no longer worth what it costs to produce.
The second is adult enrollment. The number of adult learners in postsecondary education fell from 8.5 million to 4.0 million between 2011 and 2021, according to the National Student Clearinghouse Research Center. Community colleges lost 38% of their students. Adults are rational economic actors. They noticed the credential was weakening while costs kept climbing. They did the math and walked.
The third is the knowledge monopoly. Over half of U.S. adults now use large language models. 82% of regular users employ them for learning. OpenAI alone reaches 700 million people weekly. The general-purpose AI assistant has become the world's largest learning platform, and no institution voted on it. It happened bottom-up, one question at a time, because asking an LLM is faster, cheaper, and more contextually responsive than navigating a course catalog.
What replaced these three collapsed equilibria is a reinforcing dynamic that should concern every incumbent in the space: AI commoditizes content, which drives learners toward AI-mediated informal learning, which reduces institutional enrollment, which starves institutions of revenue needed to invest in innovation, which widens the content quality gap, which makes AI look better by comparison, which drives more learners to shift. The Coursera-Udemy merger — $7.2 billion in combined revenue since 2020 but $8 billion in market value destroyed since their IPOs — is, even accounting for the broader tech market correction, a signal of structural decline rather than cyclical weakness. Both companies defined themselves by format — courses, certificates, libraries — rather than by outcome. Neither said: "We improve the quality of decisions professionals make."
That absence is not a branding problem. It is a structural one.
The Missing Architecture
I think the real problem is more fundamental than bad content or outdated formats. The problem is that learning systems and diagnostic systems have never been structurally connected.
Most learning platforms are open-loop. Content goes out. Completion data comes back. Perhaps a satisfaction survey. But nothing in the architecture connects "what you lack" to "what gets produced next." The catalog exists. You browse it. A recommendation engine suggests something based on your role or what peers consumed. If your specific gap is not in the catalog, you get the closest match — a 45-minute course that spends thirty minutes on things you already know and five minutes adjacent to the thing you needed.
To be sure, adaptive learning platforms have attempted diagnostic-to-content connections for over a decade. Knewton, Area9 Lyceum, Realizeit — these are serious efforts by serious teams. But their adaptivity operates within a fixed content inventory. They select and sequence from what already exists. They optimize catalog navigation. What they do not do — what no system in the market does — is allow the diagnostic signal to trigger the production of new content. The gap between diagnosis and resolution is still bridged by catalog search. When the catalog does not contain what the learner needs, the system offers the closest available match and moves on.
This is notable because the last decade was not short on innovation. The industry introduced communities inside learning management systems. It embraced meetups and conferences. It built war games and leaderboards. It moved to YouTube and gaming platforms. It adopted xAPI — the Experience API — which promised to track learning activity everywhere, across any platform, in any context.
These were real improvements to the experience of learning. But none of them closed the structural gap. xAPI is instructive as a case study. The specification was designed to capture learning activity statements — "learner X did action Y on object Z" — across any environment. It makes learning activity visible wherever it happens. You can record that someone spent 45 minutes in a cybersecurity war game. You can record that they completed a compliance module on their phone during a commute.
What xAPI was not designed to capture — and what its architecture does not support — is the inference from activity to capability, and from capability gaps to content production. It tracks what people did. Connecting that to what they can do requires additional inference that the specification does not itself provide. And it has no mechanism to connect what was learned to what gets produced next.
That is the difference between a bookstore and a printing press. A bookstore — however well-organized, however intelligently curated — contains a finite inventory of pre-produced works. If what you need is not on the shelf, you leave with the closest substitute. A printing press produces what is needed, when it is needed, because the need triggered the production.
Every learning platform in the market today is a bookstore. Some are beautiful bookstores. Some have extraordinary recommendation engines. Some have adaptive algorithms that walk you through the store more efficiently. But none of them prints the book you actually need because you needed it.
Peter Senge described reinforcing loops in The Fifth Discipline as the engines of both virtuous and vicious cycles. The AI-mediated learning market is currently running what he called a "Success to the Successful" archetype — a reinforcing dynamic that favors informal AI learning over institutional education, and will continue to favor it as long as institutions compete on content delivery rather than on the structural connection between diagnosis and production.
What would a closed loop look like? A diagnostic engine identifies competency gaps — not role-based assumptions, but observed gaps in what a specific individual can demonstrate. Those gaps do not trigger a catalog search. They trigger content production. A knowledge unit addressing the specific gap is produced, quality-gated by domain experts and automated validation against authoritative sources, and delivered. The individual engages with it. An assessment — which might be a simulation, a portfolio demonstration, or a structured application exercise — verifies whether the gap closed. The assessment result flows back to the diagnostic, which updates its model. The next diagnosis reflects the shift.
The architecture is designed so that each cycle refines the next. Diagnosis incorporates Bayesian priors drawn from research literature, industry knowledge graphs, and where available, company-specific intellectual property. Every assessment updates the posterior. Every demonstration either confirms or corrects the diagnostic model. The loop is designed to correct error, not amplify it — though I want to be honest about what is design intent and what is validated at scale. The architecture has these properties. Whether they hold across thousands of learners and dozens of domains is something we are testing, not something we have proven.
I also want to apply the same systemic scrutiny to this loop that I apply to the incumbent market. Reinforcing loops run in both directions. A closed loop that connects diagnosis to production could, over time, narrow its definition of "competency" to what its assessment mechanisms can measure — excluding capabilities that are real but not easily testable. It could develop false precision, where the diagnostic model becomes confident about things it does not actually know. These are real risks. The mitigation is continuous calibration against external benchmarks, human expert review, and the intellectual honesty to treat the loop as a hypothesis being refined rather than a machine that runs itself. The system does not replace human judgment about what competency means. It provides infrastructure for developing that judgment more precisely than an open-loop alternative.
What if we stopped trying to improve the delivery mechanism and reinvented the content itself?