From Designing Courses to Designing Work
How the role & output of corporate L&D is changing in 2026
Hey folks!
This week, I was in (a very hot!) London working with the L&D team of one of the world’s largest management consultancies. Over the course of two days, there was one question I got asked numerous times — and it’s a I’ve been asked a lot over the last six months:
“How do we develop senior level knowledge and skills in our juniors, when AI now does all the work that used to build that knowledge and skill set?”
In the “old world,” employees were developed primarily by doing the low-stakes work and learning on the job. The drafting, the first-cut analysis, the background research: by doing the repetitive, lower-stakes tasks and observing how others handled both these and the higher-stakes ones, we learned and grew in our roles. In effect, we were apprentices on an apprenticeship. For the vast majority of us, nobody explicitly designed that apprenticeship: it emerge organically from our work and the team around us
In 2026, the work still gets done — the deck is made, the sales call is taken, the client is consulted — but increasingly AI does some or all of it, which in practice means a lot of what we do becomes automated and hidden. And this is where the challenge emerges, because what AI changed isn’t the way humans build knowledge and skill: it’s the conditions that used to make it happen.
In this week’s post, I’ll explore both halves of that challenge — the problem (why the old apprenticeship has quietly stopped working) and the emerging solution (what we design instead, and why the unit of learning itself is changing).
Let’s dive in!
The Science Bit - how people actually build knowledge and skills at work
We build a LOT of courses in corporate L&D, but courses were never where most workplace capability came from. The most-cited rule of thumb in our field — the 70-20-10 model, originally from Lombardo and Eichinger’s work at the Center for Creative Leadership — estimates that something like 70% of what people learn at work comes from doing the job itself, 20% from other people, and only 10% from formal training. The exact split is debated, and rightly so, but the direction is not: the bulk of professional skill is built at work, through work — not in the courses we spend most of our time designing.
What that “70” and “20” actually are, when you look closely, is an apprenticeship — just an informal, unnamed one. You start on low-stakes work, get feedback, watch how more experienced people handle the harder stuff, and gradually take on more as your confidence and insight grow. The learning sciences have a precise name for this.
Lave and Wenger (1991) called it legitimate peripheral participation: newcomers learn by doing real, consequential work at the edge of a community of practice, moving inward as they become more capable. Brown, Collins and Duguid (1989) made the same case under situated cognition — that knowledge is inseparable from the activity and context in which it’s used. And the mechanism underneath both is now well evidenced. Ericsson’s work on deliberate practice shows expertise is built not by time served but by effortful attempts with feedback at the edge of current ability; the cognitive-apprenticeship tradition of Collins, Brown and Newman (1989) names the moves that make tacit expertise learnable — modelling, coaching, scaffolding, articulation, reflection — with fading, the gradual withdrawal of support, as competence grows.
Strip all of that down and you get one core loop: you watch an expert and hear their reasoning (model), you attempt it yourself at the edge of your ability (attempt), you get specific feedback on the gap (feedback), and take on progressively harder work as the support recedes (stretch, then fade).
This loop of Model, attempt, feedback, stretch, fade is the the engine through which most workplace development happened, and the old world ran it for free: drafting the doc was the attempt, the partner’s mark-up was the feedback, watching a senior take the awkward client call was the modelling, and the steady step-up in difficulty was the stretch.
The loop is what AI has disrupted. The junior now prompts rather than produces, so the attempt — and its feedback — mostly vanishes; the expert’s reasoning stays buried inside the model’s output rather than narrated aloud, so the modelling thins; and with the first rung gone, the stretch arrives all at once. The output is created, but every mechanism that used to build knowledge and skills has quietly been switched off.
Of course, the mechanisms themselves haven’t stopped working — they just no longer happen as a by-product of the work. So the emerging job of the L&D team isn’t to invent a new way to build knowledge and skills; it’s to take the same loop the work used to deliver by accident and deliver it by design instead.
From Growth as By-Product to Growth By-Design
It’s worth pausing on why the work carried that loop so reliably, because it explains why losing it hurts so much — and points at the fix. Here’s the move I want us to make as designers. That routine work was never valuable for its output. The output was always disposable — a first draft someone else corrected, a summary someone else acted on.
It was valuable because it was the primary delivery mechanism for tacit knowledge: the kind of knowing-how that, as Polanyi observed, we can’t fully articulate and therefore can’t teach directly. Nonaka built a theory of organisational learning on the same insight — tacit knowledge transfers through shared practice, through doing the thing alongside people who already know, not through instruction. The junior tasks were the shared practice. The deliberate-practice loop was smuggled in through the grunt work.
Which means the thing AI has actually done is not “automate routine work.” It has automated the carrier of the tacit curriculum. The apprenticeship wasn’t sitting next to the grunt work — it was hiding inside it. Take the work, and the hidden curriculum goes with it.
This is why this wave feels categorically different to anything we have experienced before. We’ve automated cognitive work before, but not the mechanism which leads to understanding.
Take the calculator as an example. When it absorbed arithmetic, it removed a task — long division by hand — but it didn’t remove the thing that taught number sense. Children still learned what division was, when to reach for it, how to sanity-check whether an answer was plausible, why a result felt off. The calculator took the grunt work of computation and left the conceptual scaffolding intact — in fact, the usual argument for calculators in classrooms is exactly that they free up attention for the higher-order reasoning. The judgement-building layer sat above the task that got automated, untouched. Typesetting and filing were the same: we lost the manual labour, not the understanding that surrounded it.
What’s happening now because of AI is the inverse. AI doesn’t sit beneath the understanding, handling the mechanical layer while the thinking carries on above it. It reaches up into the understandings layer itself — structuring the argument, weighing the options, making the call — which is precisely the layer the old tasks existed to develop.
The calculator automated the answer and left the reasoning; AI increasingly automates the reasoning and leaves us with the answer. So this isn’t “the calculator, again, but for knowledge work.” This time, the automation and the learning mechanism are the same activity — so removing one removes the other.
And that’s the precise thing we, as designers, now have to rebuild on purpose: not the task AI took, but the development that used to hide inside it.
A New Design Challenge
The vanishing “workplace apprenticeship” model poses a new challenge to those of us who design learning experiences. For decades, our job was essentially to design what happened around the work — the course before the role, the workshop alongside it, the e-learning to plug a gap. The apprenticeship itself, the actual development that happened in the doing, largely designed itself organically within the work. Now, it doesn’t.
The new challenge is to design the development that used to be automatic: to take the developmental loop that work delivered for free — model, attempt, feedback, stretch, fade — and engineer it deliberately back into the work itself. Put another way: we’re no longer just designing courses about the work - we’re designing the work so that doing it develops the people doing it.
If that sounds abstract, let me share a real world example. Heather Stefanski, McKinsey's Chief Learning Officer, described on a recent podcast how the apprenticeship that used to happen on its own at the firm eroded during COVID — and how the conclusion that she and team drew was that you can't simply hope it happens or expect it to happen. Her term for the response is purposeful apprenticeship: judgement transfer designed in on purpose, rather than left to proximity and luck.
Four Critical Design Principles for the Age of AI
Despite a lot of chat over the last couple of years or so, AI doesn’t demand that we reinvent the mechanisms through which humans learn; the learning science already named them, decades ago, and the evidence for each is still 100% viable and relevant.
What’s new is only that we now have to supply the primary mechanism for workplace learning deliberately instead of letting the work supply it organically. Here’s that this looks like in practice:
Our L&D instincts would likely tell us to take each mechanism above and build a training exercise for it: a course or coaching conversation that an employee stops to complete, off to the side, before returning to real work. But that’s still designing around the work. What if we did the opposite — and put the mechanisms back inside the real work?
In practice, that means engineering the real task itself so that producing it forces the mechanism to fire. Concretely:
Instead of a course on how to structure a client deck, the real deck the junior is building surfaces a partner’s annotated version — reasoning attached — at the moment they start.
Instead of an exercise on critical evaluation, the real workflow won’t let them drop in an AI-generated slide until they’ve written one line on why it belongs.
Instead of a workshop on forming a point of view, the real analysis won’t unlock AI until they’ve logged a hypothesis first.
Same deliverable, same intended outcome — but the unit you’re designing is no longer a course, or even an exercise: it’s the conditions of the task itself. This is learning in the flow of work and apprenticeship in its most literal and distilled sense: the employee never stops doing their job, but the job itself is now built to develop them while they do it.
Growth By-Design in Practice
To make this concrete, I’ll run all four principles through one ordinary task: a new starter writing a real client deck. Each principle re-engineers a different moment of that same task.
1. Make expert reasoning visible. The mechanism: observational learning — we learn judgement by watching experts think, but only if the thinking is visible, and experts have an “expert blind spot” (Nathan & Petrosino, 2003): they’ve chunked their reasoning so thoroughly they no longer narrate it.
What we need: the junior to grasp why a deck is structured the way it is, not just copy the format.
Traditionally: a “how to structure a client deck” course, plus years of sitting near partners absorbing the reasoning by osmosis.
Now: the moment they open the deck, the workflow surfaces a partner’s real deck on a similar problem — with the reasoning attached. “I led with the cost slide because this CFO distrusts growth stories.” The thinking is externalised at the point of need. Verb: externalise.
2. Force the conceptual mode. The mechanism: self-explanation — learners who ask “how does this fit what I know?” catch 9× more of their own errors than those who paraphrase (Chi et al., 1989). The Anthropic RCT confirms it: copy-pasters lost the most skill, conceptual-questioners the least.
What we need: the junior to understand the analysis, not just relay AI’s version of it.
Traditionally: an exercise on critical evaluation, done off to the side.
Now: they can generate with AI freely — but the workflow won’t let them drop in a slide until they’ve written one line beneath it: why this slide, why here, why this evidence. Pasting without understanding goes nowhere. Verb: engineer friction.
3. Differentiate friction by ability — backwards. The mechanism: the expertise-reversal effect — scaffolding that helps novices holds experts back. Strömberg’s 26,811-student study shows AI penalised the highest achievers most, because they’re the most efficient at outsourcing.
What we need: the junior to form their own point of view before reaching for the tool.
Traditionally: heavy support for strugglers, free rein for the high performers (exactly inverted).
Now: before the workflow grants AI access, it asks for a hypothesis and a confidence level — and it gates hardest for your sharpest people, the ones most at risk of never forming a view of their own. Verb: gate.
4. Require the attempt before the assist. The mechanism: the generation effect — producing an answer yourself strengthens it, but reading the answer first undermines it. Consulting AI first is reading the answer first.
What we need: the junior to be able to structure a recommendation unaided.
Traditionally: show a model answer, ask them to produce one like it (which now means prompting AI for it).
Now: the workflow withholds the AI draft until they’ve sketched their own — then shows both side by side and asks what differed and why. The gap between attempt and model is the feedback, and it only exists because they went first. Verb: sequence.
Closing Thoughts
It’s worth being clear about what is and isn’t actually changing here, because the noise of the last two years has blurred this quite a lot.
What AI does not change is how humans build capability. The foundational loop — model, attempt, feedback, stretch, fade — is as true today as it was in the 50+ years before ChatGPT existed.
What AI does change is where those mechanisms come from. For decades they were an organic by-product of doing the work — nobody designed the apprenticeship; it ran itself. AI has quietly switched that off, not by breaking the mechanisms but by removing the work that used to carry and deliver them. The result is that, if we want to build capability, we need to design and deliver the mechanisms it requires. That single shift — from capability as a by-product to capability by-design — changes our job in two major ways:
First, it changes what we design. The unit of learning is no longer the course, or even the clever exercise off to the side, because the learning was never off to the side — it was in the work. The thing we now design is the work itself: the task, and the conditions around it, engineered so that doing it develops the person doing it. That’s learning in the flow of work taken literally, not as a slogan.
And it changes our role. “Design the course, build the resources, run the survey” used to be the bread and butter of our work. The job that potentially replaces it is not just bigger but also more complex. Out jobs is not producing content about the work, but designing and shaping the conditions of the work. In practice, this pulls us into the room where the work itself gets designed, alongside the managers and system owners who control the tools, the access and the review steps. That’s a harder job, a less siloed one, and a far more consequential one than the old job description ever allowed.
None of this is settled, of course. Every mechanism here is proven, but the assembly — a whole body of work re-engineered along all four principles at once — is still in research and development. The diagnosis is solid, the components rest on decades of evidence, but the proof it “works” as a new way to develop employees’ knowledge and skills is still in the post.
Which is exactly why this is one of the most interesting problems in our field right now. AI didn’t make our job obsolete: it made the easy version of it obsolete — and handed us the potential to rebuild a version that actually matters and actually works. I’d rather work on that than on the next onboarding course the job no longer needs.
Happy designing!
Phil 👋
PS — If you want to position yourself at the bleeding edge of AI-augmented L&D, apply for a place on my AI Bootcamp for L&D.
