Hey folks 👋
If you’ve followed my work for a while, you’ll know I’ve been arguing for some time that the course is losing its place as the default unit of learning. What I haven’t been able to show you — until now — is exactly what’s replacing it.
Over the past twelve months, some of the world’s biggest knowledge-work employers have redesigned how they train people, and a clear pattern runs through every rebuild. Big Law firms are running AI role-play tools that let lawyers practise questioning witnesses. KPMG is rolling out a tax-career simulator across its US tax practice. Meanwhile, Deloitte and PwC are tearing up their graduate training and starting again.
TL;DR: the world’s leading knowledge-work organisations are dropping stop-and-learn courses as the default unit of learning, and replacing them with practice-and-feedback cycles built around the specific moments and decisions that make someone good at the job.
In this week’s post, I break down what these organisations are actually building and what it means for those of us who design learning: the three design moves that repeat across every rebuild, a zero-budget version any team can run next week, and the critical question nobody’s asking — does any of it actually work? Templates and a vendor checklist included.
Let’s dive in!
Why the Course is No Longer Our Focus
Here’s an uncomfortable truth our field has always half-known: courses were never where most workplace capability came from. The old 70-20-10 rule of thumb had the direction right even if the numbers are debated — the bulk of professional skill was built through the work itself and the people around it, not through formal training.
What actually made people good at their jobs was an informal, unnamed apprenticeship: years of routine tasks, completed in close proximity to experts who have feedback and whose experienced judgment slowly rubbed off over time.
The course survived as our default unit of learning anyway — and for an understandable reason: the apprenticeship that was actually developing people’s expertise ran itself, for free, invisibly, as a by-product of doing the job. Nobody had to design it.
So “designing learning” came to mean designing the small formal layer of development that existed around the work, and the course was a perfectly sensible container for it.
AI has broken this arrangement not by making courses worse, but by switching off the engine that was actually doing the development work. AI has removed the way employees used to build their expertise: years of routine work, powered by lots of feedback and observation of more senior colleagues doing the work well. That’s exactly the work AI now automates — and, to a large extent, makes invisible. The routine tasks juniors learned on are increasingly done by the model; the senior reasoning they absorbed by proximity now sits hidden inside its output.
To be clear, people still need knowledge in their heads — you can’t judge AI’s confident answer without it. What’s changed is that transmitting knowledge was never the binding constraint on competence, and the thing that was — the slow formation of judgment through real work — has just lost its delivery mechanism.
So the big question becomes: how do we teach expertise in the age of AI? The answer emerging from the world’s largest organisations is: by design. What the work used to deliver by accident — attempt a task, get feedback, watch how an expert handles it, go again on something harder — is being deliberately engineered back in, as practice-and-feedback cycles built around the specific decisions that make someone good at the job.
The goal is to teach the scepticism, judgment and insight that used to develop over years more intentionally and rapidly, from day one. Deloitte, for example, has rebuilt its three-year audit graduate scheme around simulated scenarios, with exams front-loaded into year one; PwC has rewritten its early-career curriculum to train employees as reviewers of AI’s work from their first week on the job.
So what the biggest employers are now acknowledging — openly, with budgets attached — is the thing 70-20-10 always claimed: courses were never actually what made people good at their jobs - the apprenticeship was. And since the apprenticeship no longer happens by default, it has to be built on purpose. As I introduced in an earlier post, for L&D this means a shift from designing courses to designing apprenticeships.
What follows is a field report on what the v1 of this this world looks like. Read it not as corporate news, but as a preview of how increasingly learning gets designed in 2026 and beyond.
What Workplace Training 2.0 Looks Like
So what does an engineered apprenticeship actually look like? Not in theory — in shipped, budgeted, rolled-out practice?
Below are four builds from the past twelve months, drawn from law, tax, audit and consulting. They differ in form — two are simulators, one is a rebuilt graduate scheme, one is a rewritten curriculum — but read them side by side and you’ll see the same design logic running through all four: find the decisions that make someone good at the job, create the practice and feedback the work no longer provides, and start on day one.
As you read, notice what was designed in each case. None of these organisations set out to build a course about the work. Each one built a way to practise doing it.
1.Big law is rehearsing depositions with AI
Six major US firms — Orrick, K&L Gates, McDermott, Littler, Taft and Brownstein — piloted DepoSim this year, an AI simulator for depositions — the formal questioning of a witness under oath before trial — built by AltaClaro with Verbit (Law.com, 2026). AI agents play the witness, opposing counsel and the court reporter; the associate takes the deposition live and gets structured feedback afterwards. The pilot logged 160+ hours across the six firms, with 97% of participants strongly agreeing it was valuable (Above the Law, 2026) — a satisfaction figure, note, not a performance one.
Look at what was designed here: not a module about deposition procedure — a place to practise doing a deposition, with feedback. And look at how firms use it: one litigation chair ran the simulator live in front of 500 attorneys, pausing to explain each decision as he made it; mentor–mentee pairs use its feedback to start coaching conversations; associates with real depositions coming up use it to prepare (Thomson Reuters Institute, 2026). Watch an expert work, practise it yourself, get feedback, apply it on a real case — recognisable learning design, arrived at by instinct.
2.Arbitration teams are rehearsing cross-examination
At Three Crowns, the international arbitration firm, junior associates use a simulator called Atelier — developed with Dr Megan Ma of Stanford Law School’s LIFT Lab — to rehearse cross-examining a witness, a skill traditionally reserved for advocates years into their careers, adapting live as the witness responds unexpectedly (Bloomberg Law, 2026).
The Bloomberg Law piece, written from inside this work, draws a boundary every designer in this space needs. Skills built on recognition — knowing what counts as a key document, spotting when a contract term is unusual — can be practised in a simulator, and learned faster there than through years of live work. The social side of this things, however, can’t. Nobody simulates their way into a profession’s culture and unwritten rules, or into carrying themselves well with a client. This is an emerging design principle: simulate the performance-critical professional decisions, but keep the “human-to-human” learning in the room.
3.KPMG is simulating an entire tax career for 10,000 people
KPMG’s build is the scale story. TaxSIM, built with Centaurian AI, is designed to give junior tax professionals the skills that used to take years of repetitive client work — and it’s expected to be made available to KPMG’s roughly 10,000 US tax professionals later this year (Tekedia, 2026). That’s not a pilot in a training suite; it’s a bet, across a national practice, that designed practice can replace the practice clients used to pay for without knowing it.
The logic is stated openly: AI absorbed the routine work, the repetitions and practice disappeared with it, and something has to replace the reps.
4.Deloitte and PwC are redesigning the entry-level job
From September, Deloitte is restructuring its three-year audit graduate programme in England and Wales around realistic scenarios — classroom, virtual and simulated — with the stated aim of getting graduates competent faster than the old learn-by-doing-tasks route (ICAEW, 2026).
PwC goes further. With AI handling data gathering and routine audit tasks, new hires are expected to work as reviewers and supervisors of AI output almost immediately — first-years, in the firm’s telling, will feel like the managers of a previous generation within three years (Business Insider, 2025).
Step back and the common thread is this: the training is changing to match a redefined entry-level job. The skills these programmes now teach from day one — critical thinking, negotiation, professional scepticism, giving and taking difficult feedback — aren't new skills. They're the mid-career skills, the ones that used to arrive in year five after the routine work had done its slow teaching. With that work gone, the timeline has collapsed: KPMG is training juniors to work as "managers of agents" — its global AI workforce lead's phrase — and PwC has expanded early-career training in exactly the human skills that used to wait, because client-facing responsibility no longer does (Business Insider, 2025).
The “Engineered Apprenticeship” in Practice
Analyse the “engineered apprenticeship” and every build has the same anatomy: a practice loop the learner goes through, a craft behind the scenes that powers it, and a target the whole thing is aimed at.
Law, tax, audit and consulting all seem to have arrived at this same anatomy independently — which is usually a sign the underlying logic is real.
#1. The experience: make the call, see the expert's call, go again
Nobody is simulating document review, reconciliations or first-draft decks — the work AI now does. Every build runs the same loop instead: put the learner in a realistic situation and make them commit to a decision — ask the question, challenge the client’s number, clear or escalate the flagged entry — before they see any answer. Then show them what an experienced person would have done, and why. Then have them go again. Attempt, compare, retry: that’s the cycle replacing the course.
The point of the loop is specific: building enough first-hand pattern experience to judge work — including AI’s work — without needing years of live cases to accumulate it. You can’t teach someone to spot a wrong answer by describing wrong answers. They have to have made some wrong calls themselves, seen what they missed, and gone again. A course starts from “what do they need to know?” These builds start from “which calls make someone good at this job — and what does the expert notice that the novice misses?”
#2. The craft behind it: expert reasoning, captured on purpose
The loop only works if there’s an expert’s call to compare against — which points at the second move, and the hardest craft behind all of these builds.
The old apprenticeship passed on judgment by proximity: you sat near the partner and, over years, something rubbed off. Every new build depends on the opposite: experts explaining their thinking so it can be captured and taught. The litigation chair pausing mid-simulation to say why he asked that question. Feedback criteria that spell out what a good answer looks like. A senior’s marked-up draft turned into a practice scenario.
Getting reasoning out of experts’ heads used to happen by accident, through seating plans. Now it has to happen on purpose — and it’s becoming the most valuable skill in learning design, because every simulator and every feedback rubric is only as good as the expert thinking inside it. It’s also a skill most L&D teams have never practised: we were trained to collect content from SMEs, not to ask them how they decide.
#3. The target: judging AI's work, from day one
The third move is about where all this practice is aimed — and it’s the clearest statement of the new entry-level job. At PwC, reviewing AI output is the work now. So the skills that used to arrive in year five — scepticism, knowing when the plausible answer is wrong, knowing when to escalate — have to be taught from the start.
This is the hardest problem of the three, because checking skill has always come from experience, and experience is what a first-year doesn’t have. A recent position paper, based on watching how a software team actually used AI, puts the stakes plainly: on routine, well-defined tasks, AI narrows the gap between novices and experts — consistent with the famous call-centre study where the weakest agents gained most (Brynjolfsson, Li and Raymond, 2025). But on complex tasks that need domain judgment, AI widens the gap: experts use it to go further, novices accept plausible-but-generic output (An, 2026). It’s an argument from observation rather than a large study, but it matches what many of us see daily.
Which makes the design stakes plain: giving juniors AI without teaching them to check it doesn’t close the gap. It compounds it. The organisations in this piece have understood that — and built the checking practice in from day one.
Implications for L&D
Let’s be clear about what does and doesn’t change here, because the noise of the last two years has blurred it.
What stays the same: everything we know about how people learn. The loop running through every build in this piece — watch an expert, attempt the task, get feedback, go again on something harder — is not new. It’s the same evidence base our field has stood on for decades: deliberate practice, cognitive apprenticeship, feedback, fading support. AI hasn’t changed how humans build capability by one inch. What it’s changed is where that loop comes from: the work used to deliver it for free, by accident, and now somebody has to deliver it on purpose. That somebody is us. Our science is not obsolete — it’s finally, explicitly, the job.
What changes: every stage of our workflow. For most of our field’s history, “designing learning” meant designing the course which lives outside the work. The skills that mattered were content skills: writing, structuring, storyboarding, building.
If the primary unit of design is now the practice-and-feedback cycle, each stage of the job changes — and each change demands a skill most of us have never practised. Here’s how I see the workflow and skillset changing:
1. Analysis: from mapping content to mapping judgment.
Before: ask SMEs what the role needs to know, and turn the answers into a topic list and learning objectives.
After: work out which decisions make someone good at the role — the calls that separate an expert from a novice — and which of those AI has removed the natural practice for. The deliverable is a map of the role’s critical decision moments, not a topic list. The craft underneath is expert reasoning elicitation: experts have automated their own judgment so thoroughly they can’t list it, but they can narrate it through real cases — which is what the decision-narration template above is for.
2. Design: from building courses to building cycles.
Before: sequence content into modules, add an activity and a quiz, deliver, mark complete.
After: design the loop — make the call, see the expert’s call, go again — around each critical decision: realistic situations, commitment forced before any answer is shown, feedback that compares the learner’s reasoning to the expert’s, difficulty that steps up, support that deliberately fades. Sometimes that’s a simulator, sometimes a ritual like AI rounds; the format is a downstream choice, not the starting point.
3. Delivery: from around the work to inside it.
Before: the course before the role, the workshop alongside it, the e-learning to plug a gap — while the actual development happened invisibly on the job.
After: the workflow that asks for your hypothesis before AI’s answer, the senior’s reasoning attached to the real deck, the monthly debrief of actual cases — so that doing the job develops the person doing it. This is work design, an emerging specialism, and it puts us in a new room: alongside the managers and system owners who control the tools and the review steps. A harder, less siloed job than the old one — and a far more consequential one.
4. Evaluation: from reporting activity to proving transfer.
Before: completions, attendance, satisfaction scores, maybe a quiz pass rate.
After: define what good judgment looks like in the real role, measure whether practice shows up there once the scaffolding is gone, and kill or fix what doesn’t. Here’s the uncomfortable part: almost every number in this piece is a satisfaction number, and nobody selling these tools has published evidence that performance inside them transfers to the work — because nobody else in the buying chain has a reason to demand it. Being the function that does may be the most valuable thing L&D can offer. It’s our assurance role — and our licence to keep the budget.
Conclusion: L&D 2.0?
To close, I’d like to be precise about the claim that “the course is dying,” because strictly speaking this isn’t true. The course isn’t declining in volume; it was only ever the small formal layer of how people got good — the 10 in 70-20-10. What’s changed is what needs to be designed. The 70 and the 20 used to come for free, but now they don’t.
It’s worth doing the maths on this. For decades, L&D’s design remit covered roughly 10% of how people actually built capability — the formal course layer — while the other 90% designed itself, invisibly, inside the work. Seen this way, it becomes clear that AI hasn’t shrunk our job; it has handed us the other 90%. The practice, the feedback, the observation of experts, the graduated stretch — everything the informal apprenticeship used to deliver by accident is now ours to deliver by design.
Put another way: AI isn’t automating L&D’s job away — it’s expanded our remit by an order of magnitude, from designing the layer around the work to engineering the whole engine of capability.
Before we get carried away, though, a sobering stat. This week I ran a rapid analysis of 400 live L&D role postings. 90% still name a course as the thing the hired person will design and build. Most employers, in other words, are still advertising the old job.
But an emerging few are already hiring for the new one — and the salary differences are intriguing. Roles focused on course design and production carry a median salary of £35,000 in the UK in 2026, while roles framed around AI, consulting and capability-building pay a median rate of £52,551. In the US, the gap is starker still: $85,493 versus $161,400 (Hardman, 2026). Read those two numbers together and you get the real story of this moment: AI is commoditising the course and - at the same time - professionalising the learning design role.
For decades we've argued that L&D deserves a more powerful and proactive seat at the table. We never quite got it, in part because our remit was 10% of how organisations actually developed. The unexpected gift of AI is that the other 90% — the part that actually builds capability — has just landed on our desk. Seen through this lens, AI isn’t a threat to our profession: it’s the biggest expansion of its remit in living history.
Happy designing!
Phil 👋
PS: If you want to learn how to use AI like a pro in your day to day work, apply for a place on my AI Bootcamp for L&D.
