The Missing Rungs: What Leadership, Lab, and Crowd Skips

Ethan Mollick’s essay Making AI Work: Leadership, Lab, and Crowd offers a framework for getting value from AI inside a company: Leadership to set direction, a Lab to concentrate the learning, and the Crowd of employees to discover uses in their own work. To his credit, he admits the framework might not be enough. He’s right. I have no quarrel with the three groups — they’re a fair sketch of who’s involved. What the piece is missing is the part I’ve spent my career on: the structure of how work actually improves. Naming who should improve things is not the same as knowing how improvement happens.

And here is the thing about how improvement happens — it doesn’t matter whether you’re talking about sports, martial arts, engineering, or AI. The progression is the same. Every field has its version of the same ladder, and most organizations are stuck near the bottom of it. That is why the individual gains everyone reports almost never add up to gains the company can actually see.

We Have Run This Experiment Before

We actually agree on a lot — including this: parachuting outside experts into a company to fix it with AI is not the answer. We can be confident about that, because we’ve already run the experiment. In the 1990s it was called Business Process Reengineering. Michael Hammer told companies to redesign their processes from a blank sheet; by 1995 it was a $51 billion consulting industry, and studies put the failure rate near 70%.¹ Hammer himself later conceded he hadn’t been smart enough about the human side. The pattern was simple: outsiders were hired to map processes the people doing the work could have described in an afternoon, and the redesign never stuck, because the people who knew the work were treated as the object of the change instead of its source.

My colleague Tyson Heaton makes this case directly in The Reengineering Is Coming, and he’s right that it’s coming again — this time wearing AI. The “forward-deployed engineer” now sweeping into large companies is the reengineering consultant with a better toolkit and the same blind spot. Heaton points to MIT research finding that 95% of enterprise AI pilots produce no measurable impact.² That number should sound familiar. It is the 1990s all over again.

So we agree on what doesn’t work. We even agree that the framework alone isn’t enough — Mollick says as much himself. Where he stops, I want to keep going, because knowing what fails doesn’t tell you what succeeds. The answer isn’t a better class of outsider or a more energetic crowd. It’s structure — a ladder you have to climb.

First, Define the Work

Before you improve anything, you have to be clear about what you are improving. Mollick raises a version of this himself. His point is about which work to do: once AI makes a task cheap to perform, the real question is no longer how to do it but whether it’s worth doing at all. That’s right as far as it goes, and it’s more than the speed story most people tell. But “do it or don’t” is the bluntest cut there is. It treats a task as a single thing you keep or kill — when most tasks are neither. Work is built in layers: tasks made of sub-tasks made of steps. And a single task usually carries value and waste together, so deciding at the whole-task level means keeping the waste to save the value, or throwing out the value to lose the waste. To do better you have to go inside the task and separate the two. Any task, looked at honestly, breaks into three parts:

Value-added work — the part the customer actually wants and would willingly pay for. The weld that holds, the answer that’s correct, the design that ships.
Incidental work — everything required to get the value-added part done that the customer would never pay for on its own: the setup, the searching, the walking, the handoffs, the checking.
Pure waste and rework — the part that shouldn’t exist at all. Doing it twice, fixing defects, producing things nobody needed.

Improvement means working on all three, and the fastest returns almost always come from attacking the pure waste first. Most jobs are mostly the second and third categories; the value-added sliver is usually thinner than anyone expects.

Directed properly, AI can improve any of the three. Left to its own devices, it simply speeds up whatever is already there — standardizing the waste, automating the rework, making a report nobody needed arrive faster. That is why separating the three has to come first: until you have, you cannot tell a real gain from faster waste.

The Ladder Is the Same Everywhere

Improving any process follows the same four-step progression, whether the tool in your hand is a stopwatch, a kettlebell, or a large language model. Each rung is a different skill, and you can’t skip one.

Rung	What it is	What it takes
1. Learning	Learn the work as it’s actually done: the major steps, the key points in each, the reasons why, and how exceptions get handled	Doing the work yourself, through every case — regular and irregular — not just watching it
2. Standardizing	Fix the current best-known method to a time standard and a quality standard	Performing it repeatedly until it reliably holds
3. Improving	Take the standard apart and improve the pieces	Analysis — Eliminate, Combine, Rearrange, Simplify
4. Redesigning	Rethink the whole flow, end to end	A customer-focused value-stream view, and the leadership will to disrupt the system

Rung one is learning, and it’s the one almost everyone takes for granted. Companies assume the work is already understood — that the tasks are defined and known. They rarely are. The real knowledge is tribal: undocumented, full of exceptions and tacit judgment carried in people’s heads, not on paper. You learn a job by doing it — working through the routine cases and, more importantly, the irregular ones, until you know in your hands what makes the result good or bad and why. You do not learn it by standing beside someone, watching, and writing down what you see. That distinction matters, because watching-and-mapping is precisely the outside expert’s mistake: an observer captures the happy path and misses every exception — and exceptions are where real work lives. If you can’t perform the job through its hard cases, you don’t yet understand it well enough to improve it, and you certainly can’t hand a piece of it to an AI and trust the output.

Rung two is standardizing: the current best-known way to do the job, performed repeatedly until it reliably holds to a time standard and a quality standard, with the edge cases accounted for. Repetition is the point — a standard you hit once by luck is not a standard. This is your baseline. Without it you have nothing to measure an AI-assisted version against; you’ll feel faster and have no way to know whether you’re actually better.

Rung three is improving — incremental refinement, what improvement circles call kaizen. You take the standardized process apart and improve the pieces. The oldest method in the trade is ECRS: Eliminate, Combine, Rearrange, Simplify — in that order, waste first. It works, and it has a ceiling: it makes a known process better but stays mostly local, inside the boundaries of the task as it already exists. Most good practitioners top out here, and many never reach it at all, because they skipped the first two rungs and have nothing solid to improve.

Rung four is redesigning — radical breakthrough rather than refinement, kaikaku in the original Japanese. You stop improving the pieces and rethink the whole flow, end to end, from the customer backward. It takes a value-stream view of the entire system and the willingness to tear it up and rebuild it rather than tune it. This is where AI’s largest leverage lives — removing or reconceiving whole stretches of work, not speeding up a single step. And it never happens on its own. Disrupting an entire system is a leadership act, which is exactly where Mollick’s framework is thinnest.

You Can Read Where a Company Is Standing

Because the rungs are a progression, they are also a diagnostic. You can read where an organization sits from what you see.

What you see	Where they stand
No improvement at all	Rungs 1–2 — it never learned its own work well enough to standardize it. (Most companies.)
Local improvement only: a faster report, a quicker step, Mollick’s “secret cyborgs”	Rung 3, on isolated tasks — real, but it never adds up, because people tune their own step, which is all they’ve been shown.
System-level gains, the kind that show up in the numbers leadership watches	Rung 4 — and almost no one is here.

Rung four is the hard one, because it takes the end-to-end view: value-stream or full-system process mapping, an understanding of how every step connects to the customer at the far end. Few people in any company hold that view at all, and fewer hold it for the whole system rather than their own corner.

Now lay the AI requirement on top. To drive a rung-four redesign with AI, a person has to understand the entire system and understand the technology well enough to see where it changes what’s possible. Almost no one is both at once. That intersection — deep system knowledge and real AI fluency in the same head — is vanishingly small in most organizations. And that, finally, is why the enormous individual gains never become organizational ones. The people who could connect the two are the rarest people in the building, and no framework has been built to develop them.

What This Asks of the Three Groups

The framework isn’t wrong. It’s empty until you put the progression inside it.

The Lab, as described, stays generic and abstract — prototype workflows, build benchmarks, share the prompts that work. None of it is wrong, and none of it is enough. The improvement the framework is after comes from system redesign, and system redesign is not a function of prompting, prototyping, or benchmarking. Those put you on the ladder; they don’t carry you up it. And the AI skills climb the very same ladder as the work:

Rung	The AI practice	What it gives you — and where it stops
1. Learning	Prompting	Necessary, not sufficient — you won’t prompt-trade your way to systemic improvement.
2. Standardizing	Reusable skill files for recurring work	Standardizes the common cases; edge cases and exceptions still need expert routing and handling.
3. Improving	Agents built with an improvement architecture in mind — ECRS: eliminate waste, combine steps, rearrange the sequence, simplify	Substantial local gains.
4. Redesigning	AI-native workflows end to end — autonomous learning and RL loops, humans and AI in the orchestration and check phases	System-level redesign, the actual prize.

Sharing prompts is genuinely useful, but in most cases it won’t leave the first rung. The Lab’s real job is to walk people up all four, on real processes that serve a real customer. Practice on toy problems is just training dressed up as transformation.

The Crowd was never a matter of permission, and treating it as one misses what is already happening. Your best people are using AI right now, with or without anyone’s blessing — it even has a name, shadow IT. Hand them a mandated tool that underperforms while the frontier models sit one browser tab away — a corporate Copilot, say, when the better model is a click away — and they will quietly pay for the good one themselves and use it anyway. It is part of why so much of this use stays hidden: people routing around tools that aren’t good enough and not advertising it. Permission isn’t the lever — it isn’t even the question.

What the crowd actually needs is to be encouraged and developed, and not through a lab or a pilot alone. A lab or a dojo can build skill, but the learning that matters happens where the work happens: in individuals and teams wrestling with their actual processes, not in a sandbox off to the side. That takes time, clear goals, and real resources — and leaders who genuinely push people to experiment and learn. Mistakes have to be allowed, on one condition: that they are caught, corrected, and folded back into a better standard. Developed this way and given a line of sight to the whole system, a crowd can climb past rung three. Merely permitted, it will hand you a thousand local optimizations and not one transformation.

And it all starts and ends with Leadership — which is exactly where the framework gives the least direction. Mollick does ask more of leaders than a green light: he wants them to paint a vivid future state and anticipate how the work will change. Fair enough — but the lever he actually reaches for is permission, making it safe to use AI in the open. That is necessary and nowhere near enough. You cannot advance a system by advancing permissions. Rung four — the only rung that yields system-level gains — takes architecture: someone has to define what customer value actually is, then strip away the bundles of waste and handoffs that have piled up across the whole system, work that cuts straight through the functional silos every organization instinctively defends. That is leadership in the full sense — setting the purpose and direction, committing the resources, naming the goals and objectives, working in the open with transparency, and then checking the results and learning from them. It also asks something leaders often resist: learning enough about the nature of AI to lead it honestly. You cannot direct what you refuse to understand.

And it cannot be handed off. Leadership is scarce, and AI is one more demand on a crowded day — but delegating a change this fundamental to a lab, a pilot, or an outside expert isn’t strategy, it’s abdication. Everything else can be built beneath a leader. This cannot.

Start Where You Stand

There is no single starting line. Inside any company you will find the whole range — in improvement skill and in AI fluency alike — from people who have neither to the rare few who have both, and the same spread runs across teams, functions, and processes. The job is not to march everyone back to the bottom. It is to assess honestly where each person, team, and process actually stands on both, and to climb deliberately from there.

What hasn’t changed is the pattern itself. Learning, standardizing, improving, redesigning is how work has always gotten better, and AI does not rewrite it. What is new is the intersection: a powerful, fast-moving technology has landed on top of two old problems most organizations never solved — not knowing how to lead, and not knowing how to improve. AI doesn’t relieve either one. It presses on both at once, and faster than anyone is used to. In most organizations this will not resolve automatically, or gracefully.

The advantage won’t go to whoever experiments fastest or stands up a lab soonest. It will go to those who can see which rung they are on and climb it deliberately — so that AI redesigns the system instead of merely accelerating the mess. Same technology. Different methods. Different outcomes.

Michael Hammer first argued for radical redesign in “Reengineering Work: Don’t Automate, Obliterate,” Harvard Business Review (July–August 1990). The roughly $51 billion market size, the ~70% failure rate, and Hammer’s later admission that he had underestimated the human dimension are recounted in Tyson Heaton, “The Reengineering Is Coming” (Lean Enterprise Institute). The high failure rate also appears in Hammer and James Champy, Reengineering the Corporation (1993). ↩
MIT NANDA, The GenAI Divide: State of AI in Business (2025), which reported that roughly 95% of enterprise generative-AI pilots produced no measurable bottom-line impact — as cited in Heaton, “The Reengineering Is Coming.” ↩