There was a time when understanding how work actually happens was a respected skill. Lean practitioners walked the floor. Six Sigma teams traced defects to their roots. Toyota Production System veterans treated waste as something you could see, measure, and design out. The job was operational: decompose the system, identify the constraints, watch where value actually flowed.
Then organizations got abstract.
Strategy floated up. Execution sank down. Process understanding got handed off to specialists or vanished into software few people could fully read. The people designing systems were no longer the people who understood them in detail. The people who still understood them found themselves explaining, more often than they used to, why their work mattered.
If you spent your career building that capability, you have noticed something lately. Every conversation about the future of work assumes the future belongs to a different skill set. AI engineers. Prompt designers. Data scientists. The phrase “process improvement” sounds, to a lot of executives, like something their parents did.
I want to suggest the opposite. The skill set you built has not expired. It has become, quietly, the missing ingredient in every AI initiative that is not working.
How did the work of understanding systems become unfashionable?
The failure pattern is common enough that it has a name: pilot-rich, transformation-poor. One global investment bank documented in a 2026 Harvard Business Review piece had deployed over 250 LLM-connected applications across its business units. Enterprise-wide performance gains: effectively zero. Every team reported productivity improvements; none of those improvements made it onto the balance sheet.
This is not a budget problem or a model problem. It is what happens when the end-to-end value chain never gets decomposed to find the actual constraints. Each team optimized its own slice, and the slices did not add up to anything.
Which is exactly the failure mode operational disciplines were built to prevent.
AI Did Not Make Process Thinking Obsolete — It Made It Urgent
Most strategy decks have not made this turn yet.
AI compresses the cost of cognitive execution. Analysis is cheap. Documentation is cheap. First drafts are cheap. The bottleneck has moved.
It has moved to defining the system correctly. Where are the inputs? What are the outputs? Where does value actually flow? Which steps exist because the work demands them, and which exist only because humans used to be the rate-limiting factor?
These are the questions Lean and Six Sigma trained people to ask. They are also the questions that determine whether AI creates value or amplifies dysfunction.
Wilson and Daugherty made the case directly in HBR in 2025. They call it “kaizen 2.0” — the principles from Toyota’s 1940s production system (worker empowerment, root-cause analysis, just-in-time, jidoka) map directly onto the requirements of effective AI implementation. The toolkit has evolved; the operating system underneath has not. A 2025 PECB analysis put a finer point on it: modern Lean Six Sigma only has a “family resemblance” to its predecessors — the practice has shifted from memorizing tools to building a systems-thinking mindset. The mindset is the durable part.
The conditions that made systems-thinking valuable did not disappear. They were hidden by a decade of comfortable abstraction. The cost of skipping the work is visible wherever AI has been deployed without it.
Faster Broken Processes Are Still Broken
The trap most organizations are walking into is to apply AI to the processes they already have.
An AI agent drafts a complex commercial contract in seconds. The contract then sits in a manual legal review queue for two weeks, because the firm’s risk and governance framework was designed for a world where contracts arrived once a week, not once a minute. The agent is, by every individual measure, working brilliantly. The system is, by every meaningful measure, as slow as before. Multiply that across a hundred enterprise workflows.
A 2026 HBR analysis frames this as the “last mile” problem. Process debt — the accumulated workarounds and handoffs from decades of acquisitions and reorgs — creates workflows so brittle that AI exposes problems faster than it can resolve them. Bolt automation onto a system designed for human limits and you get faster dysfunction, not better outcomes.
The investment bank with 250 applications is the cautionary version: its productivity gains failed to reach the balance sheet because the saved time was absorbed into low-value activities rather than reallocated. Goldratt would have recognized it in five minutes. It is the local-versus-global optimization problem in modern dress.
Theory of Constraints offers the diagnostic. A system’s performance is always limited by its bottleneck. AI applied to a non-bottleneck stage generates queues, delays, and information overload — the output of an over-resourced upstream stage piles up at the actual constraint, which has not been touched.
So the question is not whether AI makes processes faster. Of course it does. The question is whether the process should exist in its current form at all.
Toyota Did Not Automate 75 Spreadsheets — They Eliminated Them
Toyota offers the cleanest contrast.
Their resource allocation planning had grown, over decades, into a sprawling operation: per Deloitte’s 2025 reporting, roughly 50 team members worked across about 75 spreadsheets, navigating 50 to 100 legacy mainframe screens. The process worked. It was also an archaeological site. Each spreadsheet was a fossil of a constraint that no longer applied — coordination overhead from an era when batching, sequencing, and translation between specialists were the only ways to move information around.
The intuitive AI move would have been to automate each spreadsheet. Get faster at filling them in. Build a copilot for each mainframe screen. That move would have accelerated dysfunction and called it progress.
Toyota did something else. Its AI leadership team has described the founding question this way: “What problem are we trying to solve, and then what’s the fit-for-purpose technology?” They decomposed the workflow and classified each step against its underlying constraint — which steps existed because of physics or regulation, and which existed only because humans were doing the cognitive work. Then they redesigned the system from a clean sheet and deployed AI inside the redesigned process, not on top of the old one.
The planning team went from 50+ members to 6-10. Planning time dropped from hours to minutes. About 75 spreadsheets and 50-100 outdated mainframe screens were retired. None of it came through headcount cuts — the freed planners were redirected through a new “Talent & Experiences” function focused on upskilling. Jidoka applied to the transformation itself. The same AI tooling on the old workflow would have produced marginal improvement and called itself a win.
Merck shows the second version of the pattern. Pharmaceutical manufacturing has a chronic problem: false product rejects — good product flagged as defective. It is a specific form of waste, exactly the kind Lean was built to identify. Traditional approaches — manual calibration, statistical process control — had reached their limits. Merck combined the Lean diagnostic (define the waste, trace its root cause to the limits of existing detection methods) with generative AI synthetic data that trained models to distinguish real defects from false positives. False rejects dropped by more than 50% across multiple product lines. Lean defined the problem. AI executed the solution.
McKinsey put numbers on what happens when this scales. If organizations redesign workflows rather than automating isolated tasks, AI agents and robots could add roughly $2.9 trillion per year to the US economy by 2030. The conditional matters. The number does not arrive from automation. It arrives from redesign.
Where the older disciplines need to change
Experience with Six Sigma gives good reasons to question how directly these skills transfer.
The first is timing. DMAIC projects span three to six months. AI capabilities shift in weeks. The deliberate pace that made Lean reliable looks, under acceleration, like a liability. It is a fair critique. The skills are the durable part; the methodology’s traditional cadence can change. Toyota’s redesign required deep process understanding and did not take years. The mindset can operate at compressed timescales when the people doing the work already have it.
The second is more provocative. If AI can analyze workflows end-to-end, identify bottlenecks in real time, and propose redesigns autonomously, why does human decomposition still matter? AI is genuinely good at pattern recognition inside a defined system. It is meaningfully worse at the meta-question of whether the system should exist in its current form. Process debt is the proof: AI exposes brittleness in legacy workflows faster than it can fix it, because the fix requires organizational context and tacit knowledge that resist formalization.
The third objection is the deepest. Operational discipline is fundamentally about reducing variation and optimizing within existing system boundaries. Even applied to “redesigned” systems, that mindset may anchor organizations to incremental improvement rather than radical reimagination. The concern is real — and empirically inverted right now. Today’s failure mode in enterprise AI is not too much structure but too little. Pilot-rich and transformation-poor is a discipline deficit, not an excess. The 250-pilot bank is not innovating its way out; it is automating its way deeper into local optimization.
The fourth asks how much of the skill still transfers. The thesis claims operational skills transfer to AI-era system design. But when systems include autonomous agents, probabilistic outputs, and emergent behaviors, decomposition may change so much that calling it the same skill set is misleading. The transfer is real but not automatic. The cognitive orientation persists; the specific tools must evolve — which is why modern Lean Six Sigma already has only a “family resemblance” to its predecessors. The discipline has been mid-evolution for years. The risk is real for practitioners who refuse to update. The alternative — abandoning systems thinking entirely — is worse.
The Constraint Shift: A Framework for Knowing When to Redesign
You can run this on a workflow you already know.
-
Constraint Audit. Map every step. For each, ask whether it exists because of a structural necessity (regulatory sign-off, physical lead time, accountability) or a cognitive one (sequential review, translation between specialists, batch processing for human efficiency). Compute the ratio of cognitive steps to structural ones. Above 0.6 cognitive, you have a redesign candidate, not an optimization candidate.
-
Clean-Sheet Question. Define the desired outcome independent of the current process. Then design backward, respecting only the structural constraints. Every step in the current process that does not appear in the clean-sheet design is a candidate for elimination. This exercise usually fails through contamination — the people who know the current process cannot unsee its workarounds. Have someone unfamiliar with the legacy workflow review the redesign.
-
Bottleneck Reidentification. When constraints change, bottlenecks migrate. The bottleneck in a legacy process is usually cognitive — analysis, coordination, documentation. In a redesigned process it is usually structural — governance approval, data quality, organizational adoption. Optimizing a non-bottleneck stage generates queues and overload without improving throughput.
-
Structural Boundary Design. At the structural constraints you confirmed must persist, design explicit checkpoints. Not fewer gates — better-placed ones. For each, define what the checkpoint requires, who is accountable, and what AI can prepare in advance to cut the gate’s cycle time without removing human judgment.
-
Feedback Instrumentation. Build leading indicators for each structural boundary. If a structural constraint’s cycle time drops sharply, it may have shifted from structural to cognitive — and the system needs another redesign. Treat redesign as a continuous capability. The redesigned process becomes the new legacy the moment you stop watching.
What to Do Monday
Pick one workflow you know well. Map the steps. Run the constraint audit — cognitive versus structural — and compute the ratio.
If it is above 0.6 cognitive, the workflow is a redesign candidate. Ask the clean-sheet question before anyone writes a line of AI integration code for it.
If your organization’s AI productivity gains are not appearing on the balance sheet, you do not have a technology problem. You have a bottleneck identification problem. Find the constraint nobody redesigned around.
Treat operational discipline as a design capability, not an efficiency tool. Direct it at redesigning systems for changed conditions, not at optimizing systems whose assumptions stopped applying years ago.
Your skills did not become obsolete. They went dormant while organizations confused abstraction with sophistication. AI is making the dormancy expensive. The firms now running large fleets of pilots and wondering where the gains went are often the same ones that long ago stopped doing the decomposition work that makes any of this pay.
You already know how to find them.