The Productivity Stall Was Not an Accident

Focus Did Not Retire. It Got Reallocated. Why Productivity Flatlined. Why Analytics Did Not Save It. Why Copilots Will Not Either Unless We Change the System.

A supervisor sits in front of three screens that all claim to be the truth. One system says the line is running. Another says the work order is still open. A third says quality is holding material the dock is already staging. The historian trace shows a drift that looks harmless until you know what it does to viscosity, then to yield, then to scrap, then to the customer complaint that arrives two weeks later, after the root cause has been buried under the next fire. An operator is waiting. A maintenance tech is waiting. A planner is waiting. The organization is waiting, because the organization was built to wait until the right person has interpreted the signal, assembled context, and secured permission.

That scene is the only chart that matters. It contains the full story of the last fifteen years. More sensors. More dashboards. More alerts. More analytics. More insight. No proportional step change in the thing leaders actually need. The ability to convert evidence into permitted action quickly, repeatedly, and safely.

The modern enterprise is excellent at being right in retrospect. It can narrate a miss with forensic elegance, wrap it in governance language, and present it as prudence. It can explain why the shipment was late, why the yield fell, why the customer called, why the overtime climbed, why the line ran hot. It can do all of that and still be unable to change the mechanism that produces the next miss. Getting it right demands a different posture. It demands admitting that certainty is often performance, not truth. It demands building systems that reduce the distance between a signal and a verified intervention.

What would have to be true for this outcome to keep repeating.

That line is not a flourish. It is an audit. It forces a firm to stop rehearsing explanations and start tracing causality. It forces leaders to separate observation from inference, then separate inference from action. It forces the uncomfortable recognition that the same outcome is repeating because the architecture keeps reproducing it. If the architecture is the cause, then rhetoric cannot be the fix.

The easiest story to tell about productivity is also the most incomplete. Experts retired. The skills pipeline thinned. Plants became harder to staff. Complexity rose faster than training could keep up. The old hands walked out the door with the knowhow. All of that is real. None of it explains why the enterprise spent a decade and a half investing in analytics and still did not bend outcomes the way it promised it would.

Focus did not retire. Focus moved. It moved up and out of the operating system. It moved into financial engineering, transaction logic, and risk-managed narratives. It moved into the parts of the firm where rewards were faster, accountability was more diffuse, and the story could be defended with a spreadsheet. It stayed there long enough to change what the enterprise rewarded, who rose, what counted as intelligence, and what got treated as noise.

Focus did not retire. It moved. The operating system kept running. It stopped getting better.

That shift created a repeating misdiagnosis. Outcomes stopped improving in proportion to effort, so leaders treated the gap as an information problem. If the firm just had better data, better dashboards, better analytics, then decisions would improve. Decisions would speed up. Productivity would rise. So the firm added tooling that produced more information. Signals multiplied. Exceptions multiplied. Hand-offs multiplied. Verification work grew. The organization became more articulate about what was going wrong, then it stayed slow in the only way that matters.

There is a second misdiagnosis nested inside the first. When leaders do not like what the outcomes are saying, they often decide the measurement must be wrong. Sometimes it is. Sometimes it is not. The posture of getting it right does not let you dodge either possibility. It forces you to pressure-test the data without treating doubt as an escape hatch. It forces you to ask whether the measurement is missing quality improvements, whether industry composition is distorting the picture, whether indices are hiding value, and whether the macro line is masking firm-level dispersion. Then it forces you to ask the only question that matters operationally. Even if measurement is imperfect, does the organization convert evidence into action faster than it used to.

That is where the tool story ends and the truth begins. A conversion system is not a dashboard. It is an operating model. You can know more and still be slow. You can measure more and still be brittle. You can explain a miss and still be unable to prevent the next one, because the constraint is not visibility. The constraint is the distance between evidence and permitted action.

The deal room did not cause the productivity stall alone. It made it easier to live with it, because it offered an alternate way to win.

The Deal Room Ate the Plant

Cheap money did more than lower borrowing costs. It changed what counted as smart. When capital is cheap, the hurdle rate falls. Leverage becomes easier to justify. Buybacks become easier to finance. Acquisitions become easier to rationalize. The timeline of reward compresses. The spreadsheet becomes a weapon, because value creation can be claimed without touching the physics of production.

This is not a moral claim. It is a structural claim. If a firm can raise the stock price through capital structure choices faster than it can raise productivity through capability building, attention migrates to the lever that pays faster. Over time, the firm learns to treat operations as a cost center to optimize, not a capability to build. The organization rewards fluency in narratives over fluency in constraints, then it mistakes that fluency for control.

The reason this is so dangerous is that it does not look like abandonment. It looks like sophistication. A deal model is clean. A synergy number is clean. A headcount reduction is clean. A buyback authorization is clean. They all produce instant charts. Capability building does not. Capability building is slow, noisy, and local. It requires leaders to live inside the mess long enough to see what is actually true about their processes, their people, and their decision flow.

Productivity work is stubborn in a way a spreadsheet cannot tolerate. It requires repeatability. It requires disciplined operations. It requires what leaders often call culture, but what the floor experiences as standards that hold under pressure. That is why the best productivity work is politically expensive. It forces the organization to admit where it has been lying to itself, which is why organizations prefer projects that produce motion without requiring confession.

Mergers and acquisitions add another layer. They promise scale, but they also produce integration friction, and integration friction is paid in expert attention. The firm’s best operators become translators between systems, not builders of mechanisms. The people with the deepest knowledge get reassigned into reconciliation work. The organization calls that integration. In practice, it is a tax. If that tax is not measured, it becomes permanent architecture.

Drift is what happens when a firm keeps choosing moves that do not threaten existing power. It is what happens when capital and attention stay in places where feedback is fast and blame is diffuse. It is also what happens when boards demand certainty about outcomes that cannot be certain, so leaders choose the moves with the cleanest stories, not the strongest mechanisms. The firm becomes right on paper while the operating system becomes slower in practice.

This is where being right becomes a trap. Being right is a rhetorical goal. It encourages people to defend a prior decision as if defense preserves legitimacy. Getting it right is a design goal. It forces people to reopen assumptions, follow mechanism, and change the system when the evidence says the system is wrong. The enterprise has been trained to do the first because the first is safer, and safety is what drift wears as a uniform.

Risk-Free Governance Selects Safe Decisions

A second mechanism hardened the drift. Modern governance can separate power from consequence. Many leaders and boards win asymmetrically. If the stock rises, they win. If the strategy fails, the downside is spread across the workforce, customers, and time. That asymmetry does not require bad people. It requires a system that caps personal downside while amplifying upside.

The deeper issue is not ethics. It is architecture. In a functioning enterprise, authority is tied to liability. If you make a bad decision, you feel it. That fear is not a flaw. It is the discipline that keeps a business honest. It forces obsession with the customer, intolerance for waste, and the humility to stay late and fix what is broken. When that link is severed, a new class emerges. Not defined by title, but by insulation. The risk-free insider. The person who can be wrong without paying the full price of being wrong.

Under those conditions, the organization selects safer decisions. Safer decisions are rarely called safe. They are called prudent, disciplined, de-risking, protecting the franchise. The vocabulary is always responsible. The operational consequence is often drift, because the safest move is the one that can be explained without touching the operating system.

This is the hollow competence problem. People can be competent at the wrong thing. They can be fluent in governance language, portfolio narratives, and the performance of prudence. They can explain, align, and socialize. They can run the meeting. They can make the deck beautiful. They can produce agreement without producing mechanism. That is why the firm can feel professionally managed while its operating system quietly degrades. The organization becomes excellent at being right, then forgets the harder discipline of getting it right.

This is where procedure becomes a comfort object. When leaders fear consequence, they add steps. They add reviews. They add committees. They add sign-offs. Each step is defensible. Each step sounds like control. Then the steps become permanent. The organization starts paying for time as if time were free, then it acts surprised when competitors buy the future with speed.

When a risk-free system meets a real operational problem, it often chooses delegation over ownership. Bloated cost. Broken supply chain. Stale product. The reflex is to buy a strategy deck. Not because leaders cannot think, but because decks are career insurance. If the plan fails, blame can be routed outward. Accountability can be diluted through process. This is not management. It is outsourced responsibility. The enterprise spends shareholder capital to protect insider legitimacy, then calls it diligence.

Financial engineering is attractive because it is legible. It is reversible. It is explainable. Productivity transformation is messy. It threatens internal power. It exposes ignorance. It forces accountability back into the operating layer. It forces decisions to be owned, not narrated. That is precisely why it is avoided, even when it is the only move that protects long-run competitiveness.

A board can test whether it is drifting. Take the last year of major initiatives and ask one question. Which ones materially reduced the time between a deviation and stabilization on constraint processes, without increasing safety or quality risk. If the firm cannot answer with operating evidence, it is buying activity. It is paying for motion and calling it discipline.

The most revealing part is what does not get said. When leaders present an initiative, they describe its capabilities. They describe its adoption curve. They describe its governance. They rarely describe which operational loop will run faster because of it. When they do, they describe the loop in words, not in timestamps. That is how the organization protects itself from accountability. Words can be defended. Time cannot.

Knowledge Drained. Tools Filled the Gap

Now the labor story matters, but it matters for a reason many boardrooms understate. Manufacturing productivity is not an office metric. It is embodied knowledge in a physical system. It is disciplined operations, formal process knowledge, standard work, changeover mastery, constraint management, maintenance intuition, quality containment instincts. It is also informal knowhow, the pattern recognition that never made it into a manual, the small tell that precedes a failure, the sequence memory that keeps a deviation from becoming consequence.

When that knowledge drains faster than it is replenished, the system becomes fragile. The tempting response is to buy information because information is easier to buy than expertise. Over the past decade and a half, many enterprises did exactly that. They built dashboards, alerts, pipelines, and insights that demanded interpretation from fewer remaining experts.

That choice did not just add visibility. It added cognitive load. It added reconciliation work across systems that disagree. It added handoffs. It added meetings. It added verification tasks. It added the kind of administrative labor that looks like diligence right up until you audit how much time it steals from control.

Information rose for years. Authority stayed slow. The gap between them became the new cost center.

This is how analytics trapped many firms in the shallow end of the solution pool. Describing symptoms got easier. Mechanism work stayed hard. The organization got better at narration. It became more fluent, then it stayed slow. A line does not stabilize because a chart looks better. It stabilizes because someone intervenes correctly, in sequence, with permission, using a playbook grounded in process reality.

The deeper problem is that instrumentation scales faster than discipline. A sensor can be installed in a day. A dashboard can be built in a month. A disciplined response system takes years, because it must be trained, practiced, and enforced under pressure. When leaders choose the fast path, they are often choosing the path that increases variance. More signals do not reduce variance if the response mechanism is weak. They can increase it, because the enterprise reacts to noise, then it calls the reaction responsiveness.

The productivity stall looks like a macro mystery because it is a micro habit. It is a habit of describing instead of converting. It is a habit of adding surfaces instead of deleting handoffs. It is a habit of treating governance as protection rather than as a system that must be engineered for speed and safety at the same time.

The Copilot Temptation

Copilots entered this landscape as a promise of relief. They sit beside the worker and offer drafts, summaries, recommendations, and answers. In bounded language work, that promise is often real. The mistake is to assume that local efficiency becomes system throughput by default.

A plant does not win because email takes fewer minutes. A plant wins because it stabilizes faster. A plant wins because containment happens earlier. A plant wins because downtime is prevented rather than narrated. You can save minutes in the office layer and still lose hours in decision queues, because the bottleneck is not prose. The bottleneck is permission and verification under consequence.

This is where the evidence becomes a warning label instead of a headline. In one field experiment, hundreds of entrepreneurs were given access to a GPT-4 powered mentor through a familiar messaging channel. The result was not a clean average lift. High performers improved materially. Lower performers declined. The divergence appears tied less to different advice and more to how people selected and implemented advice.

In other workplace settings, assistance shows the opposite pattern. In bounded customer support work, generative assistance raised productivity on average, with large gains for less experienced workers and little change for the most experienced. In field experiments with professional knowledge workers, the pattern looks jagged. AI can raise speed and output within its capability boundary, then degrade accuracy and judgment outside it. The point is not to choose a favorite study. The point is to respect the boundary conditions. When work is bounded, feedback is rapid, and correctness is checkable, assistance can compress learning curves. When work is open-ended, feedback is delayed, and quality depends on sequencing and judgment, assistance can widen dispersion.

Manufacturing contains both regimes. Many deployments treat them as one, then act surprised when the results are uneven.

Public sector trials are useful for a second reason. They separate perceived time savings from system conversion. Large deployments have shown that users can save meaningful minutes per day on routine tasks and still remain constrained by the larger system. Separate evaluations have found that those time savings do not necessarily translate into department-level productivity improvement, in part because of overhead from low quality outputs, human oversight, and induced tasks that would not exist without the tool.

That is the missing middle. Minutes saved are not productivity gained unless the operating system converts minutes into throughput and risk reduction. A tool can reduce drafting time and still increase coordination time. It can reduce search time and still increase review time. It can accelerate output volume and still leave the bottleneck untouched, because the bottleneck is not the creation of artifacts. The bottleneck is who can act, when, with what evidence, and with what recorded accountability.

Verification is the tax that decides whether the tool is relief or load. In a plant, plausible is not a standard. Safe is the standard. Correct is the standard. Verified is the standard. If an output is not trusted, it must be verified. Verification consumes the scarce resource the enterprise is already losing, which is deep judgment. Any tool that increases uncertain output can increase cognitive load even when users feel helped.

Plausible is cheap. Verification is expensive. In high-consequence work, the bill decides the value.

There is a metric that exposes the lie. It does not care how many copilots were deployed. It does not care how modern the interface looks. It cares about time under consequence.

Event. Detection. Reasoning. Intervention. Stabilization.

How long does that chain take on the loops that pay the bills. How often is the first intervention correct. How often does the system require escalation. How many handoffs occur. How much verification work is needed before action is permitted. How many approvals are required after the evidence is already visible.

If a firm cannot measure that chain, it cannot claim control. If an AI deployment does not reduce that chain on real operational loops without increasing risk, it is assistance, not conversion. Assistance can be worth buying. It should not be confused with productivity.

This brings the story back to its root. The decisive constraint is not intelligence. It is permission.

The decisive constraint is not intelligence. It is permission.

Access is not a technical detail. Access is the architecture of power. If a tool cannot reliably reach the evidence that matters inside the workflow that matters, it becomes an overlay. It sits beside the person who still must reconcile systems, interpret signals, validate output, and chase approval across decision gates.

The best evidence that this is true is not academic. It is operational. When a system cannot access the right data, it generates generic output. When it cannot act inside the workflow, it creates suggestions that must be manually executed. When it cannot capture the audit trail, it forces humans to rebuild justification after the fact. Each limitation adds more verification work. Each limitation pulls experts back into the role of translator, which is the exact role the enterprise can no longer afford to scale.

Inference Removal Is the New Labor

The demographic reality is not getting easier. Deep tenure is declining in many settings. Meanwhile the cognitive surface area keeps expanding, because each new tool adds more signals and more reconciliation paths. When fewer experts must cover more surface area, the enterprise cannot keep demanding that humans serve as the inference engine.

This is where strategy has to change from answer engines to inference removal.

Inference removal does not mean eliminating human judgment. It means eliminating the requirement that humans be the integration layer between systems, permission boundaries, and slow governance. It means the system captures context without forcing operators to retype it. It means recommendations are grounded in governed evidence rather than generic plausibility. It means actions route through explicit permission models so safe interventions can occur without rebuilding authority from scratch each time. It means reasoning and outcomes are recorded so learning compounds rather than evaporating in meeting notes.

Inference removal also changes what leaders measure. Instead of measuring adoption, measure loop time. Instead of measuring usage, measure stabilization. Instead of measuring output volume, measure error reduction and containment speed. A firm can only get this right if it stops pretending that productivity is a feeling. Productivity is a system behavior. It is visible in where queues form, where handoffs accumulate, where exceptions concentrate, and where people wait for permission.

Education is the same story in a different room. If generative tools become answer vending machines, they can erode the very capability that determines long-run performance, which is the ability to select, frame, and implement advice under constraint. The core skill is not retrieving an answer. The core skill is asking the next better question, forming a test, spotting the wrong assumption, and knowing what must be true before action is safe.

The productivity stall was not an accident

It was a set of choices that reallocated focus, rewarded safe narratives, and treated visibility as control. Copilots will not fix that by existing. They will make the enterprise more fluent. They will not make it faster where it matters unless permission changes, verification is engineered, and decision loops are redesigned.

Until authority is re-tied to consequence, the firm will keep selecting safe stories over mechanism, and it will keep mistaking professional narration for control.

The posture that matters is not bravado. It is precision. The enterprise does not need to be right. It needs to get it right.

References

This piece draws on official productivity and manufacturing measurement work that frames the slowdown debate and its competing explanations, including Bureau of Labor Statistics analysis of economy-wide and industry-level productivity trends and the Federal Reserve Bank of New York’s analysis of the manufacturing productivity slowdown, and it incorporates evidence that measurement can understate manufacturing productivity when quality improvement is not fully captured and when sector composition dominates aggregate changes (BLS productivity slowdown analysis. New York Fed Liberty Street Economics manufacturing productivity analysis. Atalay, Hortaçsu, Kimmel, Syverson working paper on manufacturing productivity measurement). It relies on capital allocation and incentive context that helps explain why attention can shift away from productivity capability building when financial engineering is rewarded (OECD work on weakness in business investment. Federal Reserve analysis of buybacks and investment). It uses workforce projections to treat demographic erosion as a control constraint rather than a staffing headline (Deloitte and the Manufacturing Institute workforce gap projections). It pressure-tests GenAI claims through boundary-setting empirical research that shows heterogeneous impacts depending on task structure and feedback loops, including open-ended decision contexts where selection and implementation drive dispersion, bounded work where feedback is rapid and correctness is checkable, and professional task environments where capability boundaries are uneven (Otis, Clarke, Delecourt, Holtz, Koning field experiment on heterogeneous GenAI impacts. Brynjolfsson, Li, Raymond workplace evidence on generative AI in customer support. Dell’Acqua and coauthors work on the jagged technological frontier). It uses UK government evaluations of Microsoft 365 Copilot to separate minutes saved from system-level productivity conversion and to highlight the verification and induced-work overhead that appears when tools are deployed into permissioned environments (UK cross-government findings report. UK Department for Business and Trade evaluation of M365 Copilot). It also draws on reporting and advertising claim discipline analyses that surface integration friction and the gap between perceived productivity and objective conversion in enterprise deployments (Reuters reporting on adoption and integration constraints. BBB National Programs NAD decision on Copilot claims). Foundational operating doctrine shaping the conversion framing is grounded in Michael Carroll’s prior work on conversion, decision latency, permission architecture, and inference removal, including Enough Intelligence. Shaping Destiny Without Digital Gods. The Line Between First Generation AI and Second Generation AI. The Question Engine. One Degree for Everyone and Everything. The Architecture of Permission No One Admits They Are Running. The Looming CoPilot Disaster in Manufacturing.