The Intensification Trap
The pitch was simple. AI handles the routine work. You handle the creative work. The machine takes the repetitive tasks off your plate, and you use the freed-up time for thinking, strategy, the work that only humans can do. Everybody wins. The productivity gains are real and the workday gets lighter.
That was the pitch. Here is what actually happened.
In February 2026, Aruna Ranganathan and Xingqi Maggie Ye published the results of an eight-month ethnographic study in the Harvard Business Review. Ye, a doctoral student at UC Berkeley’s Haas School of Business, had embedded herself in a US technology company of approximately two hundred employees. She observed them twice weekly, tracked their communication channels, and conducted more than forty in-depth interviews across engineering, product, design, research, and operations. The study covered April to December.
The finding was not what the productivity narrative predicted. Employees who used AI tools did not work less. They worked more. They worked faster. They took on a broader scope of tasks. They extended work into hours that used to be empty — lunch breaks, evenings, the ten minutes before a meeting. The tool that was supposed to create slack consumed it.
Ranganathan and Ye identified three forms of intensification. The first was scope expansion: people began taking on work that previously would have belonged to someone else or would not have been attempted at all. The definition of “my job” widened. The second was boundary dissolution: because AI makes it easy to start and continue tasks, work seeped into moments that used to function as pauses. A prompt during lunch. A query before bed. The natural stopping points in the workday dissolved. The third was parallel processing: workers kept multiple threads alive simultaneously, running AI processes in the background while attending meetings, reviewing code, or drafting documents.
None of this was mandated. Nobody told these employees to work more. The intensification was voluntary — or appeared to be. The tool made doing more feel possible, accessible, and rewarding in the moment. So people did more. And the doing more became the new baseline.
The Jevons Paradox of Labour
This pattern has a name, although it comes from a different century and a different resource. In 1865, the English economist William Stanley Jevons observed that James Watt’s improvements to the steam engine — which dramatically increased the efficiency of coal use — did not lead to a reduction in coal consumption. They led to an increase. The engine was more efficient per unit of work, so it became economical to use it for more work. Total coal consumption rose, not fell.
The Jevons Paradox — the finding that efficiency gains in resource use lead to increased total consumption of that resource — was about coal. But the mechanism is about incentives, and incentives do not care what resource they attach to.
AI is the Watt engine of cognitive labour. It makes each unit of knowledge work more efficient. A draft that took two hours takes forty minutes. A data analysis that required a specialist can now be approximated by a generalist with the right prompt. A report that would not have been written — because the cost exceeded the perceived value — now gets written, because the cost dropped below the threshold.
The per-task efficiency is real. Ranganathan and Ye do not dispute it. The paradox is that per-task efficiency, aggregated across an organisation that measures output volume, produces intensification, not relief. The total amount of cognitive work increases. The workday gets denser, not shorter. The treadmill speeds up.
This is not a technology failure. The technology did exactly what it was designed to do. This is an incentive design failure. The organisational systems surrounding the technology reward output volume. When a tool makes it possible to produce more output, the system absorbs the increase. The freed-up time is not returned to the worker. It is reinvested in more work — automatically, structurally, without anyone making a conscious decision to intensify.
The Metric Rewards Intensification
Charles Goodhart, a British economist, articulated a principle in 1975 that has become one of the most cited observations in organisational design: “When a measure becomes a target, it ceases to be a good measure.” Goodhart was writing about monetary policy, but the principle applies wherever metrics drive behaviour.
Most organisations measure productivity as output per unit of time. Tasks completed. Tickets resolved. Documents produced. Lines of code written. Emails sent. The metric is volume.
When AI enters this system, volume increases. The metric improves. Dashboards turn green. Quarterly reports celebrate the productivity gains. And nobody asks whether the people producing the output are working harder, thinking more, or resting less — because the metric does not track cognitive load. It tracks volume. And volume went up.
The Upwork Research Institute, in partnership with Workplace Intelligence, surveyed 2,500 workers globally in 2024 — including 1,250 C-suite executives. The headline finding: 77 per cent of employees using AI said the tools had added to their workload. Not reduced it. Added to it. The sources of the additional load were specific: 39 per cent reported spending more time reviewing or moderating AI-generated content. 23 per cent reported investing more time learning to use the tools. 21 per cent reported being asked to do more work as a direct result of AI.
The executives, meanwhile, were enthusiastic. Ninety-six per cent expressed high expectations that AI would enhance productivity. The gap between the executive expectation and the employee experience is not a communication problem. It is a measurement problem. The executives are looking at the volume metric. The metric is up. The employees are living inside the volume metric. The volume is up because they are doing more work.
Goodhart’s law predicts this precisely. The organisation targeted output volume. The metric improved. And the metric ceased to be a good measure of what it was supposed to represent — which was not volume, but value per unit of human effort.
What AI Leaves Behind
There is a subtler dimension to the intensification that the volume metrics miss entirely.
When AI handles the routine tasks — the first draft, the data extraction, the template, the summary — what remains for the human is not routine. It is the hard part. The judgment calls. The ambiguity. The decisions that require context, nuance, and domain expertise that the model does not have.
This is the cognitive residue of AI-augmented work. The easy tasks are automated. The hard tasks remain. And the hard tasks are harder now — not because they changed, but because the easy tasks used to provide cognitive rest. The routine work was a break. Not an exciting break, not a restful break in the conscious sense, but a period of lower cognitive demand that allowed the brain to recover between episodes of effortful thinking.
Daniel Kahneman’s distinction between System 1 (fast, automatic, low-effort) and System 2 (slow, deliberate, high-effort) thinking is useful here. Routine tasks live in System 1. Judgment tasks live in System 2. System 2 is metabolically expensive — it consumes glucose, it produces fatigue, it has a limited daily budget. A workday that mixes System 1 and System 2 tasks is sustainable. A workday that is predominantly System 2 — because System 1 tasks have been automated — is not sustainable at the same duration.
The Workday research, conducted with Hanover Research and published in January 2026, found that 37 per cent of the time employees saved using AI was lost to rework — correcting errors, verifying outputs, rewriting content that failed to meet quality or context requirements. Seventy-seven per cent of employees reviewed AI-generated work as carefully as, or more carefully than, work done by humans. Only 14 per cent consistently reported net-positive outcomes from AI use.
The rework is System 2 work. Evaluating AI output requires judgment, comparison, verification — all effortful, all metabolically expensive, all drawing from the same limited cognitive budget that the “freed-up” time was supposed to replenish. The employee saved forty minutes on the draft and spent twenty-five minutes verifying the output. The net saving is fifteen minutes. But the cognitive profile of those twenty-five minutes is more demanding than the forty minutes they replaced, because reviewing someone else’s work for errors you cannot predict requires more sustained attention than producing work according to a pattern you already know.
The volume metric sees fifteen minutes saved. The employee’s prefrontal cortex sees twenty-five minutes of high-demand verification added to a day that was already saturated.
The Demand-Control Collision
Robert Karasek’s demand-control model, developed in 1979 and refined across four decades of occupational health research, describes job strain as the interaction between two variables: the demands placed on the worker and the control the worker has over how to meet those demands.
High demands combined with high control produces active work — challenging, sustainable, and associated with learning and professional development. High demands combined with low control produces high-strain work — the configuration most strongly associated with burnout, cardiovascular disease, and chronic stress.
AI-augmented work, as Ranganathan and Ye documented it, is increasingly high-demand. The scope is wider. The pace is faster. The boundaries are dissolved. The cognitive load per hour is higher because the routine tasks are gone and the judgment tasks remain.
The control dimension is where the trap closes. In most organisations, the individual worker does not control the volume targets. They do not set the expectations for output. They do not decide how the time savings from AI are allocated. The time savings are automatically absorbed by the system — by the next ticket in the queue, the next project that becomes feasible, the next report that someone requests because “the AI can do it quickly.”
High demands. Low control. Karasek’s model predicts strain. The prediction is not theoretical. PwC’s 29th Global CEO Survey, published in January 2026 and covering 4,454 CEOs across 95 countries, found that 56 per cent reported no significant financial benefit from their AI investments. Only 12 per cent reported both cost and revenue improvements. The investment is producing intensification at the worker level without producing returns at the organisational level. The treadmill is faster but the building has not moved.
The Incentive Nobody Redesigned
Here is the sentence I want to hold: the problem is not the AI. The problem is the incentive structure that surrounds the AI.
When an organisation deploys an AI tool and measures its impact through output volume, the organisation has built an intensification machine. Not intentionally. Not maliciously. Structurally. The tool increases per-task efficiency. The metric rewards volume. The volume increases. The employees absorb the increase. The cycle continues.
The escape from the trap is not to remove the tool. The tool works. The per-task efficiency is genuine. The escape is to redesign the incentive.
This requires measuring something different. Not output volume. Output value per unit of cognitive effort. This is harder to measure than volume — which is precisely why organisations default to volume. Volume is easy to count. Cognitive effort is not. But Goodhart’s law is clear: the easy metric, once targeted, ceases to measure what matters.
What would it look like to measure differently?
It would mean tracking not how many reports the team produced, but which reports led to decisions. Not how many tickets were resolved, but which resolutions held — which ones did not bounce back as re-opened issues within thirty days. Not how many documents were drafted, but which documents were read, cited, and used. The metric shifts from production to impact.
It would also mean tracking cognitive load directly. Not as a wellness initiative — as an operational variable. Workday’s research found that nearly nine in ten companies had updated fewer than half of their roles to reflect AI capabilities. Only 37 per cent of heavy AI users received increased skills training. The organisations are deploying tools without redesigning the jobs those tools change. The demand profile of the role has shifted — more judgment, less routine — but the role description, the performance metrics, and the expected output volume have not.
The role is new. The incentive is old. The gap between them is the intensification trap.
The Treadmill Metaphor
I keep coming back to the image of a treadmill — not because it is clever, but because it is precise.
A treadmill increases your speed without changing your location. You run faster. Your heart rate rises. Your muscles work harder. You sweat. And at the end, you are in exactly the same place you started.
AI-augmented work, under the current incentive structure, is a treadmill. The per-task speed increases. The volume increases. The cognitive effort increases. And the outcomes — measured at the organisational level — do not improve proportionally. PwC’s data is unambiguous: 56 per cent of CEOs report no financial benefit. The treadmill is running. The building is not moving.
The treadmill is not broken. It is functioning exactly as designed. The problem is that nobody asked whether a treadmill was the right machine for the goal. If the goal is to move forward — to create more value, not more volume — you need a different machine. A machine that rewards impact, not throughput. A machine that measures the quality of decisions, not the quantity of outputs.
The treadmill measures steps. The destination requires direction.
What Organisations Can Do
I am wary of prescriptive lists. They reduce complex problems to action items, and action items without context are Goodhart’s law waiting to happen — the list becomes the target, the target ceases to be useful. But there are structural interventions that change the incentive landscape rather than just adding tasks to it.
Redefine what the AI time savings are for. If AI saves a team eight hours per week, those eight hours must be explicitly allocated — and the allocation must not default to “more of the same work.” The options include: deep thinking time (unscheduled, unmonitored blocks for complex problem-solving), skill development (learning that is not directly productive but builds future capacity), recovery (cognitive rest that reduces the cumulative load and prevents the burnout that Karasek’s model predicts). The allocation is a design decision. If the organisation does not make it, the system will make it for them — and the system defaults to more volume.
Separate the volume metric from the performance metric. Volume is an activity measure. It tells you what happened. It does not tell you what it was worth. Performance metrics should track outcomes: decisions made, problems solved, quality sustained, errors avoided. These are harder to count. That difficulty is the point. Easy metrics produce Goodhart’s law. Hard metrics produce information.
Redesign roles to reflect AI-augmented work. If AI handles the routine, the role is now a judgment role. Judge roles require different skills, different training, different cognitive profiles, and different rest patterns. A role that was 60 per cent routine and 40 per cent judgment is now 20 per cent routine and 80 per cent judgment. The performance expectations, the training, and the workload targets must reflect that shift. They almost never do.
Make cognitive load visible. Not as a wellness initiative in the HR portal. As an operational dashboard alongside the volume metrics. When leaders can see that volume is up and cognitive load is up, the conversation changes. The intensification becomes visible. And visible problems are easier to address than invisible ones.
The Integration
Here is the tension I want to hold without collapsing.
AI tools are genuinely efficient. They reduce the time and effort required for specific tasks. The per-task improvement is measurable and real. Ranganathan and Ye did not find that AI fails at its intended function. They found that it succeeds — and the success, in the context of volume-based incentives, produces a paradox that is harmful to the people doing the work.
The tool is not the trap. The incentive is the trap. The tool is a treadmill. Whether it moves you forward or keeps you running in place depends entirely on the system it sits inside.
Organisations that deploy AI without redesigning their incentive structures will get intensification. Not because the AI is flawed, but because the organisational design is optimised for a metric that AI makes easy to inflate. The metric goes up. The humans inside the metric get more tired. The outcomes, at the level that matters — financial returns, sustainable performance, employee wellbeing — do not improve.
Organisations that redesign the incentive — that measure impact rather than volume, that allocate saved time deliberately, that acknowledge the cognitive shift from routine to judgment work — will get something different. Not necessarily more output. Possibly less output, by volume. But better decisions. Fewer errors. Sustainable pace. And tools that people use because the tools make their work better, not because the tools make their work more.
The intensification trap is not a technology problem. It is an incentive design problem wearing a technology costume. The technology did what it was asked to do. The question is what the organisation asked it to do — and whether anyone thought to ask the people doing the work whether “more, faster” was the right answer.
The treadmill is running. The question is not whether to step off. It is whether to change the destination.