Few productivity gains in modern knowledge work have been measured as enthusiastically, or as narrowly, as those credited to AI. Outputs, drafts and lines of code per worker are all up. Almost every metric an HR analytics dashboard surfaces is moving in the right direction, and almost none measure what is being built into the worker.
A 2026 randomised controlled trial by Shen and Tamkin found that developers using AI assistance completed tasks with measurably worse conceptual understanding and weaker debugging ability than their unassisted peers.1 The time saved on execution was offset, in roughly equal measure, by the time spent querying the model. More work shipped. Less skill formed. No net hours back.
This is not the standard productivity-cost trade-off. The cost is not money or hours; it is capability, the skill and judgement a worker carries forward across tasks. Knowledge work compounds. So does its erosion.
Most organisations adopting AI have treated the question as one of permission and tooling: which roles, which workflows, which licences. The harder question is what each delegation costs in human depth, and whether that cost is being tracked anywhere on the management dashboard. In most organisations, it is not.
The Shen and Tamkin result echoes an earlier BCG and Harvard study of 758 management consultants, which recorded a 12.2% productivity gain on tasks where AI was a good fit, and a measurable decline in accuracy on tasks where it was not.2 The authors called the boundary between the two regions the 'jagged frontier' — the irregular line that separates problems where the model is reliably competent from those where it confidently produces wrong answers. The trouble is that the line is invisible from the inside. Most workers cannot tell, in the moment, whether their current task sits inside the frontier or just outside it. Confidence in the output and reliability of the output do not move together.
The implication is uncomfortable. AI raises the floor on routine work while lowering the ceiling on judgement work, and both effects can run in the same person, on the same morning, without either being visible.
The mechanism behind the capability loss is not mysterious. Expertise is built through deliberate practice: effortful work at the edge of current ability, where the brain is forced to model the problem from scratch. Psychologists call the absorbed state of that work flow, and the consensus across decades of research is that it is the substrate on which professional judgement is built.
AI assistance, used well, accelerates output. Used reflexively, it bypasses the phase where skill actually forms. The model produces a plausible draft, the worker tidies it, and the cognitive load that would have built capability never lands. The worker has been productive. The worker has not been practising.
The closest analogy is satellite navigation. GPS gets the driver to the destination faster than a map, and leaves the driver unable to find the route again unaided. The driver completes more journeys per week. The driver becomes worse at driving.
From assistants to agents
The pressure on capability sharpens as AI moves up the autonomy ladder. Gartner's 2026 maturity model describes a progression from assistants in 2025, to task-specific agents in 2026, to collaborative agents in 2027.3 Each step enlarges the surface area of work being delegated, not augmented. Assistants suggest. Agents act. Collaborative agents own multi-step workflows that previously made up a junior's first two years in the role.
Each rung up that ladder is a rung down on the practice ladder. The work that used to teach someone how to do the work is increasingly done by something else, and the apprentice pipeline that produces senior judgement five years from now depends on the practice still happening today.
What to measure instead
The standard productivity dashboard treats output as the dependent variable, with tools, hours and headcount as the inputs. There is no slot for capability. The fix is not to discard productivity metrics; it is to widen the frame so that depth shows up alongside speed.
Four signals are worth tracking. The first is deep-focus duration: how much time, in a given week, an individual spends in extended single-task cognitive work without interruption from chat, tools, or AI queries. It is the closest practical proxy for flow.
The second is unassisted task completion rate. Organisations periodically test, on a sample of routine work, whether team members can still finish the task without AI assistance and at what quality. The metric works as a leading indicator, not a punishment device. A team whose unassisted rate is falling faster than its assisted rate is rising has a capability problem masked by a throughput gain.
The third is learning velocity, the rate at which new categories of problem are tackled and resolved inside the organisation. Teams that ship more of the same look healthy on output and stagnant on this measure.
The fourth is the ratio between AI-augmented and AI-free working blocks across the week. The number itself matters less than the deliberate choice to measure it. Teams that never schedule unassisted blocks tend not to have any.
A practical playbook
Slowing AI adoption is the wrong response. The right one is to pair adoption with structure that protects the conditions for skill formation. Three practices appear repeatedly in organisations that take the capability question seriously.
The first is to pair AI-assisted sprints with unassisted deep-work blocks, named and scheduled, long enough to support absorption (typically ninety minutes or more) and protected from incoming messages.
The second is to tier work by capability stakes. Routine tasks where the model sits reliably inside the jagged frontier are good candidates. Tasks that develop a new skill, or sit near the edge of current ability, are not. The decision should be deliberate, not a function of which application window happens to be open.
The third is to use workstyle analytics that distinguish cognitive engagement from tool activity. Most workforce-monitoring products today measure presence, application use, and AI tool adoption. None of those signals tell a leader whether the team is building expertise or transferring it to the model. Flow-state detection, of the kind WorkstyleIQ is developing, is one of the few emerging approaches that even attempts the distinction at scale. The technology is early. The measurement question it points at is not.
The interesting question for the next eighteen months is not whether AI raises individual output. It plainly does, within its frontier and on its terms. The question is whether the organisations betting on those gains have any way of telling, two years from now, what their teams have become capable of without it, and what they have quietly stopped being capable of with it.
Footnotes
-
Shen, A. and Tamkin, A. (2026). 'Is AI Productivity Degrading User Capability?' Galdren. https://galdren.com/ai-degrading-user-capability ↩
-
'The Impact of Generative AI on Knowledge Work.' Triangulate Knowledge, citing the BCG and Harvard Business School study of 758 management consultants. https://www.triangulateknowledge.com/blog/the-impact-of-generative-ai-on-knowledge-work ↩
-
'What Do the Best AI Productivity Reports Reveal in 2026?' UC Today (2026), summarising Gartner's AI maturity model. https://www.uctoday.com/productivity-automation/ai-productivity-reports-2026/ ↩
