AI productivity has a hidden capability tax

Consider an illustrative composite: a professional-services team reduces the time required to prepare a first draft from four hours to forty minutes. The first month looks excellent. More work moves through the system. The team appears responsive. Utilisation improves.

Six months later, a less flattering picture begins to emerge. Senior reviewers spend longer unpicking confident but poorly supported passages. Junior staff can assemble polished documents but struggle to explain the judgement underneath them. The team has become faster at producing an artefact and slower at establishing whether it deserves to exist in its present form. When the model, source material or workflow fails, too few people can reconstruct the reasoning without starting again.

This is a composite example, not a reported case. But it captures a management risk that is becoming visible across several very different evidence sets.

The question is no longer simply whether AI saves time. It often does. The harder question is what else disappears with the time.

The first dashboard can be right and still be incomplete

The strongest argument against a simplistic “AI makes people less capable” story is that AI can produce real productivity and learning gains.

A major study published in The Quarterly Journal of Economics followed 5,172 customer-support agents using a generative AI assistant. Access to the system increased issues resolved per hour by 15 per cent on average. Less experienced and lower-skilled workers gained around 30 per cent, and workers with two months of tenure and AI assistance performed about as well as unaided workers with more than six months’ experience.[1]

That is not an illusion created by counting keystrokes. It is a meaningful operational outcome.

The same study also found evidence of durable learning. During software outages, workers who had used and followed the AI’s recommendations retained gains relative to their own pre-AI performance. In this setting, the tool did not merely carry weaker workers. It helped transmit some of the practices embedded in the work of stronger colleagues.[1]

Yet the broader organisational picture remains unsettled. The October 2026 revision of a multi-country NBER survey covered nearly 9,000 CEOs, CFOs and senior executives in seven countries. It found 78% of firms using AI and 80% of executives regularly using it, for an average of 2.1 hours a week. Executives reported a modest 0.2% boost to annual labour-productivity growth over the preceding three years. These are reported firm-level perceptions, not a causal estimate. They underline that adoption, individual time saving and realised organisational value are different measures.[2]

A task can become faster without the surrounding workflow becoming more productive. A worker can produce more without the organisation converting the surplus into better service, more capacity, higher quality or lower cost. A team can accelerate one stage while shifting review, correction or risk into another.

The first dashboard may therefore be accurate. It may simply be measuring the part of the system where the benefit is easiest to see.

What the capability tax is — and is not

The capability tax is the future cost incurred when an AI-assisted process raises present performance by reducing the human practice, understanding, context or verification capacity the organisation will later require.

It may appear as:

• more review and correction after faster first production
• weaker ability to explain assumptions or detect plausible errors
• loss of the routine work through which junior people become experienced people
• thinner organisational memory because fewer people participate in creating the work
• slower recovery when a model is unavailable, changes behaviour or fails in an unusual case
• increased dependence on a system the organisation does not fully control

This is not a literal tax and it is not a universal rate. It is a management accounting concept: a way to recognise costs that ordinary productivity measures omit.

Nor is it an argument for preserving manual work merely because it is familiar. Organisations have always stopped teaching skills that technology made genuinely unnecessary. Few finance teams need to maintain proficiency with mechanical adding machines. The important distinction is between a skill that has become obsolete and a capability that is still required to verify, govern, adapt or recover the system that replaced it.

The tax begins when the organisation removes the work but still needs the judgement the work used to develop.

Charge one: verification becomes the work

Generative AI changes the economics of production. A first pass can be almost free. Evaluation remains expensive.

Human-factors researchers at Microsoft described this as one of the “ironies of generative AI”: automation shifts the user from production to evaluation, restructures workflows, creates interruptions and can make easy tasks easier while making difficult tasks harder.[6] A separate Microsoft Research survey of 319 knowledge workers, covering 936 first-hand uses of generative AI, found that higher confidence in AI was associated with less critical-thinking effort. It also found that critical thinking did not disappear so much as move towards verification, integration and task stewardship.[7]

That shift matters because evaluation is not a lighter version of production. To spot a subtle error, the reviewer often needs more expertise than the person who could have produced an acceptable answer unaided. The output arrives fluent, complete and apparently coherent. The reviewer must decide whether the underlying assumptions, evidence and exceptions are also coherent.

Software engineering makes the relocation unusually visible. DORA’s 2025 and 2026 research reports that AI often accelerates initial code generation while time is reallocated to auditing and verification. Higher adoption was associated with both greater delivery throughput and greater delivery instability. DORA calls the extra checking burden a “verification tax”.[5]

METR’s early-2025 randomised trial provides a useful warning about perception. Sixteen experienced open-source developers completed 246 real tasks in repositories they knew well. With the AI tools available at the time, they took 19 per cent longer, even though they had expected to be 24 per cent faster and afterwards still believed they had been 20 per cent faster.[3]

That result should not be treated as a verdict on 2026 coding tools. METR itself said in February 2026 that later data were distorted by selection effects, task selection and unreliable time measurement when people ran multiple agents. It now believes developers are probably receiving more benefit, but says its newer estimates are weak evidence for the size of that benefit.[4]

The durable lesson is not “AI slows developers”. It is that felt productivity is an unreliable measure, especially when production is pleasant and the hidden work occurs later.

A commercially interested source provides a more dramatic, historically bounded signal. Faros, which sells engineering-intelligence software, reported Q2 2026 telemetry from 22,000 developers across 4,000 teams. Its comparison of lower and higher AI adoption associated higher adoption with 51% larger pull requests, 28% more bugs per pull request, five times longer median review time, three times more incidents per pull request and ten times more code churn.[9]

Those figures are observational, reported by a company with a direct commercial interest in better measurement. They do not prove that AI caused every defect. They do indicate a credible failure mode: production capacity can increase faster than the organisation’s capacity to review, integrate and operate what is produced.

Faros's later Q3 2026 report shows that this pattern is not static: per-change incident pressure eased markedly as organisations adapted, while monthly incidents and QA time remained elevated. The two reports study different phases of adoption and are vendor-run observational analyses, not proof that AI caused every downstream cost.[13]

When that happens, the work has not vanished. It has moved.

Charge two: context can disappear faster than output grows

A person who creates a piece of work usually acquires a mental model of it. They know where the awkward decisions sit, which assumptions are fragile and which shortcuts will become dangerous later.

AI can produce the artefact without requiring the same journey. That is often the point. It is also where context can leak from the organisation.

The risk is easiest to recognise when something breaks. In July 2025, Dex Horthy’s team at HumanLayer experimented with a “lights-off” software factory in which AI produced code that humans did not read. Horthy later told The Pragmatic Engineer that the experiment was abandoned four months later after a production failure the model could not diagnose. It took days to find the fault and three weeks to re-onboard humans to a codebase nobody had read.[10]

This is one company founder’s account, not evidence of prevalence. HumanLayer also sells tools built around human-agent collaboration, so the story is commercially aligned with its position. It remains valuable because it isolates the mechanism with unusual clarity: output had accumulated while human ownership of the system had not.

The same pattern can occur outside software.

A financial model may calculate correctly while nobody can explain the assumptions inherited from a previous prompt. A compliance report may cite the right rules while the team no longer remembers why a judgement was made. A proposal may read well while the person presenting it cannot defend the evidence under challenge. A service team may rely on AI-generated responses until an unusual case falls outside the patterns the model has seen.

Context is not documentation alone. It is the distributed ability to explain, challenge, change and recover the work.

An organisation does not fully own an AI-assisted process simply because it owns the output file. It owns the process when named people can still understand its boundaries and take responsibility when the usual path fails.

Charge three: automation can remove the apprenticeship

Many organisations describe the work most exposed to AI as “junior work”. That can make automation appear almost costless. If a model can prepare the first draft, summarise the case, produce the routine analysis or write the basic code, senior people can concentrate on higher-value judgement.

The missing question is where the next generation of senior judgement comes from.

Routine work is not valuable only because it produces an output. It is also how people encounter repeated patterns, make small mistakes, absorb standards and develop the tacit sense that tells them when a case is not routine after all.

Evidence from education does not automatically transfer to employment, but it shows the mechanism. In a randomised field experiment involving nearly 1,000 secondary-school mathematics students in Turkey, access to a general GPT-4 interface improved performance during practice by 48 per cent. When the tool was removed, those students scored 17 per cent below the control group. A safeguarded tutor that offered teacher-designed hints rather than direct answers improved assisted practice by 127 per cent and largely removed the later performance penalty.[8]

The immediate performance metric therefore concealed two very different learning outcomes. The system that gave answers made students look stronger while it was present. The system designed to preserve productive effort helped without creating the same loss.

A 2026 NBER working paper reaches a related conclusion through economic modelling rather than workplace observation. In a model where people build skill through the tasks they perform, cheaper automation can either improve welfare or tip the economy into a “human-capital trap”, depending on whether learning and frontier-maintenance activity are preserved.[12]

The empirical workplace evidence remains incomplete. We do not yet have a universal measure showing how much AI use erodes or improves professional development across occupations. But leaders do not need to wait for a universal percentage before asking whether their own workflow still creates capable people.

If juniors no longer perform the work through which seniors learned, the organisation needs a replacement learning architecture. Otherwise, it is not removing low-value labour. It is consuming its future expert pipeline.

Charge four: resilience is easy to mistake for inefficiency

Traditional productivity programmes often remove duplication, spare capacity and manual fallback because these look inefficient in normal conditions.

AI creates a similar temptation. If the model is almost always available and usually useful, maintaining unaided capability can look like waste.

But resilience is not the same as routine duplication. It is the ability to continue, recover and exercise judgement when the system behaves outside its normal range.

The QJE customer-support study offers an encouraging result: workers retained some productivity gains during AI outages, suggesting that the system had helped them learn rather than merely replacing their effort.[1] The HumanLayer example offers the opposite pattern: the team had allowed understanding to decay until recovery required substantial re-entry work.[10]

The difference is not best explained by “good AI” and “bad AI”. It is explained by workflow design.

In the customer-support setting, the AI offered suggestions based partly on the behaviour of top performers. Agents retained the authority to accept, edit or ignore them, and the system withheld suggestions when training data were insufficient. The human remained responsible for the conversation.[1]

In the lights-off factory, the human was deliberately removed from reading the work. The system could continue producing until a failure demanded knowledge that had not been maintained.

A productivity measure that ignores recovery capacity will reward the second design until the day it fails.

The capability tax is not inevitable

The evidence does not support an anti-AI conclusion. It supports a design conclusion.

AI can create a capability dividend when it helps people see expert patterns, gives them useful feedback, removes uninformative toil and lets them practise higher-order judgement. It creates a capability tax when it supplies the answer while bypassing the understanding, verification and practice that the organisation still needs.

Three positive signals are worth noticing.

First, the customer-support study found both immediate productivity and durable learning, particularly for less experienced workers.[1]

Second, the school experiment found that safeguards changed the learning outcome. The underlying model was not enough; the interaction design determined whether assistance became a crutch or a tutor.[8]

Third, commercial case material suggests that organisations can absorb higher output when they redesign the whole workflow. Cursor reports that more than 30,000 NVIDIA developers use its tools daily, with committed code tripling while reported bug rates stayed flat. The case also describes AI being extended into testing, review, debugging and deployment rather than being added only to code generation.[11]

This is a vendor customer story and not an independently audited causal study. It should be read as a useful existence signal, not a universal promise. Its practical lesson is still credible: an organisation cannot safely accelerate production while leaving every downstream control designed for the old production rate.

DORA expresses the same idea more generally. AI acts as an amplifier. Strong testing, internal platforms, clear workflows and reliable feedback loops allow benefits to travel through the system. Fragmented tooling, weak data and fragile infrastructure allow debt and instability to travel faster instead.[5]

The management objective is therefore not maximum AI use. It is more useful organisational capability per unit of time, cost and risk.

The Capability Tax Test

Before calling an AI deployment a productivity gain, examine six dimensions together.

Qualitative editorial management screen, not a validated psychometric score. Apply it to a workflow or team, not to police individuals.

This is a management screen, not a validated psychometric scale. Apply it to a workflow or team, not to police individual workers. A red flag in one dimension does not automatically cancel the deployment. It tells leaders where an apparent gain may be borrowing from the future.

What leaders should ask before accepting the productivity claim

1. Which outcome improved, and which activity merely increased?
2. Where has review, correction or exception handling moved?
3. Can the people responsible explain and repair the output without asking the same model that produced it?
4. Which learning opportunities have disappeared, and what has replaced them?
5. Do senior experts still contribute original judgement, or are they mainly approving model-shaped work?
6. What happens during an outage, a model change or a low-frequency case?
7. Where does the saved time go — into more value, better quality, greater capacity, or simply a higher expected pace?

These questions should be asked of internal teams as well as vendors. A supplier may demonstrate faster production. Only the organisation can establish whether the surrounding capability is strengthening or thinning.

Productivity should include what remains

The most dangerous AI productivity claim is not necessarily false. It is incomplete.

A team may genuinely produce more. A worker may genuinely save hours. A workflow may genuinely move faster. None of those observations answers whether the organisation is becoming more capable.

That requires a wider account: useful outcomes, downstream cost, retained understanding, skill formation, human ownership and recovery capacity.

An organisation that can produce more but understand less may simply be spending capability it has not yet priced.

The practical rule is straightforward:

Do not measure AI only by the minutes it removes. Measure it by the capability the organisation still possesses when the model is wrong, unavailable or replaced.

That is the point at which an AI productivity gain becomes durable organisational progress rather than an advance funded by a hidden tax.