A task done faster creates an opportunity. It does not decide what happens to the time.
An employer might serve more customers, reduce a backlog, improve quality, shorten hours or cut staffing. Workers might learn a skill, spend more time checking output, or face a higher target for the same pay. Customers might receive a cheaper service or a larger bill. Which outcome occurs depends on the work, the market and the decisions made around the tool.
This is why a productivity result needs a second question: a gain for whom, measured how?
We should welcome real improvements. We should also refuse to treat every improvement in a task score as a completed social benefit. The path from a better tool to a better life contains decisions that research must examine and people should be able to influence.
The studies answer different questions
In the final published version of Generative AI at Work, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the rollout of an AI assistant to 5,172 customer-support agents. They report an average increase of 15% in issues resolved per hour, with gains varying by experience and skill. Some high-performing agents showed declines on quality measures. The study compares workers who received access at different times. Because access was not randomly assigned, other differences between groups could affect the conclusion. This one workplace also cannot represent every kind of work. The result does not show what happened to wages or economy-wide employment. Generative AI at Work
That is a substantial and useful result within its scope. It also gives a manager several reasons to look beyond the average. A tool may help a new worker differently from an expert. Faster service may not improve every measure of quality. A firm deciding whether to adopt should ask which parts of the result fit its own tasks and which need fresh testing.
Now consider a different setting. METR’s early-2025 randomized study allowed 16 experienced open-source developers to use AI on some of 246 repository tasks. In that setting, tasks with AI access took 19% longer. The finding challenges the assumption that access must save time. It is also a small, dated test. METR’s February 2026 update says newer estimates are unreliable because participant selection, task choice and concurrent work make measurement difficult. Neither result establishes a universal current slowdown. Developer study, 2026 measurement update
There is no contradiction that needs to be removed by choosing our favorite paper. The people, tasks, tools and study designs differ. A system can be valuable for customer support and less helpful for an expert fixing a familiar codebase. It can also improve over time while becoming harder to evaluate cleanly.
The right response is to make the question more precise. Which work is being tested? How experienced are the people doing it? Does the clock include checking and repair? Is the output judged by someone who does not know which method produced it? Who declined to take part, and might their reasons change the result?
These questions are especially important when a large claim rests on a small trial. A successful demonstration can justify a larger test. It cannot, by itself, justify reorganizing every job around the same assumption.
Saved minutes and changed earnings
A study can find changes in work before it finds changes in pay or hours.
In their March 2026 paper, Anders Humlum and Emilie Vestergaard link surveys of roughly 25,000 Danish workers in 11 exposed occupations to administrative records. Their early-period analysis, through December 2024, finds task changes while ruling out average earnings or hours effects larger than 2% under its study design. Adoption was not randomly assigned, so other differences between the groups being compared may affect the conclusion. The result is evidence about an early Danish setting, not proof that future effects will be small everywhere. Still Waters, Rapid Currents
This distinction matters for anyone promised a personal benefit from AI. Less time spent drafting may be replaced by more time reviewing, documenting or managing new tasks. Even a genuine reduction in work time does not automatically appear as a shorter paid week. An organization may use the capacity for other work.
There is nothing inherently wrong with using saved time to serve more people. A public service with a long queue might reasonably make that choice. But the choice should be stated. If workers are told a tool will reduce pressure, and the plan is actually to raise output targets, those are different promises. They should be debated as different promises.
Likewise, an increase in a company’s income does not tell us how much came from AI, how long it will last, or which people bore the costs. Benefits may arrive after investment, training and failed attempts. A fair assessment should include those costs. It should also include burdens shifted outside the organization, such as extra correction work required from customers.
The solution is not to demand one perfect measure. Use a small set of outcomes that answer the real question. Measure completed work and quality. Track time spent on review and repair. Examine pay, hours and changes in staffing separately. Ask whether customers receive better service. For a public program, include access and the ability to correct mistakes.
A dashboard becomes useful when someone has authority to act on it. If every measure improves except the one that matters most to the people affected, the project needs a response beyond explaining why the headline still looks good.
Participation must happen early enough to matter
Workers can contribute knowledge about tasks, customers and failure points. Whether participation improves outcomes is an empirical question, and the available evidence needs careful interpretation.
Lane, Williams and Broecke’s OECD research surveyed workers and firms in finance and manufacturing across seven countries in early 2022. Consultation was associated with more favorable reported AI outcomes. The authors explicitly say the analysis cannot establish causality: consultation might improve work, reflect better management, or change perceptions. The worker survey used online panels, and the evidence largely predates the current generative-AI wave. It supports taking worker input seriously, while leaving the effect of a particular participation scheme to be tested. OECD workplace surveys
There is also a moral argument for participation that does not depend on proving a productivity bonus. People whose working conditions will change have a legitimate interest in the decision. We should keep that argument separate from a claim that consultation always makes a firm more efficient.
What would meaningful input involve? Staff could help define the problem before a product is chosen. They could take part in paid trials, identify tasks a benchmark misses, and review proposed changes to workload. Their representatives could obtain enough technical advice to assess a claim. There could be a clear process for raising concerns without having to pretend that every concern is a software defect.
These are proposals for an arrangement people could test and negotiate. They do not amount to a single ready-made policy for every country or workplace. Small firms, public agencies and large employers have different resources and rules. The common test is whether the input can change a decision.
A meeting held after the contract is signed may still help implementation. It is less able to change the basic purpose of the purchase. A survey can gather opinions without giving anyone power over the result. The timing and authority attached to participation should therefore be visible.
Agree how a gain will be used
Before deployment, the parties should set out a few concrete questions. Which benefits are expected? Which costs will be counted? When will results be reviewed? Who can ask for a change? What happens if the tool fails to meet its promise?
For example, an organization could agree to use an initial period of verified time savings to reduce a backlog and improve training, then review the result before changing staffing targets. Another could test a shorter work schedule while holding service quality constant. A third might focus on improving access for people who have been poorly served. These are possible uses of gains, not outcomes established by the studies above.
The arrangement needs room for failure. If review work consumes the expected savings, the finding should be recorded. If the tool benefits only some tasks, deployment can remain limited to those tasks. If the comparison is too noisy to decide, the honest result is uncertainty. A trial loses much of its value when everyone knows it must be declared a success.
Ownership and bargaining also matter, but they should not be used as magic words. A financial stake may distribute some returns without giving workers control over daily work. Consultation may shape daily work without changing who receives profits. Consumer competition may lower prices without protecting employees from a badly managed transition. A serious proposal says which channel it is trying to change.
In my unpublished manuscript Abundant Republic, I described an aim as “making essentials affordable and freeing humans from drudgery.” That is an aim, not a result demonstrated by citing an earlier book. To pursue it here, we need to trace how a specific investment would improve access, reduce a burden or widen a person’s choices.
This gives us a clearer test of an AI gain. Do not stop at the fact that a machine can produce more. Ask whether the improvement survives checking, whether the whole service is better, and how the benefit reaches people. Then put the promised use of that benefit into a form that can be reviewed.
A society does not have to choose between ambition and careful measurement. It needs both. Ambition tells us which improvements are worth seeking. Measurement helps us discover whether we are getting them, for whom, and at what cost.