Can AI make work larger, or will it only raise the output quota?

The most important question about AI at work may not be how much faster it helps us complete today’s tasks. It may be whether it allows us to attempt work that was previously too difficult, too expensive, or simply beyond our capacity. There is a difference between doing the same work faster and making more ambitious work possible. A faster report is useful. A service that reaches people who were never served before could change the size of the opportunity. But those outcomes are not automatic consequences of installing a model.

The technology itself is uneven. In a podcast discussion, experts described AI’s capabilities as “jagged”: strong at some tasks, surprisingly weak at others. That makes the human role more consequential, not less. People still need to specify the goal, recognize where a model is useful, and help it work around what it cannot do.

This is easy to overlook when a tool produces a polished answer. Fluency can look like competence across the whole task. It is not. The model may handle one part well and fail at the judgment, context, or coordination the rest requires. The work has not disappeared; some of it has moved into defining the problem and checking the result.

That shift creates two very different possibilities for a company.

One is to use AI to raise the quota: same process, same service, more output expected from each person. The tool saves time, so the time gets filled. This may improve efficiency, but it does not necessarily improve the work or the working day. The productivity gain may accrue mostly to the organization, while the employee experiences it as a faster treadmill.

The other possibility is to use the freed capacity to enlarge the work: serve customers who were too costly to reach, explore more options before making a decision, or take on problems that the old process could not support. That can create value for customers and new forms of work for employees. It asks a harder question than “How many minutes did the model save?” It asks what the organization can now do that it could not do before.

Early economic evidence is suggestive, but it does not settle the question. A recent U.S. Bureau of Economic Analysis analysis found that state-industry groups with higher worker-reported AI use had stronger real-output paths after 2020. Employment differences were positive too, but estimated less precisely. The authors explicitly caution that the analysis is descriptive: adoption may be related to other factors, and identifying AI’s causal contribution remains difficult.

That distinction matters. The findings do not prove that AI created more work, protected jobs, or caused output to rise. They do complicate the simple story that greater AI use must show up as fewer people doing less work. At least in these comparisons, higher reported use coincided with stronger output and no clear employment decline.

But even if AI contributes to higher output, we still have to decide what “more” means. More units completed? More customers served? Better quality? More time for judgment and relationships? Those are not interchangeable outcomes. A system optimized only for volume can make an organization busier without making it more useful.

The same distinction applies to the individual employee. If AI drafts the first version, what happens to the time saved? Does the person use it to test a better idea, understand a customer’s problem, or build a skill? Or does the saved hour become an expectation to produce another deliverable? The tool does not answer that. Management choices do.

And “human oversight” cannot just mean asking someone to approve an output at the end. If the model is good at the visible task but poor at understanding the real objective, the person has to shape the objective from the beginning. They need enough time, authority, and knowledge to notice when a plausible answer is wrong for the situation. Otherwise, oversight becomes a ceremonial step attached to a decision already made.

This is also why adoption numbers alone tell us little about impact. A person can report using AI frequently while their organization has not changed its services, processes, or decision rights. The more revealing question is what changed around that use. Did the organization redesign the work? Did customers receive something better? Did employees gain room to apply judgment, or did the target simply move?

Organizations should also separate “a person’s work got faster” from “the firm can now take on more demand.” An employee may use an assistant to complete a routine draft in half the time, but that does not reveal whether the saved time goes toward mentoring new staff, speaking to more customers, or producing more units. It also does not tell us who pays for infrastructure, integration, and quality review. A serious productivity plan should measure both sides: what new outcome customers receive, and what capacity employees retain. If one grows while the other is squeezed, the system is productive by one narrow measure and costly by another.

There is a temptation to treat technical progress and economic progress as the same thing. They are not. A model can become more capable before organizations know how to use that capability well. In the podcast, participants asked when technical advances would become visible as a large, obvious economic impact. That is the right question, but the answer depends partly on organizational choices, not just on the next model release.

My view is that AI will make work larger only when leaders deliberately make room for larger goals. If the only measure is output per person, the default will be to raise the quota. If the measures include access, quality, customer outcomes, and the capacity people retain to exercise judgment, then saved effort has somewhere more valuable to go.

So when an AI tool saves your team time, decide what that time is for before the quota rises.

Otherwise, the technology may increase what gets produced without increasing what becomes possible.