The Productivity Dividend Has a Routing Problem

AI is giving some workers a day back. Two-thirds of them get little guidance on what to do with it.

By DeQuan · Predicai

A few weeks ago, I wrote about a simple question I had started asking at the end of my workday:

What did the recovered attention buy me?

I had begun keeping a short record of what AI took over, what that work used to cost me, and what happened to the time afterward. A quote check got shorter. Rebuilding the state of an account got shorter. A handoff package stopped needing to be reconstructed every time.

The useful part was not the time savings. It was what happened next.

I read the account instead of relying only on the CRM. I called instead of emailing. I walked into conversations better prepared.

Previously: Signal 015: AI Saved You 45 Minutes. Who Owns Them?

At the time, I thought of this mostly as a personal productivity problem. The research coming out now makes me think it is an organizational one.

The missing half of the productivity equation

BCG surveyed 11,749 workers across 14 markets for its 2026 AI at Work report.

Among regular frontline AI users, 42% said AI was saving them at least eight hours a week, roughly a full workday.

That sounds like the productivity story everyone has been waiting for.

Then comes the more interesting number.

Sixty-six percent said they received limited or no guidance about what to do with the time they saved. More than half said they were not redirecting it toward strategic work.

That is not an AI capability problem. It is a routing problem.

Source: BCG, AI at Work 2026

We have spent years asking whether AI can save time. Increasingly, the answer is yes, at least for some workers doing some kinds of work.

The next question is harder:

Where does the time go?

The worker can change before the work does

One of the strongest field experiments so far helps explain why this matters.

Researchers studied 7,137 knowledge workers across 66 firms in a six-month randomized experiment. Half received access to generative AI integrated into tools they already used for email, meetings and writing.

In the second half of the experiment, the 80% of treated workers who used the tool spent about two fewer hours per week on email and worked less outside normal hours.

But researchers did not detect a corresponding shift in the overall quantity or composition of their tasks.

Source: Dillon et al., Shifting Work Patterns with Generative AI

That finding is easy to overlook.

The worker became more efficient before the work system changed.

You can draft the report faster while the same approval chain remains in place. You can summarize the inbox while the organization keeps producing the same volume of email. You can prepare an analysis in twenty minutes and still sit through the hour-long meeting built around the old process.

Individual productivity and organizational productivity are not the same thing.

AI can give the worker capacity before anyone has redesigned the system around that capacity.

Time saved is not value created

A 2026 working paper based on a representative survey of Korean workers offers an unusual version of the same problem.

The researchers estimate that generative AI reduced working time by about 3.8%, but found almost no relationship between time saved and changes in output. Much of the recovered time appeared to become on-the-job leisure rather than additional production.

Source: Generative AI and the Reallocation of Time: Productivity, Leisure, and Fulfilling Work

This is a preprint based on survey evidence, not a randomized field experiment, so I would not treat the estimate as definitive.

But I like the question it forces.

Is that a failure?

From the employer's perspective, perhaps some of it is. If the purpose of an investment is greater output and the output never appears, the economics matter.

From the worker's perspective, the answer is less obvious.

If AI removes an hour of low-value administrative friction and someone ends the day less exhausted, that is still an outcome. If the returned attention leads to a better decision at 4 p.m., that may be more valuable than another document.

The mistake is treating all reclaimed time as if it naturally converts into more production. It does not.

Something has to happen between efficiency and value.

Output, capacity and capability are three different things

There is another distinction I think we need to make much more carefully.

AI can improve what someone produces today without necessarily making that person better at the underlying work tomorrow.

A September NBER working paper studied 133 practicing patent lawyers at eleven U.S. law firms in a preregistered three-month randomized trial.

Lawyers with access to a custom AI drafting assistant produced higher-quality work. Junior lawyers saw particularly large immediate benefits.

Then the AI was removed.

After three months, participants performed a core professional judgment task without AI. The treated lawyers did better overall, but the durable improvement was concentrated among senior lawyers. Their unaided performance improved by 0.45 standard deviations.

Junior lawyers showed no average improvement. Their results actually bifurcated: there were fewer mediocre scores, but more poor scores and more good ones.

Source: Autor et al., NBER Working Paper 35720

The provenance matters here. Google funded the direct costs of the experiment. Five of the seven authors disclosed full-time Google employment, and another disclosed a paid Google contractor relationship. Those disclosures belong beside the finding.

What interests me is not whether this single experiment settles the question. It does not.

It is that it forces us to separate three outcomes:

• Output: Did AI improve the thing I produced?
• Capacity: Did AI return usable time or attention?
• Capability: Am I more able to do the work when the AI is gone?

Those are different measurements.

And the patent-lawyer result is not entirely isolated.

Four experiments involving 3,562 participants, published in Scientific Reports in 2025, found that AI collaboration improved immediate performance, but that advantage did not carry into the subsequent tasks people performed alone. The researchers also found lower intrinsic motivation when participants transitioned from AI-assisted work back to solo work.

Source: Wu et al., Scientific Reports

That does not prove AI erodes expertise.

It does tell us that assisted performance should not automatically be reported as human skill development.

There is also new work created by AI

The productivity ledger has another side.

BCG found that 47% of respondents said they now spend more time managing or directing AI than performing the underlying work themselves. Forty-one percent of regular users reported greater cognitive load.

That is not necessarily a bad trade. Ten minutes directing a system may be preferable to forty minutes manually doing repetitive work.

But the ten minutes still exists. So does checking. So does correction. So does deciding whether an answer is trustworthy enough to act on.

A study published September 24 followed 642 hybrid workers at two points seven weeks apart. The researchers use the term AI delegation paradox for what they observed: handing execution to AI can reduce direct task effort while increasing the need to monitor progress, anticipate errors and remain mentally attached to work that the machine is performing.

Greater AI process vigilance was associated with more work-related rumination and poorer psychological recovery.

Because this was a two-wave survey rather than an experiment, it cannot establish that AI delegation caused those outcomes. Still, it identifies a cost that ordinary productivity metrics can miss.

Source: Frontiers in Psychology, AI delegation paradox study

The task may leave your hands without completely leaving your head.

A useful real-world experiment is hiding in a snack company

One recent corporate example caught my attention because of what the company called the program.

pladis, the company behind brands including McVitie's, ran an AI adoption program called Steal Back 10 Hours.

According to a Microsoft customer case study, 48 employees who completed the full training program and productivity tracking reported saving 580 hours per month in aggregate, a little more than 12 hours per person on average.

Source: Microsoft customer story: pladis

This is vendor-published, self-reported evidence, not an independent productivity trial. I would not put the number beside the randomized studies and pretend they carry the same weight.

What I like is the framing.

The company named the capacity it wanted back. That is already more useful than “adopt AI.”

The next question, of course, is what those hours bought.

That is the question I am keeping

My original log was intentionally unsophisticated.

• What did the system take over?
• What did that work used to cost?
• What did the returned attention buy?

The first two are relatively easy to measure. The third is where the real argument lives.

Sometimes the answer should be more output. Sometimes it should be deeper work. Sometimes it should be a customer conversation, better preparation, learning, coaching or creative work. Sometimes it may reasonably be recovery.

What I do not want is for every recovered minute to disappear automatically into another low-value task.

That is how AI can make work faster without making work better.

The next phase of AI productivity will require us to measure more than how quickly a task disappears.

We need to know what replaces it.

And we need to stop collapsing output, capacity and human capability into the same number.

AI can create a productivity dividend. Whether that dividend compounds depends on where we send it.