The Hour Is a Decision

On October 29, the University of Manchester’s Productivity Institute reported that nearly nine in ten organizations use AI somewhere in the business, and only one in four has changed how work is done to make the most of it. The report, The AI Ambition Gap, was commissioned by Bytes Technology Group, an IT reseller.

Bart van Ark, who led the Manchester report, put the gap in one sentence: “The ambition gap lies between those individual efficiencies and sustained business results.”

That gap is usually described as a technology problem. It is closer to a time problem. The promise of AI at work was time: draft faster, summarize faster, answer faster, and get hours back. That part is coming true. What gets far less attention is what happens next. An hour saved is not a result. It is a choice, and in most organizations the choice gets made by whatever is already in motion: the next task in the queue, the colleague waiting on a review, the expectation that quietly went up.

Call it the unclaimed hour: time AI gives back that no one has assigned a purpose. It does not sit empty. Something claims it. The question worth asking is whether anyone chose what that “something” is.

The hour is real

The most current measure comes from the St. Louis Fed’s Real-Time Population Survey, a nationally representative survey of working-age Americans run with economists Alexander Bick, Adam Blandin and David Deming. Each quarter it asks people who used generative AI for their job how many more hours the same work would have taken without it. In the third quarter of 2026, those answers added up to about 2.1% of all U.S. work hours, counting everyone who did not use AI as saving nothing. In the second quarter, 39.2% of employed adults said they had used generative AI for work in the past week.

Two percent of every working hour in the country is a great deal of time. The survey does not ask where it goes. To see that, you have to follow the work.

The hour moves down the line

Software is where this is easiest to measure, because every step leaves a timestamp.

Harvard economists Fiona Chen and James Stratton studied 720,542 workers at 655 software firms, using activity data from January 2021 through August 2026 supplied by Jellyfish, an engineering-analytics company. In the current version of their paper, dated November 1, 2026, AI coding agents drove large increases in code written: 44% more lines of code and 36% more commits, on top of the gains from earlier AI assistants. Finished work moved much less. Completed issues rose 13.7%, an effect the authors call borderline significant.

The reason shows up one step later. After firms adopted agents, the time from submitting code for review to getting it approved rose 56.1%. Code waited in the review backlog 59.9% longer. The number of comments on each change rose 38.6%.

A separate study by Mert Demirer, Leon Musolff and Liyuan Yang, covering more than 500,000 developers and using Microsoft’s GitHub Copilot usage data, found the same shape. (Demirer and Musolff disclose that they are paid research consultants to Microsoft.) Across three generations of AI coding tools, ending with autonomous agents, commits rose about 240% cumulatively. By the time that work reached actual software releases, the gain was about 30%. Their summary: large task-level gains “have therefore translated only partially into shipped and used software thus far.”

Thirty percent more releases is a real gain. But the saved hour did not vanish between the keyboard and the release. It moved down the line, to the person who has to check the work. Nobody assigned the hour there. The workflow did.

What it looks like from the desk

Hubstaff, which makes time-tracking and workforce software, surveyed 1,068 U.S. workers in October. It is a vendor survey and unweighted, so treat it as a signal from the desk rather than a measurement. Among AI users who reported savings, 76% save at least an hour a week. Only 26% say they take on more work with it; 39% finish earlier; 15% say nothing changes. And 45% spend an hour or more a week reviewing or correcting AI-generated work from other people.

Not all of that review is waste. Checking AI output is often exactly the work that should happen. The point is narrower: most organizations never decided that this is where the hour would go, so they cannot tell the necessary checking from the avoidable cleanup.

Same tool, two divisions

Now the encouraging part, because there is one.

Stanford’s Arvind Karunakaran, with Roshni Raveendhran of the University of Virginia and Tami Kim of Dartmouth, spent nearly two years observing a law firm that gave the same AI tool to two divisions that were virtually identical in practice area, size, and pay and promotion structure. Stanford HAI reported the research on October 7.

The difference was how each manager introduced it.

In the first division, the manager presented the tool as a way to draft documents faster and shorten turnaround times. A productivity pitch, the kind most of us have written.

In the second, the manager asked the paralegals which tasks they found most boring, and what they had always wanted to do but never had time for. Then whether the tool could take the routine work and free up time for the rest.

That division also got real backing: formal training, weekly knowledge-sharing lunches, Q&A sessions with the vendor, mentorship, and protected time to explore the tool.

Paralegals in the second division used the tool 58% more than their peers and experimented with it 70% more. In the first division, paralegals described feeling like “just another cog in the machine.”

Two caveats, and they matter. This is one firm, reported through Stanford’s account of the research, not a controlled trial. And the second division got more than a different conversation; it got training and time. Karunakaran’s point is that the framing only works with that support behind it. Without it, “employees recognize it as little more than cheap talk.”

The scorecard has to move too

One more detail from the law firm. The paralegals who used the time for deeper research and case analysis did higher-value work, and they handled fewer cases doing it. Paralegals had always been judged by case volume.

Put those two facts next to each other. The better use of the hour was invisible to the firm’s old scorecard and could easily be read as a negative.

If you count volume, you will get volume, and the hour will just create more that needs review. If you count tokens or logins, you will get tokens and logins. The researchers make the same warning about tracking AI consumption: treat heavy use as value and people will use AI for its own sake.

Here is the fair reading of all of this. The time AI saves is not being thrown away. Some of it comes back to people: in Hubstaff’s October survey, 39% of AI users who save time say they finish their work earlier. Much of it goes into other work, including review that genuinely needs doing.

What is missing in most places is not management. It is a decision someone can point to. The law firm’s two divisions had identical software. What differed was whether a manager answered the question “what is this time for?” before the time showed up. Then management backed the answer with training, protected time and a scorecard that could see it.

When nobody answers that question, the workflow answers it for you. That is the wrong default. Not the time the tool can save. The silence about what the time is for.

Try this with your team

Ask the second manager’s two questions, out loud, at your next team meeting:

1. Which part of your work is most tedious?

2. What have you always wanted to do here but never had time for?

Then decide, on purpose, where the saved hour goes: more volume, deeper work, more review, or time back. All four are legitimate. Pick the one that works for your team, write it down, and tell people.

Then check two things. Does your scorecard measure finished value, or only activity? And whose desk is the review work landing on now?

So own it

AI will keep giving time back. That part is settled. What is not settled is who decides what that time is for. Leave it unclaimed and the workflow will spend it for you, often on review queues and rising expectations. Claim it, and you get what that law firm’s second division got: people who wanted to use the tool.

If you want a quick read on where your own organization stands, the free Decision Ownership Check takes about two minutes: ten questions and an AI Governance Exposure score from 0 to 100. It is a directional read, not a validated diagnostic. fellowshipintelligence.com/decision-ownership-check.

Sources

Federal Reserve Bank of St. Louis, FRED, “Generative Artificial Intelligence, Time Savings: Employed Adults” (RPSGENAITSALL), Real-Time Population Survey, Q3 2026, updated October 20, 2026: https://fred.stlouisfed.org/series/RPSGENAITSALL

FRED Blog, “Does generative AI save time at work?”, August 27, 2026 (share of employed adults using generative AI for work, Q2 2026): https://fred.stlouisfed.org/blog/2026/08/does-generative-ai-save-time-at-work/

Fiona Chen and James Stratton, “Artificial Intelligence in the Firm: Bottlenecks in Software Production,” Harvard working paper, version of November 1, 2026: https://fion.ac/jellyfish.pdf

Mert Demirer, Leon Musolff and Liyuan Yang, “Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools,” NBER Working Paper 35275, May 2026, revised September 2026: https://www.nber.org/papers/w35275

Business Wire via AOL, “Businesses are buying AI faster than they can make it pay, finds University of Manchester research for Bytes Technology Group,” October 29, 2026: https://www.aol.com/articles/businesses-buying-ai-faster-pay-082600000.html

Hubstaff, “State of the Workweek 2026: Burnout, AI, and the Work Employers Don’t See,” October 2026: https://hubstaff.com/state-of-the-workweek

Stanford HAI, “Want Employees to Embrace AI? Stop Selling It as a Productivity Tool,” October 7, 2026: https://hai.stanford.edu/news/want-employees-to-embrace-ai-stop-selling-it-as-a-productivity-tool