The minutes we saved. The money we didn't.

The worksheet is a single printed page, the kind that circulates a few weeks after a pilot ends. Sixty employees have been using an AI assistant, and each of them, the page says, saves about twenty minutes a working day. Over a twenty-day month that comes to 400 hours. At a fully loaded labor cost of $50 an hour, the bottom line reads $20,000 a month, under the words "cash saved." The figures here are invented for the arithmetic, but anyone who has sat through the review of an AI pilot has seen a page like it. Every line above the last one is arithmetic. The last line is a claim.

Last week that claim met some inconvenient evidence. On January 19, as the World Economic Forum's annual meeting opened in Davos, PwC published its survey of 4,454 chief executives in 95 countries and territories. Only 12 percent said AI had delivered both lower costs and higher revenue. Fifty-six percent said they had seen no significant financial benefit from it to date.¹ The next day, on a Davos panel titled "Scaling AI: Now Comes the Hard Part," the chief executive of Philips, Roy Jakobs, described the biggest challenge as rethinking how the work itself is done, and suggested that companies spend at least as much effort on how people will actually use AI as on building it. "Adoption is ultimately where success is measured," he said.²

The worksheet and the survey are not really contradicting each other. They measure different things, and the space between them is where management does its work. A saved minute is a change in how work can be done: released capacity, time made available because comparable work now finishes faster. A financial return needs something else in the business to change: a payment that stops, a sale that would not otherwise have closed, a service result someone chose in advance and then measured. Call that realized value. The first does not turn into the second by itself. I am not claiming the saved minutes are imaginary, or that a tool must pay for itself in cash to be worth having; faster service and lighter workloads can justify an investment on their own. Nor does the survey tell us why returns are missing. Costs, weak tools, slow adoption and poor measurement all play their part. What follows is one mechanism worth checking first, because it is the one the worksheet hides. (The "we" in the title is the voice of that familiar page, not a report on any particular company.)

The time is real

Start with what the worksheet gets right, because the best evidence on AI at work finds genuine, measurable gains. Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed 5,172 customer-support agents at a software company as a generative AI assistant, which suggested replies during live conversations, was rolled out to them in stages. With the assistant, agents resolved about 15 percent more customer issues per hour on average. Less experienced and lower-skilled agents gained the most, in both speed and quality, while the most experienced saw small gains in speed and a slight dip in quality. Customers, for their part, became more polite and less likely to ask for a manager.³

Notice what was measured. Issues resolved per hour is a throughput rate inside one workflow: how quickly an agent can close the next ticket in the queue. It is a strong, careful result, and it is a result about the task. It is not a measure of the company's profit, its payroll or its cash flow, and the authors do not present it as one. Between a faster ticket and a better quarter lie several more steps, and each of them is a decision somebody has to make.

Follow one saved hour

Go back to the worksheet and follow a single saved hour. It does not arrive as an hour. It arrives as twenty minutes a day per person, in pieces: three minutes here because a draft reply appeared at once, five there because a policy lookup took seconds instead of a search. Spread across sixty people, the 400 hours are not a person. You cannot assign them to a new project the way you would assign a new hire, because they occur in fragments, at different times of day, to people with different skills.

Then ask where each fragment goes. Suppose the team handles requests that arrive at a steady pace. If requests come no faster than before, an agent who closes each one sooner simply waits a little longer for the next. The time is real, and it may well be welcome, but no extra work was done and no expense changed. Or suppose the faster step feeds a slower one. A team drafts refund letters, and every letter waits for a supervisor's sign-off, which happens in a single batch at the end of the day. Draft the letters in half the time and they still go out at the same hour. The queue has simply moved to the supervisor's desk. Nothing a customer can see has sped up, and the month costs the same.

In both cases the saving is genuine at the task and absent from the results. That is not a failure of the technology. It is what happens when released time meets an unchanged system: the time waits behind a constraint, or work expands to fill it, or it disappears into a slightly calmer day. Not every workflow is this stuck. Where work is already piling up in a queue, as it was for the support agents, faster handling turns into more completed work with little extra effort. The question is which kind of workflow you have, and the worksheet never asks it.

Where the hours can go

Released capacity can become realized value by a few different routes, and each needs a different decision. The most direct is to stop paying for something. If the team was working paid overtime or relying on contractors to clear a backlog, the saved hours can take their place and the spending falls. That is cash, and it shows up in the accounts. A second route is more output: answering requests that would otherwise have gone unanswered, or serving more customers with the same team. This creates value only if there is demand for the extra work and the revenue it brings exceeds what it costs to deliver.

A third route is hiring you avoid. A growing team that absorbs next year's volume without adding staff has saved money, but against a counterfactual. That claim needs a credible forecast of the growth that would otherwise have required those hires, and it does not reduce this year's payroll by a cent. The fourth route is not financial at all: a faster reply, fewer errors, fewer weekends worked by a team that was stretched thin. These are legitimate outcomes, and often the right ones to pursue. They need their own measures, such as a response time, an error rate or a turnover figure, rather than a dollar amount invented to make them fit the worksheet.

Figure 2 follows the first route alone, because it is the easiest to see. Hold the released capacity fixed at 400 hours and change only one thing: how much paid overtime the team actually stops working.

The model is deliberately narrow. It counts the overtime that stops and nothing else, so its zero does not say the time is worthless; it says this particular route has not produced a saving yet. And at the far right, where all 100 eligible overtime hours are gone, it would be a mistake to credit the other 300 hours with a return. They are still capacity, waiting for a decision.

What the gain costs

The worksheet has a second blind spot: it counts what the tool gives and not what obtaining it takes. Start with the time itself. Twenty minutes saved on drafting is a gross figure. If every draft then needs five minutes of checking and the occasional correction, the released capacity is fifteen minutes, and the honest number is that net one, after review, rework and the coordination a new way of working demands. Count the review only once, though. If it has already been subtracted from the hours, it should not be charged a second time as a separate labor cost.

Then there is the money. Licenses, the cost of running the models, integration with existing systems, training, and the people who keep it all working are real expenditures, and several of them recur every month. Recurring costs belong against recurring benefits: a tool that removes $4,000 of overtime a month and costs $3,000 a month to run is ahead by $1,000 a month. The one-time cost of getting it running is a separate question, answered by asking how many months of that surplus it takes to earn the money back. A positive monthly figure is not a full return on the investment, and it is certainly not payback.

What Davos was pointing at

This is how I read the Davos discussion. "Redesigning work" can sound like a slogan, and in many slide decks it is one. It becomes concrete when someone with the authority to do so changes a specific thing that governs how the released time is used: next month's overtime roster, the contractor agreement that is up for renewal, the response time promised to customers, the sign-off rule that held every letter until evening, the routing that decides which work reaches which team. Each of those is a decision with a name attached, and each is the point at which minutes become a result someone can check.

That is not the same as buying more AI, and it is not the same as training everyone. Both can help, but neither changes what happens to an hour once it has been freed. The survey cannot tell us how many of the companies reporting no financial benefit skipped this step; self-reported results from thousands of companies cannot carry a causal explanation that fine. PwC does note that the chief executives reporting both cost and revenue gains were two to three times as likely to say they had embedded AI extensively across their businesses, which is consistent with the argument here but is a correlation, not proof of it.¹ The interpretation is mine. What the survey does establish is that, for most chief executives, the gains visible in demonstrations had not yet appeared in the two places a finance team looks: costs and revenue.

Objections, considered

Benefits take time. This is the strongest objection. Technologies that change how whole economies work need large complementary investments in new processes, skills and ways of organizing, much of them intangible and invisible in the accounts. Erik Brynjolfsson, Daniel Rock and Chad Syverson call the result a productivity J-curve: measured productivity looks too low while firms pour effort into those intangibles, and too high later, when the investments pay off.⁴ I accept the mechanism, and it should make anyone patient with a disappointing first year. But it is an account of how investment and measurement behave across an economy. It does not promise that a particular pilot will eventually pay, and it does not tell a board when. Patience is a reason to keep measuring, not a reason to stop.

This treats layoffs as the goal. It does not. Cutting headcount is one way to turn time into cash, and often not the best one. Clearing a backlog, ending overtime for a tired team, answering customers faster and raising quality are all purposeful uses of released time. The argument is narrower: if the claim is a financial saving, some spending has to fall, and somebody has to decide which. If the claim is something else, name it and measure it.

The CEO survey misses what workers experience. Agreed. It is a self-reported, company-level view, with its global figures weighted by the size of each country's economy, and it cannot see the agent who now closes tickets faster. A controlled study of one workflow and a survey of chief executives answer different questions at different levels, and both can be right. The puzzle this dispatch is about lives precisely in the gap between them.

Measuring all this could cost more than the tool. Sometimes it would, and then it should not be done. A writing assistant used by a handful of people needs no study, and a small tool people like can be kept on those grounds. The effort should match the claim. A savings figure presented to a board, the kind that justifies a budget or appears in a plan, deserves a baseline, a defined outcome and some credible comparison, such as a team that has not received the tool yet, or the same team before it did.

The time saving is the beginning of the business case.

Five questions for the next worksheet

• What changed in the work? Which step got faster, by how much after review and correction, and for whom.
• Where did the released capacity go? Into more output, less overtime, faster service, or into waiting.
• Which outcome improved, against which baseline? A cost, a volume or a response time, measured before and after.
• What did getting there cost? Recurring costs against recurring benefits, with the one-time costs counted toward payback.
• Who can explain the conversion? A named person who can walk from the minutes to the result and show the evidence.

None of this means throwing the worksheet away. Its arithmetic is fine: sixty people, twenty minutes, twenty days, 400 hours. Keep it. Sunday's edition argued that a number needs an owner; a word like "savings" needs one too. So change the label on the last line to what it actually measures, capacity valued at an assumed labor cost, and add one more line beneath it in a different hand. That line stays blank until someone can write in a payment that stopped, an output that grew or a service that got better, and show how they know.