Time Saved Is Not Money Saved: Does AI Automation Cut Costs?

Sometimes, and less often than the sales deck implies. AI automation produces time savings reliably. It produces cost savings only when the returned hours are collected into something that has a line on your profit and loss, which in practice means a hire you do not make or a role you consolidate. Most implementations never take that second step.

The two get sold as the same thing. A vendor counts hours saved per week, multiplies by an hourly rate, and prints an annual figure. That multiplication is only valid if you stop paying for the hour.

The hours saved are real, and small

The best measurements put the saving in the low single digits of a working week. Anders Humlum of Chicago Booth and Emilie Vestergaard of the University of Copenhagen surveyed roughly 25,000 workers across 7,000 Danish workplaces in two rounds, then linked the answers to administrative payroll records. Their paper, dated 15 July 2025, puts average time savings from AI chatbots at 2.8% of work hours.

A Bank of Korea study by Donghyun Suh and Samil Oh, released in February 2026 on a weighted sample of 5,512 Korean workers, found the same order of magnitude: generative AI cut working time by 3.8%.

Put that in money. Eurostat’s 31 March 2026 release put average hourly labour cost in the EU at €34.9 for 2025, and €38.2 in the euro area, with non-wage costs making up 24.8% of the total. A 2.8% saving on a 40 hour week is about 67 minutes. At €34.9 an hour that is roughly €39 a week, or near €1,800 a year per person. That number is a ceiling, not a result.

The money mostly does not arrive

Both studies that went looking for the money found close to none. Humlum and Vestergaard’s estimates for earnings, total hours worked and wages are precise zeros, with confidence intervals ruling out average effects larger than 1%. The time saving existed. It did not reach anyone’s pay, and it did not reach the employer’s cost base either.

Suh and Oh explain the mechanism. They asked the same workers about time saved and about change in output. The correlation between the two answers was 0.008. The recovered time was absorbed as on-the-job leisure rather than converted into more work getting done.

Firms report the same result from the other side

Executives measuring their own companies land in the same place. NBER working paper 34984, published March 2026 off The CFO Survey and its roughly 750 financial executives, reports a mean labour productivity gain from AI in 2025 of 1.8%. The median was 0.0%. When the authors derived the gain implied by the revenue and employment changes those same CFOs attributed to AI, the mean fell to 0.6% and the median stayed at 0.0%.

NBER working paper 34836, from February 2026 and revised in March 2026, surveyed about 6,000 senior executives across US, UK, German and Australian firm panels. It found 69% of firms actively using AI, executives averaging 1.5 hours of use a week, and nine in ten reporting no impact on employment or productivity over the previous three years.

The CFO Survey also ranked why firms adopt AI. On a 0 to 4 scale, “reducing labour costs” scored 2.01 and “reducing non-labour costs” 1.80, the two lowest motives on the list, against 2.88 for improving production efficiency. Firms are not buying AI to cut payroll. Payroll then does not fall. Those two facts are the same fact.

The three ways an hour turns into money

An hour saved becomes money in exactly three ways: a hire you avoid, a role you consolidate, or a block of time you redeploy onto work that bills. If none of those is true, the fourth outcome is the default, and the default is that the hour is absorbed and nothing changes on the P&L.

Route	What has to be true	Where it shows up
Avoided hire	Volume grows and headcount does not	Headcount plan, flat payroll against rising output
Consolidated role	Saved hours concentrate in one or two people, not spread thin across twelve	Payroll
Redeployed to revenue	Hours arrive in usable blocks and go onto work that bills or sells	Revenue line, cost line unchanged
Absorbed (the default)	Nothing in particular	Nowhere

Note what the first two have in common. Both require the saving to be concentrated in a person, not distributed across a department. Twelve people each getting 67 minutes back is 13.4 hours a week across the business, and it is also zero roles, because you cannot make a third of a person redundant and you were never going to.

What nobody has measured

Three claims in this argument have no tier 1-2 source behind them, and we would rather say so than dress up a guess.

First, the fragmentation point above. No government agency or named research paper we could find measures the block size of the time automation returns, only the total. Suh and Oh get closest by showing the time is absorbed rather than converted, but they do not measure whether it came back in five minute slivers or two hour blocks. The fragmentation claim is our own model of why the totals do not convert. Treat it as a hypothesis with good circumstantial support.

Second, avoided hiring. Nothing in the tier 1-2 literature quantifies “we did not hire because we automated.” NBER 34984 and 34836 give expected net employment effects (aggregate employment expected to fall by less than 0.4% due to AI in 2026), which is a different measurement. Any avoided-hire number in a vendor proposal is a model output, not an observation.

Third, thresholds. No source establishes how many hours per week per person a saving has to clear before it shows up financially. Anyone quoting one, including us, is quoting a rule of thumb.

Decide where the hours go before you build

The fix is boring and it happens before any code is written. In the spec, next to the process being automated, name the person whose hours are being freed, the number of hours, and the destination of those hours. Three sentences. If the honest destination is “the team gets some breathing room,” write that down and log the project as a workload or retention measure, not a saving, and keep it out of the ROI model.

This is the same discipline as measuring the process before you automate it, applied to the other end of the pipe. The baseline tells you what the hours cost today. The destination tells you whether removing them changes anything. Without both, you get the pattern in how to measure AI ROI: a project everyone agrees was useful and nobody can defend in a budget review.

Our own agents were built against that test. The work they took over was work we would otherwise have paid people to do, so the saving landed as roles we did not open rather than as minutes returned to existing staff. The mechanics of the running fleet are on the agent fleet proof page, and getting this decision made before the build is most of what a fractional AI officer is for.

Send us the process you are thinking about automating, plus one line on where the freed hours would go. If you do not have an answer to the second part yet, say so, and that is the first thing we will work out with you. Written reply within one business day, no meetings required. Start async.