Saving ten minutes on a case does not yet mean taking ten minutes off a bill. This distinction, familiar to any team that has to absorb extra work, helps explain the debate over the profitability of artificial intelligence. A tool can speed up a task without reducing expenditure; it can also improve a service without that improvement showing up immediately in the accounts.
The press release published by Gartner on 26 October 2026 highlights this tension. Its headline says that 3% of organisations exceed their cost-saving targets through productivity gains. It also reports that 77% of board members expect AI to deliver productivity gains and cost reductions. These figures describe expectations and their reported fulfilment, not a universal measure of the value of all tools.
Read a percentage alongside its denominator
‘Exceeding a target’ is a demanding and relative criterion: it depends on the level of the original target. An organisation may achieve something and still fall short of an overly ambitious goal. Conversely, exceeding a modest target is not enough to demonstrate an exceptional result.
In its body text, Gartner’s press release uses broader wording about the desired savings than its headline. It does not specify, alongside the 3% figure, the sample size or the corresponding questionnaire. It would therefore be unwise to infer that 97% of all companies derive no benefit from AI, or that their investments are all failures.
That limitation does not make the warning pointless. It prompts a more precise question: when a company announces a gain, what exactly has it measured?
One study does measure a task being completed faster
A study published in 2025 in the Quarterly Journal of Economics provides a different kind of evidence. Erik Brynjolfsson, Danielle Li and Lindsey Raymond analyse the gradual rollout of a conversational assistant to 5,172 customer support agents. They measure an average increase of 15% in the number of problems resolved per hour.
The result varies considerably with experience: the least experienced agents improve more, while the most skilled record small speed gains and a slight drop in quality. The authors stress that their study concerns one company and medium-term effects. It does not measure the general consequences for wages or employment.
This study does not automatically contradict Gartner’s warning. The number of requests resolved per hour and the achievement of a budget target answer two different questions. Comparing them helps explain the gap; confusing them would lead to a false conclusion.
Three outcomes that are too often conflated
The first is the time spent on a task. A report may be prepared faster, but the stopwatch must include checking it and making any necessary corrections. Measuring only the generation of the first draft means stopping before the work is finished.
The second is the team’s capacity: how many completed cases can it handle at comparable quality? The time freed up may allow it to respond faster, clear a backlog or deal with cases previously left aside. These are possible benefits, even if staff costs remain unchanged.
The third is a saving actually realised. This requires comparing expenditure over a defined period, taking into account the additional costs associated with the tool. The result cannot be obtained by mechanically multiplying every minute ‘saved’ by an hourly rate.
A company should therefore be clear about what it is promising. ‘Producing more’, ‘responding better’ and ‘spending less’ are three defensible objectives. Combining them in a single figure makes it impossible to know which one has been achieved.
A simple example to avoid misleading calculations
Imagine a fictional team preparing one hundred cases a week. Each case used to take twenty minutes. With an assistant, preparation takes twelve minutes, plus three minutes for checking: the net gain is five minutes, or five hundred minutes over the week—about eight hours and twenty minutes.
This calculation is useful, but it does not amount to an accounting saving of eight hours and twenty minutes. If the team remains in place with the same hours and costs, it has additional capacity. We still need to see what it does with that capacity.
If checks take longer or errors mean that some cases have to be redone, the actual gain decreases. If the freed-up capacity is used to reduce a significant wait for customers, the value may instead lie somewhere other than a reduction in expenditure. The example does not predict the outcome for any company: it simply shows the steps in a verifiable line of reasoning.
Measure the whole process, not the demonstration
To assess a use case, one can start with a clearly defined task and an expected outcome: a case accepted, a query resolved, a usable document. Failed attempts and rework are part of the cost of that outcome, just like subscriptions or calls to the service.
The comparison should then cover cases of comparable difficulty. If the tool handles the easy requests while people take all the exceptions, comparing their average times without accounting for that distribution does not tell us much.
Finally, the assessment should retain a quality criterion. Going faster while producing more corrections later shifts the workload rather than removing it. The useful end point is work that is accepted, with an explicit error rate.
AI should therefore not be judged through a simplistic choice between miracle and disappointment. The operational question is more demanding: for which task, with what checks, do we get what net result? It is at this level that an organisation can turn a productivity promise into a proven benefit.
Illustrative photo: a computer workstation, with no established connection to the companies studied. Credit: Shixart1985 / Wikimedia Commons / CC BY 2.0.
