---
title: "How to measure returns on AI — Marsanne Studio"
description: "Usage, outcomes, change management: the three families of measures for the return on investment in AI. Translation of an article from The Economist’s Bartleby column."
url: "https://marsanne.studio/en/insights/how-to-measure-ai-return-on-investment/"
lang: "en"
---

Insights

# How to measure returns on AI

From ‘tokenmaxxing’ to something more normal.

The Economist · Bartleby · 20 August 2026 · Translated from English by Marsanne Studio

In the beginning was AI, and business leaders around the world lost their heads and decreed that everyone must use the technology, no matter whether it was useful and no matter what it cost. They set up AI-usage leaderboards and coined silly words like “tokenmaxxing”. And lo, usage rose. Then the bosses remembered very basic notions such as “budgets” and wondered whether the idea was such a good one after all. So another question echoed through the boardrooms, a question about returns, which was harder to answer.

This potted Genesis of AI adoption is not entirely fair. As long as an organisation is trying to encourage usage, it still makes sense to look at indicators such as employees’ token consumption. “I expect a software engineer to spend more than a financial analyst,” says an executive at a software publisher. “But for someone in finance, it shouldn’t be zero.”

Yet measures of usage – the inputs – say nothing about the value created. And even when AI does produce benefits, points out Mika Ruokonen of LUT University in Finland, many of them quietly accrue to individuals, in the form of scattered time savings. If an hour or two freed up in an employee’s day goes on browsing Vinted, someone gains (Vinted, mostly), but it is not your company (unless you work at Vinted).

Shifting the focus to return on investment means concentrating more on outcome measures: a team’s productivity, customer satisfaction. These still need to be designed with care. Scientists, for example, are using AI to produce more papers (hurrah!) of doubtful value (oh dear!). A recent study by Tulsi Suchak of the University of Surrey and her co-authors found that papers drawing on one health and nutrition dataset appeared at an average rate of four a year between 2014 and 2021; in the first nine months of 2024 there were 190.

Outcome measures must therefore be designed to capture AI’s positive and negative effects alike. At Google, for instance, engineers have long been assessed on three dimensions. Speed: how long a code review takes. Ease: how much friction the system imposes when, say, a new developer joins. And quality, the decisive one: “I don’t just want to go fast and make it easy to ship catastrophic software,” says Richard Seroter of Google Cloud.

On top of that, the benefits of AI often take a while to show up. It takes time to learn how to use the technology. Increased output creates bottlenecks elsewhere. A paper published last year by Kristina McElheran of the University of Toronto and her co-authors looked at AI adoption in American manufacturing and confirmed the existence of a “J-curve”: productivity dips before it improves. The effect is particularly marked in established firms and seems to be partly explained by changes to the management practices that previously allowed them to track performance. Any forecast of returns must allow for this initial period of disruption.

Leaders also have to settle a much more fundamental question: what kind of return are they after? One way to show a financial return on AI is to treat time savings as a pretext for cutting headcount. For some roles and in some circumstances, that may be justified. But the chainsaw and staff morale go together about as well as salt and slugs. And people who are no longer there cannot be redeployed more productively. A better option is to calculate what AI saves on future hiring. (Working out the additional revenue attributable to the technology is more of an art still: it takes, at a minimum, good baseline data and a culture of A/B testing to isolate the effects.)

Above all, there is total uncertainty about the direction AI is taking. That is an argument for a third family of measures, alongside those on inputs and on outcomes. These measures, which Mr Seroter calls “organisational”, aim instead to capture the quality of change management. They might cover employees’ satisfaction with how AI has been rolled out, or progress in building in-house expertise in the areas that AI will not automate. A sole focus on usage ignores the importance of returns. A rigid focus on returns ignores the need to keep learning. ■

## Source

‘How to measure returns on AI’, Bartleby column, The Economist, 22 August 2026 edition, published on 20 August 2026. Original illustration by Paul Blow.

French translation produced by Marsanne Studio for reading purposes. The original text and its rights belong to The Economist Newspaper Limited.

[Read the original article on economist.com →](https://www.economist.com/business/2026/08/20/how-to-measure-returns-on-ai)

[← See all insights](https://marsanne.studio/en/insights)