AI ROI: how to calculate it from actual usage data

Among the equipped companies Idun Group meets, 93 had told us by the end of October 2026 how they measure the use of their AI. 39 measure nothing. Another 14 rely on the statistics supplied by the vendor. In large companies of 5,000 employees or more, 18 out of 29 have no measure at all. This is a count of the companies that told us, not a representative survey.

Yet these are the leadership teams being asked for the ROI of AI. They end up calculating it without knowing which seats are used, or for what.

Where do you start when calculating AI ROI?

Calculating AI ROI starts with actual usage: what teams really do with the tool, rather than the number of licences bought or what people report in a survey. The method fits in five steps:

1. Measure actual usage. The logs of the tools you provide (Microsoft 365 Copilot, ChatGPT Enterprise, Claude…) show who uses them and how often. To know for which tasks, you need to read a sample of exchanges or ask the teams. Automation traffic is set aside, otherwise it inflates the volumes. The reading is done by team and by task, within a framework agreed with the data protection officer and, where applicable, with employee representatives.
2. Keep the uses that serve an objective. Saving time on a task with no value for the company earns nothing. Each use you keep is tied to an objective: handling more cases, answering faster, reducing rework.
3. Record the recoverable hours on those uses, with the teams and task by task: the time the task used to take, the time it takes with the tool, what is redone by hand. The logs are used to cross-check.
4. Convert them into cost at an hourly rate built from the company's own profiles, which management control knows how to establish. A generic rate produces a figure nobody recognises in a committee meeting.
5. Compare with the full cost: licences, usage-based consumption, support, internal time spent on deployment. Then do it again at renewal, with the same method.

Hours are counted first; their conversion into euros comes at the end, at the company's rate.

Why is AI ROI so hard to prove?

AI ROI is hard to prove because the available figures on its profitability are almost all self-reported. The large surveys ask executives what they observe; none of them measures usage inside the tools. You have probably seen these figures in a committee deck; for each one, look first at who was asked, and when.

• The Banque de France surveyed some 7,000 companies from late February to early April 2026 (Bulletin No. 266, 3 September 2026). 67% of companies with 20 or more employees report using generative AI. About 31% of those that use it report productivity gains, or 22% of all companies surveyed, according to the Banque de France's own calculation. It concludes that the cost-benefit balance of adopting AI will need to be better documented.
• The MIT NANDA report (July 2025) estimates that 95% of organisations get no measurable return on their profit and loss. The figure concerns custom enterprise tools, not general-purpose assistants, and rests on 52 interviews, 153 surveyed executives and an analysis of more than 300 public initiatives.
• BCG (September 2025, 1,250 executives) counts 5% of companies achieving AI value at scale, and 60% reporting only minimal gains.
• Deloitte (October 2025, 1,854 executives in fourteen countries, France included) raises the question of timing: 6% report a payback in under a year, and most place a satisfactory ROI on a typical use case at two to four years.

The Wharton survey (October 2025, some 800 decision-makers at US companies with 1,000 or more employees) seems to say the opposite: 72% report tracking ROI metrics linked to the business. The contradiction is only apparent, because "I track metrics" and "I have observed a gain" do not measure the same thing. Neither survey says what teams do inside the tool.

The "159% median ROI" so visible in French search results comes from a study self-published by a service provider on 200 French deployments; it relates a self-reported annual gain to the investment, and the study itself states that this is not a net ROI.

The same MIT report points to another gap: 40% of companies say they bought an official subscription, while workers at more than 90% of the companies surveyed report using personal tools. An ROI calculated on licences alone sees only part of the usage: the rest goes through personal tools the company cannot see. For adoption and costs across the whole company, see our guide to AI in the enterprise.

How do you calculate the ROI of generative AI?

The ROI of generative AI compares what the uses tied to a business objective bring in with what the whole set-up costs (licences, consumption, support, internal time):

ROI = (recoverable hours × rate built in-house − full cost) ÷ full cost

Each term comes from a different source:

Consumption counts in that full cost: some tools are paid per seat, then per request, like GitHub Copilot.

We speak of recoverable hours, never of hours saved. "Saved" promises a cost reduction the measure does not prove. Recoverable hours are the time freed up on a useful task when the tool takes part of it over; what the company does with that time then shows up against the objective.

The amount the formula gives is therefore the value of the time freed up, at the company's cost. It is not a saving booked to the profit and loss: it becomes a gain when the linked objective moves, with more cases handled or shorter lead times.

In a randomised trial published by METR in July 2025, 16 experienced developers expected AI to speed them up by 24%; they took 19% longer, and still believed they had sped up by 20%. METR stated in February 2026 that these results no longer reflect the effect of current tools: the trial says nothing about today's AI, only about the possible gap between perceived and measured time. That is why task time is recorded with the teams, task by task, and cross-checked against the logs.

How do you isolate the gain from AI from other changes?

You isolate the gain from AI by measuring task by task, rather than across the whole company. One executive interviewed by Deloitte put it this way: "We only managed to get a ballpark estimate of the benefits because it was hard to separate the gains from AI initiatives from those of other initiatives, like operational excellence, team reorganization or changing roles."

A company-wide indicator, such as revenue per employee, mixes AI with everything else. A task tracked before and with the tool, in the same team, isolates it better: if nothing else changed on that task, the difference belongs to the tool. When something else did change (new software, a larger team), it is written next to the figure: the difference no longer belongs to the tool alone.

What does a measure of actual usage show?

At a consulting and training firm, measuring the logs of around forty Claude seats over one month showed three things no survey would have revealed:

• 31% of the seats had produced nothing, even though real activation reached 85%;
• at the vendor's public usage-based rate, the consumption of the three heaviest users would have cost 12 to 25 times the price of their seat;
• half a day spent rewriting the description of a shared tool returned about 6 hours a month, time spent in Claude reworking unresolved requests.

The same measure gave the base for the calculation, assisted work across 13 recurring tasks, and a quality indicator: exchanges in which a shared tool is loaded have 43% fewer unresolved requests.

These figures cover a single month, because the export keeps only thirty days, and they promise nothing elsewhere. Nor do they give an ROI in euros: hours cannot be converted without the company's time baseline and rate. The details are in this case study.

Which indicators should you track each month?

Five indicators are enough to follow AI ROI without commissioning a new study for every committee meeting:

• the share of seats that produce, automations set aside;
• the cost per producing seat, rather than per paid seat;
• the number of assisted tasks tied to a business objective;
• the recoverable hours on those tasks, updated with the teams;
• the share of requests left unresolved or redone by hand.

The first two can be read in the logs and invoices; for Microsoft 365 Copilot, see why an active licence is not a used licence. The third and fifth require reading a sample of exchanges, within the framework agreed with the data protection officer. The fourth is built with the teams.

What an AI ROI calculation must not do

An AI ROI calculation loses its value in front of a CFO as soon as it does any of these six things:

• display an ROI figure before any measurement, as ROI simulators do, promising a result nothing has yet measured;
• apply a universal hourly rate the company does not recognise;
• annualise one month of measurement when activity is seasonal;
• confuse recoverable hours with hours saved;
• attribute to AI the whole gain of a period in which the team, the tools or the volumes also changed;
• count only licences and forget consumption, support, internal time and the seats nobody uses.

The sheet to send to your CFO

To frame AI ROI before the next committee meeting, gather these five lines. The first and last come from the team that runs your tools and from the invoices; the others come from the business teams and management control:

1. The number of paid seats, and the share that produced something last month.
2. The five most assisted tasks, and the business objective each one serves.
3. For each, the time before and with the tool, estimated with the teams.
4. The hourly rate used, built on the profiles concerned.
5. The full cost for the month: licences, consumption, support, internal time.

Idun Group has carried out more than forty AI diagnostics in a year. Our analysis of actual usage combines your tools' logs with interviews with your teams, to say what to scale, what to consolidate and what to cut.