A consulting and training firm had rolled out around forty Claude seats. During the month observed, 85% of them were used: that is real activation, read in the usage logs. The same reading also shows that 31% of the seats produced nothing.
Real AI adoption is the share of paid seats that are actually used, read in the tool's usage logs. It is usually counted company by company. Inside a company that is already equipped, the question gets sharper: which seats produce, and why some in-house tools take hold while others die. We measured it at this firm in July and August 2026. The firm stays anonymous; its figures are given as proportions.
Measure AI adoption in usage logs, not in a survey
A survey tells you who thinks they use the tool. Usage logs tell you who opened it, how often, for what, and with which shared tools. The firm's AI leadership wanted to stop steering adoption blind. We read its logs, with no interviews and no questionnaire.
The basis of the measurement:
- around forty seats of Claude, Anthropic's assistant;
- one working month observed, from July to early August 2026;
- close to 180 human conversation turns per active user;
- 85% of the seats used by a person at least once.
The raw count was larger. A large share of the lines came from automations, agents that send requests on their own. They were set aside: counting that traffic as human use would have inflated adoption. That is the first step to measure AI adoption in your own company: separate what people do from what machines do.
Geoffrey Harrazi and Aurélien Sabaly, co-founders of Idun Group, presented the measurement to the firm's AI leadership in thirty minutes, on 10 August 2026. It reads by use and by tool, never as an assessment of people. In the report, each user has a pseudonym, and the key stays with the client: it is handed over only on written request from its management. This kind of reading is framed before it starts, with the person in charge of data protection and, where there are any, employee representatives.
What is this firm's AI adoption rate?
At this consulting and training firm, the real adoption rate of Claude is 85% of the seats, measured in the usage logs of July and August 2026.
Three words are often confused, and they do not measure the same thing:
- access: the number of paid seats;
- activation: the seats used at least once over the period;
- production: the seats whose use produced something beyond a few tries.
Surveys mostly measure access, sometimes self-reported activation.
Some of these users stopped after a few tries, though. Added to the seats never touched, that makes 31% of the seats that produced nothing over the month, close to one in three. Production covers 69% of the seats.
These seats are a decision to make at renewal: reassign some, restart others with a first use case, remove the rest. The period matters too: the month observed runs from July to early August, and a quiet seat may belong to someone on holiday.
According to INSEE, the French statistics office (Insee Première of 21 July 2026), 18% of French companies with 10 or more employees say they use at least one AI technology in 2025, and 58% of those with 250 or more. These figures say how many companies are equipped; inside each of them, the share of seats that produce is still to be measured. Our guide to AI in business covers adoption, costs and ROI across the whole company.
How much do the heaviest users consume?
At this firm, the three heaviest users are worth 12 to 25 times their seat, in the precise sense of the calculation: what their consumption would have cost at the vendor's public rate, set against the price of their subscription. It measures consumption, not value produced for the business. It includes a small share of automated traffic.
On the same tool, a few seats consume far more than the others, which the seat invoice does not show.
The same reading shows one seat that kept hitting its usage limits: more than three times per working day over the measured period, mostly with its automations. Its subscription was too narrow for such intensive automated use. If that seat has a standard subscription while an idle seat carries a larger one, swapping the two changes nothing on the invoice.
One limit goes with this figure: people who only use the chat count as zero in this calculation, because the chat carries no cost in the export. Their use exists; it simply is not priced as consumption.
Why are some in-house tools adopted and others not?
An in-house tool gets adopted when it is shared across the organisation, precise, and well described. The firm had built its own Claude skills, in-house tools that give Claude its brand guidelines or the method for answering a tender. Some load every week; others were dropped within days. On this one fleet, three measurements set them apart:
| What was measured | What the logs show |
|---|---|
| Reach | A skill deployed for the whole organisation reaches 6.4 times more people than its personal version |
| Precision | The skills that survive carry 45 hard values each (colour codes, sizes, units), against 0.4 for the others |
| Triggering | 89% of loads are decided by Claude, not by the user |
Since Claude almost always decides on its own to load a skill, training teams on invocation commands would change little in these loads. The lever is the skill's description, the text Claude reads to know when to use it. These gaps come from a single client.
What does half a day of work on an in-house tool return?
At this firm, half a day of work returns about 6 hours a month. That half day consists of rewriting the description of the presentation tool so that it triggers when it should. The 6 hours are time spent in Claude redoing unresolved requests or restating the brand guidelines, which the shared tool reduces. It is not time saved on the work itself: the logs cannot see that.
Two measurements show why a shared tool changes the work. When it is loaded, there are 43% fewer unresolved turns and 69% fewer frustrated turns. And the base is known: assisted work covers 13 recurring tasks.
We do not convert these hours into euros: that would require knowing how long each task took without the tool, and that baseline belongs to the firm.
What this measurement does not say
The figures from this measurement carry four limits, written at the top of the report given to the client:
- a single month observed, from July to early August: the platform export keeps thirty days, for a rollout that was five months old, and the period partly falls during the holidays;
- the time measured is time spent in the tool, never the time the work would have taken without it;
- hours do not convert into euros without the client's time baseline;
- these figures describe this firm's current state and do not predict what the same work would give elsewhere.
The scope, finally, is limited to Claude: the firm's other tools were not instrumented.
Five questions to ask about your own fleet
To find out where AI adoption stands in your company, put these questions to the team that manages your licences:
- How many seats produced something last month, automations set aside?
- How many produced nothing, and how many stopped after a few tries?
- What do the three most active seats consume, on which tasks, and do they hit their limits?
- Which in-house tools get loaded, by how many people, and how many exist in several versions?
- What period does the measurement really cover, and which tools can it not see?
On Microsoft 365 Copilot, the same question comes with different reports: we covered it in Copilot usage report: active does not mean used. What the Claude console shows and what it leaves out is on our Claude for business page.
The report given to the firm, with its four limits, can be shown to you in anonymised form before anything is signed, so you can judge our analysis of real usage on the evidence.
