On September 22, we were at the Data and AI Salon in Nantes, divided between the conferences. Comparing our notes, we found the same scene from one room to the next. At the question time, someone in the audience asked what the project cost and what it brings in. When asked about the cost, the answer came without hesitation. When asked about the gain, she was more hesitant, and most often relied on usage indicators.
We encounter the same problem with our clients. It’s not a lack of sincerity: the feedback was precise, the recommendations were fair. It shows that the market has reached a plateau, where adoption is advancing faster than the ability to measure its effects. However, a few interventions have approached the answer. Their common feature lies in a decision that can be taken even before deployment.
The individual gains of generative AI are no longer really a matter of debate. Reference studies measure a An average productivity gain of 14 %, rising to 34 % for the least experienced profiles, and a reduction in professional writing time of 40 %. In the accounts, these gains are still to be seen. Macroeconomic statistics do not yet show any evidence of them, and few decision-makers report a measurable effect on their operating results.
Several rooms presented dashboards of the same type: active user rates, number of assistants created, connected tools. These indicators are useful. But a tool used by 60 % of the employees can make a lot of money or not make any at all, with exactly the same usage curves.. Adoption is a condition of value, not its proof. It can even prolong what we call the gadget. The system works, the teams appreciate it, someone defended it in committee, and yet its gains never reach the bottom line.
In Nantes, the cases presented roughly fell into two families. In the first, the most common, an assistant helps each collaborator to write their reports and emails. The gain is real: a few minutes per day and per person. But these minutes are scattered across multiple calendars and never form a line that someone could see.
The second family is more rare. AI intervenes in a process whose cost is already monitored, for example a quality control process that is activated when the defect appears, rather than several steps later when it is necessary to fix and then re-implement. The savings are then reflected in the cost of repairs, which the company already tracks with a history and a responsible person.
The difference does not come from the quality of the tool. It comes from a decision made upstream. In the second case, we knew before deployment where the gain would appear and who could achieve it. This dual reference point, a place in the accounts and a person to verify it, is what we call the point of verification.
Before deploying, know where the gain will appear and who will be in a position to observe it.
Setting a measurement point involves answering three questions before launching:
If no line answers the first question, the project is not necessarily useless. It then needs to be presented as it is., an investment in comfort or capacity, instead of defending it on the promise of returns.
The most convincing interventions of the day went further. Rather than adding a tool to the existing organization, they redesigned the process by bringing people and AI together. Economic history is well aware of this discrepancy. Electricity was available in industry as early as the 1880s, but its effects on productivity did not appear in the statistics until the 1920s, when factories stopped substituting it for steam and reorganized their workshops around it. The gain came less from the technique than from the organization built around it.
The reluctance is understandable. A rate of adoption almost always progresses. A quantified gain, however, engenders resentment in those who announce it and can turn against them. But the budgets opened in 2023 and 2024 are reaching a point where committees demand accountability, and a rate of usage will have difficulty defending a budgetary line.
If taken early, this commitment changes in nature. With a set milestone for launch, a gadget can be stopped after three months instead of being funded for three years. At the time of the arbitration, the sponsor has proof that he could not have produced it otherwise. The measure that seemed to expose him becomes his protection.
That is where we place our role. Announcing a return on investment would just be a promise. We prefer to install with our customers a measurement capacity based on their own figures, recorded at the locations they have chosen, and which continues to operate after our intervention.
Since the start of the academic year, the JEMS Consulting & Transformation team has been conducting research on the links between AI and a company’s financial performance. The goal: to project the benefits of a project before deploying it, by identifying in advance the lines and the people who will be responsible for verifying them. We will share the first lessons here.
We will return to the conference next year. We would like to hear more feedback starting with the end: which line was supposed to move, by how much, who noticed it. This milestone will not be reached with prospective thinking about what AI could become, or with the mere race to adoption. Already this year, those who spoke about avoided costs rather than deployed licenses showed the way.
Brynjolfsson, Li and Raymond, «Generative AI at Work», Quarterly Journal of Economics, 2025.
Noy and Zhang, «Experimental evidence on the productivity effects of generative artificial intelligence», Science, 2023.
Paul A. David, «The Dynamo and the Computer», American Economic Review, 1990.
Notes from the JEMS+ team, Data and AI Expo, Nantes, September 22, 2026.
Do you want to continue the discussion?
Report Entreprises & IA 2026
88 % of the IA POCs do not reach a large-scale deployment (source IDC and Lenovo). The first JEMS report analyzes the 6 factors that keep AI at the experimental stage, with the feedback from Renault, Michelin, CACIB, Artelia and Manitou.
