How to measure AI ROI before scaling it up: two real cases

Aldes estimated an 800,000-euro return on investment in the very first year of its AI project, driven by an improved sales conversion rate and productivity gains for its technical-sales teams. Selectour, for its part, deployed an AI conversational agent for online bookings, launched in November 2025, which now handles 70% of bookings made through the site, with 110,000 conversations logged since launch. Both examples, documented by Bpifrance Conseil as part of a white paper drawn from supporting more than fifty SMBs and mid-sized companies, share something more interesting than the figures themselves, measuring these results was not improvised after the fact.

That is the main difference between an AI project that produces a demonstrable ROI and one that, months after launch, still cannot prove its usefulness. In the second case, the company has almost always launched its tool without first defining the indicator that would judge its effectiveness. It then ends up hunting, after the fact, for a figure to justify the spend, a difficult and unconvincing exercise. In the first case, the indicator existed before the project, conversion rate for Aldes, share of automated bookings for Selectour, allowing a direct before-and-after comparison.

This difference in method matters more than company size. Aldes and Selectour operate at volumes above a typical ten-to-fifty-person SMB, and the figures cited are not directly transferable. But the principle holds at any scale, choose an indicator already tracked before rollout, rather than inventing a new one specifically for the occasion, which would allow no reliable comparison for lack of a prior baseline.

For an SMB hesitant to start due to the cost of a prior diagnostic, help exists. Bpifrance and France Num offer Diag Data IA, tailored support co-funded up to 40% by Bpifrance, which caps the company's remaining cost at 6,000 euros excluding tax. Two thousand such engagements are planned for 2026-2027, making this a concrete option to structure, ahead of a project, the measurement that will later evaluate its result.

A positive ROI on a first use remains, however, local proof, not a guarantee of reproducibility. An indicator that improves markedly in one department says nothing about what will happen elsewhere in the company, this ties into a broader, well-documented reality, that of data silos that often separate departments and prevent simply replicating a local success. Measuring the ROI of a first project is a necessary step, it does not exempt you from later checking the conditions for scaling it.

One last point deserves attention, ROI measurement should not stop at the most visible direct gains. Aldes factored productivity gains for its technical-sales teams into its calculation, a less spectacular benefit than higher revenue, but just as real and measurable once a time or workload indicator already existed before the project. An SMB that limits itself to direct financial indicators risks underestimating a meaningful part of its project's real return.

This is exactly what the Steering step of the IMPACT method covers, which follows deployment, defining tracking indicators before the project launches and comparing results against the baseline measured upfront, rather than improvising an evaluation once the tool is already in place. An 800,000-euro ROI or a 70% automation rate only have value, for a company drawing inspiration from them, in the method that measured them, not in the figure itself. A five-business-day diagnostic defines the indicators suited to your project before it launches.