Article
An AI gain is measured against a baseline taken beforehand, on a date set in advance, with a written method, all costs included. Without that, it is a potential, not a gain.
Key takeaways
- A baseline taken beforehand, a date set in advance, a written method: without them, no gain, only a potential.
- Three statuses: identified potential, hours recovered, savings achieved.
- A payback period is given all costs included: building, running, review, the teams’ time.
- A freed hour is only worth something if it is reassigned or avoids a hire.
Three statuses, three kinds of proof
Learn more
Every figure says where it comes from. The same number of hours is not worth the same depending on whether it is forecast, measured or confirmed.
| Status | Where it comes from | What you can say |
|---|---|---|
| Identified potential | An assessment: validated interviews, calculation shown | “We can free up about…” |
| Hours recovered | A dated before-and-after measurement, on the same tasks | “We freed up… between this date and that date” |
| Savings achieved | The client: hours reassigned or a hire avoided, only with a dated measurement, costs deducted, and the written sign-off of its finance department | “We saved…”, confirmed in writing by the client |
Take the baseline before changing anything
Learn more
A measurement only makes sense against a baseline. Before go-live, record for each task targeted: the volume (how many times a week or a month), the time per occurrence, who does it, and the period recorded. Choose an ordinary period, with no closing or holidays.
An assessment gives a first baseline, reported and validated. For the measurement, compare records of the same kind: a report with a report, a log with a log.
Write the method before looking at the results
Learn more
The method is written before go-live, in the quote or the roadmap, and does not change afterwards.
- The tasks measured, defined once and for all.
- The measurement date, set in advance: for example three months after go-live.
- The hourly cost used, by role where possible.
- What counts as a gain: an hour reassigned or a hire avoided, not a “freed” hour that gets lost.
- Who measures, and who validates the result.
The stories we tell ourselves
Learn more
- Announcing the potential as a gain: forecast hours are not hours recovered.
- Leaving out costs: building, running, review, the teams’ time. A payback period is given all costs included, or not at all.
- Measuring over an unusual period, or only among the people most at ease with the tool.
- Changing the definition of the task along the way.
- Mistaking an impression for a measurement: 68% of French small businesses using AI report a positive impact (France Num 2026 barometer); that is a perception, not a measurement.
The calculation, all costs included
Learn more
The cumulative net value at a given month: minus the one-off costs (assessment, building, the teams’ time), minus the recurring costs of each month elapsed (running, review, subscription), plus the value of the hours recovered for each month of service. The payback period is the month in which it goes back above zero.
It is the rule used by the simulator and the homepage example; the Sources and methods page details each input.
Frequently asked questions
How long after go-live should you measure?
Long enough for use to settle: three months, for example. What matters is setting the date before seeing the figures.
Can you measure without a tool?
Yes. A simple record is enough: volume and time per occurrence, over an ordinary period, before and after, by the same people.
Which hourly cost should you use?
The loaded cost of the roles concerned. Failing that, €45/h, the average loaded hourly cost of French companies with 10 or more employees in 2025 according to Insee (€44.7), rounded; it overstates a clerical task.
What if the measured gain is below the potential?
That is information, not a failure to hide: state the gap, look for the cause (adoption, exceptions, data quality) and adjust what comes next. Comparing forecast gains with measured gains is what makes the next estimate credible.
Sources
The sources (2)
- Insee, “Salaires et coût du travail en Europe” (2 July 2026): average loaded hourly cost in 2025, €44.7 (in French)
- DGE, “Baromètre France Num 2026 : quatre TPE-PME sur dix utilisent désormais l’intelligence artificielle” (28 September 2026): 68% of users report a positive impact (in French, accessed 23 October 2026)
Further reading
Keep reading
- ROI of an AI project: how to calculate the gain and the payback period
- How many hours does it take for an AI agent to pay off?
- Playbook: running a successful AI assessment in an SME
And in your organisation?
An AI agent interviews your teams about their work and an expert validates every deliverable. You know where AI will save you time, with costed use cases and a roadmap.