Observed, estimated, unknown: measuring the ROI of an AI project without making things up

Picture a sales proposal landing on the desk of a small business owner. On page 3, in large type: “Estimated annual gain: €148,000.” Below, in small print, the sum of three lines: hours saved converted into euros, additional revenue, and avoided errors.

The figure looks precise. It isn’t. It mixes time nobody measured, sales someone hopes for and errors nobody ever counted, and adds them up as if they were the same thing.

Here is the method we use to avoid that, and that you can apply on your own to any project, with or without us.

Even the most cited studies have this problem

The most widely shared MIT report of 2025, “The GenAI Divide: State of AI in Business 2025” (Project NANDA, copy of the report), states that 95% of organizations are getting zero return from their generative AI investments. The figure went viral on LinkedIn.

What got less attention is the report’s own methodology note: the figures are based on individual interviews, not official company reporting, sample sizes vary by category, and success definitions may differ across organizations. A project counts as successful when users or executives say it had a lasting impact.

That is no reason to dismiss the study. It is a good illustration of the point: a figure without a unit, a formula and a shared definition does not help anyone decide.

In France, AI is still recent for small businesses: according to Insee (Insee Première no. 2061, July 2025), 9% of companies with 10 to 49 employees reported using at least one AI technology in 2024, against 33% of those with 250 employees or more. Few companies have a track record to compare against. All the more reason to start from their own numbers.

Three statuses for every number

Before calculating anything, we put each number into one of three categories.

Observed: measured on your current data. The number of invoices processed last year, the time spent, timed on ten cases, the number of quotes sent. This is solid ground.
Estimated: an explicit, bounded assumption. “With the tool, a case would take between 5 and 10 minutes instead of 25.” It is written as a range, never a single value, and we say where the assumption comes from.
Unknown: the number that is missing to reach a conclusion. We don’t guess it: we write it down, and say how to get it (a two-week measurement, a pilot, a data extract).

An honest ROI case almost always contains unknowns. What matters is that they are visible instead of hidden in an average.

One sheet per gain

Each expected gain from the project gets its own sheet, with five lines:

The unit: hours per year, euros of margin, number of errors, days of lead time. One unit per gain.
The formula: the calculation, written out in plain words.
The inputs: each number in the formula, with its status (observed, estimated, unknown).
The range: low result and high result, varying the estimates.
The missing number: what would need to be measured to narrow the range.

A purely illustrative example, for entering supplier invoices:

Unit: hours per year.
Formula: invoices per year × (current time per invoice − time with the tool).
Inputs: 2,400 invoices per year (observed, accounting export); 6 minutes per invoice today (observed, timed on 20 invoices); 1 to 2 minutes with the tool (estimated, depending on the quality of the PDFs received).
Range: 160 to 200 hours per year.
Missing number: the share of unreadable or unusual invoices that will stay manual. You get it by running one month of invoices through the tool as a test.

Replace these with your own numbers: the structure is what matters, not the example.

The rule that changes everything: don’t add them up

Once the sheets are filled in, the temptation is to convert everything into euros and add it up. That is where most ROI cases go wrong.

Hours freed up are not euros. They only become euros if they are reassigned to something that earns money, or if they avoid a hire or overtime. Otherwise it’s comfort, which has value, just not on a cash-flow line. Our calculator Hours saved, euros earned (in French) is built exactly for that conversion, or for noticing that it produces nothing.
Revenue is not margin. If the tool helps you sign more quotes, the gain is the margin on those sales, not their amount.
Avoided errors are counted first as a number. Their cost in euros is often only a rough estimate: present it as such.

So the gains are presented side by side, each in its own unit, with its range. The owner sees everything and decides knowing how solid each line is.

The cost, in full too

The same rigor applies to the other column. The cost of a project is not just the provider’s invoice:

setup (development, configuration, data migration);
subscriptions and tool usage, per month;
maintenance and changes;
your team’s time: interviews, testing, training, and the slower pace of the first few weeks.

Our pricing is public and excludes VAT; internal time has to be estimated with you, and it is rarely zero.

How to decide with ranges

The question is not “what is the ROI?”, but: does the project pay off on observed gains alone, taking the low end of the range?

Yes: go ahead. The rest is a bonus.
No, but the unknowns could tip it: measure them first, with a short pilot on a narrow scope. A pilot that answers a precise question beats a full project launched on an assumption.
No, even in the high scenario: don’t do the project. We say that too; it’s what we call an ROI-first approach.

What the method doesn’t do

It doesn’t predict the future. It makes the assumptions visible, so they can be checked after go-live by measuring the same numbers again with the same method.

It doesn’t put a number on everything either. Some real gains fit no honest unit: less dependence on a single person, a team that is less swamped at month-end. We write them down in plain words, without attaching a made-up amount.

And it doesn’t replace data. If your baseline numbers are scattered across five tools, the first job is often to bring them together. Our free questionnaire Is your data ready? (in French) helps you find out in a few minutes.

Where to start this week

Pick one process that annoys you: entering invoices, payment reminders, quotes. For one week, note two things each time it happens: how often, and how long it takes. That is your first observed number.

Then fill in a sheet with the five lines above. If you get stuck on the missing number, that’s normal: it’s exactly what we look at together during a free 30-minute diagnostic.

For an example applied to quotes, see also our article Respond first, win the quote.