The main risk of an AI project isn't technical. It's shipping something that works… and changes nothing on the P&L. Demos impress, teams get excited, and six months later nobody can say what the project earned. The antidote is one discipline: define the value metric before writing the first line of code.
The four families of value
Almost every AI integration ROI breaks down into four families:
- Time: hours saved on a task, shorter processing times, capacity absorbed without hiring.
- Quality: lower error rates, consistent answers, less rework and fewer goodwill gestures.
- Revenue: better conversion, average order value, retention, faster commercial response.
- Risk: incidents avoided, demonstrable compliance, reduced dependency on key people.
Pick one primary family per project. A project promising all four at once will measure none of them seriously.
The five-step method
- Measure the baseline. How long does the process take today? What error rate? What cost per case? Two weeks of observation are often enough. Without a reference point, there's no proof.
- Set a quantified, dated target. "Cut first-response time by 30% within three months" — not "improve support". A quantified target forces the right conversations at framing time.
- Isolate the scope. Compare like for like: same team, same request types, same season if possible. Otherwise every variation will be contested.
- Count full costs. Development, infrastructure, your teams' time during the project, training, then maintenance and monitoring. An ROI that ignores maintenance is fiction.
- Review weekly. A simple dashboard — volume handled, quality, time saved, incidents, adoption — is enough to steer and correct early.
A worked example: document processing
A logistics team manually keys in data from 400 documents a week, at 12 minutes per document. An AI extraction pipeline reduces human intervention to an average 3-minute check.
| Item | Before | After |
|---|---|---|
| Weekly time | 80 h | 20 h |
| Monthly cost (fully loaded) | ~ €3,000 | ~ €750 + monitoring cost |
| Data-entry error rate | 4% | < 1% (anomalies flagged) |
The honest calculation includes the project and infrastructure costs, amortized over the year — and checks on the ground that the 60 freed-up hours are actually reinvested (other tasks, absorbed growth) rather than evaporated.
The classic traps
- Theoretical ROI: "15 minutes saved × 200 cases × 220 days"… which assumes every saved minute turns into value. Check what teams actually do with the freed-up time.
- The unrepresentative pilot: tested on easy cases, generalized to hard ones. Include the ugly cases in the pilot.
- The tool adopted at 20%: value comes from usage. Measure adoption as much as model performance — an accurate tool that gets bypassed earns nothing.
- The metric that moves for other reasons: seasonality, team changes, a new offer. Hence the isolated scope.
The compounding effect
The first project rarely pays spectacular multiples — and that's normal: it carries the learning costs. Its real value lies elsewhere: it puts in place the clean data, the infrastructure, the measurement habits and the confidence that make the second project twice as fast and half as expensive to deliver. AI ROI is built as a portfolio, not a one-off bet.
So the question to ask your vendor — or yourself — isn't "what can AI do?", but "which metric will we move, by how much, and how will we prove it?". If the answer is vague, the project isn't ready.