The performance of an AI project in business is not measured by the deliverables produced, but by business results: lead times, quality, rework, costs, and reallocated time. This is the thesis defended by Alexandre Ruiz, enterprise systems architect and author of La Couche IA, an essay dedicated to artificial intelligence and organizational transformation. For him, eliminating a process is sometimes preferable to automating it.
- AI projects must be judged by business results, not by deliverables produced.
- Five indicators measure value : lead time, rework, decision quality, reallocated time, and total cost.
- Individual gains can disappear when collective work practices stay unchanged.
- Eliminating an unnecessary step can create more value than automating it.
AI demonstrations share a common weakness: they often stop the moment the document is generated. A summary appears in seconds, a table fills automatically, an email writes itself almost alone. However, for the author, creating the deliverable is not the business result. To measure a transformation, one must observe what happens next: validation, error, rework, decision, satisfaction, and cost.
Five indicators beyond the deliverable
The first indicator is the complete lead time between a request and its resolution. Saving twenty minutes on drafting a note doesn’t change a file pending for three days. The second is the rework rate: a response produced quickly but corrected twice may cost more than a slow and reliable one.
The third indicator concerns the quality of the decision. Producing a hundred recommendations is worthless if they don’t inform any arbitration. The fourth concerns reallocated time: did teams truly manage to dedicate themselves to customer relations, training, or more complex cases? The fifth is the total cost, which must include licenses, integration, supervision, risks, updates, and maintenance.
An individual gain absorbed by the collective
A randomized trial by Microsoft Research involving 6,000 employees observed a decrease in time spent on emails, but no significant change in meeting time. The lesson, according to the author, is not that AI doesn’t work: an individual gain can be absorbed by unchanged collective practices. NBER research on customer support assistance also shows that benefits vary by profile and task.
Eliminate rather than automate
These indicators can lead to a sometimes counterintuitive decision: not automating a step that would be simpler to eliminate. Before building an agent, meaning an AI program capable of chaining several actions autonomously, which transfers a file between three teams, why not clarify decision-making responsibility? An AI strategy is not judged by the number of automations in production, but by the cost and quality of a sustainable business outcome.
