AI and ROI: where to start without going wrong

The most common question we hear from executives is not "should we invest in AI?" but "where do we start without making mistakes?". The difference matters: it reflects growing maturity and legitimate scepticism about the market's unrealistic promises.

The first mistake is starting from the technology. Many organisations begin by choosing a tool — an LLM, a machine learning platform, a copilot — before identifying the business problem it should solve. This approach generates expensive "showcase" projects that are hard to industrialise and create disappointment that slows future initiatives.

The Cardan-AI method reverses the reasoning. We start from the P&L: what are your largest cost items? What bottlenecks limit your throughput? What repetitive decisions consume expensive human resources? These are the questions that reveal real high-ROI use cases.

Three criteria qualify a good first AI use case: (1) the gain is measurable in euros or hours recovered, (2) the necessary data exists and is accessible, (3) adoption by teams is realistic. A use case that ticks all three boxes can deliver results in 6 to 12 weeks.

The AI maturity diagnostic — what we call the AI Readiness Assessment — is the founding step. It evaluates simultaneously the quality of your data, your technical capabilities, your change culture and your regulatory constraints. This assessment takes 2 to 4 weeks and structures the entire programme that follows.

Once the priority use case is identified, the golden rule is to start small and measure quickly. An 8 to 12-week pilot on a limited scope, with KPIs defined from the outset, lets you validate the value hypothesis before committing large-scale resources.

In summary: start with the problem, not the solution. Qualify your use cases by their business value and feasibility. Pilot fast. Measure rigorously. This discipline — sober and ROI-driven — is what separates organisations that durably transform their performance from those that accumulate POCs that go nowhere.

Concretely, how do you size a use case before committing? We build a one-page mini business case: the current cost of the problem (hours × rate, scrap, late penalties, lost sales), the expected gain at a conservative percentage, the cost of implementation and maintenance, and the payback period. A use case with a payback beyond 12 months rarely survives industrialisation: it is wiser to keep it in reserve and start with a quick win that will fund the next ones.

The mirror mistake of 'all technology' is 'all POC': accumulating proofs of concept that never reach production. The cause is almost always the same — a pilot launched without a business owner, without a planned industrialisation budget and without stop criteria. Decide from the outset what triggers scale-up and what triggers a stop. This portfolio discipline, routine in industrial project management, remains rare in AI programmes — and that is exactly where real ROI is won.