Data & AI, a magical duo: what leaders really need to steer

FCE Nice conference “Data & AI, a magical duo”, hosted by Flora Debrosses, with Didier Aït (Optim’ease), Audrey Lamy-Martinot (Data Acteur) and Arnaud Pinte (iPepper).

The room is packed. The desire for clarity prevails. Flora Debrosses opens the evening with a simple observation: artificial intelligence has no consciousness. It imitates. It computes. It assists. The future will be played out in the complementarity between humans and machines.

Three voices structure the debate.

• Didier Aït positions the leader as the conductor.
• Audrey Lamy-Martinot brings us back to reality: without a data asset, there is no useful AI.
• Arnaud Pinte shows the path use case → prototype → industrialization, with rapid returns on investment.

The tone is measured. Rigor, pedagogy and field experience.

The leader, conductor of AI

A paradigm shift. For the first time, the company is working with an imitation of intelligence. Models dialogue, summarize, classify, generate. They do not perceive the world. They feel nothing. They are assistants.

Didier insists: AI does not write strategy. It executes it. Hence the requirement of meaning. Objectives, mission, values, business model. Then prioritization: where does AI create value now, without diluting the company’s differentiator?

Key message: do not fetishize the technology. AI serves know-how. It does not create it.

Start with the forgotten asset: the data heritage

Without data, AI runs empty. First step: take inventory. Where is the data? Who is responsible for it? What is it used for?

Audrey offers a telling analogy: we monitor our bank account. We must monitor our data capital. Data is an intangible asset that gains value through use.

Three operational pillars stand out:

• Data governance. Clear roles and responsibilities. One owner per data object. Update rules. A controlled chain of impact between departments.
• Data quality. Continuous processes. Explicit quality indicators (Key Performance Indicators, KPIs). Regular monitoring: completeness, freshness, uniqueness, accuracy.
• Data dictionary. A shared vocabulary. Business terms are defined, validated, disseminated. Without a lexicon, no reliable collaboration.

Methodologically, Audrey combines top-down (starting from business need) and bottom-up (letting the data speak). The meeting of the two produces better decisions. She adds a concrete point: make models explainable. A conversational assistant can present statistical results in an intelligible way.

From use cases to prototype: rapid returns, frugality and sovereignty

Arnaud breaks a belief: AI does not distance us from our business, it brings us back to it. We start from a pain point. We prototype in weeks, not months. We measure value. We adjust.

His approach boils down to three verbs: Detect → Shape → Build.

• Detect: select a clear, quantified business pain point.
• Shape: prototype quickly, with the available data and a tight scope.
• Build: industrialize if proof of value is there.

A major point of attention: reliability. Large Language Models (LLMs) “always answer.” Even when they don’t have the right information. On figures and dates, the risk of hallucination exists. We must set the pace with confidence thresholds. Add safeguards. Sometimes accept a fine-tuning phase before reaching the expected level.

Another angle: sovereignty and frugality. Not everything requires a cloud giant. Models specialized on enterprise data can run in-house, reducing the footprint and protecting the information asset. The choice of tool comes after analyzing the data and the risk.

An example everyone can relate to: optimizing technician routes in real time. AI reallocates trips, limits back-and-forth, provides field information (nomenclatures, parts). Result: operational fluidity, time saved, reduced costs.

Mistakes to avoid and the right success metric

Common mistake: launching without an assessment. First, document. Then, target.
Another pitfall: wanting “the big bang.” The right approach: small steps, quick proofs, simple measurements.
Last trap: forgetting the teams. An AI project remains a transformation project. Without buy-in, it stalls.

How do we know we are making progress? Through objectives tied to differentiating know-how. That’s the compass. We track variances. We correct. We keep a critical mindset. We don’t fall in love with a solution. We stay pragmatic.

Fear exists. It is addressed by working side by side. Showing usefulness on the job. Reducing re-entry of data. Freeing up time for the core business.

The company benefits from recognizing a Chief Data Officer role. Their mission: governance, security, quality, vocabulary, change management. A bridging role between technical and business sides.

Tools, costs and funding: making grown-up trade-offs

The market is teeming. “Serious” LLMs abound. “Lightweight” connectors accelerate integration between payroll, point-of-sale, warehouse, business management, customer relations. We build a Data Warehouse and Data Marts. AI queries the right data sets.

On the systems side, the ERP (Enterprise Resource Planning) remains the backbone. The CRM (Customer Relationship Management) holds the customer view. The priority is not the tool. The priority is the data and its quality.

The budget question is real. Support schemes exist depending on projects and sectors, via public funding actors such as Bpifrance (the French Public Investment Bank), subject to eligibility conditions. The takeaway: seek levers suited to your size and context, while keeping governance in-house.

The audience — Fears, desires, new territories

The dialogue confirms a healthy tension. Save time, yes. Replace teams, no. In healthcare, the stakes are tangible: relieving practitioners of administrative tasks to secure care. In art, AI opens up new practices. The prompt becomes writing. Photography did not kill painting. AI does not kill art. It shifts the boundaries.

Common thread: dispelling fear through use. Show. Test. Measure. Adjust.

Practical roadmap

• Flash audit of data and uses. Map applications, flows, sources, access, risks. List three quantified pain points.
• Minimal governance. Appoint data owners. Issue update rules. Open a shared dictionary.
• Prioritized quality. Target five “gold” data objects. Define simple indicators. Set up a review ritual.
• Prototype in four to six weeks. One use case. One scope. A binary verdict: value proven or not.
• Reasoned security and sovereignty. Distinguish what can go outside the company and what must stay internal. Document trade-offs.
• Scaling up. Industrialize what works. Open the second use case. Train through on-the-job support.

Key takeaways

• The leader frames and prioritizes. Technology follows.
• Data is an asset. It is managed as such.
• Returns on investment exist when you start from the business and prototype quickly.
• Reliability is steered. Hallucinations are contained.
• Acculturation is ongoing. Trust is born from use.

Open questions for your executive committees

• Which data asset truly differentiates your company today, and how do you monetize it?
• Which business pain point deserves a prototype within six weeks?
• Where do you set the acceptable reliability bar for decision-making, and how do you measure it?
• What sovereignty/cost/performance trade-offs are you ready to commit to, in black and white?
• Is your data dictionary readable by a new employee in 30 minutes?

We had interviewed Arnaud Pinte and Flora Debrosses as part of our #EntrepreneurIA investigation. Interview by Pascale Caron.