AI Diagnostic in 4 Phases
Before deploying a tool or choosing a model, the real question is: where does AI genuinely create value in this organisation? This methodology is designed to answer that question with rigour — and avoid pilot projects that never make it to production.
Current State Assessment
Before any recommendation, a thorough diagnostic lays the right foundations. This phase aims to understand the strategic context and the reality of available data — two inseparable dimensions that prevent technically unfeasible or strategically disconnected use cases from being pursued.
Strategic framing.
Identifying the organisation's highest-value domains: which processes generate the most cost, friction or risk? Which business objectives are priority? This step establishes selection criteria grounded in the company's reality rather than in technological enthusiasm.
Data framing.
Exploring what data is actually available: its volume, quality, accessibility and degree of structuring. A brilliant AI use case on paper can collapse within weeks if the necessary data is absent, fragmented or too degraded to be exploitable.
The Comex tool illustrates this phase: it generates in minutes a structured strategic and technical framing from a business need description, simulating the exchanges of an initial steering committee.
Use Case Identification
This is the most critical step — and the most often rushed. Identifying genuinely good AI use cases does not happen in a boardroom with a post-it brainstorm. It requires diving into operational realities. Three complementary angles help move beyond surface-level ideas.
Analysis of actual process performance.
Mapping existing processes (As-Is), identifying bottlenecks, low-value repetitive tasks, and recurring friction points. The goal is not to list every conceivable AI opportunity, but to locate the 3 to 5 areas where impact would be most significant.
Confronting management's view with field realities.
Operational teams experience problems that management does not always perceive. This structured confrontation — field interviews, direct observation, cross-functional workshops — is often the source of the most relevant and best-adopted use cases, because they are owned by those who will benefit directly.
Mining digital feedback.
Analysing often under-exploited existing data: support tickets, inbound emails, satisfaction survey verbatims, error logs. These sources carry a strong signal about real user pain points — raw material directly exploitable to identify AI use cases.
The Sirene tool illustrates this capability: it enables natural-language querying of the SIRENE/INSEE database, demonstrating how to make a complex public dataset exploitable without prior technical skills.
Prioritisation by Value Added
Not every identified use case deserves to be built. Prioritisation relies on an evaluation grid crossing two axes: expected value added (time savings, error reduction, improved user experience, quantified ROI) and implementation complexity (data availability, team technical maturity, regulatory constraints, integration effort).
Rapid prototyping to validate before investing.
Before committing to full development, a lightweight prototype tests value hypotheses against real data. This approach drastically reduces the risk of delivering a technically correct tool that does not meet the actual expectations of end users.
The DVF tool illustrates this logic: built to make a dense public dataset (property transactions) exploitable, it demonstrates how to turn raw data into a decision-support tool without over-engineering.
Deliverable and Implementation Roadmap
The final deliverable is not a list of generic recommendations but an operational framing document: prioritised use cases with quantified rationale, recommended target architecture, a tech stack adapted to the context, a phased deployment plan and tracking indicators. This deliverable is designed to serve as an immediately actionable decision basis for an executive committee.
Change management.
An AI project succeeds not only through the technical quality of its solution — it succeeds because the teams who need to use it have been on board from the start. Targeted training, adapted internal communication, identification of business champions: change management is planned from the prioritisation phase, not added at the end of the project.
Long-term monitoring.
Going live is not the end of the project. Tracking indicators are defined upfront: adoption rate, measured time savings, error reduction, user feedback. These metrics enable deployment adjustments, justify investments to stakeholders and feed the next improvement cycles.
A methodology tested on my own tools
This methodology was not designed in theory. It was first applied to my own projects: Comex, Sirene and DVF all went through these four phases — current state data assessment, relevant use case identification, value-based prioritisation, production deliverable. Before guiding an organisation through this process, I ran through it myself on real projects, with the real constraints of a solo developer: imperfect data, limited resources, the need to ship something functional and usable. That practitioner experience is what gives the method its credibility.