Les trois paliers de l'ingénierie IA
A three-rung maturity ladder for placing any system built on a language model. Rung 1, building with a model: you call an API, you wire simple document retrieval, and you handle the base vocabulary (prompt, embedding, context window, tokens). Rung 2, production: the question is no longer whether it works on your machine but what happens when thousands of people hit it at once. You replace bare semantic search with hybrid search and re-ranking, you measure retrieval quality, you trace failures, you contain hallucinations, you hold latency and cost. Rung 3, agentic: the system reasons across multiple steps, calls tools, keeps state, evaluates its own output and decides what comes next. And it carries, on top, rung 2 evaluations and guardrails applied to its own behaviour.
Strengths
- Three rungs only, holdable in your head, usable in a meeting with no slides
- Each rung names concrete bricks to add, not an abstract seniority level
- Model-provider agnostic, the ladder holds for an in-house system as for an integration
Limitations
- A ladder stays a simplification: a real system can sit on rung 2 for retrieval and rung 1 for observability
- The vocabulary circulates mostly in career-change content, where it works as a sales argument more than as a working grid
Best for
- A product team that has to tell a client honestly where its AI feature stands
- A developer with a working demo who cannot name what separates it from production