Julien Talbot · activity ergonomist
Make AI observable, verifiable, recoverable.
When AI changes an activity, gestures, responsibilities, trade-offs, evidence, and recovery conditions need to become visible — before anything gets transformed.
Public talks, field work, writing, AI labs — grounded in real work.
Entry points
Choose the right path.
Three decisions, not a catalogue: buy organization support, inspect the Labs method, or prepare a public talk.
Organizations
Buy organization support.
Real-work diagnosis, situated AI integration, and supervisable agents: organizational interventions go through Ergonomia.
Labs
Explore the technical proof.
Reduced traces, evals, oracles, and product signal: Labs documents the method, not organization deployment.
Talks
Give people a shared AI language.
Talks and workshops for executives, professions, product teams, prevention, universities, and events.
Proof
Stages, essays, traces.
Every proof keeps its context: public stage, dated article, method trace, or replayable case — so you can inspect it.
UNEP, CCI, Sorbonne, Préventica, Cinov Ingénierie, Bras-Panon, French Tech Réunion: public proof, not an abstract promise.
Essay Agents are work systemsDelegation has to be judged from real activity: action, proof, observable state, responsibility, and recovery.
Labs Production-adjacent agent tracesTurn agent failures into replayable cases, oracles, and regression matrices.
Latest
Recently published.
Recent writing gives dated material for deciding, checking, or recovering a situation.
An XML system block that forbids intent narration: one repro (10 min → 33.9 s), corpus measures, an honest A/B, and what the patch cannot fix.
Jun 4, 2026 The Enterprise Agent Problem Is BeliefAn agent is not enterprise-ready because it can act. It is enterprise-ready when the belief it creates about its action is calibrated to evidence.
Jun 2, 2026 An AI talk should talk about real workWhy a useful AI talk should start from what AI changes in work, decisions, responsibility, evidence and human recovery, not from tools alone.
May 29, 2026 Raw Traces Are Not EvalsThe missing layer between real agent failures and measurable model progress: reducing messy traces into replayable eval seeds without laundering the signal.
Contact
Email Julien
For an intervention in your organization — diagnosis, AI and agents in real work — go through Ergonomia. For a talk or a conversation, share the situation, the decision at stake, and the constraint you can see.