Amy Edmondson: the failure researcher your AI roadmap needs.

The QSP Summit lineup this year leans heavily on macro narrative — AI and the future economy, Europe's position, the usual keynote altitude. Amy Edmondson is the exception: the one speaker whose research translates directly into how a product team should run on Tuesday morning. If you ship software for a living, hers is the session to plan around.

Who she is

Edmondson is the Novartis Professor of Leadership and Management at Harvard Business School and has twice been ranked the world's most influential management thinker by Thinkers50. She is best known for the concept that made her famous and is now routinely flattened into a buzzword: psychological safety — the shared belief that a team is safe for interpersonal risk-taking, that you can flag a problem, admit a mistake, or ask a naive question without it being held against you.

The origin story is worth knowing because it inverts the intuition. In her 1990s research on hospital teams, the better-led teams appeared to make more medication errors. The data wasn't measuring errors — it was measuring the willingness to report them. The best teams didn't fail less; they surfaced failure faster. That single inversion is the foundation of everything she has published since, including The Fearless Organization (2018) and Right Kind of Wrong: The Science of Failing Well (2023), which won the Financial Times Business Book of the Year.

Why this matters more in the LLM era, not less

Deterministic software fails loudly: the build breaks, the test goes red. LLM products fail quietly and probabilistically — a grader drifts, a model handles a customer segment worse than another, a prompt change degrades an edge case nobody is watching. The only reliable detection mechanism for this class of failure is people choosing to say "this output looks wrong to me," early and often, against social pressure not to. Edmondson's hospital finding maps onto AI products with almost no translation required: a team whose error dashboards look clean may simply be a team that has stopped reporting.

Her more recent taxonomy of failure is just as operational. In Right Kind of Wrong she separates basic failures (preventable, in known territory), complex failures (system breakdowns from compounding factors), and intelligent failures — undesired results from a genuine experiment in new territory, run small, where the outcome could not have been known in advance. Shipping LLM features is intelligent-failure territory by definition: nobody, including the model vendor, can fully predict behavior before contact with real users. The discipline she prescribes — hypothesize, size the bet small, instrument it, harvest the lesson — is uncomfortably close to a job description for an AI product manager.

What to listen for in Matosinhos

Three things worth carrying into her session. First, how she distinguishes psychological safety from comfort — her standard correction is that safety enables candor and high standards together, not lowered ones; expect her to push against the soft-HR reading of her own work. Second, how the failure taxonomy holds up when the experimenting agent is partly a machine: who owns an intelligent failure when the hypothesis was generated by a model? Third, anything she says about learning velocity as a competitive variable — for European companies told daily that they are behind on AI, "fail well, faster" is a more actionable strategy than most of what the macro keynotes will offer.

Perpenda exists on a related premise: that judgment about AI systems is trainable, and that honest calibration — knowing which claims are settled, contested, or unsettled — beats confident fluency. Edmondson has spent thirty years making the organizational version of that argument. Worth the seat.