The Form Wearing a Chat Interface: When the LLM Is Pure Attack Surface

We designed the LLM out of a flagship AI feature and it got better. This talk is the argument for that decision, made honestly. When a decision rests on server-side evidence rather than prompt nuance, a free-text input field adds an untrusted component inside your trust boundary without adding intelligence. This session breaks down how to identify when you are in this scenario, what gets eliminated when you remove the model (prompt injection, classification gaming, non-deterministic evals, inference costs), and when a form legitimately falls short. Finally, we explore the hybrid model that won: the model proposes, the human ratifies, and a deterministic engine decides.
Format: TECHNICAL TALK, BUSINESS TALK
Location: ROOM A
Speakers: Maciej Bil
Level: Intermediate to Advanced
Audience: Registered participants