January 2025, Agentic AI was in its infancy — Gartner put significant enterprise investment at 19%, Manus hadn't shipped yet, and there was no playbook for Agentic AI. We started building one for a regulated government entity anyway.

Six months later, we had our 1st agentic app in production. Most don't get there. Only 15–20% of enterprises move agents past pilot, and Gartner expects 40%+ of agentic projects to be canceled by 2027.

This talk is about what was actually hard. Not the model. Not the agents. The users. The productization. The gap between a demo that dazzles and a product people trust on a Monday.

You'll leave with:

[] Ship the least AI you can get away with, first — why restraint is the rational strategy
[] Deterministic UX is not polish. It is the product — designing around non-determinism, not pretending it away
[] The competition you can't see will kill you — the habit your users won't tell you about

A first-person field report, not a framework. Bring your scars.

SPEAKER BIO:
Abhijit Dutta is a product builder, currently shipping AI products at WorkSafeBC, where the multi-agent system in this talk lives in production. Fifteen years across SaaS, B2B, e-commerce, and government (SALUS, APOLLO, Fatigue Science, Trulioo, BuildDirect) — building products that needed to work on a Monday, not demo well on a Friday