Decagon CEO Jesse Zhang argues AI agent companies win or lose on training loops and KPIs, not the model itself.
Decagon CEO Jesse Zhang argues AI agent companies win or lose on training loops and KPIs, not the model itself.
- Oura Ring reduced human-escalation rate from 1-in-3 to 1-in-20 using Decagon’s agent training loop.
- The core pitch: non-engineers on the customer’s team must be able to train and improve the agent themselves, continuously.
- Key KPI to prove to customers is human-escalation rate, not resolution rate or CSAT alone.
- Post-sales motion for AI agent startups is longer and more expensive than traditional SaaS, requiring ongoing iteration.
- Decagon prices on labor-cost displacement: the baseline is what a human support headcount would cost.
- Moats in AI agents come from proprietary training feedback loops and customer-specific data, not the underlying model.
- Jesse advises against over-indexing on FDEs (forward-deployed engineers) as the improvement mechanism — the loop must be owned by the customer.
2026-02-16 · Watch on YouTube
