Seven lessons for customer-facing AI agents with screens

Drawing on over a year of building and researching customer-facing AI applications, the author argues that language models should never act as gatekeepers for sensitive actions. Instead, models should only propose actions while deterministic code enforces security rules, permissions, and business logic against system records. The article outlines seven key lessons for deploying customer-facing AI agents safely, covering topics such as retrieval methods, single-agent versus multi-agent architectures, silent failures, probabilistic guardrails, and generative user interfaces. Open article →