Descriptions:
Chris Lovejoy (Member of Technical Staff, forward-deployed engineer) and Saul Howard (VP of Engineering) from Anterior present at AI Engineer on the gap between successful AI agent proof-of-concepts and production enterprise deployments. Drawing from firsthand experience embedding with large US health insurance companies, they identify the specific blockers that appear the moment a POC ships its benchmarks and moves toward productionization.
The talk walks through a concrete scenario: a two-engineer team builds a POC in four weeks that hits every performance metric, only to face a cascade of requirements at the stakeholder meeting — compliance audit trails for every agent action and data access, PHI lifecycle controls governing what data agents can read and when, role-based access control that mirrors human HIPAA requirements, clinician approval workflows for certain decisions, and on-premises data residency constraints that prohibit data from leaving a customer VPC. Each of these requirements has deep architectural implications that weren’t built into the POC.
Lovejoy and Howard present specific primitives they’ve developed in response: event sourcing (append-only event logs) as an architectural pattern that makes auditability fall out naturally from storage design rather than requiring retrofit; schema-driven object storage for PHI with strict RBAC enforced at the data layer; and structured human-in-the-loop interfaces that integrate clinician approval into agent workflows. They argue these patterns generalize directly from healthcare to any regulated industry — finance, defense, government — where process compliance is non-negotiable and the hard problem of AI deployment is infrastructure, not model capability.
📺 Source: AI Engineer · Published August 19, 2026
🏷️ Format: Keynote Launch
