Automation with visible boundaries

I led state controlled-substance reporting automation, coordinating product scope and exception handling. This was deterministic reporting work, distinct from the AI assistant work.

The reporting workflow prepares state-standard dispensing files and delivers them securely, with a clear exception path for incomplete work.

Traceability and reviewable requirements help make reporting automation dependable. These public product principles do not expose internal implementation rules.

The product work included reporting requirements, failure handling and a phased rollout.

I launched the customer-care console and personally led duplicate-account search, embedded chat, packet workflows and clinical-ticket routing.

Customer-care contacts often arrive with relevant history that the representative must reconstruct before helping. I envisioned assistance that brought useful context into view sooner and worked with customer-experience and contact-center partners to bring it into the product. Source-linked context mattered: a representative needed to inspect the basis for a suggestion, not take polished text on faith.

Around that core, I worked on self-service assistance, representative support and conversation review. Retrieval from approved knowledge can help with routine questions; suggested answers must remain distinguishable from actions; a human needs sufficient context when a conversation exceeds the assistant’s scope. Product evaluation included quality, latency, cost and correction effort. Clinical authority remained with qualified people.

The shipped billing assistant explains claim rejections and routes work; its chat scope does not authorize mutations. Separately, prescription data-entry automation paired quality evaluation with operator feedback. I defined net automation to account for verification send-backs, because a higher automation rate without its error rate is an incomplete product claim. The billing-before-verification workflow is a separate design study.

The shipped billing assistant explains claim rejections and routes work; its chat scope does not authorize mutations. Separately, prescription data-entry automation paired quality evaluation with operator feedback. I defined net automation to account for verification send-backs, because a higher automation rate without its error rate is an incomplete product claim. The billing-before-verification workflow is a separate design study.

And a failure worth more than most of the successes. An assistant stopped handing callers to humans. Every component was healthy; every dashboard was green; the handoff between them had broken. We had monitored the nodes and not the seam. The fix added transfer monitoring and a callback fallback. The lesson has taken up permanent residence: in a system made of systems, instrument the joints.