TL;DR
Teladoc's unified homepage was leaking conversion and driving costly support calls. I led an agentic AI Health Assistant that understood open-ended member intent and completed care tasks — reversing the decline with ~$23M in added revenue, ~70K calls deflected, and NPS up from 7 to 55.
The Problem
When several acquired companies were merged into one product, the homepage became the single front door — and conversion started falling. The root cause wasn't one thing; it was that the product couldn't meet members where they were. People arrived with real-life, messy intents — “refill my meds and book today,” “what's my copay for therapy?” — and a static UI plus a rigid, rule-based assistant couldn't translate that into action. Members either bounced or called support at roughly $14 a call.
The Approach
I made the case for agentic AI in early 2024 — before it was an obvious choice in regulated healthcare — by framing it as a capability decision, not a feature. I structured the pitch across three lenses: strategic alignment (reframing the homepage as a care-orchestration surface tied to the shift to utilization-based revenue), financial justification (a 0.1% conversion lift was worth $20–30M; the MVP cost under 10% of that upside), and execution readiness (a PRD, a layered architecture, phased rollout, and clinical/compliance guardrails co-designed up front). I built cross-functional advocacy so it landed as the team's bet, not just mine.
The Solution
I partnered with engineering and data to break the system into a Knowledge layer (RAG over benefit policies and FAQs — chosen over fine-tuning for freshness and cost), a Reasoning & Orchestration layer (LLM intent detection with deterministic tool-calling so actions always run through real APIs), and an Evaluation layer (task resolution, trustworthiness, risk exposure, hallucination, latency). I deliberately scoped the MVP to three high-frequency, high-tolerance use cases — scheduling, excuse notes, and pharmacy refills — and held the line against scope creep. A PHI redaction layer kept protected data away from the model, and a hybrid model strategy cut cost per task by half.
The Impact
In about seven months: ~$23M in conversion lift, ~70K support calls deflected, ~$2M in savings, and NPS from 7 to 55 — at roughly 20 cents per task versus $14 per call. The deeper win was reusable infrastructure: the orchestration framework, RAG patterns, and eval harness mean the next AI product doesn't start from zero.