Refills.com is a DTC telehealth platform. Like most DTC health companies, its growth engine depended on converting paid-acquisition leads into first orders — work handled by a human sales team texting and calling leads one at a time.
New-order conversion ran through human reps: expensive, inconsistent from rep to rep, and limited to business hours. Leads that came in overnight or stalled mid-checkout went cold before anyone reached them. Scaling revenue meant scaling headcount — the unit economics didn't improve with volume.
The question wasn't whether AI could send messages. It was whether an agent could actually sell — qualify a lead, handle objections, know when to send a checkout link, and know when to stop — inside a regulated telehealth context.
We built an SMS agent that engages leads across the acquisition funnel — new sign-ups, abandoned checkouts, stalled intakes — and carries the conversation through to a completed first order. The design work was less about the model and more about the boundaries: what the agent may say, when it follows up, and when it hands off.
Early drip-style follow-ups performed like drip campaigns. The step-change came from letting the agent hold a real two-way conversation — answering questions, handling price and legitimacy objections, and timing the checkout link to expressed intent.
The agent sells; it does not advise. Medical questions are detected and deflected to providers, opt-outs are honored instantly, and escalation paths route edge cases to humans. Compliance constraints were designed in as product requirements, not bolted on.
The agent ran against the human sales team on live leads, measured on the same conversion and revenue numbers. It earned the funnel by beating the baseline — the cutover decision was a chart, not a leap of faith.
The agent now drives 25–30% of Refills.com's total daily initial-order revenue. After it matched and then outperformed the human team on live-lead conversion, Refills retired the human sales function entirely — every sales conversation is now AI-run, around the clock, at near-zero marginal cost.
The system prompt turned out to be the product spec. Most of the iteration that mattered happened in the sales strategy encoded in the prompt — objection handling, tone, timing — not in the surrounding infrastructure. On the next agent, I'd treat prompt iteration as the core product loop from day one and build the eval harness before the first conversation ships.