A trust-first benefits agent: it explains a member’s plan in plain language, grounds every answer in their actual plan document, and treats the handoff to a human as a designed feature, not a dead end.
View the live site →In five years as a Care Advocate, the same handful of questions drove most inbound contacts: what a deductible is and why it applied, why an in-network visit was billed a certain way, what benefits exist, how to find a provider. Each costs a member 15 or 30+ minutes for something that should be instant.
The category has a documented adoption failure: AI care assistants sit under 15% engagement. Only 4% of Americans can define deductible, copay, coinsurance and out-of-pocket max together, and 60% are uncomfortable with AI in their care. Capability was never the barrier. Trust and literacy are. My constraints: solo build, plan-level documents only, no real claims data, no medical advice.
Members ask for a human because the AI doesn’t feel trustworthy, not because it can’t answer, the failure is relational, not technical. So Sage leads with informational empowerment: it defines every insurance term on first use, quotes exact dollars only when it can stand behind them, and treats the human handoff as a warm transfer with full context carried over, not an error message.
Generic “most plans cover…” answers are exactly what makes members distrust these tools, because their plan is not “most plans.” Each conversation is bound to one member and their actual Summary of Benefits and Coverage. The agent behaves like it knows you, and structurally cannot leak another plan’s numbers, because it never receives them.
In healthcare, what the agent must never do matters more than what it does. I wrote the hard rules first: no clinical advice, exact dollars only for copays, coinsurance routed to a provider estimate, an explicit 911 trigger list, and abuse and bereavement protocols. Every later flow inherited safety instead of retrofitting it.
A blank input tells an anxious member nothing. Sage opens by naming the member and their actual plan (“I have your Clinic S Plan C100 open”), surfaces the human handoff before the first question, and offers four tappable starters mapped to the core use cases. The free-text field is still there for anyone who wants it.
I shipped a prototype that's live and runs end to end across four use cases: benefits and coverage, claims and billing, provider search, and ID card and documents. The handoff to a Health Navigator is a visible UI state, a distinct card that names the three things being carried over, so the member never repeats themselves. Provider search runs a real tool over an openly-synthetic in/out-of-network directory, with ZIP and network read server-side so the one-plan grounding holds.
Same question, two correct and plan-specific answers. The agent isn’t retrieving a generic definition, it’s reading this member’s plan and answering for them.
I tested safety instead of assuming it. I ran 15 scripted conversations against the constraint layer; live testing caught the agent quoting out-of-network coinsurance against the in-network deductible, understating exposure by thousands. That surprise bill is exactly what this product exists to prevent, and it only appeared under a live run. The fix is now a hard rule: in-network and out-of-network never mix.
Real, deployed, and reachable across all four use cases, grounded in the member’s actual plan. I backed it with 18 real plan summaries and 12 stress-test personas, and ran it through a scripted safety and grounding pass, with one rule guaranteed in code. Success metrics are defined and ready for V2 usability testing:
Share of conversations resolved without a human handoff.
From question to actionable information.
Whether members can correctly restate what they were told.
Member-reported confidence in the handoff itself.
“The document that defines what the product must never do is worth more than the one that defines what it does.”
Writing the constraint layer first was the highest-leverage decision, it felt slow, but it meant safety was designed in, not bolted on. And I trusted the prompt more than I should have until I ran it: exercise the behavior, don’t just author it, and move a rule into code when the outcome has to be guaranteed.
View the live site →