Voice AI that handles patient calls, safely.
Advising on a state-of-the-art voice AI that makes and takes patient calls, tested for clinical safety so clinicians get their time back.
01 · The problem
What Quadrivia AI actually needed.
Clinicians spend a huge share of their day on the phone and on admin: booking appointments, chasing results, following up on chronic conditions. That task tax is time not spent on care, and it caps how many patients ever get seen, hardest in the places where clinicians are already scarce. The obvious fix is to let AI handle those calls. The reason it has not happened is that healthcare has zero tolerance for an unsafe answer, and a voice that talks to patients has to clear a clinical bar, not just a product one.
02 · Context and insight
The reframe that set the direction.
I came in as a product and strategy advisor at the intersection of my two specialties applied to healthcare: AI and consumer products. The reframe: the hard part is not whether the AI can hold a patient conversation, it is whether you can prove it is clinically safe before it ever speaks to one. So the work is framed around provable safety in the call, not raw capability. An AI that can both make and receive patient calls at clinician-level quality only matters if every one of those calls is safe by default.
03, The approach
The decisions that mattered.
An AI that calls patients, and answers when they call
The product handles both directions of the conversation: outbound calls for appointment reminders, chronic-condition follow-ups, and results outreach, and inbound calls during the peaks that overwhelm a front desk. Each call returns to the care team as a clean transcript and summary, so nothing is lost and a human can step in at any point. I advised on which workflows to automate first and where the AI should always hand back to a clinician.
Clinical safety, proven in the call
Capability is the easy half; safety is the product. I focused on how the AI is held to a clinical-safety bar on every call, so it knows when it can answer, when it must escalate, and when it must stop. Designing for that bar from the first conversation, rather than bolting checks on afterwards, is what makes a clinic willing to put it in front of real patients.
Reducing the task tax to widen access
The point of automating routine calls is not cost-cutting, it is access. Every hour of admin the AI absorbs is an hour a clinician can spend on care, which means more patients seen with the same staff. I advised on framing the product around that outcome: give clinicians their time back, and make decent healthcare reachable for more people.
04 · How it's built
Close to the stack, not above it.
An advisory engagement on a stealth, state-of-the-art product, so I am deliberately light on internals here. My leverage is product judgment and a focus on clinical safety in the call. The advice carries weight because it comes from someone who founded an AI company and ships AI-first product directly, applied to the one domain where getting it right is not negotiable.
The engagement is current and intentionally low-profile, so this is contribution, not a closed result. My focus is the part that makes patient-facing voice AI adoptable at all: proving it is clinically safe in the call, and framing the product around giving clinicians their time back so more people get care. The throughline matches everything else I work on, AI applied to a consumer problem, here with the highest possible bar for getting it right.
What I’d carry forward
In healthcare the question is never can the AI answer, it is can you prove it answers safely, every time. Building for that bar from day one is what separates an impressive demo from something a clinician will actually let near a patient.