Patient-facing AI, governed from day one.
AnyBio is the layer health AI runs on. These are the agents we run on it: one that keeps patients on pace, one that asks what a coordinator would have called to ask. Each installs into an existing RPM or care-management platform with one webhook, runs over SMS, and logs everything it proposes, sends, blocks, and escalates.
The layer between the reading and the phone call.
Adherence Agent
Keeps every patient on pace for their device-day thresholds, and leaves the on-pace ones alone.
Most missed monitoring months are not refusals, they are drift. The Adherence Agent watches each patient's pace against your program's thresholds, including the 2-to-15-day and 16-plus-day supply windows, and reaches out only when a patient is drifting off pace. Patients on track never hear from it. Opt-in and opt-out are built in, and every message is checked against your program's rules before it is sent.
Sends templated messages. The gate checks each one before it goes.
Symptom Agent
When a reading lands out of range, it asks what a coordinator would have called to ask.
An out-of-range reading, starting with cellular blood pressure, opens a short structured interview over SMS. The agent asks what your protocol says to ask, records the answers as structured symptoms rather than a wall of text, and hands off when it is done. Your clinicians author the protocol, the opening message, the red-flag rules, and where an urgent case routes. What it collects lands in the record as a FHIR questionnaire response alongside the reading that triggered it.
Holds a bounded conversation. A deterministic red-flag layer can escalate at any turn.
Already have an agent, or a validated algorithm you trust? It runs on the same rails, under the same governance, and the logic stays yours. See what is yours to bring and what is ours to supply.
Two agents, two different safety mechanisms.
A one-way nudge and a two-way interview are different shapes of problem, so the same control does not fit both. Claiming otherwise would be the easier sentence and the less true one.
Outreach that goes one way can be templated, so the control is the words themselves. Every message the Adherence Agent would send is evaluated against your program's policy first, and allowed, held, or routed to a person, with the decision recorded next to what was originally proposed.
An interview cannot be templated, so the control moves from the sentences to the task. The Symptom Agent has a small fixed set of things it is able to do: ask, record a symptom, escalate, hand off, finish. And a deterministic red-flag layer reads every inbound message and can route to your care team regardless of where the conversation had been going.
The point of the second one is that the safety-critical behaviour sits in deterministic code rather than in a model's judgement, which means it can be tested the way any other rule is tested.
What they will not do.
Never pretending to be a person
A patient is told what they are talking to, and how to reach a human, before the conversation starts. California AB 3030 and Texas SB 1188 both ask for this, and it is enforced at session start rather than left to whoever configured the program.
Not an emergency line
Every conversation opens by saying so, in your clinician's words. An urgent case routes to your care team with the standing guidance your protocol specifies.
No clinical determination
They collect, they surface, they escalate. What any of it means for a patient is a licensed clinician's decision, and nothing here is built to substitute for one.
Not live until your clinicians say so
The protocol, the red flags, and the routing are vetted and signed off by your clinical team before an agent runs on a real patient. Working code is not the same as cleared to use.
Every message is checked against your program's rules before it is sent.
Governance layer
PHI-Safe Model RoutingHuman-in-the-Loop ReviewLLM Cost and Usage TransparencyCompliance GatesAgent Types with Scoped AuthorityFull Audit TrailInside your program's rules. It sends.
Brought back inside your language boundaries, then sent.
Held and raised to a person, with the reason on the record.
Agents sit on the rails, not beside them.
Signal comes in from the device catalog, lands on rails that carry ingestion, identity, storage, BAAs, output governance, and an audit log, and the agents run on top of that. What they produce goes to your coordinator's queue and into the record as EHR-ready FHIR. Any model, any agent, same governance.
YOUR DEVICES
CLINICAL SYSTEMS
DEVICES CELLULAR + BLE
ANYBIO RAILS GOVERNED + AUDITED
AGENTS ON THE RAILS
COORDINATOR QUEUE
EHR-READY FHIR
YOUR DEVICES
CLINICAL SYSTEMS
Panel growth without phone-time growth.
The outreach and the context-gathering are the minutes that scale with your panel, and they are the minutes an agent can absorb. The design goal is that a coordinator carries more patients without carrying more calls. The question worth asking first is not how many minutes you billed: it is what fraction of your enrolled panel cleared its device-day threshold last month, and whether you know why the rest did not.
Talk to us about your monitoring program.
For RPM and care-management teams who want patient-facing AI with the governance already built in, and a record of everything it proposed, sent, blocked, and escalated.
