AI agents can connect every American to better care, but they need to answer three questions

12 hours ago 10

Rommie Analytics

By PRIYANKA AGARWAL

Soon, clinic front desks will field a new kind of call: an AI agent asking to reschedule a patient’s cardiology follow-up and send over her latest labs. The person answering has no way to know whose agent this is, whether the patient actually asked for this, or whether the medication list it’s offering is current. So the careful answer is no. The patient is back on hold, and a follow-up that could have been booked in seconds slips another week.

As a hospital medicine attending at UCSF, I saw where those slipped weeks can lead: a woman whose breast cancer had advanced until it broke through her skin, leaving hospice as her only option; a patient with diabetes who could never get their blood sugar under control and now faced an amputation. Neither became that sick overnight. Somewhere along the way, a referral went unscheduled, a lab result never reached the right doctor, or a medication list fell out of date. That is why I co-founded HealthEx, so that people, and the agents they trust, can act on their full health history.

Today a person can ask an AI model a health question and get a thoughtful answer in seconds. But the moment they want to see their own doctor, find out what a procedure will cost, or get a specialist the history behind a new diagnosis, they are back on the phone or in a portal. AI agents can carry people across those gaps: booking the follow-up, confirming coverage, gathering the records, filling out the intake form. The models can already do this work. Healthcare isn’t ready to let them, because a doctor’s office or health plan has no reliable way to know who an agent speaks for, what it’s allowed to do, or whether what it says is true. Agents will be adopted in healthcare at the pace of trust, and today trust is the bottleneck.

That bottleneck matters more every month. KFF polling shows nearly three in ten U.S. adults now turn to AI for health information at least once a month, up from one in six just two years ago. The records those tools need are moving, too: national health data exchange through TEFCA has grown roughly a hundredfold since early 2025, according to HHS/ONC. Demand is here, and so is the data. What’s missing are the rules that let an agent act on both.

The promise of agents in healthcare

I believe agents are the future of how Americans manage their health and engage with the healthcare system.

They can connect people to every part of a system that is hard to navigate but delivers some of the best care in the world, and help each person reach their own health goals without the paperwork and hurdles that stand in the way. The agents I’m most excited about connect people to the doctors and care teams they already have, or help them find the best specialist for their condition. Today, much of AI in health stops at the virtual visit; the bigger opportunity is the rest of the system. For someone managing a chronic condition or caring for a parent, continuity is the care, and that connection is the difference between intending to get care and actually getting it. Multiplied across the country, it is what health abundance looks like: more people getting the care they need, sooner. Getting there is work for the whole industry.

Three questions every request has to answer

In healthcare, there can be no anonymous agents. Before a provider or plan acts on an agent’s request, it needs answers it can verify to three questions.

Who is this person? A verified individual, at a known level of identity assurance, rather than whoever happens to be logged in. The agent acts openly on that person’s behalf, so it never needs their passwords or has to sign in as them to a patient portal or claims system.

Did they authorize this? Consent that names the organization, the data and the action, and that the person can revoke at any time. An agent asked to book a mammogram shares what the imaging center needs for the appointment, and nothing more. When a request goes further, the agent comes back to the person.

What does the agent know, and can it be relied on? The person’s whole record, current and complete. Most people’s history is spread across places they have received care. An agent filling out a new-patient intake form without that history starts the clinician off with a gap, or with a wrong answer. An agent that can’t yet do that reliably is not ready for healthcare, however capable it is elsewhere. A medication list typed from memory is a starting point; one drawn from the person’s records is something a clinician can act on.

I believe a trust framework has to come first. Get it right, and every health system and plan has a reason to say yes to an agent, without building a separate integration for each one, and every person can choose whichever agent they want to act for them. Skip it, and each organization has to be convinced one at a time.

For a health system or plan, that yes is worth having. An agent that arrives verified, authorized and carrying the person’s records can mean fewer staff touches per request, intake that arrives complete, and patients who make it to the follow-up. The concerns are real, too: an agent impersonating a patient, an agent that is confidently wrong, unclear accountability when something goes wrong, and staff fielding a flood of agent calls. A shared framework answers each one, with verified identity, records the organization can trace to their source, a log of every action, and requests that arrive in a standard form the organization can review on its own terms.

What a shared framework needs

Answering those questions consistently, across thousands of organizations, takes a shared framework with five elements:

A required level of identity assurance, IAL2 identity proofing and AAL2 authentication, bound to every agent session. Consent scopes mapped to specific clinical data and actions, consistent with HIPAA and state privacy law. A standard package of the person’s context, built on FHIR and USCDI, so every organization receives the same shape of information. A hand-off that lets a person’s everyday assistant pass a task to a health agent without ever holding the person’s medical records itself. A person can start a task with the assistant they already use, while their records stay with an agent held to healthcare’s rules. A record of every action an agent takes, which the person can review, with permissions they can revoke and no sale of their data.

Through all of it, the organization receiving the request stays in control. It keeps its relationship with the person, decides what to share or do, and keeps its own record of each request. The framework carries the request and the organization’s answer, and changes neither.

The framework should also be open, built as a profile on the standards agents and health IT already use, so any agent and any health organization can adopt it. The Personal Agent Protocol, which Sierra and Meta introduced this fall, gives healthcare a strong place to start: a shared way for businesses to recognize a person’s agent, which a health profile can extend with the identity, consent and context that care requires. A trust framework works only when the people who rely on it help write it: health systems, health plans, clinicians, patient advocates and the teams building agents, together.

Getting time back

Nobody should have to spend hours on hold, carry CDs of images from office to office, or guess what their care will cost. When an agent a person trusts can reach their doctor, their plan and their records, with their permission and on their terms, they get that time back, and their care team can work as one. The work in front of healthcare now is agreeing on the rules of trust, so that when agents arrive at the front door, the answer can be yes.

If you lead a health system or a health plan, ask every agent you plan to let in three questions: Who is this person? Did they authorize this? Can I rely on what it’s telling me? Expect a verifiable answer to each. If you build agents, build them to be recognized at the front door. And if you’re helping write the standards, bring health systems, plans, clinicians and patient advocates in from the start.

At HealthEx, we help people bring their health context into the AI tools they already use. Now we are forming a working group with health plans, providers and EHRs to develop this framework in the open. If you want to help shape it, our team would welcome the conversation.

Priyanka Agarwal, MD, MBA, a former associate professor of medicine at UCSF, is co-founder and CEO of HealthEx.

Read Entire Article