The Companion Layer

I'm a believer in what Steven Johnson calls the “slow hunch”¹: an idea that forms over years, swirling in the background as something not yet fully shaped. Darwin is his example — very close to natural selection for decades after the Beagle came home, the thing assembling itself long before he wrote it down.

Mine has to do with population health outreach and the coordination burden we hand to patients and families.

I've had the chance to run a large health system's vaccine program, which came with ownership of quality metrics like Medicare 5-Star flu and Pediatric Combo 10. That work put me next to some genuinely world-class population health outreach: the medical assistant who offers you a flu shot while you're in for something else, extended walk-in access during flu season, and the one I want to focus on — the emails, texts, and letters we send to people who are due for something.

Here is the structural problem with that last one. Every quality measure has its own cohort, and our systems run outreach per measure. The same person potentially gets a mammogram reminder, a flu text, a statin adherence call, and a colonoscopy letter, from four programs on four schedules, none of which know the others exist.

Sources: CMS Innovation Center (2022); Pham et al., NEJM (2007); Kern et al., JGIM (2019). Medicare fee-for-service populations.

Visits are only part of it. Add the messages, the calls, the labs, the refills, the prior authorizations, each generated by a different silo and processed by a team handling only what falls in their in-basket. That's grounded in scope of practice, which is appropriate. But it's also about managing burden on any one team, our aperture problem.

So the slow hunch is about a different way to manage outreach and virtual interaction. Breaking the silos. I think AI is the first thing that makes it buildable.

Not a portal

The idea I keep converging on is a personal health companion.

This is not a wellness app, and it is not an AI layer bolted onto a patient portal, though it may grow out of one. MyChart and the interfaces like it exist to manage flow and access for the health system. They were built as secondary systems to EHRs, often provide a minimum level of user functionality, and are behind the curve of changing user preferences for multichannel engagement.

The first real moves towards a companion layer are already here. Epic's Emmie is the AI assistant inside Hello World, their patient communications platform² — the SMS, voice, and email engine health systems use to reach patients at scale. OpenAI launched ChatGPT Health in January and opened it to all U.S. adults in July³.

That last one deserves a number attached to it. In the only independent evaluation published so far, ChatGPT Health under-triaged 52 percent of gold-standard emergencies, and its crisis safeguards fired inconsistently.¹¹ OpenAI has also released a separate, more rigorously validated version restricted to clinicians — which means the carefully evaluated product went to professionals and the less evaluated one went to everyone else. This is a system handling 230 million health questions a week.

Both are useful. Both are also shaped by whose problem they were built to solve, and with Epic you can see it in what gets measured. The published outcomes for Hello World² are a thirty percent drop in no-shows, five thousand appointments rescheduled, thirty percent of balance notifications paid in full, forty staff hours saved a month. Every one is real. Every one is operational or financial. Not one is a health outcome.

That's not a criticism of Epic. It's a description of where the assistant was placed. Emmie lives inside the outreach machine, now with a conversational front end. It makes that machine work considerably better. It does not change the siloed approach to the outreach. ChatGPT Health comes at it from the other direction: a general assistant that now has your labs, with no relationship to your care team at all.

There is a harder version of this problem underneath. When health systems choose AI tools, most of them pick whatever their EHR vendor offers rather than an independently validated alternative, and the deciding factors are integration convenience and institutional trust rather than performance.¹² Epic's own sepsis model went into wide deployment before external validation and turned out to require clinicians to work through 109 alerts to find a single true case. So the concern is not that the incumbent builds something bad. It is that the incumbent builds something adequate, ships it already integrated, and wins on convenience regardless of whether anything better exists.

Which is the strongest argument I know for putting the companion on the patient's side of the wall. A patient-side companion is not competing in that procurement cycle. It competes on whether a person finds it useful enough to keep opening — a different selection pressure entirely, and one where good enough to protect a moat is not a winning position.

What I'm describing moves from reactive — use it when you need care — to relationship-first. An AI that advocates for the patient, helping them manage their own health, their own data, and their own coordination burden.

The person I picture most often is somebody in their fifties taking care of a parent, keeping up with their own screening, and trying to build decent habits for their kids. For a large share of the population, managing coordination for themselves and then for the people they love is genuinely overwhelming.

That's a world of unmet need — understanding, answers, guidance, coordination — that healthcare and insurers keep out of provider channels out of necessity. We have a good sense of how large it is, because people are already asking somewhere else. OpenAI reported that 230 million people ask ChatGPT health and wellness questions every week⁴, close to a third of its weekly users. That demand didn't appear when the tools did. It was always there. It just had nowhere to go.

And it is not only consumer behavior pointing at this. Joel Vengco, chief information and digital officer at Hartford HealthCare, describes the agents his team has been building for two years as companions — for patients, clinicians, and administrative staff. He names the gap directly: patients often wait days or weeks for an answer, and sometimes the answer is just what do I do next, or remind me again what this is.¹⁰

We've been trained by scarcity. People self-ration on cost, on access, or on the sense that they shouldn't bother anyone. There's a whole category of questions that never get asked. What was my last A1c? I'm making dinner — what's a good option for my numbers? Is there aluminum in vaccines, and are they safe for my kid? The questions we don't want flooding our virtual channels are exactly the ones a trusted, well-bounded resource should handle, with clear escalation when a human is needed.

Getting to daily use takes four things: trust built through consistent accuracy, real work done on your behalf, a relationship that knows your life, and support for your curiosity rather than deflection of it. Those come first, and the order matters. Most visions start with clinical integration and assume the relationship follows. I think it runs the other way.

What it takes to build one

Inbound, many separate campaigns arrive and leave as one conversation. Outbound, one request fans into every system that has to answer it.

These are functions, not products. Any of them could be built several ways and a few exist already. What doesn't exist is the assembly.

The one worth pressure testing is the membrane — the bi-directional boundary that sits on the patient's side. It takes in what the health system sends and carries out what the person needs.

That's the answer to the problem I opened with. Today four programs send four messages on four schedules and none of them know the others exist. With a membrane, that outreach arrives at your companion first, which knows what else came this week, what you already did, and when you're willing to be interrupted. The health system still sends. You stop being the integration point.

Before going further I should state the objection in its strongest form, because it is the real one. A recent Nature Health analysis makes the point precisely: what matters is not depth at any single layer but depth occurring across several at once.¹³ When the interface, the data, and the ability to act all deepen together, advice and action and payment finally connect, and care stops being abandoned at the handoffs. That is the entire benefit I have been describing.

And the same co-occurrence concentrates the risk. When the system that interprets your question also routes your care and is paid for that care, the advice sits inside a closed loop with a structural conflict of interest. Whoever controls the interface controls triage. The benefit and the harm come from the same mechanism, which is why the answer cannot be less integration. It has to be integration with the control sitting somewhere else.

It is worth being precise about what the companion actually holds, because two records matter here and they are not the same thing.

The trajectory record is slow: who you are, where your health is headed, what you have said matters to you. The coordination record is fast: what is open right now, who owns it, what is overdue, across every service line you touch.

The companion maintains the second one as its daily work, and stays grounded in the first. That second part is what separates a coordinator from a coach. Without the trajectory, a companion can chase open items competently and still never tell you whether any of it is moving you toward the life you said you wanted.

And underneath everything sits identity, consent, and data. That's where the architecture becomes a position. If those sit inside the health system, what you've built is a better portal. If they sit with the person, you've built an advocate. That decision gets made early and quietly, and it determines everything after it.

Why it has to be relational

The evidence for relationship keeps growing. Google's AMIE, in a randomized study of simulated video consultations⁵, scored on par with primary care physicians on diagnostic accuracy and management — and was rated favorably against them on empathy, rapport, and confidence in care. How we make patients feel isn't adjacent to the medicine. For adherence, for disclosure, for whether someone comes back at all, it may matter as much as what we prescribe.

Set that against what we currently offer. You call, press through a menu, talk to two or three people, get bounced between departments, and still don't get your answer. Or you send a secure message, then a second one to a different department, because primary care and specialty care don't share a thread.

Now imagine that instead of gathering the energy to navigate that, you just talk to your companion. It's on demand. You can talk as long as you want. It has the same voice every time, never gets tired, never runs out of time. It knows your name. It knows you went to the doctor last week. It knows you have a dog, and it knows about your kids.

It isn't your front door to the health system. It's your problem solver. It takes what you need and negotiates with the system on your behalf. The relationship reduces the energy it takes to start a conversation, and then the effort it takes to finish the task. Two different costs, and we've never really counted either.

Trust also unlocks something practical. People are far more willing to connect a wearable or accept autonomous check-ins if they feel they have control of their information. Trust is the precondition for the data, not the other way around.

The clearest example isn't in the United States. South Korea is aging faster than any country on earth — the share over sixty-five doubled in about fifteen years to more than a fifth of the population, and there aren't enough doctors, social workers, or family caregivers to keep up. All three shortages I described in the aperture piece, already arrived.

What they built is a care-call service from Naver Cloud called Talking Buddy⁶, now used by cities and counties to check on tens of thousands of seniors living alone. It began during the pandemic as a single scripted question about fever. Welfare officials came back asking for something that could actually converse, because there were too many isolated people and not enough hands.

In late 2024 it called a seventy-seven-year-old woman outside Seoul who had woken in severe pain and could barely speak. She managed a few words and hung up. The system flagged it and alerted a social worker. She was in surgery for an acute hernia within hours.

Two details matter more than the story. The system remembered — on a later call it asked about her recovery. And social workers monitor the interactions and step in, because nuance and background noise still defeat it. That's the architecture working as designed: a relational agent holding the thread, with a human on the other end of the escalation.

That's the shape of what I'd most want here. Post-discharge check-ins. The days after a hospitalization when someone is home, unwell, and suddenly responsible for everything that was being managed for them the day before.

The real constraint

Right now our health records belong to health systems. Patients are siloed from their own information, across multiple systems, behind multiple logins, in fine-print PDFs, in language written for someone else.

There's a real opportunity in a companion that manages what it can of that within appropriate safety limits, then gives you a warm handoff to your care team. It carries the coordination and builds trust over time. I don't think I can overstate the value of reducing the energy it takes just to get something done. Built well, it internalizes your goals and helps you work toward them, which is what makes hard, long-term behavior change possible at all.

Not all of these components have been researched, tested, and built. That isn't an argument against building. John Gall's observation applies: a complex system that works is invariably found to have evolved from a simple system that worked, and one designed whole from scratch never does⁷. So it's an argument for starting now — small, on specific use cases, assembling the components, the interfaces, the clinical libraries, and the evaluation methods, one working piece at a time.

Maturity varies sharply across the stack. Placement reflects the state of published evidence, not the difficulty of the engineering.

That last one is the real constraint, and it deserves a number. As these systems have grown more capable, we are still struggling to design evaluation frameworks: in fact, the share of studies conducting clinical efficacy testing fell from about half for rule-based chatbots to roughly one in six for those built on language models⁸. Across more than a hundred patient-facing studies, evaluation has been driven largely by user-experience measures rather than clinical outcomes⁹. That's not a reason to wait. It's the reason to define what good looks like now, while the answers are still being set.

Last, and most importantly, adoption of AI based tools by care teams and patients will take time. A PHC won't be a fit for many patients. For the right ones, it could be a game changer.

As I say in all of these: this is one vision for how AI could change health for the better, and specifically how we might chip away at the unmet need in care coordination that nearly all of us struggle with every time we need help with our own care or the care of someone we love. I'd like to hear which parts resonate — and more than that, your own ideas about how emerging technology can reduce the burden of coordination and preventive health.