The Three Clocks

In the aperture piece I mentioned that when I establish care with a new clinician, I tell them I want to be managed as a lifelong athlete, against the ninetieth percentile rather than against my age cohort. Don't hold me to what is good enough. Hold me to the best.

That longer lens is not missing from healthcare exactly. It lives in secondary systems and practice variation.

I should say plainly where I am standing. I am not a clinician. Twenty-two years of my career have been spent behind the scenes: in clinical operations, population health, strategy, and the systems that support our care teams, explore efficiencies, and maximize throughput. I have delivered patient care through systems rather than at the bedside. That vantage point sees different things, and it is the one I am writing from to hopefully augment our community of clinical voices. For today I’m mostly speaking from an ambulatory medicine perspective as that is where the bulk of care is received.

One of those observations is that our health record systems are transactional by design, and largely by necessity. You come in for a thing, the thing gets handled, and we send you to hopefully go about your business. If something else can be accomplished while you are in the exam or virtual room, we will address what time allows.

Structurally it is a state-to-state process. A team member enters your chart, gets reacquainted with you, builds a checklist for the visit. During the visit the patient is talked to, the record is updated, orders are placed, follow-ups are queued. When the tasks are done the record goes dark again until the next time you have more care needs. The highest-touch version of this state-to-state model is a care manager working from a registry, who might look at your chart daily. Otherwise, your information sits in a repository waiting for the next query.

The opportunity I see is a member of the team who does that checking continuously, and who is not only checking what is in the chart.

Before that, the model.

Clock one runs the whole life. Clock two is a handful of episodes. Clock three, which our system is built for, holds acute visits.

The full model, layer by layer All four layers across five life stages, with actor roles and the information flowing between clocks. Open the detailed model →

The Clocks

Clock one is the slowest. Years to a lifetime. On the front end it holds your genetics profile, your early life, childhood illness, family history, the preventive care you received. That is your starting place, and it sets an initial path. From there it ticks slowly, shaped mostly by behavior over time, adjusted by a chronic condition, weight gain, an accident.

The most important thing clock one holds is your goals, which evolve across your life. Maybe you want to play sports. Maybe at twenty-five you are a professional athlete, or maybe you just want to hike on weekends. Maybe you want children. Maybe a condition requires a certain diet, and staying with it is the goal. Toward the end of life, your goals become the things you want to remain possible. Walking the dog. A dignified ending.

Goals should be the real driver of your trajectory, within the parameters of your health profile. They are the vision of where you want to go, and once they are explicit the conversation changes: alongside the medical plan, what behavior gets me onto the path I want? The further medicine moves in this direction, the more patients will feel heard in their own care.

Your race, ethnicity, income, and neighborhood shape this clock too, not through biology but through exposure and access. And access is critical. Our current systems do plenty of clock one work — vaccinations, screenings, health maintenance, risk stratification. The concern is that they exist as a list of things to complete, not a trajectory to steer. Each item is a gap to close at the next visit.

Clock two runs days to months. This is destabilization. A chronic condition drifting out of range. A hospitalization and the weeks after it. The stretch where something has come loose and hasn't been put right yet. The care team's job here is to carry you through the episode.

This is where most suffering and most cost accumulate, and it is where the system is the most manual and disjointed. The coordination burden is enormous, and it spreads across pharmacy, specialty, primary care, rehab, home health, palliative care, sometimes hospice. The default quarterback is the patient or a family member.

Clock three is the fastest. Minutes to days. The acute encounter. You have a cold. Your knee has been bothering you for a while and you finally decide to do something about it, so you try a virtual visit, or primary care, or urgent care. This is real care and it matters, but in the grand scheme of your life it's the grains of rice.

By necessity and by incremental evolution, our systems are most developed for this clock. It is the tractable one: a defined task, a defined visit, a defined bill. Our EHRs have gotten better at clocks one and two, and programs like Meaningful Use and Value-Based Care accelerated that but also increased the data entry burden on care teams. Regardless, the architecture is still built for efficient, reactive deep dives into near-term needs.

Patients live on clock one, deteriorate on clock two, and are seen on clock three. We have structures that try to manage clocks one and two, but they're anchored inside clock three.

Who does what

If we shift to a more longitudinal, trajectory based care model, the obvious question is what happens to the care team. The short answer is that we must add digital, AI members to the team. That is not a proposal to automate the work of the current care team, in fact I think it's likely that we create jobs. The key is to think through the lens of what we’re not doing today that is now possible and not focus on maximizing our current models.

So let me be specific about what changes and what does not.

The three-clock model becomes possible because AI plays one key role: the always-on, always-organizing member of the team. Consider care managers today working from a registry. Even with a small panel, they get to each chart roughly once a day, and they have to be efficient about it. They set registry alerts. They rotate through charts in queue. They prioritize the medically complex and the socially high need. They are fully engaged and very good at this, and they still have to manually query, manipulate, and click into every chart to find out anything.

The model I am describing runs on an orchestration layer coordinating a team of agents doing many automated tasks at once: curating data packages, managing specialty agents behind the scenes, capturing context, pulling in outside information, triaging requests, and supporting a multimodal front door. It extends the care team rather than replacing it.

AI also addresses something we have never solved well: manual error and gaps in protocoled care and in hand-offs between teams and systems. It does that by reaching out consistently as a multimodal chat agent, watching for escalation, and monitoring the devices already around you, so that vitals monitoring continues between visits rather than only inside them.

Finally, it also helps the care team model a trajectory, drawing on physiological models, chronic disease datasets, social context, clinical pharmacology, and research.

Here is what that leaves for the people on the team, and I would argue it is the better half of the work. Less data collection, less synthesis and fewer manual checks. More listening, removing barriers, supporting behavior change, setting goals, understanding what someone needs right now and wants out of their life. The scope shifts toward judgment, uncertainty, and trust, which is what clinicians trained for and what the current system gives them the least room to do.

The visit changes shape too. Less asking questions and typing. More conversation.

The adjustment for patients and families is larger than it looks. We have all been trained by a model of scarcity, where people self-ration based on cost, access, or the sense that they should not bother anyone. In a model where asking is easy, the questions people currently swallow have somewhere to go. That is a piece worth its own treatment, and I will come back to it.

I do not think AI takes healthcare jobs. I think it moves our people toward relationships, judgment, and trust. That seems like a good trade.

How information moves

The biggest structural change is the shift from a series of snapshots to continuous movement.

Context travels down; consequence travels back up. Outside signals enter only with permission.

At the center of clock one is a trajectory record: your medical identity, held as a small file with the pertinent facts about you. That is the starting point. From it, a near-term and long-term trajectory get drawn, and then layers accumulate.

Which layers, and how many, should be the patient's decision. That matters more than it might appear. A record that holds your sleep, your nutrition, your adverse childhood experiences, and the content of your conversations can do considerably more for you. The honest version of this model is that the depth of the record should be a choice, made by the person it describes, revisited as they go. Some people will share everything. Some will share nothing beyond what their claims data already reveals, and their care should be no worse for it.

With permission, the layers can include social determinants, nutrition, sleep, wearable data, labs and imaging, preventive needs, and context from conversations. Over time it becomes a living version of who you are, medically.

This is not a hypothetical data structure. Published work on longitudinal health agents is converging on the same shape¹: a persistent record separating shared medical knowledge and safety rules from private, individual memory, with explicit rules governing what gets promoted into the durable profile and what stays episodic or is discarded. In one recent architecture, that structure raised accuracy on longitudinal questions from under one percent to about forty-six, while sharply reducing how much context had to be exposed to answer them¹. The retention decision and the permission decision turn out to be the same design problem.

Clock two needs its own record, and it is not the same one. Where the trajectory record holds who you are and where you are headed, the coordination record holds what is open right now: which referral is pending, which prescription has not been filled, who owns the next step, what is overdue, and what is waiting on somebody else. It spans service lines and channels, because that is exactly where coordination fails.

These two get conflated, and conflating them is a reliable way to build the wrong thing. The trajectory record is slow, clinical, and genuinely hard — ontology, projection, goals. The coordination record is fast, operational, and far more tractable. It is closer to the work-tracking systems every other complex industry already runs, and healthcare mostly does not, at least not across organizational boundaries.

Most of the burden we hand to families comes from the absence of the second one.

There is a useful word for what the coordination record is actually for, and it is completion. A recent Nature Health analysis argues that health outcomes are lost not only when care is unavailable but when people fail to finish the sequence of steps required to obtain it — the referral that never gets booked, the coverage question that stalls, the prescription never picked up.⁶

I would add the other half. Completion fails on the way in, and it fails again on the way out. The discharge instructions given to somebody medicated and exhausted. The medication list that no longer matches what is in the cabinet. The follow-up nobody scheduled. The plan agreed in a room that then has to survive an ordinary week. Getting to care is one sequence. Acting on it afterward is another, and it is longer, less supervised, and where the burden mostly lands on families.

Both are clock two. Both are coordination record problems. We measure neither one well.

Then it starts driving things.

When you are seen, your medical identity is the first thing the care team looks at. Who this is. Who they are medically. What their goals are. What care they are currently receiving. That is the baseline, and care plans are set from it. In clock two, the plan is a near-term trajectory informed by clock one, carrying you through the episode and toward a new normal, which then updates clock one.

In clock three, the visit stops being only about the thing you came in for. It becomes grounded in your goals and identity on clock one.

The clearest example of why this matters is a hospitalization. Inpatient, you are continuously monitored. People are near you at all times. You are on new medications. You are doing physical therapy that is hard, and you are supported through it. Then you go home, and you are asked to be independent, and to remember everything you were told while you were medicated and exhausted. We try to manage that across primary care, care management, pharmacy, specialty care, and insurance. Suddenly you are at home, not feeling well, leading the management of everything that was being managed for you a day earlier.

The opportunity right in front of us is handing the continuous part of that inpatient experience to an AI member of the care team. I want to build this because I have seen how disjointed those systems are. I have tried more than once to map and improve post-discharge care. Every time, the limiting factor has been our ability to close the gaps between manual processes managed across multiple, siloed teams.

The end of life is a gradient

I have spent time in the parts of medicine where this model matters most: a fragility fracture program, building a post-acute care partner network, running an inpatient and outpatient palliative program. What I took from all of it is that our systems are built around heroics and the preservation of life, by which I mean the prevention of death rather than the elevation of the time remaining.

The statistic everyone reaches for is that about a quarter of Medicare spending happens in the last year of life, usually offered as evidence of waste. It is worth knowing what the research actually found when it looked closely. A team at Stanford, MIT, and Harvard built a mortality-prediction model on Medicare claims and asked how spending is distributed by predicted risk of death rather than by what happened. Death turned out to be highly unpredictable. Less than five percent of spending went to people whose predicted one-year mortality was above fifty percent1. Even at the moment of hospital admission, the ninety-fifth percentile of death probability was around two-thirds.

Read that carefully, because it cuts against the easy conclusion. We are not, for the most part, spending heavily on people we know are dying. We are spending on people who are sick, some of whom recover and some of whom don't, and we usually cannot tell which in advance.

Which means you cannot fix end-of-life care at the end of life. There is no moment where the system reliably knows to switch modes. The only place left to work is the trajectory.

That is what the three clocks offer here. Not a switch from curative to palliative, but a continuous gradient governed by goals, functional trajectory, and disease burden. Clock two's objective function shifts progressively: from numbers to function, from optimization to shared plan. Care teams work with patients and families to define what they want to keep being able to do, what makes them happy, and what a dignified ending looks like.

What I would like to see is that when someone begins their Medicare journey, we start talking about goals early. Prevention and chronic disease management are best in class. End-of-life wishes are part of every visit. Frequent, additive, focused on maintaining physical and cognitive function, and involving family. The optimistic take on AI is that for the first time, we could have systems that bridge the space between those visits, so that what someone said mattered to them stays present.

Two questions worth asking yourself:

“What are the things I want to spend my time on?”

“How would I build a health system that helps me get there?”

The learning layer

I have talked about three clocks, and there is arguably a fourth. It does not run on the patient's timeline at all. It runs underneath, continuously, and it is what makes the whole thing tick.

The orchestration layer moves information between the clocks and manages the team of agents. The learning layer watches the orchestration layer.

AI in healthcare is a new frontier, and in a risk-averse industry there is real gravity toward tried and true technologies. Each health entity needs to start to get their feet wet because AI requires entirely new competencies and first-person learning. The near-term work is unglamorous and specific: following Gall’s Law³ we have to start building the sub-components. Simple RAG chat agents, agents that are good at multimodal interaction with people, ontologies, agent graphs, librarian agents, continuous monitoring and escalation agents. With each we get better at how we build them and we’ll start to understand how to pull them together into an organized team.

Four things the learning layer has to do:

Modularity. Every agent, model, and harness has a shelf life. The system should be designed so that components can be replaced without rebuilding around them, and so that the learning layer helps specify the successor.

Continuous evaluation. Watching for drift in agent behavior, clinical error, outcomes, and bias. Not an annual review. A standing function.

Experience. How this is actually going, for patients and for the care team. Whether the balance is right. This is the measure most likely to be skipped and most likely to matter.

Security, escalation, transparency, and discoverability. When something goes wrong, we need to be able to find the root cause inside the system. That is an architectural requirement, not a policy statement, and it is worth noting that the emerging standards for connecting agents to health data handle access but not accountability⁴. Compliance remains something you build, not something a protocol grants.

Closing

This has of course been a theoretical discussion, one approach that comes out of years of process improvement work and operational leadership inside good health systems. And with that what strikes me, over and over, is how often we solve the same problems. The same gaps show up for patients. The same items sit on the experience survey year after year: whether the team worked well together, whether my physician listened to me. We keep reinventing the same wheel to manage the misses between manual systems.

As we work with each new AI model and each new harness, doors keep opening that were not there before. There is an Ethan Mollick line I will reiterate here: today's model is the worst one you will ever work with⁵. We’re seeing the truth in those words revealed over months not years, creating an unusual moment. A chance to define something rather than inherit it.

The three clocks are one way to think about a different future. They are not the only way, nor likely the one we will land on. If I had to name the single most buildable thing in all of it, though, it would not be the trajectory record. It would be the smaller, faster one underneath — a list of what is open, with owners and due dates, that follows a person across every place they get care. I am putting it out here to spark the adjacent possible.

So I would like to know what you would build. What model would make healthcare work better for you, for patients, for the people on your teams? And more than that: how do we use this technology not to replace what we have, and not simply to get more out of the resources we already stretch, but to seek new boundaries and address the needs we have been placing in the hands of others?