digital health

The Last Mile to Clinic: Why Democratized Software Means Something Different in Healthcare

~ cmonsen ~ Leave a comment

Last Friday, as I left the house for clinic, I dictated two prompts into my phone as an experiment.

The first:

In a single shot, provide me a dashboard of the most recent Massachusetts Health Policy Commission total cost of care data.

The second, verbatim, ums included:

In a single shot, create a Street Fighter game that I can play on my phone that has Elon Musk versus, uh, Mark Zuckerberg and maybe a few other Titans of Industry as selectable characters. Should be single player, and it will randomly select a opponent, um, kind of, uh, yeah. I I don’t care too much about the details, uh, but maybe mainly emphasizing that it… I… yeah. Don’t ask for additional design decisions. Just make them.

(Yes, years of experience in clinical dictation, on full display.)

Both were done before I parked. The dashboard quietly corrected my agency mix-up — the Health Policy Commission sets the benchmark; CHIA publishes the data — and tracked down CHIA’s March 2026 Annual Report, the freshest official read on Massachusetts health spending, assembling the CY2024 story into a single self-contained interactive page: total health care expenditures of $83.3 billion, $11,663 per resident, growth of 5.7% against a 3.6% benchmark, the fourth consecutive year above target. The payer detail was in there too — commercial spending up 10.8%, MassHealth per-member costs up 20.4% as eligibility redeterminations shrank membership and left a sicker risk pool.

The game agent imposed its own guardrails, declining to build a game where real people beat each other up. It invented four parody tycoons — a rocket mogul with a flaming dash, a hoodie-clad algorithm lord who throws spinning “like” projectiles — wired up touch controls, health meters, and a round timer, and handed me an arcade fighter that runs in my phone’s browser.

On paper, these are deliverables from two different professions. The dashboard is business intelligence — in a health system, the work of an analytics team: source data, KPIs, benchmark tables, a citation trail. The game is software development — a game loop, collision detection, input handling, character state machines, the work of engineers. Two realms of expertise, two backlogs, two degree paths — and they collapsed into the same act: a paragraph of dictated intent, a single HTML file, less time than it takes to drive ten miles. That collapse has got me thinking.

The prompt is not a spec

That second prompt is no one’s idea of a requirements document. It’s a transcript of a person thinking out loud on a commute, false starts and all, and its only unambiguous sentence — “Don’t ask for additional design decisions. Just make them” — hands over the specification itself.

In early 2025, Andrej Karpathy gave this workflow a name — vibe coding: give in to the vibes, “forget that the code even exists.” It’s easy to hear that as flippancy, but it is really about abstraction. Code hasn’t gone away — both of my artifacts are hundreds of lines of JavaScript that I will probably never read — but it now sits where assembly language has sat since the compiler arrived: a layer below where the author works. I supplied intent; the rest was translation.

For two decades, business intelligence promised something it called self-service: semantic layers, drag-and-drop chart builders, governed data marts. What it mostly democratized was consumption — anyone could view the dashboard someone else built. Creation still lived behind a queue: file the ticket, wait for the analyst, review, revise, wait again. The unit of BI delivery has always been the request — and it’s the queue, more than the craft, that this new way of working removes. Applications have their own version of the gate: the cost of design, development, and distribution has meant software gets written only when enough people share a problem to justify the investment. Both gates crumbled in the same ten-mile drive.

Yes — the demo is the easy half

Anyone who has run an analytics shop will object here, and the objection is fair. A single-shot dashboard has no refresh schedule, no lineage, no access controls, no agreed definitions behind its numbers. Mine froze the moment it was built and is already aging toward obsolescence. Healthcare data punishes casual handling more than most domains, and a wrong number in a benchmark hearing matters.

But the objection, taken seriously, concedes the interesting point: everything still on the worry list is about trust and maintenance rather than construction. Building the thing has dropped off the list of scarce resources.

The most instructive part of the dashboard, in fact, was its caveat. Unprompted, it flagged that one historical figure — the 2018 growth rate — came from prior-year reporting rather than the current chart, and was worth verifying against the source databook before citing. Building the artifact took minutes; deciding whether to trust it is the work that remains. Whether MassHealth per-member growth of 20.4% is meaningfully comparable across a redetermination year is not a question about pixels. It’s a question about definitions, denominators, and context — and settling questions like that, not rendering charts, is where BI budgets actually go.

The video game, meanwhile, made a judgment of its own: it declined to stage fights between real, named people and invented parody characters instead. Software embodies opinions — I wrote that about the EMR a decade ago — and now the thing that writes the software arrives with opinions too. Both artifacts came back with values baked in — one about data provenance, one about real persons — neither of which I asked for. Are they the right values? That’s a far harder question than whether the code runs. I think, perhaps, my new game lost some of its edge (at least until its new character — Rex Orbital (haha, Elon Musk, get it?) — becomes a household name).

What this means for healthcare IT

So if the ticket backlog were to disappear — and the cost gate on applications with it — what gets unlocked? I am sure there are industries where the answer is simple: abundance. Healthcare’s answer, though, runs through what telecom engineers call the last mile: the final connection to the home was always the expensive one, long after the long-haul lines got cheap. Our long-haul to the clinic just got cheap. In 2009, Mandl and Kohane argued in the New England Journal of Medicine that health IT needed substitutable applications — an iPhone-style app economy where a clinician could download a tool, reject it, and try another. Seventeen years on, that vision arrived only partially through EHR app stores and APIs. The gap remains wide enough that even today, the 2021 call to action I made alongside smart collaborators to enable integrating web services and applications into EHRs is still an opportunity. We are getting closer.

It seems that the rest of the vision will arrive through prompts and agents. Robin Sloan once wrote that an app can be a home-cooked meal — software made for a household, with no ROI and no roadmap. Maggie Appleton extended the idea to barefoot developers: domain experts building local software for communities that the industry will never serve. Healthcare is full of these unserved communities. The nurse manager’s staffing grid. The pharmacist’s prior-auth tracker. The quality team’s registry view that exists today as a heroic spreadsheet (v27_FINAL_final.xlsx). These are the small data tools — operationally vital, too small to justify taking a project number from the queue, which is precisely why they never got built. I am already living the home-cooked version of this: when I couldn’t buy the household assistant I wanted, I built it. These two took a commute. How many commutes are there in healthcare? How many to a clinic still thirsting for easier access to technology & data solutions?

The implication is uncomfortable for those of us who steward health systems’ information environments. Our de facto governance model for software creation has been scarcity: if building required developers, and developers required budget, then “who may build” was controlled by default, and the backlog could pass for a safeguard.

When anyone can build, “who may build” stops working as a control point. The real questions surface: What do we trust? How do we validate it? Where does it run, and what data can it see? What happens when the pharmacist’s tracker quietly becomes the source of truth for others as well? These are harder questions than triaging a ticket queue — but they are better ones, because they interrogate the actual risk instead of a proxy for it. A health system that answers them well gets something the request queue never delivered: the people who understand the problems helping build the solutions. Governance targeting is aimed more at meaning and trust rather than at construction.

On one drive to clinic, the dashboard and the video game cost me the same — a paragraph of dictated intent apiece — and both were waiting before I’d seen my first patient. But the ten miles I covered were never the hard part. The distance that matters in healthcare is the last mile, from a working artifact to the point of care, and it’s paved with exactly the things a prompt doesn’t generate: agreed definitions, validated numbers, workflows that carry a tool into the visit, and trust. Building was just the expensive part, and now it mostly isn’t. We can go farther faster, and — I hope — cast more of our attention to that last mile that still must be solved in carbon.

When You Needed to Pay Attention

~ cmonsen ~ Leave a comment

Since ChatGPT launched in November 2022, there have been hundreds of AI releases. If you tried to follow all of it, you drowned. If you ignored all of it, you missed something important.

I wanted to find the middle ground — to identify the moments where the capability ceiling seemed to jump, at least for me personally. These were genuine step changes that required me to update my mental model of what was possible and led to a lot of experimentation.

By my count, roughly eight qualify. ChatGPT made AI conversational. GPT-4V gave it eyes. GPT-4o made the interaction real-time and multimodal. o1 showed it could reason through multi-step problems. Deep Research meant it could investigate a question autonomously for hours. Claude Code meant it could write and maintain real software — not snippets, but systems. Gemini 3 commoditized reasoning. And Cowork made AI a genuine collaborator in professional-like workflows (I’m looking forward to when it’s available at work).

Reasonable people could draw the line differently. But when I laid them out on a timeline, the pattern that emerged is instructive: the gaps between step changes are compressing. Eleven months between the first two. Then seven. Then four. The most recent — two months. Part of what’s driving the acceleration is that there are more serious entrants now — OpenAI, Anthropic, Google, and others are effectively passing the ball to each other, each leap prompting the next within weeks. My own usage reflects this — I started with ChatGPT, moved to Gemini for a stretch, and now spend most of my time with Claude. I expect that will keep shifting as the capability lead changes hands.

In healthcare — where carbon is harder than silicon and the opinionated systems we built over the last decade are deeply entrenched — the compression is especially disorienting. We spent fifteen years deploying EMRs. AI is forcing us to reimagine what we do with them on a timeline measured in months, even if at the same time, things may move a little more slowly.

I don’t know how long the cadence keeps compressing or at what point it will feel that things have plateaued, but I hope the visualization helps to separate signal from noise.

Carbon Is Harder Than Silicon: Why the Future of AI Depends More on Human Systems Than Hardware

~ cmonsen ~ Leave a comment

The one about the socio-technical model applied to AI.

Much of today’s AI conversation centers on tangible constraints — chip shortages, compute capacity, and the growing energy demands of model training. Those are real limits, and they’ll shape what’s technically possible for years.

But anyone who has lived through an EHR rollout or a “clinical decision support” pilot knows that the harder part often begins after the technology is installed. Once the silicon is humming, we still have to integrate it into the living fabric of care — how clinicians think, how teams coordinate, how patients experience the system.

That’s where the carbon-based agents come in. Not as obstacles, but as the essential medium through which any digital innovation actually becomes care.

The Sittig and Singh (2010) socio-technical model still provides one of the clearest guides here. It lays out eight interdependent dimensions — from technical infrastructure and clinical content to workflow, organizational culture, and external environment. It’s a reminder that safety, quality, and adoption emerge not from the technology itself, but from how these layers interact in practice.

The same logic applies to AI. A model can perform flawlessly in validation but fail to add value at the bedside if it doesn’t align with clinical priorities, decision rhythms, or accountability structures. These systems don’t just plug into existing teams; they subtly reshape how roles are defined, how authority flows, and how judgment is shared.

So yes, the field will need more chips, more power, and more scalable infrastructure. But the real breakthroughs will come from designing AI that supports the people who deliver care and ultimately the carbon-based agents central to their mission: the patient.

Switches, Boxes, and Steam — Three Questions for AI in Healthcare

~ cmonsen ~ Leave a comment

The one where 3 books answer 3 questions about AI in healthcare

Every few years a new wave hits healthcare IT. Some reshape the shoreline; some barely ripple. Lately I’m leaning toward a simple view: Gen AI is a real step forward—but the changes that endure will come from infrastructure and incentives. In as regulated and high-attention a system as health care, the roads matter more than the horsepower.

1) Who will benefit?

History says advantage concentrates around control points—the places where compute, data, and distribution meet. In today’s terms, that means model providers with scale (e.g., OpenAI, Anthropic, Google/DeepMind), the clouds that host them (eg AWS, Azure, Google Cloud), platforms already embedded in clinical workflows (eg Epic, Oracle Health/Cerner), and companies that own the last mile to clinicians and patients (eg large health systems). As though they control the “master switch,” these players have significant influence in supporting winners and losers.

But the circle does widen. A second group tends to capture durable value: the people and teams who complement those control points—clinical data stewards, evaluation and safety engineers, product integrators who turn models into reliable steps inside prior auth, triage, charting, imaging, and revenue cycle. As individuals, entrepreneurs who see where the network is going (and get there early) tend to do well: they make the glue, the adapters, the “boring” parts that let many models work safely across many contexts and control points.

As general models compete and compute capital fuels greater availability, the answer to “who benefits” may not depend as much on who has the smartest model as we thought. In healthcare, the answer is probably closer to: “where do you sit relative to distribution and standards?”

2) How will they benefit?

Consider what happened when shipping settled on a standard metal box—a 20- or 40-foot container with identical corner fittings. That one decision let cranes, ships, trains, and trucks handle cargo without repacking. Risk fell. Costs fell 90%. And the work moved: less muscle on the pier; more planning and throughput management at distribution centers and rail yards. Jobs followed the flow.

AI may trace a similar trajectory. As models ship in consistent packages—stable interfaces and licenses, companion evidence (safety & efficacy, provenance, evaluation coverage, known hazards)—risk drops across the chain. This is a chain, though, of bits and not atoms. Workflow behavior—increasingly digital—and model swaps achieve more predictable outcomes. Capital becomes willing to fund further scale because components are modular, productized, and auditable.

As for the work, I expect this to move from first-pass busywork to the “inland” roles that plan, do, study, and act towards the learning healthcare system. Technical roles for platform/orchestration, evaluation & red-team, data lineage & governance, enablement & change will blossom. Existing roles will morph. For example, on the admin side, copy-paste, phone calls, and status-chasing give way to flow coordination, exception desks, audit/QA, and patient navigation (think schedulers → access-ops coordinators; coders → utilization & compliance analysts). On the clinical side, keystrokes give way to judgment—ambient draft review, rare-case adjudication and roster reviews, care-plan design, patient counseling, and system-of-care safety & stewardship. And on the patient/caregiver side, the role shifts from passive data source to co-steward of context; from recipient of transactions to more of a navigator. There will be more need for controlling consent and sharing, supplying high-signal inputs (PROMs [patient-reported outcomes], home-device streams, life context, and values), correcting records and attaching verifiable documents, flagging errors or preference mismatches, and (for caregivers) supporting therapy reconciliation, adherence, and follow-up.

3) What is our responsibility?

I see two obligations emerging from this and running in parallel.

First: transparency as a design choice. The industrial age didn’t compound due to the steam engine alone; it compounded because we paid for disclosure. Blueprints were made public, and then builders turned them into businesses. In AI, the equivalent is releasing portable, trustworthy manifests with every meaningful update—lineage, test coverage, failure modes, guardrails—so others can evaluate, integrate, insure, and, when appropriate, improve. Procurement and reimbursement will prefer systems that come with real evidence as much as this has become the case for conventional therapeutics (ie medications) in trial and pharmacovigilance.

Second: we have to uplift the people, not just the pipes. Standards don’t only move information; they move jobs. Containerization made ports safer and faster, but it also displaced longshoremen and pushed opportunity inland. Healthcare will feel a similar migration as routine drafting and triage shrink. We will move only as fast as we develop the workforce a glide path. This looks like making the new roles visible; creating portable credentials for evaluation, operations, governance, and enablement; and retraining with intentionality. We won’t succeed if we think the players are frozen. We need to help the team skate to where the puck is going.

Bringing it together

If advantage tends to form at the control switches, and if standards are what turn demos into networks, then the next phase winners are 1. the builders who make AI reliable, swappable, and evidenced and 2. the organizations that invest in the people who run that learning system well. Gen AI “thoughtpower”—like the horsepower that came before—opens the door. The plumbing and the social compact (transparency + worker mobility) portends our trip through it.

Want to go deeper?

These ideas are an amalgam of a few books that have been highly influential in my own thinking. Please let me know if you have encountered others for this collection!

Tim Wu, The Master Switch — Why open eras often consolidate around control points, and what that means for innovation and competition.

Marc Levinson, The Box — How a universal container standard reshaped costs, jobs, and geography—useful for thinking about AI packaging and “inland” roles.

William Rosen, The Most Powerful Idea in the World — The case that incentives for disclosure (the patent bargain) made progress compound—and why entrepreneurs matter for carrying blueprints into the world.

Digital Health: “What’s taking so long?” Part 2 of N

~ cmonsen ~ Leave a comment

For all the ways that technology has visibly transformed our lives as consumers over the last decade, it has seemed like just a matter of time before the excitement of big data, social, local, mobile, process automation, artificial intelligence, and blockchain (nb. use of buzzwords intentional) will make their way into helping us meet the aims of precision medicine and population health. Though I am quite convinced that health care as an industry can be one of the most rapidly changing, I think it is fair to say that the health care consumer (ie patient) experience has remained fundamentally unchanged during this period. It feels, if anything, that the gap is only getting wider. What’s taking so long?

Continue reading →

Herald: Solving the Data Overload Problem for Doctors

~ cmonsen ~ Leave a comment

Problem

Logo_3x2Doctors are overwhelmed with data. They spend 12% of their time looking up clinical data when they could be seeing patients and still information gets missed. In fact, the IOM has identified untimely access to clinical data as a leading contributor to the 3rd leading cause of death in the US: medical errors. Existing information systems and electronic medical records are better optimized for billing and documentation than they
are for making care safer. There has to be a better way.

Continue reading →

Reach: Text-Message Appointment Reminder System

~ cmonsen ~ Leave a comment

Problem

reach exampleIn aggregate, community health centers account for the care of about 20 million people in the US. Over half of these patients represent racial or ethnic minorities and over a fifth (22%) prefer to speak Spanish rather than English.

Most CHC revenue comes from fee-for-service reimbursement paid by Medicaid (40%), private payers (7%), and Medicare (6%). This has led CHCs to pursue many of the strategies for maintaining solvency as other care centers across the US, including increasing patient visit volume and improving operational efficiency.

One problem all clinic sites face is the incidence of no-shows, patients for which an appointment is scheduled but that do not show up. It is estimated that no-shows account for 5-30% of appointments scheduled across the US and it is typically higher at CHCs. No-shows risk failing to deliver appropriate care to patients for whom they are scheduled in a timely or continuous manner, reduce access to scarce healthcare resources for those waiting for appointments, and represent up to 15% of lost revenue for the clinic.

Continue reading →

Seeking a diagnosis on the Internet: survey results

~ cmonsen ~ Leave a comment

Testing design assumptions with users is a critical ingredient in user-centered design. In Symcat’s early stages (ca 2012), we thought, for better or worse, that we would identify some eligible test users through Craigslist NYC. We were surprised by just how many people were willing to participate and collected some pretty interesting data in the process. I just stumbled upon it and I suspect much of it is still relevant, so I thought I would share. Get ready for some graphs.

Continue reading →

What Does the Massive Health Acquisition Really Mean?

~ cmonsen ~ Leave a comment

If you’re like me, you’re extremely excited about the potential for design to reshape health care. I’m not talking health care system redesign (ACOs and such–though that’s great too), I’m talking about the type of design you see on Dribbble: the focus of a recent (awesome) HHS-sponsored competition.

One of the promising upstarts of health care re-design was a 2-year-old-or-so startup called Massive Health founded by ex-Mozillite Aza Raskin. Though I tend towards the skeptical, there was a part of me that thought that not only were they on to something, but they clearly had managed to aggregate real design talent. And in health care, no less! Apparently, I was not the only one as they convinced a number of investors to throw $2.25 million in to test out what they could do.

Continue reading What Does the Massive Health Acquisition Really Mean? on the Symcat blog.

Blueprint Health Startup Accelerator: Was it Worth It?

~ cmonsen ~ Leave a comment

blueprint-cover-page-david-craigAs exciting as the digital health space is right now, there is still little guidance or validated path to getting off the ground. As part of an effort to help aspiring health care entrepreneurs, I’ll be writing a series of posts explaining some of the decisions we made for Symcat. It hasn’t been a year since we’ve started, but my hope is that our few months of experience can help those who are just getting started themselves.

One of the questions I’m most frequently asked is if our time at Blueprint Health, a health start-up accelerator, was worth it. To participate, the program requires 3 months of relocation to the NYC offices in SoHo and the forfeiture of a nearly 6% equity stake in the company. The program basically offers $20k, mentorship from its network, and office space. A few other health start-up accelerators (ie Rock Health, Healthbox) have some variations but basically the same theme. They are all very selective accepting 3-5% of applicants. While it’s nice to be accepted, there’s still the important matter of deciding if it is right for you.

Continue reading Blueprint Health Startup Accelerator: Was it Worth it? on the Symcat blog.