AI is everywhere. So why isn't it where it's needed most?

ChatGPT was released in November 2022. By the time you read this, we are about three and a half years into the GenAI era. Three and a half years is enough time to build a company, a product, and a category. Plenty of teams have done exactly that.

If you spend any time at HLTH, ViVE, or HIMSS, you already know what the loud version of the AI story sounds like. AI-native scribes. AI-native call centers. AI-native prior auth. AI-native everything. The companies on those stages are mostly founded after 2022, built around models from day one, and pitched to investors as "we couldn't have existed two years ago."

That's a real and exciting part of the story. But it isn't the whole story, and honestly it isn't even the most interesting part for me.

The more interesting question is the one you don't usually hear on a panel: what about everyone else?

What about the digital health products that already have thousands of users, a board, a roadmap, and a compliance posture they cannot break? What about the legacy operations sitting under every clinic in the country, where someone is still printing a referral and feeding it into a fax machine in 2026?

That is where most of healthcare actually lives. And it is where most of the AI conversation isn't.

Table of Contents

1. The fax machine is not a joke
2. The two stories
3. The misconception I keep running into
4. What "already in production" actually changes
5. The other half: workflows nobody is celebrating
6. Replacing work, not people
7. Where the two stories converge
8. A short, opinionated checklist
9. A closing note

The fax machine is not a joke

In US healthcare, roughly 70% of providers still rely on fax to exchange medical information as of the most recent industry surveys. Some studies put it higher when you include EHR-linked faxing. Whatever the exact number, the order of magnitude is clear: the dominant mode of clinical document exchange in 2026 is technology from the 1980s.

I bring this up not to dunk on the fax. The fax persists for real reasons (HIPAA-friendliness, universality, no dependency on the other side's stack). I bring it up because it is the cleanest possible picture of the gap I want to talk about.

In the same building where someone is faxing a referral, there is probably a screen showing an AI-native scribe pilot. Those two things coexist. They don't talk to each other. And the second one tends to get all the airtime.

The two stories

I split my week, more or less, between two kinds of conversations.

Conversation one is with founders and product leaders inside digital health companies. Their product already exists. It already has paying users. They have a board, a roadmap, a security review process, and a clinical advisory group. Now they need to figure out where AI fits, without breaking what works.

Conversation two is with operators inside clinics, practices, and health systems. They don't sell digital products. They run one. Their pain isn't a feature gap. Their pain is twenty-three browser tabs, a fax inbox, a billing queue, and a staff member doing prior auths by hand at 8pm.

These two conversations look different. The stakeholders are different. The risk profile is different. The vocabulary is different.

But the underlying question is the same: how do you bring AI into something that already exists, without pretending the existing thing doesn't matter?

The AI-native crowd doesn't have to answer this question. They get to start clean. Everyone else has to answer it every single day.

The misconception I keep running into

When I talk to non-AI-native teams, the misconception isn't usually "AI is too risky" or "AI is overhyped." Both of those exist, but they aren't the main blocker.

The main blocker is a misread of what the work actually is.

Teams come in assuming the hard part of AI integration is the model. Which model, which version, which provider, which benchmark. They want to debate GPT-5.5 vs Claude vs Gemini for six weeks before doing anything.

The model is not the hard part anymore. It hasn't been for a while.

The hard part is everything around the model. The data plumbing. The compliance posture. The trust model with end users. The fallback when the model is wrong. The audit trail for the regulator. The change management for the team that has to live with it. The integration with the EHR that does not want to be integrated with. The governance conversation with the board that has read three articles about AI hallucinations.

I wrote about this in the CompliantChatGPT post a year back: the value isn't "we have an LLM." Everyone has an LLM. The value is the system you build around it so a clinician can actually use it without breaking HIPAA, breaking workflow, or breaking trust.

For a product already in production, this is a particularly sharp problem. You are not building a greenfield AI feature. You are inserting AI into an existing user flow, in front of users who already have expectations, on top of a codebase that was not designed for it. That is a very different exercise than "let's ship an AI assistant."

What "already in production" actually changes

A few things change the moment you have real users.

You can't break trust. New AI-native products get to set expectations from day one. Existing products inherit trust they have to protect. A hallucination from your three-month-old AI startup is a bug. The same hallucination from a tool a clinician has used for four years is a betrayal.

Stakeholders multiply. A startup has a CEO and an engineering lead. An existing company has a board, a CISO, a clinical advisory group, a head of customer success who knows every account that will be upset, and a legal team that has been on edge since the start of 2023. None of them can be skipped. None of them should be skipped.

The integration surface is real. Greenfield products integrate AI into a clean architecture. Existing products integrate AI into whatever was there before. That "before" usually includes at least one system the original engineers don't fully understand anymore.

Pilots are real money, not slide money. When AI shows up inside an existing product, you can measure the delta against the previous experience. That's a gift, because you can actually prove value. It's also a curse, because if the new version is worse on any dimension users care about, they will notice immediately.

None of this is a reason not to do it. It is a reason to be honest about what the work is.

The other half: workflows nobody is celebrating

Now let's talk about the other half of my week.

Inside almost every clinic, practice group, and operations team I see, there is a layer of manual work that has nothing to do with software product strategy. It is the back office. The paperwork. The phone calls. The faxes. The intake forms that get retyped into three systems. The prior auth that takes a human two hours and produces a one-line answer.

This work is not glamorous. It does not show up in the keynote deck. But it is enormous in aggregate, and it is the largest pool of clearly automatable work in healthcare.

A few numbers that I find hard to ignore:

• The AMA's 2024 data has US physicians working a 57.8-hour week, of which 13 hours are indirect patient care (documentation, order entry, results review, referrals) and 7.3 hours are pure administrative tasks. That's more than 20 hours a week per physician of work that is not, in any meaningful sense, the practice of medicine.

• One 2024–2025 Google Cloud / Harris Poll survey, put clinician administrative load at about 28 hours per week, with office and claims staff spending even more.

• Physician burnout has come down from its pandemic peak, but 42% of physicians still reported at least one burnout symptom in 2025, and the consistent finding across multiple studies is that the dominant drivers are administrative, not clinical.

• The economic cost of physician burnout in the US is estimated at around $4.6 billion per year, through turnover, lost productivity, and errors.

Layer on top of this that healthcare is now about 18% of US GDP, projected to reach 20% by 2033, and that productivity growth in the sector has been persistently sluggish , compared to almost every other industry. The biggest, most expensive, most labor-intensive sector of the economy is also one of the worst at converting human effort into output.

This is the actual market. Automating, augmenting, and removing the operational drag that everyone in healthcare quietly accepts as the cost of doing business.

And it is the part most legacy organizations (clinics, payer ops teams, mid-sized digital health companies that grew before AI was credible) are least equipped to tackle on their own.

Replacing work, not people

Here is where I want to be careful, because the framing matters. When I talk to operators about AI workflow optimization, the first question is almost always some version of: "Are we going to lay people off?"

The honest answer, in most of the cases I see, is no, and the reason is structural. US healthcare has a chronic labor shortage, an aging population driving demand, and a productivity problem. There is more work than there are people to do it. The bottleneck isn't headcount. The bottleneck is what each person spends their time on.

The clinics I’ve talked to don't have a "too many medical assistants" problem. They have a "our medical assistants spend half their day on tasks a piece of software could do" problem. Take that half back, and the same team can serve more patients, give more attention to the hard cases, and stop burning out.

That is the right framing. AI workflow optimization gives operations teams their time back. The repetitive work no one signed up for, like the phone tag, the retyping, the faxing, or the form that exists only because it always has, gets handled by automation. The judgment calls, the strategic thinking, and the human problem-solving stay with the humans.

The same logic applies inside digital health products. We are not replacing the clinician using your platform. We are removing the click-heavy, friction-heavy parts of using it so the clinician can do what the platform was originally supposed to help with.

Where the two stories converge

Step back and the two halves of my week start to look like the same problem.

The digital health product already in production needs AI to integrate without breaking what works. The clinic operations team needs AI to remove manual work without breaking what works. In both cases, the existing system has earned the right to be respected. In both cases, the AI is not the centerpiece. The system is.

This is what I think most people miss when they look at AI in healthcare in 2026. The interesting work is not on the demo stage. The interesting work is the careful, unglamorous integration of AI into things that already exist, by teams who already have users, already have stakeholders, already have a HIPAA posture, and already have a fax machine somewhere in the building.

It is slower than building from scratch. It is also more durable, because the system around the AI is what makes the AI useful, and the system was already there.

A short, opinionated checklist

If you are inside a healthcare organization (digital health company, clinic, ops team) and trying to figure out where AI actually goes, a short version of how I think about it:

1. Find the work that is repeated, structured, and currently done by a person under time pressure. That is your starting line. Documentation. Intake. Prior auth drafting. Pre-charting. Inbox triage. Refill workflows.
2. Don't start with the model. Start with the workflow. Map the current state honestly, including the parts nobody likes to admit.
3. Decide what "good" looks like before you build. Time saved per task. Error rate vs the human baseline. User satisfaction. Pick two or three and measure them.
4. Keep a human in the loop, on purpose, where it matters. Clinical decisions, anything that touches PHI in a new way, anything that goes out to a patient. The point isn't that AI can't do it. The point is that trust is earned slowly and lost quickly.
5. Plan for the boring stuff. Audit logs. BAAs. Role-based access. Rollback. The unsexy compliance work is most of the work.
6. Pilot small, measure honestly, expand only when the numbers and the users agree. Internal teams who use the tool every day are a better signal than any metric.

A closing note

If I had to summarize the thesis of this post in one line, it would be this: the AI conversation has been dominated by the products that started with AI, and the more interesting question is what happens to everything else.

Everything else is most of healthcare. Most of the users. Most of the spending. Most of the operational drag. And, for the next few years, most of the value is still on the table.

Next week I'll be in NYC Tech Week on a panel, talking about exactly this: how AI gets integrated into products when it already exists, what trust actually looks like in production, and where the human keeps belonging. If you are in town, please come say hi.

And if you are inside a digital health product, clinic, or operations team and any of the above sounds like your week: that's the work we do at Light-it, both on the product side and now as a dedicated AI workflow optimization service.