A forty-year-old problem,
briefly available.

A pharmacist I know told me last summer that she has not taken a real lunch break in three years. The reason wasn’t that her store was busy, though it was. The reason was prior authorization.

Prior authorization is what happens after a doctor prescribes a drug and before the insurance company will pay for it. In theory it’s a cost control. In practice it’s a forty-year-old paperwork system that costs the US about $35 billion a year, takes 45 minutes per submission of skilled labor that’s already in short supply, and delays care for millions of people, some of whom die waiting. The numbers aren’t disputed. The AMA has been publishing them for years. Ninety-four percent of physicians say it delays care. Everyone in healthcare agrees it’s a problem.

And yet for forty years no one has fixed it.

That’s the part that interested me. When a problem is this expensive, this obvious, and this old, the explanation is almost never that no one has tried. Lots of people have tried. Clearinghouses tried in the nineties. EHR vendors tried in the two-thousands. There’s a wave of AI startups trying right now. The explanation has to be something else.

I think the explanation is simpler than it looks. Prior authorization is unpleasant work in five different directions at once, and almost everyone who’s tried to fix it has shrunk from the parts that matter and gone after only the parts they were willing to touch.

You can see this in what’s been built. The clearinghouses route paperwork without reading it. The AI tools read clinical notes but won’t submit anything to a payer. The EHR modules integrate but don’t automate. The fax services deliver but don’t close the loop. None of these is a bad product. They’re all reasonable responses to an unreasonable problem. But none of them actually does the thing the pharmacist needs done, which is to take a prescription and a patient chart and an insurance card and an unanswered fax sitting on her counter and turn them into an approved authorization without involving her at all.

The reason no one has built that is that building it requires doing five hard things at once. You have to read clinical notes well enough that a pharmacist will trust your answers. You have to pull data from systems designed never to talk to each other. You have to make phone calls to PBMs still running on IVR systems from 1995. You have to be HIPAA-compliant from the first commit. And you have to know, for each combination of payer and drug and plan, which channel is most likely to actually result in an approval this week.

Any one of these is a project. All five together is a company. Most people who look at the problem pick one of the five and hope the others somehow resolve themselves. They never do. In a regulated workflow, a half-solution doesn’t reduce cost. It adds a new layer of it.

You might ask, fairly, why this is suddenly possible if it wasn’t before. Three things have changed.

The first is that LLMs got good enough, in the last eighteen months or so, to read a clinical note and produce an answer to a payer’s question that a pharmacist can verify in seconds rather than minutes. This sounds like a small thing and it’s a very large one. Before this, every attempt to automate the reading part produced answers no one could trust. After this, the bottleneck moves.

The second is that voice synthesis got good enough, around the same time, that you can call a PBM, navigate their phone tree, talk to a human agent, and gather what you need without the agent realizing they aren’t talking to a person. I’m not sure how to feel about this in general. Voice agents indistinguishable from real people are a thing the world is going to have to figure out, and I don’t think the answer is obvious. But for prior authorization specifically, it removes the last excuse for not closing the loop.

The third is regulatory. CMS finalized a rule in early 2024 that requires payers to support electronic prior authorization through a FHIR-based standard, and it takes effect in January 2027. Every payer in the country is rebuilding their PA infrastructure right now. A follow-up rule extending the mandate to drugs is already proposed. This means that during the window in which I’m building, the entire industry is also being rebuilt — and the new infrastructure will be the kind a startup can plug into instead of work around.

None of these three by itself would be enough. You could read clinical notes well in 2024 and still not be able to submit anything. You could submit electronically in 2027 and still not be able to read the notes. You could have both and still not have a regulatory environment that rewards you for showing up. It’s the combination that makes this possible, and the combination is new.

The unpleasantness is, I think, the main reason this is still available. It has kept most ambitious people away from the problem for a long time. When a smart engineer thinks about healthcare startups, they tend to think about consumer wellness apps, or AI scribes, or something diagnostic. They don’t think about prior authorization, because prior authorization sounds like something you’d do if you couldn’t think of anything better.

But the things that sound boring are often the things that matter most. Look at what happened with online payments before Stripe. For years, every developer who’d ever tried to take a credit card on a website knew the experience was miserable — merchant accounts that took weeks to open, gateway integrations that read like they’d been written in a different decade because they had, fraud rules that changed without notice. Thousands of developers had hit that wall. None of them wanted to be the ones to fix it, because fixing it meant a long walk through banking relationships, fraud systems, and financial regulation. So the field sat there, obviously broken, for years. When the Collison brothers finally took it on in 2010, they had almost no competition for one of the most valuable problems on the internet.

Prior authorization is in roughly that position now. It is one of the largest sources of administrative waste in American healthcare, which makes it one of the largest sources of administrative waste anywhere. The companies that have tried to fix it have either gone after the wrong layer or stopped short of the part that actually changes things. The big healthcare companies don’t want to disrupt their own clearinghouse revenue. The big tech companies aren’t going to spend a decade getting SOC 2 Type II so they can sell $5-per-PA software to a pharmacy in Iowa. And until very recently, the technology to do it right didn’t exist.

So the field is, briefly, empty. That won’t last.

ParaHealth started a year ago as a tool that could read a patient’s chart and fill out a Zepbound prior authorization form. Zepbound, if you don’t know it, is one of the GLP-1 weight-loss drugs. Almost every commercial insurance plan requires prior authorization for it, and the form has 166 questions. A pharmacist or medical assistant who fills one of those out by hand spends thirty to forty-five minutes on it. Our first version did it in about four. The answers were good enough that the pharmacist who tried it said, in a tone I’ve thought about many times since, “wait, that’s it?”

That was the moment I realized this was a real company.

Since then it’s grown into something larger, but the principle hasn’t changed. The product reads the chart, fills out the form, calls the PBM if it has to, submits through whichever channel is most likely to work, tracks what happens, and writes the appeal if the answer comes back wrong. Each of those is a piece of work that someone in a pharmacy or a doctor’s office is doing right now, by hand, and would much rather not be.

We charge per authorization. Five to fifteen dollars depending on complexity. The first ten are free. Per-transaction pricing means we only make money when we save someone time, which is the only honest way to charge for a thing like this.

The beachhead is GLP-1s, because that’s where the volume and the pain are both highest right now. After that, the same architecture works for specialty biologics, oncology, cardiology, mental health, and basically any other drug class with the same broken process behind it. The reason this generalizes is not that the product is general but that the problem is. Once you’ve built the thing that turns a chart into an approved authorization for one drug class, the work of doing it for another is mostly question sets and prompt tuning.

I don’t want to oversell what’s there. We have a working product, real interest from real pharmacies, and a roadmap I think is right. We do not yet have the data moat that comes from processing a million authorizations, or the EHR integrations that come from getting through Epic’s app review, or the SOC 2 certificate. Those are coming. They take time. The thing that takes time more than anything else is just doing the work, in the order it has to be done in, without skipping any of it.

Prior authorization is the wedge, not the ceiling. The reason I’m planning to spend the next ten years on this is that the same pattern that makes prior authorization broken makes a dozen other things in American healthcare broken, and most of them are broken for the same five reasons. Eligibility verification. Benefits determination. Referrals. Claims. Appeals. They are all forms of the same problem: a clinical decision on one side, a patient on the other, and a fog of forty-year-old friction in between.

If you build the layer underneath that solves prior authorization properly — the reading, the calling, the submitting, the tracking, the audit trail — that same layer can solve the other ones, because the primitives are the same. So the company that gets prior authorization right is, in effect, building the operating system for the administrative side of American healthcare. Most people don’t think of that as a category yet. I think in five years they will.

But I want to be careful about that. The big mistake people make with ambitious ideas is to lead with the ambition. The pattern that actually seems to work, when you look at the companies that have done the largest things, is to start with something small and specific that obviously works, and let the larger thing become visible over time. Amazon started by selling only books. Airbnb started with three air mattresses on the founders’ apartment floor. Shopify started as the checkout for a single snowboard shop. The way to build a big thing is not to build a big thing. It’s to build a small thing that is the right small thing, and to keep going.

That’s what I’m trying to do. Right now, the small specific thing is helping pharmacists get GLP-1 prior authorizations done in four minutes instead of forty. People like the pharmacist I mentioned at the beginning are who I’m building this for. She still hasn’t taken a proper lunch break. That’s the fight.