In order to moderate (moderately) well during tonight’s panel, I’ve been trying to wrap my head around how AI is being deployed in healthtech. Anecdotally, I’ve noticed a pretty broad pattern: People in the culture industry loathe AI; people in tech are excited but paranoid; and people in healthcare are like, Give it here I want it right now. No, more. Now. More. This includes my GP, who was grumpy that he can’t use ChatGPT at work and instead has to use the lousy AI they make available inside his hospital system. 

So do doctors want to be replaced by AI? Not really. If you’re under 40 you may not know that doctors used to love their jobs, to the point of unbearable smugness. They hardly ever complained except about golf. Now they’re miserable drudges who fill out forms. They call this “pajama time” because you’re doing it in your jammies after a day of seeing the worst things imaginable at the office. Google, for its part, provided what I believe to be one of its all-time dumbest AI overviews on the subject, something so stupid I needed to include it here in full:

“Pajama time” refers to two main things: a popular children’s bedtime book and song by Sandra Boynton (and related music), and a colloquial term for the excessive administrative work physicians do at home in their pajamas, contributing to burnout. Both involve wearing pajamas, but one is a fun bedtime routine for kids, while the other highlights a serious issue in medicine.

God I hope we survive this era. Anyway. There are tons of exciting, gee-whiz things going on in healthtech—AI-assisted drug discovery, robot-assisted gallbladder surgery, exoskeleton shoes and personal airbags. Some of the work being done in cancer detection is thrilling, and will pay off for decades.

But to me, the really, truly exciting stuff turns out to be…paperwork reduction. One term to look for is “ambient documentation,” which is sort of transcription-plus; products from companies like Snowflake promise to save you hours of pajama time. This is very interesting to me, because most (non-medical) AI products don’t really make concrete promises like this; they just sort of slop AI around on the floor with a big digital mop and say, “Looks clean to me!” AI in healthtech is mature enough that “ambient documentation” is old news, and you can market an actual benefit instead of waving your arms in the air and dropping hints in your newsletter. I want what they have!

Another big opportunity is dealing with insurers. As anyone who gets health care in America knows, we have amazing hospitals and doctors, but accessing them is more depressing than postwar Italian cinema. One term used here is “prior authorization automation,” i.e. letting a bot burst into tears on the phone to Aetna so you don’t have to. Emme (see below) is one swing at this—it does a lot more besides—and you’ve also got your Cohere Health, which sells to health plans, or Infinitus, which uses voice agents to call insurers, or SamaCare, which helps with specialty drug approvals.

I absolutely get why these businesses are happening, but my co-founder Rich describes this sort of process as “procedural debt,” and this may be the most profound case of it on earth. When your company provides (1) voice agents that (2) talk to the voice agents of an insurer to (3) negotiate whether someone can get their utterly standard colonoscopy, there’s something really dumb going on.

Then again, healthcare is a $5-trillion industry in the United States (roughly $2 million of which goes to actually keeping you healthy) and—ah, you know what? America made a guy who killed a health insurance CEO a folk hero, so I’m not going to pretend the status quo is working. But once again, here we are, and maybe software can help with the suck, even if it sucks to further institutionalize the suck with swarms of suck agents doing suck work. Maybe accelerating the suck is the only way to minimize it. Maybe AI is a hypersuck technology.

So we’re entering an era where millions of agents are shrieking at each other about who pays for your weird medical creams. This implies a lot of new systems coming online as we metabolize this change—and, given everything that makes healthcare what it is, means that we’re facing perhaps the greatest IT migration challenge on Earth. As much faith as I have in AI’s ability to help manage data, health-data migration is regulated, complex, and just plain hard. (We’re doing some right now at work, and whoosh.) As summarized in Becker’s Health IT (we all read it, right):

Achieving interoperability starts with aggregating data from the EMR and other sources into a data repository and enabling data to flow seamlessly using methods such as API-to-API data exchanges. Once data is aggregated and is interoperable, it can be mined and care gaps can be identified so actions can be taken to close these gaps.

That sounds super nice, but unpacking that paragraph might take a few years or decades, with some of the worst meetings you’ve ever imagined, especially when an “API” is an LLM-powered bot talking to another bot over a landline. OpenAI for Healthcare, Claude for Healthcare, and Google’s MedGemma are all useful LLM tools that can access vast troves of health information, and we’ll see amazing lurches of progress as new tools get deployed, or systems get better at identifying cancer, or diagnosing your horrible bunions. But it looks like a lot of the real work is going to be in making system A talk to system B, and finding ways to cut down on pajama time. And it’s going to hypersuck.

I guess it’s pretty typical that I looked at the thrilling world of healthtech and got excited by paperwork reduction, but it is a way to help, and I like the idea that you could give doctors fewer reasons to complain, since they all spend half of each appointment whining about MyChart. The tech industry will have to put in some pajama time to help out, but that’s okay. We like being useful.