Deploying AI in Healthcare

We live in a world where the demand for health care is just rising so quickly.

We have 10,000 people aging in the Medicare every single day,

and we just can't train doctors fast enough to take care of all these people.

The practice surface area for the clinician

will look fundamentally different in the next three to five years.

When I was deploying my own software back in the day,

doctors would grow and they'd be like,

oh, yet another tool?

Like, why are you stuffing this down my throat?

The delta between the magic of the tool that they're experiencing

and their consumer lives and what they do in their work

has for the first time, narrowed.

Just even a little bit to a point where it's just fun way

change the nature of how they view technology.

I think this is the first time where there's hope.

There is a pathway to doing more with less.

There is a pathway for the job of being a clinician, being a nurse,

to be a filling one.

There is a pathway for the experience of a patient,

not being as confusing and full of despair as sometimes it is.

When Nikhil Bhutama was a PhD student at Stanford,

he lost a mentor to a medical error.

Instead of finishing his MD PhD,

he dropped out to work on fixing healthcare with technology.

He spent years in the early deep learning community,

including time with the researchers who would go on to found open AI.

He watched Transformers Emerge in 2017

and saw the scaling loss start to click.

Then he did something unusual.

He and his co-founders started a medical practice,

not to deliver care forever,

but to understand what it actually feels like to run one,

to work with doctors, implement an EHR,

and see where technology falls short.

That experience became the foundation for ambience.

Today, the company works with some of the largest

academic medical centers in the country.

Over 75% of clinicians use the product daily,

and one health system is projecting $30 million

in net new margins from the platform.

A16Z General Partner Julie U.

talks with Nikhil Bhutama, CEO and co-founder of Ambience Healthcare.

Super excited to have you here, Nikhil Bhutama,

who is a CEO of Ambience, one of the co-founders.

Ambience, I feel like you guys were founded, what, 2020?

2020.

I think we were the first investment you made on Zoom.

Yes.

On Zoom, that's right.

Right when COVID was hitting,

everyone was going crazy.

But even crazier is what has happened in the last five plus years

on the AI front, and you sort of had a front row seat

to the whole thing.

So you should share a little bit about your background,

and how you got involved in this whole crazy AI space,

and what's it been like to be at the helm of one

of the most cutting-edge companies in the clinical AI space

as the market has evolved as rapidly,

as it has over the last few years.

Yeah, I mean, whoof, the last 12 years in AI

have been kind of insane.

And there's almost like a few different arcs to the story.

But you know, I started my career thinking

I was going to be an MD-PhD, and then I ended up dropping out,

mostly because I lost a mentor to a medical error,

and you know, thinking about the net benefit to the world

for another MD-PhD versus thinking about

how do you solve these problems systemically?

And I grew up here in the Bay Area,

and sort of was embedded at Stanford,

and so you might remember when Andrew Ang was building

deep relief nets, and scaling up those models at GPUs.

This is like 2010.

I think after that, I started writing a book on deep learning.

And I think at a certain point got called out to Silicon Valley

to hang out with a lot of the early researchers

in this space.

We were all hanging out in Greg's apartment for a little while,

and that group became OpenAI.

In that time, I think there was a belief that in general,

unsupervised learning and reinforcement learning

was going to give rise to general intelligence

and general reasoning models.

But no one really had a clear sense of what that was going to look like.

You had a group of people working on OpenAI Jim,

trying to get RL agents to walk in simulated environments,

you had a bunch of people working on variational autoencoders,

and I think the hard part of doing any kind of work

in that time is every six months something would happen

and an entire branch of deep learning

would just collapse.

We went from thinking,

belief nets were going to be the architecture of choice

to no one cares about deep belief nets in like about 12 months.

And I think what was interesting is, you know,

my co-founder Mike and I took a step back,

and we said, we're most excited about thinking about

how do we take these technologies and apply them to healthcare.

And so we actually did the crazy thing, which is,

we started a care delivery asset,

and we started to not only run that practice,

work with doctors, implement an EHR,

but at the same time we were taking all of the techniques

that we were seeing at the research labs

and bringing them into the practice, right?

So 2017 transformer came out,

we were using the transformer in production

to ingest claims data, predict risk hospitalization.

And I think during this time, sort of post transformer,

we just saw the entire research community sort of just collapse

on this architecture because it was so clear

that it solved many of the challenges around language modeling

and reasoning in a way that we had quite see

the previous architectures work and all of a sudden scaling laws

and RLA Jeff, we saw a demo of GPT-2 over the dinner table.

And I think Mike and I were coming to this conviction

that we were on the precipice of some sort of exponential here.

And when you see the people you respect the most

and some of your closest friends go risk their reputation

to raise billions of dollars to scale up these architectures,

something is happening here.

In those early days, when you saw the performance of those early transformers,

today when GPT-3 first came out, everyone was like,

oh, it sucked the healthcare.

There was a ton of hallucination risk.

It wasn't trained on a lot of the proprietary data,

sort of not on the internet, so to speak about medical practice

and guidelines and all that.

And then obviously it's improved drastically over the last few years.

But back then, was it super bad or how did it actually perform

in a healthcare setting and when you were applying it to your first startup?

Like how long was the pull of postdoc training

and fine tuning that you had to do to get it to actually work in a clinical setting?

The iteration cycles were insane.

So today people are like, well, I could pull it all out of the box

and it kind of works out of the box for certain use cases.

And then the moment you move into really deep domain verticals

is when you have to really invest in post-training.

Back then, honestly, you had to rethink how you did pre-training,

how you did post-training.

The iteration cycles work a year.

And you're building deep custom bespoke data sets

and all the sort of like ML operations infrastructure

that you had to build for self-driving cars you were pulling over.

So it was a very different world to get these architectures to work.

And also very different scales, right?

We were talking about tens of millions,

if not maybe a hundred million parameters

or some of these models in 2017 and 2018,

versus now we're talking about trillion plus parameter models.

And you guys kind of did the reverse thing

where a lot of startups start by building technology

that gets sold into the healthcare markets in particular

and then realize how hard it is to sell technology to providers, for instance,

and then decide to go full stack and say,

we just want to eat our own dog food.

We want to capture more values associated with the services delivered

on top of a technology platform.

You guys did the opposite where you started full stack

and then decided that you wanted to do a platform company

in your next play.

Talk to us about how you got to conviction that that was the right move.

Yeah, I think when we started our previous company, Remedy,

we were just extremely sober to recognize

that we had very little empathy for what it's like

to sit in the shoes of an operator of a health system.

And to truly understand the entirety of the context

and the job to be done in the opportunities,

we felt very strongly that we had to hold the responsibility

ourselves first and foremost.

And I think that normally did give us the flexibility

to be able to have the rapid iteration cycles inside of the company

to see what worked and what didn't work.

But I think it also set the foundations for,

if we ever did start a platform company,

what would it like feel like to sit in the shoes of the CEO

looking at 1% to 3% profit margins,

looking at a workforce in burnout crisis amongst your staff,

having to navigate the complexity of an IT stack

and having to work with Epic

and thinking about how do you even make decisions in that world?

And so in many ways, I think going through the experience

of running a care delivery asset

gave us a ton of intuition for now

is we're building a platform company.

How do you actually do that successfully?

And how do you build something that can resonate not just

with the end user on the ground,

but also resonate with the economic buyer, the CEO,

the CFO, the chief operating officer of an organization

that has to make sure that this system

is financially sustainable too.

And just to get into a little bit of the history of ambience,

you guys in some ways have traversed a fairly large surface area

of the provider market, right?

You've worked with, obviously today,

some of the most prominent health systems in the world.

You also have experience working with smaller medical groups,

even digital health companies at the beginning.

Talk to us about, what's your view on the segmentation of the market?

And also, what's the stage of maturity that we're at today

with respect to AI adoption?

Because I think that's one of the most surprising things

that's occurred in the last few years is that providers

used to be the last bastion of technology adoption.

They were always the laggers.

They were always behind every other segment

like the payers and life sciences and what have you.

And then all of a sudden in the last couple of years,

they are arguably amongst the fastest adopters of AI

and not just in the administrative setting, but like your product

like at the point of care for doctors.

And so how do you sort of like rock all that

and where do you see the pockets of like sort of highest adoption rates

amongst that set of providers that I described?

Yeah.

Maybe we take a step back when we think about why

this is so compelling to the organizations we work with.

The reality is like, you look at a doctor's work day today.

There's not a lot of joy in the practice of medicine anymore.

You can't look your patient in the eye.

You're constantly feeling like you're running behind.

It feels like you went to school to take care of people

and to actually grapple with the clinical medicine in the room.

But most of your time is spent doing all sorts of other stuff, right?

You're like searching through the electronic medical record

to find information, you're writing notes,

you're navigating these thousands and thousands

of coding and billing rules that are different by type of pay

or different by region.

They change year over year.

And you're like, I didn't go to school to do this.

And so I think in many ways,

the organic pull from the market was there for a very, very long time.

And it wasn't a question of,

hey, can is there a demand for technology that can do this?

The question really was for a long time.

Can technology even serve this need well enough in the first place?

And I think as the market evolved,

there's almost a bifurcation between what I think

are largely the high complexity, high value set of use cases

and the low complexity, low value set of use cases.

And so for us, we found a home specifically

in some of the largest ideas in the academic medical centers.

And part of why we think that's a very interesting place to play

is just the vast breadth of medicine

that's being practiced in those environments

and the depth across that breadth

is extremely challenging to go tackles.

You think about, if you're trying to build,

our view is that the practice surface area

for the clinician will look fundamentally different

in the next three to five years.

Like today, even if you compare the experience today

using something like ambiance versus what it was a year or two years ago,

it already looks so so transformatively different.

But the challenge is that when you're operating

in these really complex clinical settings,

the job of a primary care physician

who's primarily trying to manage multiple chronic diseases

versus the needs of a sub-specialized oncologist,

they're so different.

They're making different decisions.

They have different workflows inside of the EHR.

They're looking at different data sources

and so to be able to build the kind of infrastructure that serves,

the broad range of medicine is really, really hard.

On the other hand, the moment you move to sort of the mid-market

or the small sort of three, five, ten, dock practices,

the complexity drops dramatically

and so it's much easier to serve.

And so my guess is, in general,

you'll see a proliferation of lots of players

trying to compete over the mid-market.

For instance, I think you've got the EHRs

in that space that are trying to reinvent themselves

to be AI first companies.

You've got AI scribes going after that space

and I think they're quickly finding out that to create enough value,

they have to own more and more of the stack

for these organizations.

And then I think the enterprise segment of the market

is starting to shake out now where I think the reality is

there's only a couple of players

that have even had a right to play in that market.

And most of them outside of ambience

have really, really struggled to actually meet these organizations

at the complexity at which they need to practice.

So for instance, we'll have so many organizations come to us,

they've rolled out something and only 15, 20% of their doctors

actually use it.

And even those who do use it, they use it.

It's just not good enough.

It's just not good enough.

And even the ones that you use it,

they're using it for 20, 30, 40% of their visits.

On the other hand, I think for us,

we work with several large academic medical centers

and we're at the scale now where 75% of all their clinicians

use ambience every single day in clinic.

They're using it for 80% of all their visits,

which is a completely different sort of opportunity in scale.

And so I think that's sort of how the market has shook out

in many ways, which is there's a high complexity

part of the market that's really, really hard to serve.

But if you can serve it, it's hard for others to compete in.

One of the tropes, when people talk about this space in general,

let's call it clinical intelligence,

AI for clinical intelligence in general,

one of the tropes is that eventually,

the generalist foundation models will just get so good

at a global set of intelligence that they're going to win

these categories where all you have to do is basically

synthesize information and then make a recommendation.

And then really where the game is going to be

is kind of the action layer,

like the actual work that is done agentically.

What do you do with that information once you have it?

How do you actually, you know,

critically create clinical utility basically

for the patient and the provider?

Number one, do you agree with that premise that, you know,

the sort of the cliff that is clinical intelligence,

like that is coming up soon and that like any company

that's only doing the intelligence layer will soon become

very commodified and therefore the real competition

is going to be on the action layer.

Or do you actually think like to your point that the,

the length of the pool on just clinical intelligence alone,

it's so long that you still have, you know,

many years before any of the generalist foundation models

will be able to crack the code comprehensively

across the full surface area of medical practice.

So it's a fascinating question.

I might answer it a slightly different abstraction,

which is that from our experience,

AI clock speed is fundamentally different

from product clock speed.

And, and part of the reason I think about it that way

is there's several aspects to intelligence

that do get better with every generation of foundation model.

And in many ways, I think what's kind of frightening

about building in this world is the capabilities

are evolving so quickly that,

and we use this word constantly inside of ambiance,

we're building a world with the floor is lava.

And you have to have the kind of organization

that can respond to and on a dime,

be able to reinvent themselves itself

as capabilities continue to evolve, right?

So we spend an insane amount of time predicting

what the capabilities will look like over the next 18 months

and then building for that future

as opposed to building for the capabilities now.

That being said, I think what we find

is that there is still a massive last mile problem

for these models to be effective in healthcare.

And you can break it up into a couple of different categories.

And it's going to start,

it's going to start with,

do the models even have the right context to begin with?

And I think you probably have a deep appreciation for this,

just how messy it is to even build out the right infrastructure

to be able to pull context out of systems of record,

hidden behind fire APIs and proprietary APIs.

The data models are so messy and inconsistent,

you'll get specific standards

where it's sort of like a concept of a specification.

And so you have to like...

People will just like stuff like free text into a random field

and all that.

100%.

So you're like even just the problem

of across EHR instances,

being able to pull out the context from systems of record

in and of itself was an unsolved problem

and we started building ambience that we solved, right?

And so the ability to read out of any part of an EHR instance,

including the data warehouse underneath it,

and then using that and having a groomed layer

to be able to then build intelligence on top,

that was an unsolved problem.

I think another thing that a lot of folks don't fully appreciate

is that the most valuable thing for AI companies is decision traces.

Most EHRs are built on mutable data structures,

which means that you inherently destroy the decision traces.

And so fundamentally, you have to rethink the architecture

of how you actually collect this data in the first place,

even make intelligence be specific to the domain.

And so that's another sort of big problem in the data layer,

but there's a whole sort of range of data layer problems

that you have to solve.

Then in the middle, I think we have a really big challenge

from a clinical intelligence standpoint around defining quality.

And-

Where's the fine quality in which and many ways?

Exactly. And it's especially challenging

because most of the use cases are open-ended use cases.

And they start as trivial as if you have multiple pieces

of contradictory information in the chart.

You got a patient where there's no indication of thyroid problems

and the problem list, but this person is on thyroid medication

that was prescribed six months ago.

What does that mean about the state of this patient?

Even just resolving truth at that level is tricky.

Then what if you're in the inpatient setting

and all of a sudden you've got four different clinicians

doing a physical exam on this patient

and you've got one very specific physical exam

that happened about 48 hours ago

that came in through a specialty consult?

How much do you index on that versus the physical exam

that happened two hours ago by generalist?

That's a hard reasoning problem.

Criticity matters a ton.

Criticity and credentialing matters a ton.

And then there's also the realization that so much of what ends up

and needs to end up in the medical record

is never verbally sort of explained in a visit.

So for instance, an oncologist may be walking through a decision tree

on a care pathway in their brain,

which is then expressing to the patient

and words that they can understand what the next step is.

But what goes into the clinical note is that trace of the decision tree

not the words that were spoken to the patient.

And so defining quality as to what does that

specialised oncologist expect is fundamentally complicated.

And then you've got certain use cases where, for example,

ICD-10 coding, you put two doctors in a room,

they agree 60% of the time.

You put a doctor on an ICD-10 coding problem,

you ask them a question, you put a coder on that problem,

you ask them a question, versus you have both of them

in the room with 10 minutes to debate

before you ask them the question, you get three different answers.

And so defining quality is actually just a fundamentally hard problem.

And one where it has to be solved with the intelligence layer

and it's not being solved by the foundation models today.

And then I think the last piece is that

for these companies to be successful

as capabilities are evolving really, really quickly,

you've got some companies in vertical industries

where iterations speed is naturally fast.

For us, you have to have really, really deep relationships

to be able to go from prototyping something,

to deploying something in TST, to turning it on in prod,

to then being able to actually make live learning loops

with end users.

And I think one of the things that we've solved as an organisation

is how do you build extremely deep relationships

with marquee organisations to be able to go from concept to live

and deploy learning with users within like less than 30 days,

which is unheard of in our category.

So I think there's like all these things that you have to solve

for companies to be valuable in this category

that are just far beyond clinical intelligence,

so that makes sense.

So the floor will stay lava for much longer is what you're saying.

Yeah, 100%, 100%.

Maybe a question for you is, imagine you lived in a world

like if you were to rebuild Kyrus today, right?

How long would that take?

And then if you had an environment where someone had solved

the integration problem, had solved the data integrity problem,

had solved the sort of relationship with the health system problem,

how long do you think that would take for free?

Yeah, I mean, this is definitely a question I think about all the time.

And I see lots of companies that are actually doing, you know,

forms of what I would have done.

But it kind of gets just to, you know,

rehash some of what you said.

I think there's two components that I think a lot about.

One is, what is the form factor of the product that we would have built?

It would have been fundamentally different, right?

Like we were a legacy enterprise SaaS product that got deployed

into a call center setting where we had to train the humans

to use the software workflow in the right way, you know,

to get the right outcome.

And as we know in those environments, the turn rate of employees

is just naturally super high.

So as soon as you train one batch of people,

they leave and then you got to retrain them again

and you just have a huge area,

sort of range of compliance rates.

Let's call it with which people are using that software.

So if you can actually make an agent do the work instead,

you know, that removes a ton of the potential drift and outcome.

And so that's, you know, one thing that we see a lot of these days is like,

instead of just giving tools to a call center,

why not be the call center and build, you know, voice AI agents

that are effectively doing the work based on the rules of the road

that you want, you know, kind of systematized across the entire healthcare system.

The second thing I think about is exactly what you said.

And I have a question back to you on this is, you know, the data layer, right?

Like one of the biggest impediments to being able to do scheduling efficiently,

which is what we were trying to do, is that there is no source of truth.

There's no standardized way for representing clinical schedules.

And not only that, it's like individual physicians, like, you know, to their credit,

they obviously all have very different styles, very different, you know, preferences,

different ways they want to practice.

And there's no, no, no health system has an incentive to tell every doctor

to, you know, systematize it in one way because, you know, these doctors are very scarce.

They want to retain them.

They want those doctors to attract the patients, you know, that they want to see.

And so there's, you know, the heterogeneity of the underlying data

was like a huge impediment to getting it right at the system level.

And so, you know, if we were to build that company today,

or even, you know, what ambience is doing,

you're effectively creating the new system of record, right?

Like you're, you have a full texture of data that has never existed before

in the healthcare system ever, right?

Conversation, grade, resolution on what's being said between a clinician and a patient.

And then to your point, how does that translate to what actually gets documented

in an esoteric fashion and creating those links, like that's never existed before?

Same thing with scheduling, right, where, you know, as appointments are being booked

and as you hear, what was the preference of the patient?

And then how do you bump that against the preferences of the doctor

and then, you know, create a semantic that links those two things together?

Like that's a de novo set of information that like never existed before.

So I think there is another opportunity said for any company that's building in this day and age

is, you know, how do you not only think about the work that needs to be done?

But, you know, creating an entirely new set of data that not only trains your models better

to like, you know, perform better in the future,

but effectively could become an entirely new system itself, right?

And I mean, that would be the question to you as you think about the future of the EHR.

Like to your point, the EHRs are trying to vertically integrate into these workflows

at themselves, you know, I think obviously I'm betting against them being the ones who win that game.

But, you know, is this the moment, like all of us have been waiting for decades

for the moment where we can finally disrupt, you know, that incumbent EHR layer

and have a shot at potentially introducing new tools that, you know,

number one, bring those systems into the modern era,

but also enable startups to have a shot at, you know, kind of owning that layer.

Do you think this is the moment or, you know, what needs to be true for us

to start to kind of eat into the layers of the stack that are owned today

by these kind of monolithic European players?

I think this is the moment and it's the moment for two reasons.

The first is deeply related to what you were talking about, which is

the organizations that figure out how to build the practice surface area of the future

and the administrative stack of the future have to be able to unlock this sort of level

of AI to product clock speed that is bottlenecked by the fact that today

we've built all these legacy systems on a legacy architecture

that's not actually optimized for AI to product clock speed.

And so one of the innovations of ambience is we've actually built out a layer

that sits on top of the EHR that pulls all the data out of sort of the systems of record,

puts it in a form that makes it easy to build AI products on top

so that the incremental cost of building a net new use case dramatically drops.

Because you can imagine every single use case in application area

is leveraging the same underlying systems of record

and you don't want to recreate that same infrastructure over and over and over again.

And so for us to go from two products to 12 products to next year,

24 products out in the market,

I think once you've had the infrastructure layer that we've created,

it fundamentally changes your product clock speed as an organization.

So that's probably the first insight.

And it's not a trivial piece of infrastructure to create.

Like for us, it took several several years of deep R&D

to be able to even do this in the first place.

I think the second is the following.

The physics law will have fundamentally changed with the set of capabilities.

And so you were talking a little bit about how if you were to build a chyrus today,

like with the form factor even be the same, right?

And I think that's happening across every use case,

across the span of the entire system, right?

Reve cycle is probably one that you're really, really deep in as we are too.

And one of the questions is like today when you have a Reve cycle problem,

how to system solve it.

And it's really twofold either they invest in training,

to teach clinicians who don't want to learn coding,

to try to get them to do the right thing,

or you've got massive back office teams

and try to like review as much as you possibly can,

and then try to correct a clinician afterwards,

which obviously clinicians hate as well,

and is extremely inexpensive and inefficient.

And so the question for us as well,

if you could distill that expertise into a model,

which is extremely hard to do, but if you could,

and then you could distribute it at the cost of software

across every interaction where Reve cycle expertise is relevant.

And you then own the window in front of that user.

How do you then leverage that set of capabilities

to make doing the right thing easy?

How do you make doing the right thing obvious?

And all of a sudden, that starts to disintegrate

a lot of the underlying assumptions

for why Reve cycle is constructed the way that it is.

So like building a product for pre-build.

Does that make sense any more in a version of the world

where you can work closely with the clinician

to create a more accurate source of truth

and automatically adjudicate all of it in real time?

Yeah.

The next obvious step from everything that we're talking about

is autonomous AI doctor, right?

So you guys are today a co-pilot.

Every use case I believe that is still being used in the wild

requires a physician to basically sign off on the note

or the documentation, et cetera.

Why shouldn't your platform be an autonomous AI doctor?

Like what's between the capabilities today

and the future reality that we will likely have

a lot of this being done in a fully automatic fashion,

doing clinical, actually making clinical judgments.

We live in a world where the demand for health care

is just rising quickly.

We have 10,000 people aging in the Medicare every single day.

And we just can't train doctors fast enough

to take care of all these people.

And on top of that, the sort of cost of adding

one more patient to a doctor's panel today is so painful.

And so I think one of the promises of these technologies

is how do we do more with less?

How do we make it painless for a clinician to say,

I do want to see more patients.

I do want to increase access to the care and expertise

that I have.

Do it in a way that is long-term sustainable.

I think most organizations want to increase access

to patient populations.

And the question is how do you do that

without asking your clinicians to do more?

Well, a big part of that is how do we start to offload

some of that work to sort of virtual care team members

who can sort of take the next step on behalf of the,

on behalf of the clinician.

And so already we're starting to experiment with

before the visit even starts.

So before, before a visit today,

Ambience has sort of anticipated everything

this doctor would need to know before they even see this patient.

So imagine a clone of yourself here,

you're an endocrinologist.

A clone of yourself has poured through all the data

for hours and hours and hours.

And it's put together a summary that of everything

you need to know for this patient.

Now imagine you move that one step further upstream,

which is you've got an agent that has access to all this context

that can also anticipate, well, what are the most likely

questions that are top of mind for the clinician?

And you start asking those questions

before the visit even happens.

And then you've loaded that also up

into the summary for the doctor.

And then on the flip side,

you had a great conversation with your physician,

the clinician sort of wrapped up with you

with an after-visit summary that gets sent back

to you through the patient portal.

What happens if you turn that into an agent

that can have a continuous conversation with you?

Answer questions.

Double check.

Did you pick up your medication?

Actually make sure that you got that sort of,

that lab test done at Quest Diagnostics.

Did you pick up your Xanax for your claustrophobia

because you have a CT scan before seeing the oncologist

and you're anxious about that CT scan?

All these things, what does it look like for now

there to be a virtual care team member

to actually help quarterback all those things on your behalf?

That is what the true promise of these capabilities is.

So let's say that I'm the CEO of a hospital system.

And like, when you're describing something incredibly exciting

but today I've got five different vendors

who are all buying for, you know, that like that pie essentially.

So you got your, I have my EHR.

I've got, you know, the foundation model companies

who are all doing, you know, launching healthcare products

and, you know, claiming that they're going to get into this space.

I've got ambience, you know, kind of starting with AI scribe

but expanding rapidly.

I've got my AI revenue cycle players.

I have my AI clinical distance support players.

You know, those are kind of like to me

the five major categories of players

that are kind of all converging on this vision.

What's your pitch to me?

Like, tell me how I grok this whole space

and how do I think about making investments

against those categories in such a way that is durable for me?

You know, over multiple years, like a five-year time horizon

as opposed to me being in a situation where I need to rip and replace vendors

like every, every six months.

Yeah.

The way that I would think about it from the shoe of the operator

is probably two-fold, two lenses.

The first is it's a lot of, a lot of excitement in noise around AI.

But there's a big difference between marketing and noise versus

if I bought this, will my clinicians actually use it?

Because if you don't have the right level of adoption,

nothing really matters, right?

And so I think that's probably where right now,

the jury is looking really negative for a lot of AI use cases,

which is these companies are very good at pitching a vision

but then when it comes to brass tax,

the adoption utilization of these technologies is extremely underwhelming.

And that's one of the cool things about this era

is that you can literally just let the doctors use it

and kind of put with their own feet, right?

100%.

And I think that already is,

it's so clear for any organization that either works directly with ambience

or just talks to an organization that works with ambience.

And most of these have tried every single one of the players

in the category that we've done an incredible job

of actually owning the window of care in front of every single clinician

set at the enterprise.

And if you can't do that, that's table stakes.

And so that's the first lens is,

can you actually get the level of adoption utilization?

I think the second is,

look, health systems are not traditional enterprise-based BSAS enterprises.

You don't have hundreds of millions of dollars of cash lying around

to be able to invest in toys, in cool toys.

And so the question then becomes,

how do I actually fund these things?

And so the second lens that we take is,

at the end of the day,

one of the reasons why AI is exciting is because

this is the first time that we think is an industry

that there is a class of technologies

that can fundamentally change operating margin.

And I think if you get this right,

the organizations that unlock their AI product clock speed,

they'll be able to adopt AI,

they'll be able to unlock new forms of operating margin,

that operating margin allows them to invest more in tools

that attract better talent.

That better talent means more volume,

more volume equals more revenue,

more revenue equals the ability to invest in more AI.

You get sort of unlock this crazy flywheel.

And I think from the CEO's perspective and the board perspective,

the organizations that figure out how to do this effectively,

they compound and become the destination of choice for patients

and the ones that don't, they're at risk of consolidation.

And so then the question really becomes,

who can I actually work with to change operating margin?

Are the EHRs going to work with me in a partnership model

where they're committing to change operating margin?

I think it's a thing that's so scary for people to do

because it requires your products to actually work.

You have to actually be good at change management.

You have to be good at measurement and attribution.

But I think because we've built out all those capabilities,

we're willing to go to a health system and say,

we will actually put our money where our mouth is

and our success is going to be dependent on your ability,

our ability to help you change operating margin.

And so I think the combination of those two things

creates a set of proof points in the beginning of a relationship

where it's almost like a religious conversion that happens

where now you walk into an organization that leverages the ambiance

and you can't go from room to room without being stopped in the hallway.

And that builds the level of trust and excitement for the future

that earns you the right to do more over time.

So I think within six months of working

with the Marquee Academic Medical Center customers we have,

we go from, hey, we're working on the scribe

to we want everything on your roadmap as quickly as possible.

And I think it's because of that sort of like excitement

from seeing it actually work and then the CFO

for the first time looking at AI implementation and saying,

this clearly pays for itself many times over

and we can take the new margin we didn't have.

One center we work with, they're projecting

over 30 million dollars of net new margin,

post attribution debate, that's just free by ambiance.

Because they don't have to hire human scribes anymore.

What's the attribution there?

A big part of it is our CM,

a big part of it is improving throughput and access.

But our CM is a big part of it.

And now you're like, well, I have all this margin

to invest in more AI.

And that's sort of the beginning of unlocking that flywheel.

And so once you see that, I think it just fundamentally changes

your perspective on the category.

It's interesting too, because I know like even just a year ago,

I remember the first wave of AI scribe adoption.

When I talked to C-suite executives at a hospital systems

and I said, you know, what's the ROI?

Like, what made you purchase these products?

They basically said, listen, Julie, actually,

there's not that much financial ROI.

We're just doing it for retention.

We want employee happiness.

We want our positions to feel like we are looking out for them

and giving them a reason to come to work every day.

And now it seems like it's very much shifted, you know,

to actually heart ROI in like phase two.

Because people didn't know it was even possible at that point in time, right?

To do this, you have to be able to track

where user behavior is happening inside of the EHR,

map that user behavior down to a code that code then being submitted,

a CDI query prevented, a denial prevented,

and then putting that all together to new cache

you're collecting in reduction in cost to collect

and doing that in a way that's going to actually

meet the master of the CFO's office.

And for us to do that requires us to basically download

the data warehouse and build an entire analytics stack

that would actually work for the CFO's office.

And so like in the absence of doing that,

of course, the CFO's never going to believe it.

But now we're seeing the real proof points across customers

where CFO's will actually tell other CFO's

how much value this category can create.

But it takes a lot of work and thoughtfulness

on getting the right use cases,

actually moving coding performance, actually doing change management

to move them, move the marker on operating margin.

Yeah. We've talked solely about the provider set of the market,

but obviously some of the things that you're speaking about

have implications for, certainly for payers at a minimum.

Yeah.

And the sort of steel cage death match of providers implement AI,

payers implement counter AI,

and then we have at least bought on bot crime around RCM.

That's actually playing out, as we know.

And people are actually talking about it on their earnings calls,

even with some of the large national payers.

Where do you see that playing out?

Like how do you see?

And if you guys are doing any work along those lines,

it would be great to hear just like real life case studies

of how do we solve that challenge?

You probably know we work with a lot of organizations

that are integrated.

They have a plan, close relationships with plans.

Yeah.

I think what we're finding is that the world may,

our view is more optimistic,

which is if you've got a system that actually can understand

source of truth,

and understand it really, really well,

not just ambient listening.

It's also deeply understanding all the past contexts as well.

So it's like you need that sort of like layer

on top of the systems record.

But if you deeply understand source of truth,

and you know exactly what happened in the visit,

and you can answer any question with high levels of fidelity

with clear audit trails,

it's not just a win for the organization

on the health system side,

but I think long-term it ends up being a win for the payer as well,

because on the flip side,

we're talking about an AI versus AI arms race.

We've already seen the labor versus labor arms race

because you've got payment integrity teams being built out

to sort of work with the RCM teams on the health system side.

But I think what ends up happening

is once you have a shared source of truth,

then the ROI of RCM becomes potentially negative

over the next five years,

and there's a very real likelihood that the ROI of payment integrity

also becomes negative over the next five years,

and it just makes sense to collaborate.

So I think I'm a long-term optimistic on this one,

even if the short-term seems a little bit tricky to navigate.

As the bar has gone up on the things that AI can do,

there's a lot of things that I remember when we first invested.

We were kind of like, oh, that's a pipe dream.

We'll have to do additional development and or wait until AI models

come along far enough for us to do XYZ,

and then here we are.

Those things are actually very feasible.

Are pitch decks about the same?

The pitch deck doesn't really change.

Yeah, yeah, but the capabilities have definitely evolved.

Is there any remaining?

What's hard today?

What's still hard to do?

What are you hearing from your customers that they want to do?

That we're not actually able to do yet with the capabilities

that are out there?

Is it AI?

That's the bottleneck?

Or is it all the other stuff that you talked about with respect

to last-mile workflow integration, RCM, all that kind of stuff?

Yeah, I think we're at a place where it's hard to disentangle.

What do you consider a bottleneck at the foundation model layer,

versus a bottleneck in post-training, versus a bottleneck in product?

I think some of the use cases that are still tricky and hard for us,

are as we're thinking about cascading context across care settings.

So how do you anticipate what's going to happen

when you've got a patient who's admitted to the ED

and now they're upgraded to the inpatient setting

and now you're trying to make predictions on what next-spass action really is?

Some of that work is still a little tricky.

I think generally predictive modeling is not particularly well solved

yet by this class of models.

But I think in general, what we're finding is that

if you've built the right team with the right applied R&D expertise

and the right internal clinical subject matter expertise

and RCM subject matter expertise to pair really, really well

with applied R&D teams, we're in a world where so much is,

there's just so much to build that we don't really feel that bottleneck.

We're almost just like our ability to understand the problem

and have teams that can go tackle the problem

is the bottleneck more than anything else.

Do you ever envision ambiance becoming like a true platform

in the sense that you open up your capabilities to a third party

to then develop on top of your system

and you become kind of that back-end layer that the Asia plays today?

It's an active conversation with most of our customers.

A lot of the academics have internal product and engineering teams

that want to build all sorts of stuff.

Sometimes that's on roadmap for us

and we talk about how do we want to think about

making the right shared investments over time?

There's a lot of great ideas that are likely just not on our roadmap.

And so this ability for us to make it easy for others to build on top

is this supernatural and that could extend to our customers

but it could also extend to the broader ecosystem.

Well, and one last question for me that's more internal facing

of your experience as building ambiance as an AI native company

and obviously, again, like being an employee at an AI company

has obviously also evolved quite a bit over the last few years.

What are some of the things that you're doing fundamentally different today

based on the availability of AI tools for your employees

that you weren't doing two or three years ago

that you think has been a huge game changer?

There's probably a number of experiences that are wild to folks

if you almost rewind back time two years ago

but on the engineering side,

like the amount of work that an individual engineer can get done now

with Opus 4.5 is quite insane.

I think what we're finding is that you just need really smart thinkers

and you don't necessarily need as many people anymore

to get lots and lots of work done.

Because it has a change like the profile of engineer that you hire?

It does. It does.

I think generally on the platform side,

you want people who can think really deeply

about long-term architectural choices in C-Ran corners

and on the product, on the product engineering side,

what you're really looking for is the kind of person

who can embed deeply in clinical environments,

spend time with customers, work closely with subject matter experts

and really get it requirements gathering.

And that's actually the bottleneck to building great products and features.

Internally, we use AI ton for research,

for sharing context.

One of the things we oftentimes think about is when a new employee onwards,

how do we take almost our internal decision traces

on how we make decisions as an organization

and make it really, really easy for a new employee

to be able to stand under the shoulders of giants?

Because if you join a new company oftentimes,

you're like, you have no idea how this place works.

You have no idea the historical context of decisions that were made.

You have no idea if you're about to make a new decision.

How do you even begin thinking about what the right framework is?

Right before and yeah.

100%.

And so I think still early on some of these use cases,

but we're starting to build internal teams to think about

if we were to internally make ourselves an AI first company

what that would look like.

Because our thesis is in the same way that you wouldn't,

if you were building a health system from the ground up today,

you would not do it the way you were building 10 years ago.

If you're building a company today,

you would not do it the way you would be doing it two years ago.

What did we not cover that you want the world to know about ambiance?

That's a good question.

There is a specific level of just humility and honor

and being able to work on this problem.

I know you've got your own personal experiences with the health system.

I sure have my own and every person ambiance comes in

because they were a loved one or had an experience with the health system

or works inside of the health system.

It's just one of those industries where the outlook

bleak for quite some time.

Folks have been talking about the system being at a breaking point.

I think this is the first time where there's hope

that, hey, there is a pathway to doing more with less.

There is a pathway for the job of being a clinician being a nurse

to be fulfilling one.

There is a pathway for the experience of a patient

not being as confusing and full of despair as sometimes it is.

I think in some ways that makes this a special time

to be working in health care, whereas in the past,

I think we didn't necessarily attract the best people

and the best talent because it was unclear.

Could you even have an impact if you wanted to on health care?

This moment is very special and I think there's just a lot of gratitude.

I think we all have to be in a position to even be able to contribute to the problem.

100% agree with that point about just the exceptional talent

that's coming into health care and the fact that I think health care is now earning its right

to be compared to best-in-class broader tech.

As opposed to this weird niche that just always looks crappier than everything else.

I would also translate that to the physician side.

My last story here is just, I remember it was probably three years ago

when one of your customers asked me if they could share my email address

with a doctor and I was like, sure, what's going on?

Two days later, I got an email from that doctor saying I specifically asked

for a contact information of one of the investors of ambience

because I just wanted to know, thank you,

for investing in this company because I've never experienced the type of joy

that I've experienced using this tool in my day-to-day job

that has made me now want to remain a doctor.

This was a person who had been considering quitting their job basically

after all the terror of COVID and everything else compounding on top of that

that you're talking about.

I 100% agree with you that after nothing but technology being a burden to this whole industry

and even when I was deploying my own software back in the day,

doctors would grow and they'd be like, oh, yet another tool.

Why are you stepping this down my throat all of a sudden to an era

where they can actually see sunshine

and I think the delta between the magic of the tools that they're experiencing

and their consumer lives and what they do in their work

has for the first time, narrowed just even a little bit to a point

where it's just finally changed the nature of how they view technology.

It's kind of wild.

When we even hint that we're about to release a new product,

the energy we feel from the clinicians is almost like lining up

for the next new Apple product.

Like you've just never seen that kind of energy before

and I think for us, it feels great that every time we create something,

there's almost at this level of magic that's created for the clinician

that sort of builds up this anticipation.

And I think we understand that comes with a bunch of responsibility too,

which is what then once we do deliver that the products actually

meaningfully change the lives of the clinicians we serve.

And so that's a responsibility on us to keep doing that quarter after quarter after quarter.

Yeah, amazing.

Well, always a pleasure to talk to you about these things to kill

and congrats on all the progress that ambience.

Thank you, Joey.

And thank you for believing us since the very beginning.

Yeah, absolutely.

Thanks for listening to this episode of Raising Health.

If you liked this episode, be sure to like, comment, subscribe,

leave us a rating or review and share it with your friends and family.

Follow us on X at a16z and subscribe to our substack at a16z.substack.com.

Thanks again for listening and I'll see you in the next episode.

As a reminder, the content here is for informational purposes only.

Should not be taken as legal business, tax, or investment advice

or be used to evaluate any investment or security

and is not directed at any investors or potential investors in any a16z fund.

Please note that a16z and its affiliates may also maintain investments

in the company's discussed in this podcast.

For more details, including a link to our investments, please see a16z.com

forward slash disclosures.

Thank you.