Super Staffing Healthcare, Codifying Compliance, & Scaling Services

0:00:04 – Upwards of 300,000 clinicians left the workforce
0:00:05 in the year 2021.
0:00:09 – The large banks right now up to 15% of their staff
0:00:10 is just in compliance.
0:00:12 – You go from having 5% net margins
0:00:14 to 30% net margins.
0:00:18 – We’ve just never seen the rate of technology adoption
0:00:20 that we’re seeing now ever in the history of health care.
0:00:21 – If you don’t do this right
0:00:24 and you don’t figure out like a very repeatable motion
0:00:25 for these smaller acquisitions,
0:00:27 it’s gonna be really challenging to build the company.
0:00:29 – How long are we willing to wait
0:00:30 to solve this problem?
0:00:33 The everyday person should really care about this.
0:00:36 – Hello everyone and welcome back to our big idea series.
0:00:38 We’re here hours away from 2025,
0:00:41 which will mark nearly 22 years
0:00:43 since the Human Genome Project was completed,
0:00:46 21 years of Gmail, 19 years of Twitter,
0:00:51 15 years of Instagram, and even 17 years of the App Store,
0:00:54 which released 500 apps when it first came out
0:00:55 and now has millions of apps
0:00:59 that have generated developers over a trillion dollars.
0:01:03 It’s been a wild ride and that ride is far from over.
0:01:05 And that’s why our partners who meet daily
0:01:07 with the folks who are building our future
0:01:11 are sharing their big ideas for 2025.
0:01:12 Last year, we predicted
0:01:15 – A new age of maritime exploration.
0:01:17 – Programming medicines, final frontier.
0:01:20 – AI pro schemes that never end.
0:01:22 – Democratizing miracle drugs.
0:01:24 – And so far in this year’s series,
0:01:26 in part one, we covered the growing hardware
0:01:28 software intersection, and in part two,
0:01:30 how large companies and startups
0:01:32 are fighting for AI dominance,
0:01:34 but also how countries are thinking
0:01:36 about its impact on their sovereignty.
0:01:39 Today, we talk applied AI
0:01:41 and how this platform shift changes the game
0:01:44 in healthcare, fintech, and even opens up a new opportunity
0:01:47 to rethink growth in traditional service industries.
0:01:49 Now, if that wasn’t enough,
0:01:51 you can check out the full list of 50 big ideas
0:01:54 over at a16z.com/bigideas.
0:01:59 As a reminder, the content here
0:02:01 is for informational purposes only.
0:02:02 Should not be taken as legal, business,
0:02:04 tax, or investment advice,
0:02:06 or be used to evaluate any investment or security,
0:02:08 and is not directed at any investors
0:02:11 or potential investors in any a16z fund.
0:02:13 Please note that a16z and its affiliates
0:02:14 may also maintain investments
0:02:17 in the companies discussed in this podcast.
0:02:19 For more details, including a link to our investments,
0:02:22 please see a16z.com/disclosures.
0:02:30 First up, AI applied to healthcare,
0:02:33 an industry fraught with dissatisfaction,
0:02:35 too slow, too expensive, too clunky,
0:02:37 but also too few staff.
0:02:39 – Healthcare is dealing with the mother
0:02:41 of all clinical staffing crises.
0:02:43 We’re short hundreds of thousands of doctors and nurses
0:02:45 relative to the level of rapidly growing demand
0:02:47 for clinical services that is projected
0:02:49 in the next five years.
0:02:51 That was Julie Yu, I’m a general partner
0:02:54 on the Bio and Health team here at Andreessen Horowitz.
0:02:57 – So can AI in the form of specialist models
0:02:58 realistically turn this around?
0:03:00 Here is Julie’s big idea.
0:03:04 – My big idea is what we call super staffing for healthcare.
0:03:06 And what that means is we’ve talked obviously
0:03:08 about just the rampant challenges
0:03:09 that exist in our healthcare industry.
0:03:11 So much is broken about it,
0:03:14 but really one of the sort of underlying fundamental issues
0:03:17 that we face is this supply-demand mismatch
0:03:20 between the amount of clinical supply that we have
0:03:22 in the forms of doctors and nurses
0:03:24 and other clinical service providers
0:03:27 versus the level of demand that we as patients
0:03:28 represent in the industry.
0:03:32 And it’s not just that the demand itself is high,
0:03:33 it’s that the nature of the demand
0:03:34 keeps getting more and more complex.
0:03:37 We’ve talked a lot about how the chronic disease burden
0:03:39 in America is increasing over time.
0:03:40 The number of diseases that any individual has
0:03:42 is growing and growing.
0:03:43 The number of drugs that we’re all taking
0:03:44 is growing and growing.
0:03:47 And so again, not only is it that there’s just
0:03:48 a shortage of labor relative
0:03:50 to the number of humans who need it,
0:03:52 but also the nature of the work
0:03:54 is also getting more complex over time.
0:03:55 – So let’s dive into that.
0:03:58 Just how short are we in this clinical staff?
0:04:00 – There’s a lot of different reports and stats
0:04:01 that have a wide range of numbers here.
0:04:03 But I think generally speaking,
0:04:05 when you look at, for instance, the number of doctors
0:04:07 that we have estimated a shortage in,
0:04:12 it’s anywhere between, let’s call it 60 to 100K doctors today
0:04:14 in terms of the relative demand that we have in our system.
0:04:17 And roughly about 75K to 150K nurses
0:04:19 in a commensurate sense.
0:04:21 But I think probably one of the more interesting issues
0:04:24 to unpack here is what do we actually mean by shortage?
0:04:26 So there’s largely two schools of thought
0:04:27 of how people define this.
0:04:30 Simply the number of human clinicians that we have
0:04:32 is one way to look at it.
0:04:33 The other way to look at it though
0:04:36 is given the number of nurses and doctors we do have,
0:04:39 are they actually fully performing
0:04:40 to the full extent of productivity
0:04:44 that number of laborers actually represents?
0:04:45 And I think it’s probably, at the end of the day,
0:04:47 a combination of both,
0:04:48 but those are two different ways
0:04:49 that people have cut that pie.
0:04:50 – And how did we get here?
0:04:52 What is leading to this shortage?
0:04:54 – If you zoom out the lens entirely
0:04:56 on this highly regulated industry,
0:04:59 there’s actually a structural and regulatory constraint
0:05:02 on the amount of supply that we allow into the industry
0:05:03 on any given time basis.
0:05:04 – Right, through education you’re saying?
0:05:06 – Education is one, certainly the number
0:05:08 of accredited medical schools is constrained,
0:05:10 both in terms of actual accreditation,
0:05:11 but also in terms of funding.
0:05:13 Standing up of medical schools is actually funded
0:05:16 by federal dollars and there’s just a physical limit on that.
0:05:18 The other version of that is licensure.
0:05:20 So once you get through your training,
0:05:22 you have to be licensed at the state level
0:05:24 to be able to practice medicine.
0:05:26 And again, there’s only a certain number of licensures
0:05:27 that are approved on a given basis.
0:05:29 You have to renew that every few years.
0:05:31 So there’s a structural constraint in the industry level
0:05:33 that constrained the supply side.
0:05:35 In addition to that, once you are a doctor,
0:05:36 there’s all sorts of challenges,
0:05:38 some of which you alluded to,
0:05:41 of the job itself that have led to constraints.
0:05:43 One is certainly the aging population.
0:05:45 There’s a ridiculous stat that in the last couple of years,
0:05:48 the percentage of doctors that are over the age of 60
0:05:51 has reached like 45%, which is crazy.
0:05:54 And so there’s this huge silver tsunami dynamic
0:05:56 of what happens when all those doctors retire.
0:05:57 There’s gonna be a huge gap just purely
0:05:59 on the basis of age.
0:06:00 There’s also the burnout issue.
0:06:01 Just given everything that happened
0:06:03 during the pandemic in particular,
0:06:05 we put so much burden on our frontline staff
0:06:07 and what we were asking of them to do
0:06:08 was obviously very challenging.
0:06:11 So doctors and nurses have left the workforce actually
0:06:14 in droves because of that burnout issue.
0:06:15 There’s lots of surveys that show
0:06:18 that upwards of 80% of doctors and nurses
0:06:20 will say the number one issue in their lives
0:06:22 is burnout due to their jobs.
0:06:24 And again, another ridiculous stat,
0:06:28 upwards of 300,000 clinicians left the workforce
0:06:29 in the year 2021.
0:06:30 – This is in America specifically?
0:06:31 – In America alone.
0:06:33 And even in the last couple of years,
0:06:37 about 7% of the country’s total active physician workforce
0:06:40 left their day jobs as being clinicians.
0:06:43 Even if we’re able to train like 100% of the capacity
0:06:45 of our medical system, of our licensure system,
0:06:46 the folks are just not sticking around
0:06:49 because the job itself has so much overhead.
0:06:51 And the last thing I’ll say is that doctors
0:06:52 are increasingly taking on more and more
0:06:55 administrative functions and academic functions.
0:06:58 So another amazing stat is I used to live in Boston.
0:07:01 Boston is actually the highest density of physicians
0:07:02 in the entire country,
0:07:04 but they have the worst access issues
0:07:07 because the vast majority of doctors there
0:07:09 are working in academic environments.
0:07:11 So more than half of their time might actually be used
0:07:15 on research, not on actual clinical activity.
0:07:17 So I can go on and on, but these are some of the reasons
0:07:19 why even if you were to look at the number
0:07:21 of doctors and nurses, there’s a lot of other factors
0:07:23 that play into the actual capacity
0:07:25 that’s represented in that pool of labor.
0:07:26 – Maybe we can also just quickly speak
0:07:28 to what the patient feels.
0:07:31 Is it longer wait times at the hospital?
0:07:33 Is it more expensive procedures?
0:07:36 Is it not being able to even get healthcare in some cases?
0:07:38 Can you speak to the second, third order effects
0:07:40 that ladder down to the everyday person?
0:07:43 – I think the most obvious issue is wait times.
0:07:45 And there’s studies that show that the average wait time
0:07:46 for let’s say a specialist appointment
0:07:49 across the country is like 50 days.
0:07:51 And it can range from 27 days on the bottom end
0:07:53 and up to 90 days on the high end
0:07:55 and even worse for certain subspecialties.
0:07:57 And so that alone is obviously an issue.
0:08:00 The other related issue is that even when you’re able
0:08:02 to book an appointment, let’s say a month out,
0:08:04 there’s data that shows that after the second week,
0:08:07 like after 14 days, basically the no show rate
0:08:08 goes way up, right?
0:08:10 Because your likelihood of actually committing to that date
0:08:12 and not canceling or not just showing up
0:08:13 goes through the roof.
0:08:15 And so one of the exacerbating issues
0:08:17 is that these slots are booked
0:08:20 and therefore the doctor is saving that time.
0:08:21 But a lot of that just goes unused
0:08:23 simply because the patients aren’t showing up
0:08:26 because they’ve waited so long for those appointments.
0:08:28 To your second point, if I have an issue
0:08:29 and I need to see a doctor immediately,
0:08:32 the longer I wait, the longer the chances is
0:08:33 that it gets worse, right?
0:08:34 I might end up in the emergency room,
0:08:37 which obviously can be 10, if not 100 times more expensive
0:08:41 to our healthcare system than that initial physician visit.
0:08:43 Another sort of issue, especially just given
0:08:44 how much is happening around the consumer
0:08:46 healthcare landscape these days.
0:08:49 The other challenge is that the less easy it is
0:08:50 for you to see an actual doctor,
0:08:52 the more likely that you’re going to go online
0:08:55 and try to self diagnose and get advice from folks
0:08:56 who might not be medically positive.
0:08:57 – Doctor Google. – Exactly.
0:08:59 And so those are some of the major issues,
0:09:00 but I think the wait time piece
0:09:02 is the underlying fundamental challenge
0:09:04 that leads to all these negative outcomes.
0:09:06 – Yeah, and I mean, you just spoke to this idea
0:09:08 of Doctor Google and that’s where technology maybe
0:09:09 is not playing the right role,
0:09:13 but it sounds like technology maybe can play a better role
0:09:14 in fixing this whole thing.
0:09:16 And I know there’s many different facets to this,
0:09:17 but I just love to hear your take on that.
0:09:20 Is it realistic that technology can actually come in
0:09:21 and curb some of these issues?
0:09:23 – I mean, I think the way to think about it
0:09:24 is absent technology.
0:09:26 How would we solve this problem?
0:09:27 And I think it boils down to like,
0:09:30 how long are we willing to wait to solve this problem
0:09:32 through just simply growing the supply side of our market
0:09:35 and also at what cost to take those seven to 10 years
0:09:37 to like fully train a doctor is not a cheap endeavor.
0:09:39 And so that’s really where I see the opportunity
0:09:40 for technologies to say, listen,
0:09:44 what if we just assume that there are a fixed number
0:09:45 of clinicians in our system?
0:09:47 And if we just have to go with the number of doctors
0:09:48 and nurses that we have today,
0:09:51 how can we increase supply and capacity?
0:09:52 Well, there’s two ways.
0:09:55 And this is where this super staffing concept comes in,
0:09:58 is how do we super staff by enabling
0:10:01 every individual provider to be able to be more productive
0:10:02 in their day to day work?
0:10:05 So that could come in the form of things like co-pilots.
0:10:07 How can I ensure that the amount of time
0:10:10 that a given doctor is taking to make any given decision
0:10:13 at the point of care in real time can be shrunken down
0:10:15 such that they are just far more efficient
0:10:16 with their decision making
0:10:20 and that the caliber of the decision that’s being made
0:10:23 can be at parity with, again, the best of their peers
0:10:26 on the basis of great data and great decision support.
0:10:28 So I think co-pilots is one way that you could imagine
0:10:30 unlocking additional capacity
0:10:34 within the current existing pool of clinicians.
0:10:35 And then a second vector,
0:10:37 which is a different form of super staffing is,
0:10:40 how do you augment the current labor pool
0:10:42 with autonomous agents?
0:10:45 And largely the form that’s taken today is,
0:10:48 if you were to unbundle the job of a doctor or a nurse,
0:10:50 there is a very significant portion of what they do
0:10:52 on a day to day basis that has nothing to do
0:10:53 with patient care,
0:10:54 that has nothing to do with the use
0:10:56 of their unique clinical judgment.
0:10:59 So can you take those tasks largely administrative in nature
0:11:02 or just communications-based and unbundle those
0:11:05 and give those to an autonomous agent
0:11:09 to then do a much better scale with much better accuracy
0:11:11 with much better reliability, such that, again,
0:11:13 the clinicians can actually focus on what they do best.
0:11:15 – And do you have a sense of what portion that is?
0:11:18 I know it will differ based on specialty and doctor,
0:11:21 but are we talking that this administrative portion
0:11:24 is like 3% or is it 30%?
0:11:24 Do you have a sense?
0:11:26 – Yeah, there’s actually studies that look at high motion
0:11:29 of what percentage of a clinician’s time is being spent
0:11:31 on non-clinical patient-facing care,
0:11:33 and it can be upward to a 50%.
0:11:37 Like half of their job is mired in these clerical tasks
0:11:38 that, again, take their time away
0:11:40 from actual care delivery.
0:11:42 Imagine them was doubling the capacity
0:11:44 of a given clinical labor pool
0:11:46 simply by applying some of these unbundling tactics.
0:11:48 – Yeah, and are we seeing tools like this already
0:11:49 on the market?
0:11:53 Where are we in the trajectory of LLMs or AI in particular
0:11:56 entering this industry and actually reshaping it?
0:11:57 – Yes, and this is what’s so exciting right now
0:12:00 is that we are truly in like a golden era
0:12:02 of healthcare technology adoption.
0:12:05 And one of the dynamics that we believe is driving
0:12:08 this rapid adoption cycle of these new AI tools
0:12:10 is what we call this kind of leapfrog dynamic
0:12:11 that we’ve talked about in the past,
0:12:14 where healthcare historically has been a laggard
0:12:16 with respect to software-based technology.
0:12:18 And we’ve always thought that was to the detriment
0:12:20 of our industry and everyone else has these great workflow
0:12:22 tools, these SaaS products.
0:12:23 Because we didn’t do that as an industry,
0:12:26 we don’t actually have the sunk cost bias
0:12:28 that other industries are facing right now
0:12:30 where they say we just spend billions of dollars
0:12:32 laying out these workflow tools.
0:12:33 We trained an entire generation of our workforce
0:12:35 to use these workflow tools.
0:12:36 And now we’re being faced with the decision,
0:12:39 okay, gosh, do I need to spend yet another billion dollars
0:12:42 on these new AI products completely rip and replace
0:12:44 this last generation of technology capabilities
0:12:47 and retrain my workforce on a whole new paradigm
0:12:49 of how they do their work versus in healthcare
0:12:51 where we’re just simply leaping from,
0:12:53 we can’t throw enough bodies at the problem
0:12:55 to how do we actually augment our bodies
0:12:58 and extend through these AI capabilities.
0:13:00 And so we believe that’s driving
0:13:01 a lot of the rapid adoption cycle
0:13:02 that we’re seeing right now.
0:13:04 There’s a few examples of this.
0:13:07 One is, so I mentioned co-pilot versus autonomous.
0:13:10 On the co-pilot side, there’s this class of products
0:13:12 called ambient scribe products,
0:13:15 where it’s a way for a doctor and a patient
0:13:16 to be having a visit.
0:13:19 And there’s an ambient scribe literally listening
0:13:21 so the doctor can actually make eye contact with you.
0:13:22 They’re not sitting there on their laptop
0:13:25 and that AI scribe is specifically tuned
0:13:28 to be able to translate that conversation
0:13:30 into medical documentation that is compliant,
0:13:33 that facilitates billing in the appropriate way,
0:13:35 that completely captures the clinical context
0:13:37 of that particular encounter
0:13:38 and then can be saved directly
0:13:39 into the electronic health record.
0:13:41 And so that is a class of products
0:13:43 that has really taken off.
0:13:45 Doctor friends that I talked to are all using,
0:13:48 one of these products is completely transformed their lives.
0:13:50 They’re getting back hours of the day
0:13:53 that are no longer being spent doing manual documentation,
0:13:55 which is a requirement of their job.
0:13:59 Another example is a class of autonomous communication
0:14:01 platforms where one of the biggest ways
0:14:03 by which the healthcare system communicates with us
0:14:05 is through call centers, right?
0:14:06 I think certainly in senior populations,
0:14:09 you tend to see that folks want
0:14:10 some degree of personal connection
0:14:12 and obviously they’re much more used
0:14:13 to using the phone modality.
0:14:15 And so there’s companies like Hippocratic AI
0:14:19 in our portfolio that has built an army of super staff,
0:14:21 of folks who can actually play the role
0:14:23 of either an outbound call center agent
0:14:26 that is trained on very specific protocols
0:14:28 that are infinitely patient
0:14:30 and can actually execute a number of tasks
0:14:32 that otherwise would rely on humans to do.
0:14:34 And what we’re seeing is it’s remarkable.
0:14:39 Number one, most hospitals have about 10 to 20%
0:14:42 of their headcount open at any given time.
0:14:44 And they simply just can’t find enough people
0:14:46 to fill those roles, namely in the call center setting.
0:14:47 And so what they’re saying is, gosh,
0:14:49 if we can’t hire people fast enough,
0:14:51 like how amazing is it that we could spin up
0:14:54 not just one or two, but 10 or 20
0:14:55 of these autonomous AI agents
0:14:57 that are actually again doing the same job
0:14:58 as these humans would do.
0:15:00 And what’s also incredibly interesting
0:15:02 and part of the why now for why we think this trend
0:15:05 is gonna be at a tipping point is they’re starting to say,
0:15:07 okay, why can’t we just pay for this
0:15:11 out of our labor budgets instead of our IT budgets?
0:15:12 And one of the historical constraints
0:15:14 with healthcare technology has been
0:15:18 that the portion of the budget of let’s say a hospital
0:15:20 that’s spent on IT has been a tiny fraction
0:15:22 of what it is in other industries.
0:15:24 The average is like two to 5% of budget
0:15:28 goes towards IT and healthcare versus like 15 to 30%
0:15:30 in like banking or financial services.
0:15:33 And so now we have this huge unlock opportunity to say,
0:15:36 rather than be constrained by that tiny IT budget,
0:15:39 what if we tapped into the 60, 70% labor budget
0:15:42 and actually be able to create additional leverage
0:15:44 by deploying AI in that fashion.
0:15:46 So these are some of the fascinating trends
0:15:49 that we’re seeing already happening in market
0:15:51 that demonstrate the sort of rapid adoption cycle
0:15:53 of these new AI products.
0:15:55 – Is it just that there is such clear demand
0:15:58 to fill these roles that we’re seeing something different
0:16:00 today or why does it seem like there has been this adoption
0:16:03 or like jumping to participate in this AI wave
0:16:05 when we didn’t really see that in the past wave?
0:16:08 – There is just a very high burden of integration.
0:16:09 At the end of the day, if you’re gonna be introducing
0:16:12 a new software tool into someone’s workflow,
0:16:15 it has to deeply integrate with an EHR type system
0:16:16 or whatever other practice management system
0:16:17 they might have.
0:16:19 And those systems are 40, 50 years old,
0:16:21 they don’t always have great APIs.
0:16:23 And so just the amount of work and time
0:16:26 and customization that’s necessary to get those in
0:16:29 has been a huge barrier to adoption in the past.
0:16:30 The second thing I would highlight
0:16:32 is that these are just magical products.
0:16:33 We have a company called Ambience Healthcare
0:16:34 in our portfolio that creates
0:16:35 one of these ambient scribe tools.
0:16:37 And when you look at the testimonial videos
0:16:38 that they have from their doctor users,
0:16:40 it’s unbelievable.
0:16:42 I mean, these folks are like literally in tears
0:16:44 saying that they’ve been a doctor for 30 years
0:16:47 is like the first time that they’ve actually enjoyed
0:16:48 the job of being a doctor
0:16:51 because it’s just allowed them to focus on the patient.
0:16:53 The last waves of technology products
0:16:55 that have been given to them
0:16:58 have just increased the amount of complexity of their jobs,
0:17:00 whereas these are doing exactly the opposite.
0:17:02 – Are there challenges that you’re seeing on the ground,
0:17:03 whether it’s the tools today
0:17:05 or the things that you see on the horizon?
0:17:06 We were talking about healthcare.
0:17:08 So it could be regulation or some other thing
0:17:11 that you think maybe actually producing friction
0:17:12 for us to implement this.
0:17:14 – I think the obvious thing is just
0:17:15 the stakes are much higher,
0:17:16 obviously in the healthcare setting
0:17:18 and the hallucination issues that you see
0:17:20 with kind of the generalist foundation models,
0:17:22 like that is just a non-starter
0:17:24 if you have issues like that in this particular setting.
0:17:27 And so what we have seen and perhaps why there was
0:17:28 maybe a little bit of lag relative to some of the
0:17:30 generalist foundation models
0:17:31 in terms of adoption and healthcare,
0:17:33 but these companies that I’m describing
0:17:36 have built specialist models in healthcare.
0:17:38 One of the issues with the generalist foundation models
0:17:40 is that, I mean, not an issue.
0:17:41 It’s obviously a huge feature
0:17:42 that they are trained on internet data,
0:17:45 like the entirety of all publicly available human knowledge.
0:17:48 But the reality of healthcare is that the majority
0:17:51 of the esoteric protocol specific,
0:17:53 compliance specific information
0:17:55 tends to not just be online.
0:17:57 It tends to be in these proprietary systems
0:17:59 behind the firewall,
0:18:01 embedded into paper-based binders and things like that.
0:18:03 And so because of that,
0:18:04 it is necessary for these companies
0:18:06 to build specialist models
0:18:08 that are trained on a different set of data
0:18:10 than they might extend some of those generalist models,
0:18:13 but they do need to ingest a different set of information
0:18:15 and specifically be tuned for safety.
0:18:17 And so people do need a specialist approach
0:18:19 to tuning the foundation models
0:18:21 to perform appropriately in this setting.
0:18:23 Another one would be exactly what you said,
0:18:26 the regulatory compliance issues associated with
0:18:28 the use of autonomous AI in particular
0:18:32 and how that bumps up against FDA approval criteria,
0:18:35 under what circumstances do you need to get a AI tool
0:18:37 approved by FDA or not?
0:18:38 – Do you?
0:18:40 – Yes, in fact, one of the benefits that healthcare has had
0:18:43 is that we are the only industry that actually has
0:18:47 an existing regulatory framework for approving products,
0:18:48 especially in the clinical setting.
0:18:52 So there are already hundreds of AI radiology models
0:18:55 or other models that sort of help automate the assessment
0:18:58 and diagnosis of a given set of patients
0:19:00 based on a certain area of medicine.
0:19:02 But how does that apply to a gen AI model?
0:19:04 There’s a lot of work being done right now by the FDA
0:19:06 to really extend those models to be appropriate
0:19:08 for these far more dynamic systems
0:19:10 that just have a lot more uncertainty
0:19:13 than your standard ML predictive model modality,
0:19:16 which is really what that system was designed for.
0:19:18 The last thing I’ll say is reimbursement.
0:19:21 There are some early signs that insurance companies
0:19:25 are creating actual billing codes for AI,
0:19:29 where they now say that when AI enabled visit
0:19:31 or an AI enabled diagnostic
0:19:33 can now be reimbursed differentially.
0:19:35 And you can get credit for using these AI systems
0:19:37 in a different way than if you were just not.
0:19:38 And so that’s really exciting
0:19:41 to see these new paradigms persist.
0:19:42 But I think there’s still very much a long pull
0:19:45 until we have the ability to say, okay,
0:19:48 if I have a fully autonomous like AI therapist,
0:19:50 Slingshot AI, which is one of our portfolio companies,
0:19:54 how can we ensure that those will be systematically reimbursed
0:19:57 at least that parity with what a human therapist
0:19:58 might get reimbursed for,
0:20:00 that has not yet been defined in our industry.
0:20:03 So I think there’s a lot of good early signs
0:20:06 that there’s a willingness by the system
0:20:08 to embrace these new paradigms,
0:20:09 but still a lot of wood to chop
0:20:11 in terms of defining a comprehensive set of ways
0:20:12 to get that in.
0:20:13 – Do you think that this has the potential
0:20:16 to even reshape the ecosystem that already exists
0:20:17 and not just augment it?
0:20:19 – Yeah, I think there’s a lot of really amazing things
0:20:21 to consider like what this unlocks
0:20:23 in terms of both the existing clinical paths
0:20:25 and then things that we don’t even do today
0:20:27 that are now possible to do.
0:20:29 So one really interesting one is just part of the constraint
0:20:31 of our current clinical supply
0:20:33 is that we do have state-by-state licensure, right?
0:20:35 So if I become a doctor,
0:20:37 I’m basically constrained by my geography.
0:20:39 And if there’s a poor patient
0:20:42 who happens to need my services in Montana
0:20:44 and I’m not licensed there, but I do have capacity,
0:20:45 I’m not able to serve that person.
0:20:48 I think what this whole kind of AI world unlocks
0:20:52 is imagine that you had national scale capacity
0:20:54 and the sort of the routing challenge
0:20:56 becomes a really unique one
0:20:58 that actually could leverage AI, right?
0:21:01 So you have this sort of national scale routing opportunity
0:21:03 to just simply take advantage of the latent capacity
0:21:05 in the system at national scale.
0:21:06 There’s this new concept
0:21:08 that actually one of our portfolio companies introduced
0:21:11 called asynchronous medicine, which means, okay,
0:21:13 the typical way we think about seeing a doctor
0:21:16 is you show up at their office, it’s a 30 minute visit,
0:21:17 it’s face-to-face and live,
0:21:19 and it’s highly transactional though, right?
0:21:22 So you’re in and you’re out and then months might go by
0:21:24 or years might go by between visits,
0:21:26 whereas asynchronous care is something
0:21:28 that completely flips that paradigm on its head
0:21:31 and says what if your doctor was in the cloud
0:21:34 and anytime of the day, 24/7, you could text them,
0:21:37 you could send them a message and they would respond,
0:21:40 but the level of ongoing engagement
0:21:42 that you can have like the friction to anyone encounter
0:21:44 with your doctor goes way down
0:21:46 because you can text them at any time of day.
0:21:49 And that’s a new paradigm that again is purely enabled
0:21:52 through these AI-powered clinic models
0:21:55 where you can have these doctors who are available 24/7,
0:21:58 who are using AI as co-pilots to enable them
0:22:02 to be hyper responsive to any question that comes in.
0:22:04 And so I do think that the nature of maybe how doctors
0:22:06 do get trained and like what specialties actually exist
0:22:09 that don’t exist today that are purely available
0:22:11 in the post-LLM world.
0:22:13 I think a lot of those contours
0:22:15 are very dynamically changing as we speak.
0:22:17 – Maybe just looking to 2025 in particular,
0:22:19 since this was your big idea for next year,
0:22:21 is there anything that you’re specifically looking out for
0:22:23 or excited to see?
0:22:26 – We’ve just never seen the rate of technology adoption
0:22:29 that we’re seeing now ever in the history of healthcare IT.
0:22:30 We’ve been waiting for these moments
0:22:32 to take what we know to be the problems
0:22:34 and we know what the solutions can look like,
0:22:37 but there’s just been again like this sort of a barrier
0:22:39 between the problem space and the solutions.
0:22:41 And I think what we’re seeing now are signs
0:22:42 that perhaps that wall has been broken
0:22:44 and we’re at a breaking point
0:22:45 where the industry needs these solutions
0:22:48 and therefore we’re gonna see a much higher adoption
0:22:49 than we’ve ever seen before.
0:22:51 I think it will come in the form of these super staff.
0:22:53 I think we will increasingly see adoption
0:22:56 of these AI tools in the form of labor
0:22:59 and coming from completely different budgets,
0:23:00 from completely different buyers
0:23:02 than we’ve historically seen.
0:23:04 I think this is all to lay down the groundwork
0:23:06 to truly unlock the latent capacity
0:23:08 that we have on the supply side of our market
0:23:10 that’s been sitting there all along,
0:23:13 but just hasn’t had the right technology modality
0:23:14 to unlock it at scale.
0:23:19 Julie just talked, AI applied to healthcare,
0:23:22 spawning a fundamental rewrite of buyers and budgets,
0:23:26 but AI may also have the potential to rewrite compliance.
0:23:29 Companies in banking, insurance, and healthcare industries
0:23:31 spend countless hours in millions of dollars
0:23:32 staying in compliance.
0:23:34 Today, banking and insurance regulations
0:23:36 spend tens of thousands of pages.
0:23:40 SBA landing documentation alone exceeds 1,000 pages.
0:23:44 That was Angela Strange, general partner at A16Z.
0:23:46 Here’s Angela’s big idea.
0:23:50 – In 2025, regulation is gonna become code.
0:23:52 Regulation is one of the biggest costs
0:23:54 to existing companies, banks and insurance companies
0:23:56 and healthcare companies specifically,
0:23:59 and a huge burden to new entrants coming in
0:24:00 and to innovation.
0:24:03 And I think in this country, we have a system
0:24:05 where we efficiently add regulation,
0:24:06 but we never take away regulation.
0:24:08 And so this problem keeps getting worse.
0:24:11 And finally, with the advent of AI,
0:24:13 technology can help solve this problem.
0:24:15 – You mentioned the fact that we add regulation every year.
0:24:18 Just how lofty are these regulations?
0:24:21 – If you look at just federal banking regulation,
0:24:25 right now there are 50,000 different rules and regs
0:24:26 across different agencies,
0:24:29 and that’s up from 10,000 in 1970.
0:24:31 – Has compliance always been this burdensome
0:24:34 or like what’s actually driving that increase?
0:24:35 – Here’s how it’s done today.
0:24:39 You’ve got a large thousand page PDF documentation.
0:24:44 A software company will make a fairly brittle workflow tool.
0:24:46 And then you will have to hire many people
0:24:48 to work inside brittle workflow tool
0:24:50 and figure out is this compliant or not compliant.
0:24:53 So for instance, anti-money laundering.
0:24:54 We’re gonna compliance monitoring tool.
0:24:57 It sends up alerts if they think Stephanie is laundering money.
0:25:00 And then somebody who works in that compliance department
0:25:01 is gonna have to do a bunch of research
0:25:02 across a bunch of web sources
0:25:04 and other different databases
0:25:06 and write up a small report, yes, no.
0:25:09 When that goes wrong, it goes really wrong.
0:25:11 And most recently, big headline,
0:25:14 TD Bank was fined $3 billion
0:25:17 for having a long backlog of these compliance alerts.
0:25:19 – If you actually break down
0:25:21 what they were failing to monitor,
0:25:25 we’re talking about tens of thousands of detection alerts,
0:25:26 thousands of investigation cases.
0:25:28 These are really big numbers.
0:25:30 And to your point, the kind of things
0:25:32 that are really hard for startups to manage.
0:25:34 So maybe we can talk about that.
0:25:36 How does that impact the economics?
0:25:38 Whether it’s the compliance officers they need to hire
0:25:39 or just like dealing with these fines,
0:25:41 how does this actually work out?
0:25:44 – The large banks right now up to 15% of their staff
0:25:46 is just in compliance.
0:25:48 You went into the fourth fastest growing job
0:25:50 in the United States in the last 20 years.
0:25:51 – Compliance officers?
0:25:52 – Compliance, surprise, I glad the witness said, right?
0:25:53 – Yeah, yeah, yeah.
0:25:55 – So you can’t hire enough people.
0:25:56 They’re completely underwater, right?
0:25:59 I’m sure the compliance team at TD was working very hard.
0:26:01 So we need a technology solution.
0:26:03 So how is this going to happen?
0:26:05 I believe that every white-collar worker
0:26:07 eventually is gonna have a co-pilot.
0:26:10 And some of those roles will be replaced by agents,
0:26:11 but we’re talking regulation.
0:26:13 So we’re not just gonna let LLMs out into the wild.
0:26:15 We’re gonna use them to turbocharge
0:26:17 very hardworking compliance people.
0:26:20 And that’s exactly where some companies are entering.
0:26:22 So for instance, you would get an alerts.
0:26:24 We think so-and-so is money laundering.
0:26:25 They can go do the web searches.
0:26:26 They can check the databases.
0:26:29 They can bring all of this information together.
0:26:31 And now instead of a compliance officer
0:26:33 having to do hours of research,
0:26:35 they can look at this report generated
0:26:37 by their friendly compliance co-pilot.
0:26:39 Yes, no, that’s a five-minute task
0:26:41 versus a several-hour task.
0:26:43 Why do you think LLMs in particular
0:26:45 are uniquely suited for this?
0:26:50 ‘Cause they’re able to apply some form of judgment.
0:26:52 And so let’s take a very specific example
0:26:53 near and dear to our hearts.
0:26:55 Anytime we want to tweet something,
0:26:57 we are a registered investment advisor.
0:26:59 We have a great compliance team.
0:27:00 We have to send our tweets over
0:27:02 and they will tell us if we are compliant
0:27:04 with the SEC marketing role.
0:27:06 I imagine nobody listening has actually read that rule
0:27:07 is 400 pages.
0:27:09 So our choices are you could have a software
0:27:14 that creates an if, then, this, that type decision tree
0:27:19 or you can have a LLM ingest that rule
0:27:21 and they’ll be able to make those judgment calls
0:27:23 based on what’s in that document
0:27:24 if my tweet is compliant or not
0:27:27 and then it could come back with specific modifications.
0:27:29 And there’s companies, for instance, Norm AI
0:27:31 is one that’s doing exactly that.
0:27:33 It is codifying specific regulation
0:27:34 such that you can send it things
0:27:36 and it’ll come back and tell you the answer.
0:27:39 – Do you feel like from the types of products
0:27:41 that you’ve seen that they are at the point
0:27:44 where they can sufficiently interpret those gray areas
0:27:47 that sometimes do exist in compliance?
0:27:50 – So LLM’s hallucinating for my consumer colleagues,
0:27:51 that’s a feature.
0:27:53 It comes out with great creative fun
0:27:55 and enhancing things in the world of compliance
0:27:56 that is definitively a bug.
0:27:58 Smarter entrepreneurs are getting around this
0:27:59 in two different ways.
0:28:03 One, it’s much more of a co-pilot versus a full agent.
0:28:07 And then they’re able to test enough edge cases around
0:28:11 that the likelihood of hallucination is very, very, very low
0:28:12 based on how they fine tune the model
0:28:15 to that specific regulation.
0:28:17 The other piece is a big part of compliance
0:28:20 is being able to show that you’ve gone through
0:28:22 all of the steps to be compliant.
0:28:25 And so versus a black box coming back with,
0:28:28 yes, we’ve onboarded this business in a specific way,
0:28:30 it’s gonna show exactly all of the checks that you’ve done,
0:28:32 all of the documents you’ve gotten,
0:28:33 what you’ve pulled out of it.
0:28:37 And so it’s a very clear delineation of what’s been done.
0:28:39 But again, it saves the officer having to do all that work,
0:28:40 they can just check the box.
0:28:41 – That makes sense.
0:28:43 And are you seeing these tools on the market?
0:28:46 Give us some examples of what already exists.
0:28:48 – Notoriously banks, insurance companies,
0:28:51 even fintechs can be slow to sell into.
0:28:52 I think there’s a couple things going on.
0:28:56 One, this has become such a burdensome problem
0:28:59 that buyers are very open-minded to new tools.
0:29:01 And two, companies of all sizes have woken up
0:29:03 and said that AI can make a significant difference
0:29:04 to our business.
0:29:06 And so the sales cycles are moving much more quickly.
0:29:10 And so they range from thinking of co-pilots
0:29:12 that go along with your compliance officers.
0:29:14 I mentioned a couple of examples there.
0:29:19 The other thing that’s happening is these older system
0:29:22 of record tools, so take transaction monitoring
0:29:26 that are built pre-AI, they used to be seen
0:29:28 as too risky to rip and replace.
0:29:32 And now what we’re seeing is the new Gen AI first tools
0:29:34 are 10x better versus 2x better.
0:29:36 So for instance, Sardine is a company
0:29:39 that’s got a much more modern transaction monitoring system
0:29:42 and they are seeing customers rip and replace,
0:29:44 optimize, it’s been around for 20 years.
0:29:46 And that’s just not something that we would have seen
0:29:50 five years ago, but it’s because it’s just that much better.
0:29:52 – I assume, as you said, many of these industries
0:29:55 are a little slow to adopt technology.
0:29:57 What stands out in terms of what may actually
0:29:58 stop this adoption?
0:30:02 – It is moving a lot more quickly than I would have thought.
0:30:03 I’ve been investing in and around the space
0:30:06 for the last 10 years and the momentum
0:30:09 and even just the ACVs that some early companies
0:30:11 are getting is quite fast.
0:30:13 I think the two things that come up with the blockers
0:30:16 where the smart founders have gotten around are one,
0:30:18 how do I prevent this from hallucinating?
0:30:20 And I think the two ways are one, showing the work,
0:30:22 and then two, proving that you have fine tuned your models
0:30:24 close enough to this regulation
0:30:25 that that actually doesn’t happen.
0:30:29 And a backup plan as a human does actually resolve it.
0:30:33 But I think the most important can be wedging in
0:30:36 at a pain point that is very acute with the right buyer.
0:30:39 And personas that weren’t sold into that often before
0:30:41 have become very popular.
0:30:41 – Oh, like what?
0:30:44 – Chief compliance officers, like BSA officers.
0:30:47 Like I imagine all of these potential buyers
0:30:48 are now soon getting a lot of interest
0:30:50 from very technical solutions.
0:30:51 – That’s great.
0:30:53 If we think maybe a little more long term
0:30:54 about opportunity here, right?
0:30:57 Given the overhead that exists today,
0:31:01 what are we unable to do because of all of that compliance?
0:31:05 – So think of just a consumer’s life in financial services
0:31:08 and how much better some of these new entrants
0:31:09 have made your life.
0:31:12 So trading now is free.
0:31:13 I would argue that’s in large part
0:31:16 because Robinhood came along, made trading free
0:31:19 and put pressure on all the large brokerages.
0:31:22 For a large portion of the US population,
0:31:25 it was very difficult to get a free checking account.
0:31:27 Time came along, enabled that to be possible
0:31:29 and put pressure on the incumbents.
0:31:32 And so it is good for everyone, for new entrants
0:31:34 to be able to come into this space.
0:31:37 And one of the biggest barriers is just regulation.
0:31:39 And so the easier that we can make it
0:31:43 for smart people of all resources
0:31:44 to be able to start new companies,
0:31:45 it’s gonna be better for consumers,
0:31:47 it’s gonna be better for businesses.
0:31:49 – And as we looked at 2025,
0:31:51 what are you looking for in particular?
0:31:54 – Yeah, I love two categories.
0:31:58 One are old systems of record
0:32:00 that you can do just infinitely better
0:32:03 and especially ones that have an extreme economic impact.
0:32:05 So I mentioned before, small business lending.
0:32:09 Many businesses get loans from the SBA.
0:32:11 It takes on average 90 days.
0:32:12 Why?
0:32:15 Part of that is there’s a 1,000 plus page documentation
0:32:17 about how do you make this loan compliant, right?
0:32:20 So what is that LOS in that process?
0:32:22 It’s very much gonna streamline that.
0:32:24 And it is great for the economy for our small businesses
0:32:26 to be able to get the credit that they need.
0:32:28 So systems of record, 0.1.
0:32:31 And then 0.2 would be any area
0:32:35 where a company needs to hire dozen plus people
0:32:37 to do a task where those people would rather
0:32:39 do something a little bit more strategic.
0:32:41 That can be a very interesting wedge in
0:32:43 for a co-pilot type opportunity.
0:32:44 And then you can start expanding
0:32:46 and doing my task from there.
0:32:47 – So as we look to these changes,
0:32:49 it might be obvious for a big bank like TD
0:32:51 how this can improve their operations.
0:32:54 But why should the everyday person care about this?
0:32:56 – The everyday person should really care about this.
0:32:58 And there’s three main reasons.
0:33:01 One, it’s a big barrier to new innovative companies
0:33:03 coming into the space.
0:33:05 And I’ve mentioned examples like Robinhood, like Chime,
0:33:08 like others that really upend some of the business models
0:33:10 of the larger banks and provide free trading, free checking,
0:33:13 just better consumer services.
0:33:17 And then it’s a hidden cost to a businesses
0:33:19 or consumers everyday life.
0:33:21 It is longer for a business to get a loan
0:33:22 because there’s a lot of regulation
0:33:24 and it’s very difficult to comply with it.
0:33:27 It is more difficult to get a loan modification
0:33:29 that you desperately need to be able to pay your mortgage
0:33:31 because somebody needs to be trained
0:33:34 in the thousand plus any main mortgage servicing guide
0:33:35 so that they’re able to decide,
0:33:37 can you get that loan mod and how will they do it?
0:33:40 So improving regulation, making it code,
0:33:42 making it easy to be compliance with
0:33:45 will improve everyone’s lives and economy.
0:33:46 But if you just look at the chart,
0:33:48 it just keeps growing and growing and growing.
0:33:50 And so this problem is going to get worse, not better.
0:33:52 That’s just federal.
0:33:55 Every state is following exactly the same trend line.
0:33:59 And the opportunity for AI to disrupt
0:34:03 how businesses operate does not stop with compliance.
0:34:06 – AI is automating jobs across traditional service industries
0:34:09 like insurance, law, real estate, and IT.
0:34:11 Though many of these businesses have historically been
0:34:14 low margin and difficult to scale,
0:34:16 some are now leveraging LLMs,
0:34:19 particularly to automate roles involving voice,
0:34:22 email, or messaging to transform
0:34:25 into high margin, scalable models.
0:34:25 – That was.
0:34:27 – Joe Schmidt, I’m a partner on the apps team
0:34:29 here at Andries and Orwitz.
0:34:30 – And here is his big idea.
0:34:33 – My big idea is romanticizing organic growth.
0:34:35 And really what that means is I see the next evolution
0:34:37 of private equity being possible
0:34:40 where small businesses can leverage technology
0:34:43 to basically automate parts of their services organization
0:34:45 and then go and use a better business model
0:34:47 to acquire other small businesses.
0:34:49 – Obviously LLMs, the talk of the town,
0:34:51 penetrating every industry.
0:34:52 But tell me what you’re seeing on the ground
0:34:55 in terms of how these LLMs are reshaping industries.
0:34:58 – It’s just amazing to see how fast adoption has been
0:35:00 across whether it’s horizontal applications
0:35:02 or vertical applications.
0:35:03 And just for this conversation,
0:35:05 maybe we’ll focus on the vertical.
0:35:07 Obviously recently announced our investment in 11x,
0:35:09 which is automating the sales function.
0:35:11 And I don’t think I’ve ever really seen demand
0:35:14 like what we’re seeing for basically automating
0:35:15 the role of an SDR.
0:35:16 And so you’re already seeing it
0:35:17 within the sales organization.
0:35:20 If you think even on a more vertical specific basis,
0:35:21 you know, I see it in like freight
0:35:23 with companies like happy robot,
0:35:25 basically automating the call center operations
0:35:27 around what a freight broker does,
0:35:29 whether it’s like calling and booking a load
0:35:31 or dealing with a new carrier
0:35:33 that a freight broker has to deal with.
0:35:34 We see it in healthcare with another one
0:35:36 of our portfolio companies, Tenner,
0:35:38 which is automating the back office processes
0:35:39 associated with a small health core practice.
0:35:41 And so I think we’re seeing a bunch
0:35:42 of these applications pop up
0:35:43 and frankly have growth rates
0:35:45 on anything I’ve ever seen again.
0:35:46 And it’s, I think it’s just a really exciting time
0:35:49 to be building software and to be building startups.
0:35:50 – Yeah, and I love those examples
0:35:52 ’cause you’re touching so many different industries.
0:35:54 And maybe if we zoom out just a little bit
0:35:56 and think about maybe like a framework for disruption,
0:35:59 how do you think about where maybe this could be applied?
0:36:00 – If you’re interested, go read our blog
0:36:02 because we’ve got a bunch of ideas on it.
0:36:03 – Yeah, and we’ll link to that.
0:36:03 – Yeah, we’ll link in it.
0:36:05 But I think that like the way to just zoom out
0:36:06 and think about this opportunity is like,
0:36:08 where are there a bunch of humans
0:36:10 that are basically dealing with whether it’s voice
0:36:12 or paper processes
0:36:14 where there’s a bunch of unstructured data
0:36:16 that they need to basically synthesize
0:36:17 and then make sense of
0:36:19 and then generate some sort of output.
0:36:21 Obviously that’s what LLMs are super good at.
0:36:23 And so like the very common way of thinking about that
0:36:25 might be like, hey, there’s a call center somewhere, right?
0:36:27 That’s like the healthcare or the insurance example
0:36:28 or the sales example
0:36:30 or maybe you have some sort of,
0:36:31 we call it the messy inbox problem here
0:36:33 where there’s a bunch of different pieces of paper
0:36:35 that are coming in and you need to synthesize them
0:36:37 and put them into some database.
0:36:39 But oftentimes that’s like an internal team.
0:36:41 But then also you could look for wherever there might be
0:36:44 DPO spend like business process outsourcing spend
0:36:46 where maybe a consulting firm is doing this for you.
0:36:47 So those are the hot buttons
0:36:49 is like looking for work like that
0:36:51 where it’s basically something that is,
0:36:53 a bit’s not Adam’s industry,
0:36:54 like you’re dealing with paper,
0:36:55 not physical world things
0:36:59 and something that is conducive to be automated by LLMs.
0:37:00 – Yeah, I have to admit,
0:37:01 when I was reading your big idea,
0:37:03 I thought, isn’t this just like private equity?
0:37:04 So convince me otherwise,
0:37:07 how is this different from traditional private equity?
0:37:09 – First, like what is private equity?
0:37:12 To oversimplify, like private equity is a buyer,
0:37:15 purchases a company using some combination of equity and debt,
0:37:17 traditionally like a good deal of debt
0:37:19 and then basically tries to improve earnings
0:37:21 in the relatively short term.
0:37:23 Normally private equity has like a three to five year
0:37:25 whole period and then sell it to the next buyer.
0:37:28 Private equity firms typically are optimizing for IRR
0:37:29 in general, we’re gonna return
0:37:31 and they’re less focused on ultimately
0:37:34 like a multiple on money return,
0:37:36 basically saying they’re not focused on the biggest
0:37:38 possible version of the company that they can build,
0:37:41 rather they’re focused on making incremental improvements
0:37:42 in this time period,
0:37:43 ’cause that’s like how they think about liquidity
0:37:45 and looks how the model works.
0:37:46 The problem with that is it creates
0:37:48 like a little bit of a perverse incentive structure
0:37:50 where typically they’re focused on one of two ways
0:37:51 of making money.
0:37:53 First is like the most obvious,
0:37:54 like when people hear like private equity,
0:37:56 it’s like, oh no, people are gonna get fired.
0:37:57 But obviously what they’re trying to say is,
0:38:00 hey, this hasn’t been run in the most lean and efficient way,
0:38:02 maybe it possibly could have.
0:38:03 And so, hey, the accounting department
0:38:04 doesn’t need to be this big
0:38:07 or maybe we can consolidate these two different teams
0:38:10 or whatever it might be, that’s the first bucket.
0:38:11 The second bucket normally is saying,
0:38:14 hey, we’re gonna go acquire other small businesses
0:38:15 or large businesses, doesn’t matter,
0:38:16 we’re gonna acquire other companies.
0:38:18 We’re gonna use more equity and debt
0:38:21 and we’re gonna basically buy the next incremental company
0:38:22 for a lower multiple,
0:38:25 roll it into our company and have multiple expansion.
0:38:27 So these are the two very common ways
0:38:28 that private equity firms make money.
0:38:31 These aren’t designed to have these businesses
0:38:32 become the best versions of themselves,
0:38:34 it’s designed to make money.
0:38:34 – In the short term.
0:38:35 – And it does.
0:38:37 So I think what’s interesting about this model
0:38:39 is that it’s actually addressing
0:38:41 like a third leg of the stool in my opinion
0:38:43 where traditionally private equity
0:38:45 isn’t focused on like fundamentally shifting
0:38:47 the way that the business works.
0:38:48 What I’m talking about is actually building
0:38:51 like a technology business and the technology core
0:38:54 to a traditional services company, right?
0:38:55 So in the past, the private equity firm would say,
0:38:58 hey, like we don’t need that many people in accounting,
0:39:00 I’m talking about automating like a large swath
0:39:01 of the company.
0:39:03 And to do so, you need to have a very different view
0:39:06 on returns and you need to have a very different view
0:39:09 on short-term earnings because what will end up happening
0:39:12 is it turns out engineers pretty expensive, right?
0:39:14 And it requires like significant investment
0:39:17 to get this business to a more automated state.
0:39:19 And so you have to not think about IRR
0:39:20 over the next two or three years.
0:39:21 And instead you need to be thinking about
0:39:25 what’s the largest long-term valuable business
0:39:27 that I can build and how do I invest in that future?
0:39:30 So that’s like the very different way of building businesses.
0:39:32 I think you have to structure these businesses
0:39:34 from the start, not as private equity companies,
0:39:37 but as like long-term permanent,
0:39:39 like we’re gonna build the best version of this company ever
0:39:41 and like a company that can go to public.
0:39:43 And frankly, I think the best versions of this company,
0:39:45 like there’s a couple of examples on the public markets
0:39:48 that are a little bit more holed-co
0:39:49 traditional software versions of this,
0:39:53 but one is a very well-known company called Danerher.
0:39:54 There’s another one called Tyler Technologies
0:39:56 based out of Texas.
0:39:57 These are like probably the best examples
0:39:59 of these businesses historically.
0:40:01 But again, they’re not AI native
0:40:04 and they’re basically focused on the first two buckets
0:40:06 of this, which is like financial engineering
0:40:08 and/or like operational efficiency,
0:40:11 not building an AI to their businesses.
0:40:12 – That’s super interesting.
0:40:13 And I think just for the listeners,
0:40:15 let’s try to be really concrete here and paint a picture
0:40:17 of what this could look like,
0:40:19 this idea or opportunity of quote,
0:40:22 AI-powered vertical specific service startups.
0:40:23 – Yeah, that’s a melt ball.
0:40:24 – Yeah, I was gonna say, it’s a lot of Ss,
0:40:26 but what would that look like?
0:40:27 Let’s try to really paint a picture here.
0:40:30 – Okay, so let’s imagine that staff has an insurance agency
0:40:32 in Columbia, Missouri.
0:40:34 Okay, so let’s call it a small insurance agency.
0:40:37 You do three or $4 million in gross written premium.
0:40:39 That generally will mean you do a few hundred thousand,
0:40:42 call it half a million dollars in revenue, that revenue.
0:40:44 Let’s say you have single-digit margins on that revenue.
0:40:46 So you’re living a nice life in Columbia, Missouri,
0:40:48 but this is a small business.
0:40:49 There are thousands, tens of thousands
0:40:51 of these businesses across America.
0:40:54 You’re selling, imagine home, auto, life insurance,
0:40:55 and a little bit of commercial insurance.
0:40:57 At your staff insurance agency,
0:41:00 maybe you have one or two or three people in the back office.
0:41:01 Maybe there’s a CSM that’s dealing
0:41:03 with like net new business that you’re writing.
0:41:05 There’s someone in the back office
0:41:06 that’s dealing with all the paperwork, right?
0:41:08 ‘Cause when you write an insurance policy,
0:41:09 it’s like, hey, I gotta get a bunch
0:41:10 of information from the customer.
0:41:13 I gotta go get the quote from my legacy system,
0:41:14 and I gotta write all this stuff down,
0:41:17 and I gotta deal with all these back office processes.
0:41:20 It’s painful, and it’s a lot of people moving paper around.
0:41:22 That’s the reason your margin is so low,
0:41:23 and you’ve tried to fix this, right?
0:41:25 You’ve tried to buy software,
0:41:26 but the only software in this industry
0:41:28 is really legacy, right?
0:41:30 There’s these two big businesses in insurance.
0:41:32 You don’t really have a great system here,
0:41:33 and so you’ve just been hiring these people,
0:41:35 and you’ve been living this, frankly,
0:41:37 hell for your entire career.
0:41:40 And so what I’m talking about now is I come to you
0:41:42 and I say, hey, Steph, wouldn’t it be great
0:41:44 if I could give you like basically automation
0:41:46 that could take all of that paperwork
0:41:50 that you’ve been paying five, 10, 15 plus percent
0:41:53 of your net margin for, and we can automate it?
0:41:54 And what does that require, right?
0:41:57 It’ll require you going and getting deeply integrated
0:41:58 from the technology side,
0:42:00 deeply integrated into the core systems of record,
0:42:02 leveraging AI agents to basically have
0:42:05 some of those conversations with the customer,
0:42:07 whether it’s on new business, on renewal,
0:42:09 or on client servicing,
0:42:12 and then actually resolving these claims in real time,
0:42:14 or like resolving these service issues in real time,
0:42:16 which is a super interesting opportunity
0:42:18 because then you start to think about, okay,
0:42:20 now Steph’s insurance agent doesn’t need four people,
0:42:23 it might need two people, it might need one person,
0:42:25 and you go from having 5% net margins
0:42:26 to 30% net margins,
0:42:29 and you go and you put a pool in, or whatever it might be.
0:42:31 Or frankly, what I think is really exciting
0:42:33 is you use that extra cash,
0:42:35 and you buy John’s insurance agency that is street,
0:42:37 and now you see what the opportunity is
0:42:38 to run this playbook.
0:42:39 – Yeah, because basically what you’re saying is that
0:42:41 what you put into the technology
0:42:44 can be replicated for insert other business here.
0:42:46 Tell me a little bit more about why you need
0:42:48 to buy the company though.
0:42:49 Could I not create a tool that just sells
0:42:52 into John’s insurance company, Steph’s insurance company?
0:42:53 – Yeah, and I think that’s,
0:42:54 you actually made a really good point
0:42:55 at the end of what you just said,
0:42:58 which is, hey, this can be applied to the next business.
0:43:00 And if you build this software, I think in the right way,
0:43:02 in the right industry vertical,
0:43:04 once you’ve built those core integrations,
0:43:06 in the example that we were just doing,
0:43:07 it’s into the applied epic system,
0:43:09 or whatever core system of records
0:43:12 exists across the small business you’re thinking about,
0:43:14 those systems are, in many cases,
0:43:16 in these legacy industry are pretty ubiquitous.
0:43:18 And so you can repurpose this technology
0:43:20 across each one of these small businesses.
0:43:22 I’m not saying it’s gonna be simple or easy,
0:43:23 it’s all required implementation,
0:43:24 but that’s the opportunity here
0:43:26 is to stay vertical specific
0:43:29 and build this deep kind of understanding of the systems.
0:43:31 But it’s a great question on why not
0:43:34 just sell software to these companies.
0:43:36 Frankly, if you’ve ever talked to a small business
0:43:38 across middle America, like these mom and pop companies,
0:43:41 they are not the most technology savvy customers.
0:43:43 And it’s not that they’re not great businesses
0:43:45 or super educated, it’s just they just aren’t accustomed
0:43:47 to running their companies this way.
0:43:49 If I’m a software vendor, and frankly, I’ve done this,
0:43:51 I’ve gone and I’ve tried to sell software
0:43:54 to mom and pop insurance agency in Columbia, Missouri,
0:43:56 trying to drive adoption is super tricky.
0:43:58 And so in buying the business,
0:44:01 instead of trying to do business development exercises
0:44:02 and convince someone that like,
0:44:04 “Hey, the pot of gold at the end of the rainbow
0:44:06 “is worth it for you.”
0:44:07 Instead, we can now say,
0:44:10 “Actually, this is what we’re gonna run the business.”
0:44:12 Operationally, this is our new cadence and our new workflow,
0:44:14 and here’s how we’re gonna get to the next stage.
0:44:16 – So you’re saying basically, in purchasing the business,
0:44:19 it gives you the leeway to actually shift things fundamentally,
0:44:22 which is really what this is articulating.
0:44:23 Maybe another question is just,
0:44:26 you seem to think this is a really big opportunity.
0:44:28 PE itself is a huge industry.
0:44:29 So why do you think this is actually
0:44:32 maybe a bigger opportunity than traditional private equity?
0:44:35 – So I think that this all comes down to capital efficiency.
0:44:36 And so for example,
0:44:38 buying one of these small companies that I just described
0:44:41 isn’t actually like that expensive of an endeavor.
0:44:42 And if you can do what I’m talking about,
0:44:44 you impact earnings and all of a sudden,
0:44:45 like the first one, two or three,
0:44:48 maybe you don’t immediately pay back using cash.
0:44:51 But after the first few of these transactions,
0:44:52 you start to throw off cash
0:44:55 and that cash can then be used to acquire more
0:44:56 and more of these small businesses.
0:45:00 And so that becomes this beautiful like flywheel effect
0:45:01 to the acquisitions that you can make.
0:45:03 And then that new business is you acquire.
0:45:06 And so unlike traditional private equity
0:45:07 where you’re constantly raising debt,
0:45:09 maybe you need more equity.
0:45:11 And again, it’s much more financial engineering.
0:45:13 Instead, you can actually just compound your investment
0:45:16 over a long period of time and build value that way.
0:45:17 So I think that’s the opportunity here.
0:45:18 – Yeah. And I guess what you’re getting at is
0:45:22 the margin change is so significant where you can do that.
0:45:24 But instead of you’re doing some of the financial engineering
0:45:27 that traditional P relies on,
0:45:30 you’re often not changing the margins from 5% to 30%
0:45:32 or to 50% even in some cases.
0:45:34 – Yeah. And I think there’s a nuanced point here where,
0:45:36 again, if you do this right in the right vertical
0:45:38 and the businesses that you end up acquiring
0:45:41 are worth more to you than to any other buyer,
0:45:43 that’s when you can start to do really interesting things.
0:45:44 Right? So I don’t know if this makes sense,
0:45:46 but let’s say I’m doing something
0:45:48 and Mark is doing something in the same space
0:45:49 for those insurance businesses.
0:45:50 We’re both running the same playbook.
0:45:52 He’s a private equity firm.
0:45:53 I’m running a technology business.
0:45:54 And I go to you and I say,
0:45:57 hey, I have this wonderful technology platform
0:45:58 when you like to come and grow with me
0:46:01 and I can take your earnings from 5% to 20%
0:46:04 versus Mark says, hey, I’ll just pay you a bunch of money
0:46:05 and you’re going to stay at 5%
0:46:07 and maybe I’m going to fire Susan
0:46:08 who runs your accounting, right?
0:46:10 Whose value prop sounds better?
0:46:12 And so in this way, you can one,
0:46:15 like your business is actually worth more to me, right?
0:46:17 Because I can impact earnings in the short term
0:46:18 based on my technology.
0:46:20 And two, you want to work with me
0:46:22 because you see, hey, I’m not going to have to go in
0:46:24 and fundamentally do a bunch of these things
0:46:25 that are really cutthroat.
0:46:26 Obviously, maybe there’s certain people
0:46:27 that are a little redundant.
0:46:29 So I’m not saying this won’t happen,
0:46:30 but I do think that it’s a much better story.
0:46:31 And it’s a growth story
0:46:33 versus just a traditional efficiency story.
0:46:35 – Yeah. And speaking of that growth story,
0:46:36 it sounds like a lot of the growth
0:46:38 is inorganic in this case.
0:46:40 At least that’s what this story is about.
0:46:40 So tell me more about that.
0:46:43 Why the inorganic growth strategy
0:46:45 versus maybe a more organic approach.
0:46:47 – Yeah. So I think that the way to think
0:46:49 about these more traditional services businesses,
0:46:51 oftentimes they’re local enterprises, right?
0:46:53 So again, like in Columbia, Missouri,
0:46:56 you have a reputation as being like a trusted source
0:46:57 of insurance services.
0:46:59 And so people want to come and work with staff
0:47:00 for that reason.
0:47:02 And it’s actually traditionally pretty hard.
0:47:05 It’s not just insurance, but think about real estate, law,
0:47:08 like all of these industries, they’re traditionally local.
0:47:10 And so it’s pretty hard to go add new customers
0:47:11 and grow organically.
0:47:13 If you’re doing that, like it’s pretty special.
0:47:14 In my opinion, the best version
0:47:16 of this inorganic growth strategy,
0:47:18 it’s like you’re acquiring a very sticky,
0:47:21 a very local, geo-specific customer base.
0:47:23 And then you’re improving the business that’s serving them
0:47:25 and freeing up that person who already has
0:47:27 like all of these wonderful connections
0:47:31 in whatever space they are to go do more of that, right?
0:47:32 So now you’re not focused on,
0:47:34 hey, shoot, I’ve got to deal with all of this stuff
0:47:35 in the back office.
0:47:37 I can go to the little league field
0:47:40 or wherever the heck you go to sell insurance.
0:47:41 And basically get to know people
0:47:42 and do business development.
0:47:45 And so now not only are we impacting like earnings,
0:47:48 we’re actually impacting the net new organic growth rate,
0:47:49 which is pretty clever.
0:47:50 The way to think about this in my mind
0:47:51 or the way that I think about it is,
0:47:54 hey, if you were running an enterprise software company,
0:47:56 right, and you were in the public markets
0:47:59 and you’re saying, hey, I could have 18, 20, 24 month paybacks
0:48:01 on a great enterprise customer logo,
0:48:03 that’s the way to think about these small businesses.
0:48:07 Can you find a way to have like strong paybacks
0:48:09 on these new business acquisitions
0:48:11 and treat them much more as just like
0:48:13 an enterprise software sales motion
0:48:15 rather than a traditional private equity firm
0:48:16 acquiring some small business
0:48:18 and doing a bunch of financial engineering on it.
0:48:20 – But obviously this is not easy.
0:48:21 – No.
0:48:21 – This is not operationally easy.
0:48:22 So tell me more about that.
0:48:23 Like as you-
0:48:25 – As I sit here in my venture capital ivory tower,
0:48:27 all these small businesses are simple to automate.
0:48:30 Yeah, I can’t wait for my tweets at me.
0:48:31 – Well, tell me more about that.
0:48:32 So as you think through the challenges,
0:48:35 what is most stark, what’s most clear
0:48:38 as the thing that someone pursuing this would need to solve?
0:48:40 – Yeah, I think the biggest one is just understanding
0:48:42 that like there’s gonna be a bunch of human problems
0:48:42 when it comes to this.
0:48:44 And so you have to be like eyes wide open.
0:48:45 You have to be a great operator
0:48:48 and you have to deal with scaling a human driven business.
0:48:49 So I think that’s the first one
0:48:51 and all the implementation problems around that.
0:48:53 There’s gonna be a bunch of hairy legacy systems.
0:48:55 You’re gonna need to navigate like all the process mining
0:48:57 and figuring out what do I need to build here
0:48:59 to actually drive incremental value.
0:49:01 So that’s the second piece, human level issues.
0:49:02 Once you get through those,
0:49:04 then you can start to do the process mapping,
0:49:05 which is also exciting.
0:49:06 And then the third piece is,
0:49:08 I’ve talked about this a little bit before,
0:49:10 but a lot of these legacy services businesses,
0:49:12 while they may be bits oriented,
0:49:15 like at some point there’s like an atom involved,
0:49:17 someone has to go do the business development.
0:49:19 Like it’s not like you can just fully automate these companies,
0:49:20 but it could be even more complicated.
0:49:22 Imagine you bought saying in the insurance example,
0:49:24 maybe you bought a third party administrator,
0:49:27 which is someone who deals with like insurance claims.
0:49:29 So you could deal with all the back office stuff around
0:49:32 handling insurance claim, accepting all the documents,
0:49:33 figuring out what’s what,
0:49:35 but AI is not gonna drive the truck to the crash.
0:49:38 AI is not gonna figure all that stuff out.
0:49:40 And so I think that’s the other tricky part of like
0:49:42 of these spaces is like identifying
0:49:45 where is the right kind of combination of bits and atoms,
0:49:46 primarily hopefully bits,
0:49:48 and then trying to figure out how to navigate
0:49:50 some of these human problems and implementation problems.
0:49:52 But it’s definitely gonna be hairy.
0:49:53 I think if you don’t do this right,
0:49:56 and you don’t figure out like a very repeatable motion
0:49:58 for these, again, smaller acquisitions,
0:49:59 I think it’s gonna be really challenging
0:50:00 to build a big company.
0:50:01 So that’s the opportunity here.
0:50:03 – What are you looking out for in particular?
0:50:04 Is it a certain type of founder?
0:50:07 Is it a certain motion that you see someone doing
0:50:08 that you think is really intriguing?
0:50:10 And you think, okay, they’re on the right track here.
0:50:12 What are you really looking for?
0:50:12 – First things first,
0:50:15 I’m looking for an entrepreneur or entrepreneurs
0:50:17 that have bottoms up understanding
0:50:18 rather than tops down understanding.
0:50:19 And what does that mean?
0:50:21 I wanna find entrepreneurs that have earned secrets
0:50:22 in these end markets, right?
0:50:24 So somebody who has maybe sold
0:50:26 to the thefts insurance agency
0:50:28 or someone who has run thefts insurance agency
0:50:30 or whatever the corollary is
0:50:31 for each one of these markets.
0:50:34 It’s very hard to come in and know the shorthand.
0:50:36 You have to trust me that I’m gonna come in
0:50:37 and help you bring your business
0:50:39 and take your business to the next level
0:50:40 if I’m gonna be that acquirer.
0:50:42 And so I think you need to be a very special person
0:50:43 on that front.
0:50:45 The second piece is an understanding of,
0:50:48 hey, can this industry be automated?
0:50:51 Isn’t it sufficiently bits oriented
0:50:53 to the point where AI can actually have an impact on that?
0:50:53 – Like what percentage?
0:50:54 – Right, exactly.
0:50:55 Is it the right kind of work
0:50:58 that is conducive to being automated by AI?
0:50:59 That’s another piece.
0:51:00 And then I think the third piece
0:51:04 is hopefully there’s a very large fragmented market
0:51:07 of acquisition targets such that like you’re not
0:51:09 basically going in and having to compete
0:51:11 with the larger mid to large size
0:51:13 traditional private equity deals.
0:51:15 And so if you can find one of these end markets
0:51:18 where there’s a significant number of mom and pop businesses
0:51:20 where you can go and run that cack-minded playbook
0:51:22 of acquiring and implementing
0:51:24 and building a repeatable motion around that,
0:51:27 I think there’s just insane room to run
0:51:29 and massive businesses to be built.
0:51:30 And so those are the combination
0:51:31 of factors that I’m looking for.
0:51:34 And hopefully someone out there is considering this
0:51:35 the way that I would run it
0:51:37 is I would go find a design customer,
0:51:39 start building software and then show yourself
0:51:40 and frankly everyone else
0:51:41 that you can actually start
0:51:43 to move the needle on earnings.
0:51:44 And that’s when I think you know
0:51:45 you got the tiger by the tail
0:51:47 and it’s time to go and run fast.
0:51:49 – All right, I hope these big ideas
0:51:52 got you geared up and ready for 2025.
0:51:55 Stay tuned for our final segment part four
0:51:56 where we dive into the big ideas
0:51:59 straight from our A16Z crypto team
0:52:03 who submitted 14, yes, 14 big ideas for the new year.
0:52:04 We’ll see you then.
0:52:07 (gentle music)
0:52:09 (gentle music)