May 2024: AI eats OpEx
And the second order effects
In healthcare, AI charting is eating scribes. We had scribes in only a handful of clinics because even though they reduce burden on providers, they end up being too expensive to deploy in all visits, particularly if a provider or clinic has low volume.
AI Charting helped unlock this at Carbon in 3 key ways.
First recorded audio is typically less lossy than a human scribe (eg you can go and re-listen to the audio if you want to, and then clarify what was said yourself).
Second, because it’s software, it’s financially feasible for all providers to have a scribe, not just the most busy ones. This is great because the cognitive load that drives provider burnout doesn’t only occur when a provider is busy seeing patients. The benefit of lower cognitive load and more complete documentation can be felt regardless of how busy the provider is.
Third, the quality and completeness of our documentation increased because providers were vocalizing their thought processes to ensure they got captured by the scribe.
The first order effect of this is that we’re less likely to need human scribes (scribes did more than transcribe, so it’s not a 1:1 trade). One second order effect is that our documentation review process becomes far less onerous because ultimately you can read the entire transcript of the visit rather than interpret the provider shorthand, and generally you can tell exactly what happened during the visit. This means that you also save provider expense (and providers can focus more on patients than on going back to a chart for any reason) because they are less cognitively burned out for the same number of visits. Overall, we’re now adding voice/audio as a new type of data in healthcare that didn’t exist before, and that has a lot of implications we’re only just wrapping our minds around:
Patients (and maybe even payors) will want the audio
AI scribes will eventually be incentivized (or even required)
Reviewing + recreating charts will be a lot easier to automate
We’ll be able to use models to provide a second (or 3rd or 4th) opinion
It’s happening everywhere else
Healthcare happens to be what I am close to, and I’ve seen the effect on our practice so I have high conviction about how it will play out (outside of Carbon Health, I think services like Abridge will deliver some of the benefits we’ve seen in our business, to large health systems). I believe in many other industries, a similar dynamic will play out (In full disclosure I’ve invested in a bunch of these companies, so they’re easy to use as examples because I more intimately understand what they do).
In multiple parts of healthcare revenue cycle, companies like Red Scissor are automating prior auth. Substrate is deploying autonomous agents to collapse the sheer number of different web apps that billing specialists have to log into to retrieve eligibility, correct CMS 1500 forms etc. Recon Health is utilizing LLMs to break down provider contracts, process correspondence letters and help reconcile claims to cash. Small independent practices often have one person doing all these things - these tools will give them superpowers. Even outsourced billers will benefit and become far more efficient as a result.
In financial services everything from transaction categorization, to financial statement generation (and the generation of footnotes and disclosures), all the way to financial audits are incredibly human intensive bodies of work, that require expensive accounting talent that there’s less of everyday. Companies like Inscope and Puzzle are flattening the manual work of accounting and bookkeeping and reducing the documentation burden, and by extension reducing the cycle time required by humans to do these tasks. Companies like Heldaway take the manual work of reading and understanding statements for private vehicles and translating them into something an RIA’s system of record can ingest, and Casap is automating Reg-E disputes and exception management. Attesto is interviewing candidates and reviewing resume’s for recruiting teams.
My instinct is that many industries that have heavy documentation burdens + a talent shortage + high turnover will benefit from AI being deployed.
Implications
Software buying decisions will start to be funded from headcount budgets
From Satya Nadella on the Q2 2024 Microsoft earnings call:
“I think what you’re going to start finding is, whether it’s Sales Copilot or Service Copilot or GitHub Copilot or Security Copilot, they’re going to fundamentally capture some of the value they drive in terms of the productivity of the OpEx, right? It’s like 2 points, 3 points of OpEx leverage would go to some software spend. I think that’s a pretty straightforward value equation.
And so, that’s the first time, I mean, this is not something we’ve been able to make the case for before. Whereas now I think we have that case. ”
In many mature corporations the way software is bought looks roughly like; there’s an annual planning process, and in that process there’s a budgeting process that considers what exactly is being used, corporate goals, revenue goals, seat growth etc, and all that is boiled into a plan for how much will be spent on specific packages of software (often depending on how mission critical they are). The process varies from place to place, but one consequence is that software purchasing is a separate decision from hiring.
My synthesis of watching this play out at Carbon Health, watching companies build this stuff, and listening to Satya’s comment is that Microsoft is for the first time seeing this discussion happen inside their customers. Said plainly, in 2024 a lot of corporations will start funding software spend from headcount budgets.
Services Businesses 👉 Tech-enabled Services Businesses 👉 Saas Businesses
This isn’t an original thought; as AI eats opex, margins will expand. As margins expand it’s inevitable that some businesses that we currently consider services will shift to have tech-enabled services margins, and some that we currently consider tech enabled services will shift to have SAAS-like margins. My instinct is the higher the amount of desk/browser/webapp work in the industry, the faster the shift (and maybe the more complete the shift? Can’t tell yet.)
I also don’t have a ton of conviction around voice work/human interaction, but companies like Center Health, Superdial, Watto, Assort Health and others are definitely making the case that voice workflows will also collapse to being done by LLMs.
Some systems will improve more slowly
Some of the pressure to make software and processes better is driven by user complaints, and AI doesn’t complain. As more of the grunt work shifts from humans to machines, the pressure to improve processes will probably ease. We’ll complain less, while getting basically the same outcome. We’ll probably have a couple of cycles of productivity gains that happen with basically zero improvement in the underlying software or administrative system or process. As an example, basically every regulatory filing that’s mostly submitting paperwork (Eg re-registering for certain types of CLIA certificates). I only have anecdotes but I have a suspicion this will happen with a lot of government website processes.
The other edge (of the double edged sword)
The nice way to say this is that a lot of peoples jobs will get better and easier as they integrate AI into their workflows and have models doing gruntwork. The not nice way to say this is that certain jobs will go away. I’m generally a believer that technological progress creates opportunity and that there’s a high probability that more jobs get created because of AI. But that’s not a law of physics. Time will tell.