To a layperson (including almost all VCs) there is a belief that healthcare administration is ripe for AI disruption.
Most people believe that this disruption will be good. Everyone knows that healthcare administration costs too much money, too much time, and prevents too much innovation.
It's natural to think that if individual healthcare administrative jobs could be automated by AI, then the entire healthcare industry post AI utopia will spend less on admin and more on care.
What people fail to realize is that the nature of healthcare administration is adversarial, and exists to balance naturally different incentives.
The core problem that must be solved is trust and incentive alignment and if anything AI makes these two things harder to achieve. Bureaucracy, the enemy AI supposedly solves, is actually just a sort of "proof of work" method of establishing trust with strangers.
AI does not fundamentally improve the ability for payers and providers to trust each other, and thus will just lead to more overhead.
Individual Jobs In Healthcare Administration Can be Enhanced with AI
Any individual healthcare administrator (on the provider or payer side) can easily have their job automated by AI because these jobs are:
Repetitive.
Rules-based decision making across unstructured and textual data.
There are many processes that are very uniform across the entire US healthcare market.
These such jobs can be automated / improved in efficiency with AI.
The trillion dollar question is whether enhancing all of these jobs with AI will make the whole system smoother.
Examples of such healthcare admin jobs disruptable with AI:
Checking insurance eligibility for patients and confirming things like their deductible spending, plan exclusions, etc.
Completing prior-authorizations and confirming the medical necessity of services from medical documentation.
Ensuring that billable services are tracked and submitted in a claim.
Reviewing healthcare claims for accuracy and proper documentation.
Filing appeals.
Responding to appeals.
But does this actually make everything more “efficient”. The answer is no. Just because each individual job in the healthcare administrative state can be improved with AI does not mean that the system will get any more efficient overall.
To see why, we must ask more than "what are healthcare administrative costs". We must ask "why are there healthcare administrative costs?".
Everyone who has done some thinking about healthcare will realize that healthcare is all about incentives, and the extent to which our system is chronically inefficient has to do with conflicting incentives across various stakeholders.
Healthcare administration exists to mediate these incentives through a system of rules and procedures that were designed piece by piece to find more ways to double check the trustworthiness of payers/providers.
This is the why of healthcare admin.
If payers and providers knew they could trust the other to be honest and non-greedy, then many burdensome admin processes would not exist.
There would be no engineering of how much to bill by optimizing different ways of recording and documenting patient visits.
There would be no insurance plans reviewing the medical necessity of something ordered by a provider before approving a claim.
These (and virtually all) healthcare admin tasks are inherently zero-sum.
In the long run, healthcare admin costs are proportional to the number of zero sum procedures added to the stack.
The fundamental problem healthcare admin is trying to solve is "how do I trust a stranger with different incentives and more information than me?" Adding zero-sum processes of review, certification, and other admin processes are effective solutions but themselves become a problem perhaps worse than the original problem of having trouble trusting strangers.
The best mental framework I have for thinking about healthcare admin is that it actually does solve the trust problem between adversarial information-asymmetric parties.
If both payers and providers add more admin and optimization to their processes by collecting information that they need from each other and consulting common standards, they will eventually get to a decision about a claim that is from a legal perspective "objectively correct" and thus a weak form of trust can be formed by performing and completing this process.
The provider perspective
The legitimate goal of the provider is to get paid for services they render at a fair price.
Providers naturally lack information about a patient's health insurance plan, including exclusions, pre-cert requirements, whether the member is active on their insurance, what the payers rates or policies on procedures are, etc.
Providers trust payers more when the get confirmation of eligibility or benefits coverage.
The payer perspective
The legitimate goal of the payer is to pay claims that are covered and deny claims that are non-covered.
Payers naturally lack information about events and circumstances that actually occurred during a medical visit and their medical necessity.
Payers trust providers more when they get medical records for pre-cert and know what codes will be billed.
Both parties can effectively threaten the other party with having to do a bunch of admin work to be trusted.
Providers can ensure that they get paid a fair amount by billing for everything under the sun and hoping that the payer at least pays for all the services that are covered from among the list of services recorded, rendered, and billed. This puts admin work on the payer to find reasons certain services need to be bundled or denied.
Payers can ensure that procedures are medically necessary by inserting utilization management programs like pre-certification or medical claim review into their processes before paying claims.
In general the party that feels like the claim is unfair is more willing to do admin tasks to get to a more fair outcome.
The "two steps ahead of you" thinker here realizes that the hack is to actually just always require the other party to do painful admin work. If they are willing to do this painful work, then they can likely be trusted because otherwise if they were not being honest they would just skip the admin work and accept your terms.
Imagine for illustration a byzantine system where a payer requires all pre-certs to be handwritten by the provider and sent by a physical courier to the payer in order to be approved. Given the high cost imposed by this, a payer can be pretty sure that any providers willing to go through this process are likely to be filing honest claims since otherwise a rational actor would not have expended so many resources in filing the appeal. It almost doesn't matter if the payer even reads the documentation submitted by the provider at this point -- the mere fact that the provider was willing to do so much work is indirect evidence that the procedure was necessary and should be approved by the payer.
This is the "proof of work" model of healthcare administration.
The trust comes not in the end result of the admin work but the in act of the counterparty spending admin resources towards that result.
After a year of paying claims and dealing with the payer side of this equation rather manually, I completely subscribe to the "proof of work" model of healthcare administration being the underlying mindset of the people that design claims processes for both payers and providers.
This model is incredibly inefficient completely by design. This model includes people sacrificing their lives and sanity in order to produce essentially meaningless reports just to prove that the reports were worth spending effort to produce.
The only solution: move to "proof of stake" administration.
The only way to solve the problem is to ensure incentives are aligned by ensuring all parties have a stake in desired outcomes.
If payers and providers had a stake in the same outcomes, then they would be able to trust each other simply by understanding each others incentives.
The system would thus not need most of the admin processes described in the sections above. Is deletion better than automation??
There are many healthcare AI products trying to skip straight to automation in this design system without even questioning the requirements of the system they are automating. They fail to realize that many of the processes they are trying to automate with AI could have been automated by simpler software yet were not.
The better question to ask and the one that might actually make a difference long-term is how to make the requirements of trusting a stranger in healthcare less dumb.
Sidebar: Risk contracting and HMOs
Did we just take the long road to risk based contracting (RBC)? Or to health maintenance organizations (HMOs) with employed physicians?
(sigh). Not necessarily. Without expanding too much on those models which do solve some of the above problems by aligning incentives, I will say the following:
RBC reduces disagreements over claims discussed in this post, however the model introduces new adversarial games between payers and providers like risk-measurement.
HMOs do indeed have aligned incentives and I like their model a lot from an incentives point of view. I just believe large organizations are naturally inefficient and stagnant.
Perhaps I will write more on HMOs or RBC in another blog post.