For years, one of the strongest arguments for artificial intelligence in healthcare has been remarkably simple: healthcare is drowning in administrative work, and AI could finally make some of it cheaper.
Now we are beginning to see a much less comfortable possibility.
What if AI makes each participant in the healthcare payment system more efficient at fighting everyone else?
An analysis from the Blue Cross Blue Shield Association claims hospitals' use of AI tools when submitting insurance claims contributed to an additional $942 million in healthcare spending over a two-year period. According to BCBSA, its analysis found a sharp increase in patients being documented as having complex conditions without evidence of a corresponding change in the care those patients received.
The finding comes from the insurer side of a long-running payer provider dispute, so it should be interpreted in that context. More complete documentation does not automatically mean inappropriate coding. Hospitals have long argued that accurate documentation ensures they are reimbursed for the complexity of patients they actually treat.
But the broader issue raised by the analysis deserves attention.
AI is arriving inside a healthcare payment system where nearly every participant already has competing financial incentives.
Instead of removing those conflicts, AI may make everyone dramatically better at pursuing them.
There is a legitimate reason hospitals are interested in AI-assisted clinical documentation and coding.
Healthcare documentation is enormously complicated. Clinicians may treat patients with multiple chronic conditions while documenting only the problems most relevant to that particular encounter. Coding teams then translate those records into standardized codes that affect reimbursement, quality measurement, risk adjustment, and other financial and operational processes.
AI can analyze enormous amounts of clinical documentation and identify diagnoses or complexity that might previously have been overlooked.
From the hospital's perspective, that can look like accuracy.
From the insurer's perspective, it can look like inflation.
That difference matters.
BCBSA's argument is essentially that coding complexity has increased without a corresponding change in treatment. If patients suddenly appear significantly more complicated on claims while the underlying care remains relatively unchanged, insurers naturally want to know why.
But there is another possibility worth considering. AI may be exposing clinical complexity that was always present but inconsistently documented.
Those are two very different explanations for the same dataset.
And this is why the $942 million figure should not simply be interpreted as proof that AI created $942 million of unnecessary care. The analysis raises questions about coding and spending. It does not, by itself, establish that the underlying diagnoses were clinically inappropriate.
What it does demonstrate is how quickly AI can change the economics of healthcare documentation.
This is where the story becomes much bigger than coding.
Hospitals are not the only organizations adopting AI.
Insurance companies are using increasingly sophisticated technology for claims review, utilization management, fraud detection, prior authorization, payment integrity, and administrative decision making.
That means we may be entering a healthcare system in which AI operates on both sides of almost every transaction.
Imagine what that could look like.
A hospital's AI analyzes the medical record and identifies additional diagnoses that could legitimately be captured. Another system generates or optimizes the claim. The insurer's AI analyzes that claim and flags inconsistencies or questions whether the documented complexity is supported by the treatment delivered.
The hospital receives the challenge.
Another AI prepares supporting documentation or generates an appeal.
The insurer's system evaluates the appeal.
Then another agent responds.
Technically, every step could become faster.
But the overall system may not become simpler.
Abridge founder Dr. Shiv Rao captured the uncomfortable version of this future when he described the possibility of “bots fighting bots, agents fighting agents.”
That phrase deserves more attention than it might initially receive.
Healthcare has spent decades building administrative processes around mistrust between different parts of the system. Providers want appropriate reimbursement. Insurers want to control unnecessary spending. Each side creates processes to protect itself from the other.
AI can automate those processes.
But automation does not automatically resolve the underlying conflict.
In fact, it may accelerate it.
This is one of the most important distinctions healthcare leaders need to understand about artificial intelligence.
Efficiency at the task level does not necessarily create efficiency at the system level.
Suppose AI reduces the time required to generate an insurance appeal from 30 minutes to 30 seconds.
That sounds like a huge productivity improvement.
But what happens if the lower cost of producing appeals causes organizations to generate ten times as many of them?
Now imagine the insurer responds by deploying another AI system capable of evaluating ten times as many appeals.
Both organizations have technically become more productive.
Yet the healthcare system may simply be processing dramatically more administrative transactions.
The same dynamic could occur with coding, prior authorization, documentation queries, claims reviews, denials, payment disputes, and utilization management.
AI makes administrative actions cheaper.
When something becomes cheaper, organizations often do more of it.
That creates the possibility of an administrative version of Jevons paradox, where increased efficiency leads to greater overall consumption of the resource being optimized.
Healthcare could therefore end up with fewer humans manually processing paperwork but vastly more paperwork being generated, evaluated, challenged, and appealed by machines.
That would be an extraordinary technological achievement.
It would also be a deeply disappointing outcome.
The better future is still possible.
The same technologies creating this arms race could eventually reduce it.
Imagine AI systems capable of identifying documentation gaps while the clinician is still completing the encounter. Instead of waiting for a claim to be submitted, challenged, denied, and appealed, potential discrepancies could be resolved earlier.
Shared standards could allow providers and insurers to determine whether specific documentation requirements have been satisfied before claims are submitted.
AI could identify straightforward cases that do not require additional review.
It could surface the small percentage of claims that genuinely need human attention instead of forcing every transaction through increasingly complex administrative workflows.
In that scenario, AI would not simply make both sides better at fighting.
It would reduce the number of things they need to fight about.
That is a fundamentally different measure of success.
Healthcare AI vendors often advertise minutes saved, clicks eliminated, documentation generated, claims processed, or tasks automated.
Those metrics matter.
But healthcare organizations may eventually need a more important metric:
How much unnecessary work disappeared entirely?
That is the difference between digitizing bureaucracy and redesigning healthcare.
As a physician, I see tremendous potential in AI reducing the administrative burden surrounding clinical care.
Clinicians should spend less time documenting.
Patients should spend less time waiting for authorization.
Hospitals should spend less money navigating payment complexity.
Insurers should not need armies of people reviewing information that machines can reliably evaluate.
But simply placing AI on top of every existing workflow will not necessarily produce that future.
We could instead create an extraordinary computational arms race in which hospital agents optimize claims, payer agents challenge them, provider agents appeal those decisions, and insurer agents evaluate those appeals.
The machines would become faster.
The models would become smarter.
The administrative infrastructure would become more sophisticated.
And healthcare could remain just as complicated underneath.
The BCBSA analysis should therefore provoke a broader conversation than whether one particular coding tool increased reimbursement.
It raises a much more consequential question about the direction of healthcare AI.
Are we using artificial intelligence to fix healthcare's administrative problems, or are we giving every participant better weapons for fighting the same old battles?
The answer will determine whether AI ultimately reduces healthcare costs or simply automates the machinery that keeps pushing them higher.
For years, some of the biggest names in artificial intelligence have made extraordinary predictions about what AI could do for medicine. Faster drug discovery. Personalized trea...
