Artificial intelligence entered healthcare with an attractive promise: automate paperwork, reduce administrative costs, help doctors process information faster and ultimately make medical care more efficient.
But a new analysis from the Blue Cross Blue Shield Association suggests that one application of AI may be producing a very different result.
According to BCBSA, changes in hospital coding associated with the growing use of AI contributed an estimated $942 million in additional healthcare spending for Blue Cross and Blue Shield plans between 2023 and 2025. The association says the increase occurred without evidence of a corresponding change in the care patients received.
That does not mean AI itself “caused” $942 million in healthcare costs in a scientifically established sense. The figure comes from an insurer's analysis of claims data, and hospitals have disputed the interpretation.
But the finding exposes something much bigger than a disagreement over one number.
AI is entering a healthcare payment system where better documentation can translate directly into more money.
And that creates an unusual incentive problem.
The AI is not necessarily treating the patient. It is reading the chart.
One of the fastest-growing applications of AI in hospitals isn't a robot performing surgery or a model diagnosing cancer.
It is something much less glamorous:
medical documentation and billing.
Hospitals increasingly use software that can scan electronic health records, laboratory results and doctors' notes to identify diagnoses that may need to be documented or coded.
Some systems also use ambient AI, which listens to conversations between clinicians and patients and automatically produces clinical documentation.
More than 60% of hospital systems were using AI-enabled technologies capable of identifying these kinds of conditions, according to BCBSA.
That can be useful.
A physician may overlook a secondary condition buried somewhere in a patient's medical record. An AI system can flag it.
Better documentation can also mean that hospitals receive appropriate reimbursement for care they actually provided.
The problem emerges when the payment system rewards the documentation of complexity more strongly than the actual delivery of additional care.
How the $942 million figure emerged
The BCBSA analysis examined de-identified claims data from Blue Cross companies.
The association found that the proportion of inpatient cases classified as medically complex increased from roughly 37% at the beginning of 2023 to 40% by the end of 2025.
That may sound like a small change.
In a healthcare system processing enormous numbers of hospitalizations, however, small changes in coding intensity can translate into substantial financial consequences.
BCBSA estimates that the increase amounted to approximately $942 million in additional costs over the two-year period relative to its 2023 baseline.
Approximately $653 million was associated with secondary diagnoses that moved more than 55,000 cases into higher-severity, higher-reimbursement categories, according to the analysis.
In other words, the argument is not that hospitals suddenly began treating millions of additional patients.
It is that more patients were being classified as medically complex.
And classification matters because classification affects payment.
The economics of a diagnosis
Hospital reimbursement in the United States is heavily influenced by diagnosis-related groups, commonly known as DRGs.
A patient's hospitalization is assigned to a category based partly on the primary condition and additional diagnoses or complications.
Certain secondary diagnoses can push an admission into a higher-severity category.
That means two patients who undergo broadly similar procedures can potentially generate different reimbursements depending on how their conditions are documented and coded.
This system was created for legitimate reasons.
A hospital treating a genuinely more complicated patient generally consumes more resources.
A patient with serious complications may require additional monitoring, medication, nursing care, procedures or longer hospitalization.
The payment system is therefore designed to compensate hospitals differently for different levels of complexity.
But that also creates a vulnerability:
If the classification becomes more aggressive without a corresponding increase in actual care, spending can rise without improving outcomes.
That is the concern BCBSA is raising.
AI makes the old problem much faster
This is where AI changes the equation.
A human coder can examine a patient's records and identify relevant diagnoses.
An AI system can potentially scan enormous amounts of documentation continuously.
Laboratory results.
Physician notes.
Previous diagnoses.
Medication histories.
Discharge summaries.
Clinical conversations.
The technology can search for patterns that humans might miss.
That capability is valuable when it captures a legitimate condition that should have been documented.
But the same capability can also expose every potentially billable condition contained within a patient's record.
The machine doesn't necessarily understand the economics of the system in the same way a human does.
It is optimizing the task it was designed to perform.
Find the relevant diagnosis.
Document it.
Code it.
And if that code changes reimbursement, the financial consequences follow automatically.
The anemia example
BCBSA highlighted an example involving patients undergoing major bowel procedures.
The association said anemia diagnoses increased substantially at hospitals using these technologies, but the increase was not accompanied by a corresponding rise in blood transfusions—the treatment it would expect to see more frequently if the additional anemia diagnoses reflected a genuine increase in clinically significant cases.
This is central to the insurer's argument.
If patients are becoming materially sicker, hospitals should generally be doing something different to treat them.
More transfusions.
More interventions.
More monitoring.
More resources.
But if the diagnosis increases while treatment remains relatively unchanged, BCBSA argues that the data may indicate a documentation phenomenon rather than a genuine change in patient severity.
That is an important distinction.
Because a diagnosis appearing in a database does not automatically prove that the patient became sicker.
But there is another side to the story
This is where the debate becomes more complicated.
The BCBSA analysis does not establish that hospitals deliberately manipulated diagnoses.
And it does not prove that AI-generated coding is inherently inappropriate.
Hospitals and healthcare organizations have argued that patients may genuinely be more complex and that improved documentation can reveal conditions that were previously missed.
That distinction matters.
Suppose a patient actually has anemia.
If the physician failed to document it five years ago but an AI system correctly identifies it today, the resulting increase in coding could represent better documentation, not fraud or abuse.
The fact that treatment did not change does not necessarily prove that the diagnosis was wrong.
Patients can have clinically relevant conditions that do not require a specific intervention.
So the question is not simply:
“Did AI increase coding?”
It clearly can.
The harder question is:
“Did AI increase inaccurate or financially driven coding?”
That requires much more than claims data.
It requires examining patient charts, clinical evidence, treatment decisions and the actual medical justification behind individual diagnoses.
This is not simply an AI problem
There is another layer that makes this story particularly interesting.
Hospitals aren't the only side using AI.
Insurers are using it too.
Hospitals can deploy AI to identify diagnoses that support reimbursement.
Insurers can deploy AI to examine claims and identify diagnoses or procedures they believe are unsupported.
The result is an emerging technological arms race.
Hospital AI says:
Find everything that supports the claim.
Insurance AI says:
Find everything that challenges the claim.
Both sides have increasingly sophisticated machines examining the same medical records.
Reuters reported earlier this year that U.S. hospitals and insurers were already using AI in this longstanding dispute over billing and reimbursement.
That changes the nature of healthcare administration.
It is no longer simply:
Doctor → hospital → insurer.
Increasingly it becomes:
Doctor → AI → hospital → AI → insurer → AI → patient.
And every additional algorithm introduces another layer of automated decision-making.
The “bots fighting bots” problem
This could create a bizarre future in which healthcare organizations spend enormous resources building increasingly sophisticated AI systems to defeat one another.
The hospital's AI searches for legitimate reimbursement opportunities.
The insurer's AI searches for reasons to reject or reduce those payments.
The hospital upgrades its system.
The insurer upgrades its system.
The cycle continues.
The technology becomes more sophisticated.
But the patient may see very little improvement.
That is the paradox.
AI could theoretically reduce administrative waste.
Instead, it could simply automate the administrative conflict.
TechCrunch reported the emerging debate, including concerns that AI could intensify the longstanding conflict between hospitals and insurers rather than resolve it.
The bigger question: who actually benefits from AI?
This is the question that healthcare executives, regulators and investors should be asking.
AI adoption is often justified using a simple equation:
Automation → efficiency → lower costs → lower prices.
But markets do not always work that way.
If AI allows a hospital to capture more revenue from the same patient, the hospital may benefit.
If an insurer uses AI to prevent unnecessary payments, the insurer may benefit.
If both sides spend billions building competing AI infrastructure, technology vendors benefit.
But that doesn't automatically mean the patient benefits.
The ultimate test should therefore be much more difficult:
Does AI reduce the total cost of delivering healthcare?
Not:
Does it reduce paperwork?
Not:
Does it increase hospital revenue?
Not:
Does it reduce insurer payouts?
But:
Does it make high-quality healthcare more affordable for patients?
Those are very different measurements.
AI could also make healthcare cheaper
It would be a mistake to interpret the BCBSA analysis as proof that AI is bad for healthcare.
There are enormous opportunities.
AI can automate administrative tasks.
It can summarize medical records.
It can help clinicians search through enormous amounts of information.
It can support medical imaging.
It can reduce documentation burdens.
It can help identify patients who need follow-up.
It can potentially accelerate research and drug discovery.
And properly implemented clinical documentation tools can help physicians capture legitimate conditions that would otherwise be missed.
The technology itself is not inherently inflationary.
The incentives surrounding the technology determine what it optimizes for.
The real problem may be the payment system
This may ultimately be less about artificial intelligence and more about the structure of American healthcare reimbursement.
AI is extraordinarily good at finding patterns.
If the payment system rewards certain patterns, companies will naturally build systems capable of finding them.
That doesn't necessarily require malicious intent.
It is basic economics.
If moving a patient from one reimbursement category to another can generate thousands of dollars, the ability to identify qualifying conditions becomes economically valuable.
And AI dramatically reduces the cost of searching for those conditions.
The technology therefore doesn't create the incentive.
It amplifies it.
Why $942 million matters—but also needs context
The headline number is enormous.
But it needs to be interpreted carefully.
The $942 million figure refers specifically to estimated additional spending for BCBS companies based on the association's analysis and methodology. It is not a measurement of total U.S. healthcare spending caused by AI.
BCBSA represents independent Blue Cross and Blue Shield companies collectively covering roughly one-third of Americans, and its analysis uses de-identified claims data.
The association itself is also an interested party in the healthcare payment system.
That does not invalidate the analysis.
But it means the findings deserve independent examination.
A stronger evidence base would include:
- patient-level chart reviews;
- independent audits of AI-generated codes;
- comparisons between AI-assisted and human-only coding;
- treatment intensity data;
- hospital-level AI adoption timelines;
- clinical outcomes;
- insurer payment records;
- and analysis from independent researchers.
That would help determine how much of the increase comes from genuine clinical documentation versus changes in billing behavior.
AI needs an audit trail
There is an important regulatory lesson here.
When an AI system influences a medical diagnosis, treatment recommendation or reimbursement decision, there should be a record of what happened.
Which data did the system examine?
What diagnosis did it identify?
Why did it recommend that diagnosis?
Did a physician verify it?
Who ultimately approved the code?
Was the diagnosis associated with actual treatment?
Was the patient informed?
These questions become increasingly important as AI moves from being a passive assistant to an active participant in healthcare administration.
If an algorithm can change the amount of money attached to a patient, its decisions should be auditable.
The dangerous assumption about AI
For years, the dominant narrative around healthcare AI has been that the technology will eliminate inefficiency.
But AI doesn't automatically eliminate inefficiency.
Sometimes it simply moves it.
A hospital may save hours of administrative work while generating additional reimbursement.
An insurer may save hours reviewing claims while increasing denials.
A doctor may save time documenting a visit while creating more detailed records for algorithms on both sides to analyze.
The technology becomes more efficient.
The system does not necessarily become cheaper.
That distinction is going to become increasingly important.
The future of healthcare may be an algorithmic negotiation
The most interesting part of the Blue Cross story isn't actually the $942 million.
It is what the number tells us about the next phase of healthcare.
Healthcare is becoming increasingly computational.
Doctors use AI.
Hospitals use AI.
Insurers use AI.
Pharmaceutical companies use AI.
Patients increasingly use AI.
Eventually, these systems will interact with one another across nearly every part of the healthcare economy.
That could produce extraordinary efficiency.
But it could also create a new problem:
an automated healthcare economy in which every participant has an AI optimized around its own financial incentives.
The hospital wants maximum legitimate reimbursement.
The insurer wants minimum unnecessary expenditure.
The technology company wants more software adoption.
The physician wants less administrative work.
The patient wants affordable, high-quality care.
Those objectives do not always align.
AI won't magically resolve that conflict.
It may make the conflict faster.
The real AI healthcare test
The industry should stop measuring healthcare AI primarily by how many documents it processes, how many minutes it saves or how much revenue it generates.
The more important measurements are simpler:
Did patients receive better care?
Did outcomes improve?
Did administrative costs fall?
Did unnecessary spending decline?
Did patients pay less?
Did physicians spend more time treating patients and less time fighting paperwork?
If the answer to those questions is no, then an AI system can be extremely efficient while still making the healthcare system worse economically.
That is the lesson hidden inside the Blue Cross analysis.
AI has not suddenly become the villain of healthcare.
But neither is it automatically the hero.
Artificial intelligence is an amplifier. Give it the right incentives and it can reduce waste. Give it a reimbursement system full of financial loopholes, and it may become exceptionally good at finding them.
The next battle in healthcare may therefore not be between humans and machines.
It may be between machines optimizing competing financial interests—and the patient caught in the middle.
Comments
Post a Comment