Akasa CEO Malinka Walaliyadde describes his company's purpose in a phrase that has become standard across revenue-cycle AI: using large language models to help capture the patient story as comprehensively as possible for health systems, and then tell that story to payers.

That sentence is worth examining closely, because it describes an activity that has two names depending on where you stand. From the provider's side it is accurate documentation. From the payer's side, the same activity, pursued aggressively enough, is upcoding. The technology performs identically in both cases.

This is not an accusation against Akasa, which appears to take compliance seriously and counts compliance officers among its advocates. It is a structural observation about what the entire category does, and why the results are so contested.

The problem is real, and larger than most people realize

Start with the case for the technology, which is stronger than skeptics usually allow.

A single patient record can average 60 documents and 50,000 words. A human coder reviewing that record under production pressure will not read all of it with equal care. Clinically significant details, a documented comorbidity, a complication noted in a nursing entry, evidence of complexity buried in a specialist consult, get missed. When they are missed, the hospital is paid less than the care it actually delivered.

That is a genuine harm, and it runs the opposite direction from the fraud narrative. A health system that treated a complex patient and coded it as a simple case is not defrauding anyone; it is absorbing the loss.

The labor problem compounds it. Walaliyadde describes an emerging crisis in coding and billing in which the highly trained workforce is aging out and providers cannot find new people fast enough to replace those who leave, even as medical complexity grows. Medical coding requires years of training and certification, and the pipeline is not keeping pace. That makes automation less a cost-cutting choice than a response to a shortage.

There is also a genuine market insight in the company's positioning. Walaliyadde notes that 85 to 90% of health system revenue is on the inpatient, hospital facility side, while many attempts to apply technology in this space targeted the professional arena instead. Aiming at facility billing is aiming at where the money actually is.

What is technically different here

Two design choices distinguish this from generic AI-in-healthcare claims, and both deserve attention.

The first is institution-specific tuning. Akasa tunes its LLM for each institution rather than deploying one model everywhere, on the reasoning that every organization documents differently, every provider documents differently, and every service line has its own patterns, so a generic model cannot address that variation.

That is technically sound. It is also where the most interesting risk lives, and it is a risk that applies to any per-institution tuned system, not just this one. A model trained on an organization's historical documentation and coding patterns learns that organization's habits. If those habits were conservative, the model inherits conservatism. If they were aggressive, the model learns aggression, and then applies it consistently across every chart at machine scale. Tuning transmits whatever was already there, and neither the vendor nor the customer necessarily knows which they had.

The second is the scale of review. Some clients have all hospital billing reviewed by the AI model, with humans continuing to review results.

That is the actual change, and it is easy to underrate. Human coding review has always been sampled, because reviewing every chart in full was impossible. Full-population review means every encounter gets examined for documentation opportunities, every time. If the average review finds even a small amount of additional supportable complexity, applying that across 100% of encounters rather than a sampled subset produces a large aggregate revenue shift, without any individual determination being wrong.

The uncomfortable symmetry

Here is where the industry-level evidence complicates the vendor-level story.

The Peterson Health Technology Institute found that provider deployment of AI is increasing billing intensity and inflating medical spending, and that payers are responding with across-the-board downcoding and broad reimbursement reductions. Both sides are automating, transaction volume rises, and system costs go up rather than down.

Now hold that against the underdocumentation argument, which is also true. Both can be correct simultaneously, and that is the crux: AI that finds genuinely missed documentation and AI that maximizes billing intensity are the same product, doing the same thing, distinguishable only by whether the clinical facts actually supported the code.

Neither the vendor nor the health system nor the payer can fully adjudicate that at scale. The vendor's model surfaces evidence in the record. Whether that evidence reflects care genuinely delivered at that complexity is a clinical judgment. And critically, an AI trained to find support for higher-complexity coding is not symmetrically trained to find reasons to code lower. The tool has a direction, because that is what it is bought for. That asymmetry is not a flaw in any particular product; it is inherent in the commercial purpose.

What would actually settle it

The most useful thing in Akasa's own materials is not a revenue claim but a compliance one: health systems are trying to protect audit defensibility in an environment of rising scrutiny, and the platform surfaces evidence-backed documentation with explanations.

That points at the right test. The question for any revenue-cycle AI is not whether it increases revenue, which it will, but whether every code it recommends is traceable to specific documentation a human reviewer and a federal auditor would independently accept. Evidence-linked recommendations with visible reasoning are auditable in a way that a black-box score is not, and an industry analysis of this space concluded much the same thing: site-specific tuning, measurable coding accuracy, human review thresholds, and auditable billing decisions matter more than generic GenAI claims.

Two further things would help resolve the ambiguity, and neither currently exists at industry scale. First, published accuracy data validated against independent audit rather than against the institution's own historical coding, since matching your own past behavior proves consistency, not correctness. Second, transparency about directional balance: how often does the system recommend coding down? A tool that only ever finds reasons to bill more is doing something different from a tool that finds documentation errors in both directions, and the ratio would be genuinely informative.

The honest position

Revenue-cycle AI is not a scam, and the companies building it are solving a real problem. American medical billing is genuinely, absurdly complex, patient records are genuinely too long for humans to read exhaustively, the coding workforce is genuinely shrinking, and hospitals genuinely lose money on care they delivered but documented poorly. Every one of those justifications is true.

It is also true that the aggregate effect of the whole category, measured across the system, is rising billing intensity, a payer counter-response, and higher costs overall. Those two facts sit uncomfortably together, and anyone selling or buying this technology should hold both rather than choosing the convenient one.

The phrase to watch is the one at the top. "Capture the full patient story" is doing enormous work, and the word carrying the weight is full. If full means complete and accurate, the technology is fixing a real deficiency. If full means maximal, it is an engine for the cost growth the sector says it wants to solve. The tools cannot tell you which they are doing. Only independent audit can, and the industry has not yet built the evidence base to answer it.

Further reading