Across the world, health systems are under the same two pressures at once: spend less, and prove that the money spent reflects real care. Nowhere is that tension sharper than in how patient conditions are recorded and paid for.
A new generation of AI is reshaping that work, and the most important feature is not speed. It is the ability to explain itself.
The problem with recording health
When a health plan is paid based on how sick its patients are, the accuracy of the record becomes everything. Every condition reported is supposed to be backed by a real clinical note. When it is not, the system pays for a version of the patient that does not exist.
This is the discipline of risk adjustment coding, and it is harder than it sounds. A single patient can generate hundreds of pages across visits, labs, and specialists. Buried in there is the evidence that does or does not support each diagnosis. Finding it by hand is slow and uneven.
What audits revealed
The cost of getting this wrong became clear in 2026. In the United States, federal auditors reviewed a group of private Medicare plans and found that 80 to 91 percent of the diagnoses they sampled were not fully supported by the medical record. A government advisory panel estimated the broader pattern at roughly 22 billion dollars in excess payments in a single year. One major insurer paid 117.7 million dollars to settle related claims.
The lesson travels well beyond America. Any system that pays based on recorded conditions faces the same risk: records that drift from reality, and a bill that follows the drift.
Why the old automation fell short
Earlier attempts to automate this work leaned on tools that produced an answer without a reason. They would suggest a diagnosis, but could not point to why. In a setting where every code may be audited, an answer no one can explain is close to useless. It cannot be trusted and it cannot be defended.
That is the gap explainable AI is closing.
Intelligence that shows its work
The approach gaining ground is called Neuro-Symbolic AI. It combines the pattern-spotting strength of machine learning with explicit clinical rules. The practical result is that for every diagnosis it suggests, it can point to the exact line in the chart that supports it, and flag the ones that look unsupported so a human can remove them.
This matters in two directions. It helps add conditions that are genuinely documented but were missed. It also helps strip out conditions that are no longer valid, which is the half of the job older tools ignored. A trained coder still makes the final decision. The AI makes that decision faster and better grounded.
Where this is heading
For health systems anywhere modernizing how they document care, the direction is set. The tools that last will not be the ones that simply produce the most codes. They will be the ones that can stand in front of an auditor, a clinician, and a regulator and show, line by line, why each decision was made.
Speed got the attention first. Explainability is what will earn the trust. In a field where the record is supposed to mirror a human life, that is exactly the right thing to optimize for.
