Coding intensity is climbing where AI documentation landed, and payers have started downcoding

Two analyses published in March 2026 looked at different claim populations and found the same shape: charts are describing sicker patients, and the treatment recorded in those same charts is not moving with them. Neither analysis claims AI caused it. One payer stopped waiting for that question to be settled. Since October 2025, according to a policy brief in npj Digital Medicine, Cigna has automatically reduced many level 4 and 5 E/M claims by one level unless the documentation clearly supports the higher complexity.

If your practice runs an ambient scribe or an automated chart-review tool, the person who decides whether a fuller note is also a supported note is still you.

What the three 2026 analyses actually measured

These are three different datasets on three different populations. Treating them as one study is the fastest way to get a fact wrong, so here is what each one covers.

Read the fine print on the Trilliant paper, because it argues against the headline you would expect. Its own title says the increase "warrants examination but likely reflects enhanced rules-based documentation." Trilliant's position is that better capture, not upcoding, is the likely driver. BHI, for its part, says its objective is to inform stakeholders about coding trends "rather than to make determinations about clinical appropriateness or provider intent."

So nobody here has proven wrongdoing. What they have established is a measurable gap between documented severity and documented treatment. That gap is what a payer's algorithm sees, and it is what an auditor pulls the chart on.

The discordance signal, in one maternity example

BHI's clearest example is postpartum anemia in maternity admissions. At the fastest-growing hospitals, the share of maternity admissions coded with it rose from 4.0 percent in Q2 2022 to 12.3 percent by Q1 2025. At low-growth hospitals it barely moved, from 7.9 percent to 8.2 percent.

Transfusion is the treatment those claims would ordinarily show. Transfusion rates stayed flat at both groups. BHI puts it plainly: an 8.3 percentage point rise in coded anemia came with a 0.4 percentage point rise in transfusions.

One BCBS plan audited a hospital system that was a national outlier. BHI quotes the plan's senior clinical executive saying that, working with a board-certified OB-GYN, "less than 20% of cases met established clinical criteria for postpartum anemia." That is one plan's audit of one system, reported by the plan, not an adjudicated finding. It is also exactly the kind of review a coder can run first.

Payers are not waiting for the causal question to be settled

The npj brief reports that Cigna's automatic one-level reduction of many level 4 and 5 E/M claims began in October 2025, and that Aetna Better Health has applied similar reviews. A default downcode does not care whether the note got longer because a scribe captured more or because someone leaned on a template. It responds to the claim.

None of this is new territory for federal auditors either. OIG found the number of Medicare inpatient stays billed at the highest severity level rose almost 20 percent from FY2014 through FY2019, that average length of stay for those cases actually fell, and that such stays are "vulnerable to inappropriate billing practices, such as upcoding." That report came out in February 2021, years before ambient scribes were common. The tooling changed the speed and the scale. It did not create the pattern.

Where a human coder is still required

An ambient scribe transcribes and structures a conversation. An automated chart-review tool scans the record and surfaces diagnoses that could be coded. Neither one establishes that a condition was evaluated, that it was treated, or that it was medically necessary to treat. Those are judgments made by reading the chart against a standard, and they are the coder's judgment.

The npj authors land in the same place from the clinical side. They want organizations to "disable auto-accept" and to run "random audits comparing audio to signed notes" that "test medical necessity."

On a coding desk that means one thing. A diagnosis surfaced by a tool is a candidate, not a code. The evidence that promotes it is MEAT support in the record for that date of service, which is the same standard that keeps a diagnosis defensible when someone pulls the chart later. A tool that mines the problem list produces exactly the pattern that problem lists alone do not validate. And a fluent, well-organized note is not the same thing as a supported one, which is why AI-generated notes still need coder QA.

The discordance check to run before the claim goes out

This is not in any of the three sources. It is the operational form of what all three are measuring: for each severity-driving diagnosis on the claim, ask what treatment or evaluation the record should show if that diagnosis is real, then go look for it.

Run this on a sample of AI-drafted encounters before it runs on you. If a row fails and you cannot fix it with a query, the answer is to not code the diagnosis, and to document that decision in your evidence worksheet or whatever your practice already uses. Confirm the official description of anything you are unsure about in the Code Book before you commit to it.

What this does and does not mean for risk adjustment

Be precise about scope. BHI's analysis is commercial inpatient claims. Trilliant's is hospital outpatient E/M. Neither is a Medicare Advantage risk-adjustment dataset, and neither reports a RAF impact.

The npj brief does draw the connector: it observes that in Medicare Advantage, "richer documentation primarily increases the plan's risk-adjusted capitation payments by raising members' risk scores." That is a statement about incentives, not a measurement of MA coding intensity, and it should be read that way.

What carries across cleanly is the standard. A diagnosis is supportable when the record shows it was evaluated, assessed, monitored, or treated, and no documentation tool changes that. It is the same failure mode that put unsupported diagnoses at the center of the OIG stroke-coding overpayment findings: the code was on the claim, and the chart did not back it.