Can AI chart review prevent claim denials?
Yes, for the denials that start in the chart: codes that can't be billed together, a missing left or right, two codes that should be one, and a visit level the note doesn't support. Here's what to check and how to test a tool on your own denied claims.
By Jake Schwartz, CEO & Founder
Yes, for the denials that start in the chart. AI chart review reads the doctor's note before it's signed and catches codes that can't be billed together, a code missing its left or right, two codes that should be one, and a visit level the note doesn't support.
Those mistakes are common. In Experian Health's 2026 survey of 204 billing and claims staff, a quarter put inaccurate coding among their top three causes of denials, and 4 in 10 said 10% or more of their claims get denied (Medical Economics).
Coverage problems, prior authorization and filing deadlines belong to the front desk and the billing office. Coding mistakes belong to the chart, and the chart is where they're cheapest to fix.
Which denials start in the chart?
Five kinds of coding mistakes come straight from the codes picked while the chart is open. Each has a published rule behind it, so each can be checked against the note before anyone signs.
| What goes wrong | What the check looks for | Example |
|---|---|---|
| Two codes that can't be billed together | Every pair of diagnosis codes against the ICD-10-CM Excludes1 notes, the official list of codes that never go together | R73.03 (prediabetes) left on the list after E11.9 (type 2 diabetes) was added |
| A code with no side | A code that has right and left versions when the note names a side | M25.569 is knee pain with no side given. If the note says right knee, the code is M25.561. |
| Two codes that should be one | Conditions the coding rules presume are linked, coded apart | E11.9 and N18.32 listed separately. If the doctor confirms the diabetes caused the kidney disease, the claim takes E11.22 with N18.32. |
| Kidney disease from high blood pressure with no stage | I12 or I13 with no code for the stage | I12.9 billed with N18.9, which doesn't name a stage |
| A visit level the note doesn't support | The medical decision making written in the note | A 99214 billed on a note that shows a 99213's decisions, or a 99213 on a note that supports a 99214 |
The first row trips up a lot of practices because the problem list is old and the new diagnosis is right. Prediabetes was true last year. Type 2 diabetes is true today. Both codes sit on the claim, and the R73 category's Excludes1 note says they can't. See how Excludes1 works.
Insurance companies already run software that looks for several of these. UnitedHealthcare's Medicare Advantage plans started enforcing Excludes1 on claims from doctors' offices in March 2026. Anthem told its California Medicare Advantage providers that claim lines using an unspecified side, when a sided code exists, will be denied. Here's what the payers' automated checks look for.
Why doesn't a claim scrubber catch these?
A claim scrubber is software that checks each claim against insurance rules before it's sent. It reads the claim, and a claim carries codes, never the note.
A scrubber can tell your biller that R73.03 and E11.9 can't go on the same claim. Only the note says which one is true today. A scrubber can flag M25.569 as missing a side. Only the note says which knee.
By the time the scrubber runs, the note is usually signed. Changing it means an addendum, a dated entry the doctor adds after signing. So the biller catches the doctor between patients, waits for the addendum, then fixes the claim and sends it. A billing lead at a family medicine practice told us a corrected claim pays in 30 to 45 days, compared with 7 to 14 for a clean one.
When billers run short of hours, those corrections pile up, and the ones nobody reaches before the insurance company's deadline get written off. See where that money goes.
Chart review runs the same check before the signature. The doctor sees "M25.569 has no side, and the note says right knee" while the visit is fresh, picks M25.561, and signs a chart that's ready to bill.
Should AI chart review suggest higher codes?
It should flag in both directions. A good review tells the doctor when a code isn't supported or a level is set too high just as plainly as when a combination code or a side is missing. It shows the rule and the line in the note behind every flag, and a person decides every change.
One-way review is the pattern regulators go after. In March 2026, Aetna agreed to pay $117.7 million to resolve allegations that it used chart reviews to add diagnosis codes while ignoring results that showed codes were unsupported.
For visit levels, the question is always what today's note supports, up or down. Here's what a 99213 note and a 99214 note each need to show.
How do you test an AI chart review tool on your own charts?
Use the denials you already have. A vendor's demo runs on charts the vendor picked. Your denial report shows the mistakes your practice actually makes.
- Ask your biller for 20 open denials from last quarter where the reason was a coding problem, such as codes that conflict, a missing side, or a visit level the insurance company lowered. Pick ones nobody has fixed yet, so the chart still holds the mistake.
- Add 20 charts from the same weeks that paid the first time. These show whether the tool raises false alarms on good charts.
- Run the tool on all 40 charts and have your biller mark each flag right or wrong.
- Count three things: right flags, wrong flags, and denied charts the tool said nothing about.
Then compare tools on how often each is right when it flags, before you compare how much each one finds. Here's why, with two made-up tools run on the same 40 charts.
| Tool A | Tool B | |
|---|---|---|
| Flags raised | 40 | 16 |
| Right flags | 20 | 15 |
| Wrong flags | 20 | 1 |
| Right when it flags | 1 in 2 | 15 in 16 |
Tool A found five more real problems. It also stopped the doctor 20 times for nothing, one false alarm for every real catch. Every flag pulls the doctor's eyes off the chart they're trying to close, and after a week of coin-flip flags most doctors stop reading them. The real catches get ignored along with the false ones.
Tool B speaks up less and is right almost every time. The doctor keeps reading its flags, so its 15 catches turn into 15 fixed charts. A miss leaves that claim where it would have been anyway, with your biller's usual review. A false alarm costs the doctor's time and, after enough of them, the doctor's trust.
We score Cortex the same way. On our benchmarks page, a wrong flag counts twice as much as a miss. Every score comes from de-identified real primary care charts, and each one is published with its range.
Cortex Lens runs these checks on every Practice Fusion chart while it's still open: the full Excludes1 list, a missing left or right, combination codes coded apart, a missing kidney disease stage, and the visit level the note supports, up or down. It's read-only, and your team makes every change in Practice Fusion. The impact page counts the denial risks Lens caught and the ones your team fixed, so you can set them next to your own denial report. Book a demo.
Sources: Denial causes and rates: Medical Economics, October 30, 2026, reporting Experian Health's fourth annual State of Claims survey of 204 billing, claims and revenue staff, fielded August 5 to 23, 2026. Excludes1, laterality and combination-code rules: CMS and NCHS, ICD-10-CM Official Guidelines for Coding and Reporting, FY2027 and the FY2027 tabular list. Aetna settlement: U.S. Department of Justice, March 11, 2026. Payment times come from a billing lead at a family medicine practice. The two tools in the test example are made up for illustration.