When provider AI meets payer AI, your data is the evidence
This week Providence’s Chief Health Information Officer, Maulin Shah, put words to something many revenue cycle leaders have felt: “Unfortunately, what we’re seeing is AI fighting AI.”
He was describing two sides using similar tools against each other. Providers use AI to improve clinical documentation and coding accuracy. Payers use AI to review claims and deny them at scale. Both work from the same clinical data. Very few people are talking about what this means for the data underneath.
A loop nobody designed
On the provider side, Providence says its AI tools help make sure procedures are accurately recorded, which leads to more precise reimbursement. In practice, models read clinical notes, suggest more specific ICD-10 codes and flag documentation gaps before a claim goes out.
On the payer side, large insurers have used automated review in prior authorization and claims for years. Those models look for patterns tied to upcoding, flag unusual claims and issue denials faster than any human reviewer could.
Put the two together and you get a feedback loop. Provider models learn to anticipate denial patterns. Payer models learn to spot the patterns provider models produce. Each side keeps adjusting to the other’s output.
Data quality now affects what you get paid
Here’s the part most revenue cycle AI pitches skip. The quality of your clinical data pipeline is now tied to your reimbursement.
When a payer’s model denies a claim, the appeal depends more and more on detailed clinical evidence. That means structured FHIR resources, timestamped notes, lab results with correct LOINC codes and procedure documentation with precise CPT mappings. If your clinical warehouse has inconsistent terminology, broken links between encounters and duplicate patient records, your coding AI is working from a weak foundation. So are your appeals.
That moves clinical data engineering out of the “someday” column on the analytics roadmap and into revenue cycle planning. The organizations in the strongest position will have four things:
- Clean, longitudinal patient records, with each patient matched correctly across systems
- Clinical data that reaches documentation tools at the point of care, rather than in an overnight batch
- Shared meaning across the EHR, clinical documentation integrity and billing systems, instead of CSV exports and manual reconciliation
- Structured audit trails that can show medical necessity when a payer’s model flags a claim
If you run dbt models on a Snowflake clinical warehouse, ask a plain question. Are these transformations built for clinical accuracy, or only for the numbers on a dashboard?
The real gap is lineage
AI on both sides also raises a governance question. When your coding assistant suggests a more specific diagnosis code, who checks it? When a payer’s model flags that code as an outlier, what structured evidence backs it up?
Model accuracy matters less here than lineage. You need to trace every AI-suggested code back to the clinical documentation that supports it, through a chain of custody that holds up under audit.
Most health systems can’t do that today. Their AI tools sit on top of scattered data and make suggestions from notes that were never structured for a machine to read.
That’s a compliance risk. If CMS or a state attorney general starts asking about AI-driven coding patterns, and I expect they will, “the model suggested it” won’t be an acceptable answer. A complete line of evidence from the clinical encounter to the submitted code will be.
Shah is right that payers and providers will have to adjust. I’d add that much of the outcome will be settled by data infrastructure rather than by negotiation.
For data teams, that means a few concrete changes. Clinical documentation integrity workflows should produce structured output. The claims pipeline needs lineage that gives the same answer every time you run it. And the clinical warehouse needs to hold evidence you can put in front of a payer, in addition to feeding reports.
Here’s where I’d start. Pull a sample of recent claims where an AI tool suggested or changed a code. For each one, try to trace the code back to the specific note, lab or procedure record that supports it, using only your data platform. Time how long it takes, and count how many you can’t finish. That gap is your roadmap, and it’s worth closing before an auditor asks you to do the same exercise.
Discussion