The quote that should stop every healthcare data leader in their tracks came from Providence's Chief Health Information Officer Maulin Shah this week: "Unfortunately, what we're seeing is AI fighting AI."
He's describing the emerging reality where provider AI systems optimize clinical documentation and coding accuracy while payer AI systems automate claim denials at scale. Both sides are deploying machine learning. Both sides are training on the same clinical data. And neither side is talking about what this adversarial dynamic means for the data architecture underneath.
The Adversarial Loop Nobody Designed For
Providence Health System reports that AI tools now help ensure procedures are "accurately recorded, leading to more precise reimbursements." Translation: NLP models parse clinical notes, suggest more specific ICD-10 codes, and flag documentation gaps before claims go out the door.
On the other side, UnitedHealthcare, Cigna, and Anthem have been deploying AI-driven prior authorization and claims adjudication systems for years. These models are trained to identify patterns associated with upcoding, flag outlier claims, and automate denials at a speed no human reviewer could match.
The result is an adversarial feedback loop where each side's model trains against the other's outputs. Provider AI gets better at anticipating denial patterns. Payer AI gets better at detecting optimization patterns. The data engineering implications are enormous — and almost nobody is talking about them.
The Data Architecture Problem
Here's what most revenue cycle AI vendors won't tell you: the quality of your clinical data pipeline is now directly tied to your reimbursement rate.
When a payer's AI denies a claim, the appeal process increasingly requires granular clinical evidence — structured FHIR resources, timestamped clinical notes, lab results with proper LOINC codes, procedure documentation with precise CPT mappings. If your clinical data warehouse is a mess of inconsistent terminology, broken encounter linkages, and duplicated patient records, your AI coding assistant is building on sand.
This means clinical data engineering is no longer a "nice to have" for analytics teams. It's a revenue-critical function. The organizations that will win the AI-vs-AI revenue cycle game are the ones with:
- Clean, longitudinal patient records with proper entity resolution across systems
- Real-time clinical data pipelines that feed documentation AI at the point of care — not batch processes that run overnight
- Semantic interoperability between EHR, CDI, and billing systems — not CSV dumps and manual reconciliation
- Structured audit trails that can prove clinical necessity with evidence when a payer AI flags a claim
If you're running dbt models on top of a Snowflake clinical data warehouse, you should be asking: are my transformations optimized for clinical accuracy, or just for dashboard metrics?
Why "AI Fighting AI" Is Actually a Data Governance Problem
The adversarial dynamic between payer and provider AI creates a new category of data governance risk. When your AI coding assistant suggests a higher-specificity diagnosis code, who validates that recommendation? When a payer's model flags that code as an outlier, what structured evidence exists to support it?
This isn't a model accuracy problem. It's a data lineage and provenance problem. 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 healthcare organizations don't have this. They have AI tools bolted onto fragmented data stacks, generating recommendations from clinical notes that were never structured for machine consumption in the first place.
The organizations deploying clinical AI without investing in the underlying data architecture are building a compliance time bomb. When CMS or a state attorney general starts investigating AI-driven coding patterns — and they will — "the model suggested it" is not a defensible answer. A complete data lineage from clinical encounter to submitted claim code is.
The Real Strategic Question
Shah is right that the payer-provider relationship needs to adjust to this new reality. But "adjustment" implies negotiation between humans. What's actually happening is an escalating algorithmic competition where the data infrastructure determines the winner.
For healthcare data engineering teams, the mandate is clear: clinical data quality is no longer an abstract goal on your analytics roadmap. It's the foundation of an AI-driven revenue cycle that is already here. Your CDI workflows need structured output. Your claims pipeline needs deterministic lineage. Your clinical data warehouse needs to be an evidence engine, not a reporting layer.
The question isn't whether to deploy AI in your revenue cycle. It's whether your data architecture can survive what happens when both sides do.