By Tracy Delgado, CSPO, RHIA, Sr. Product Manager, Payment Integrity

A VP of Payment Integrity sits down with the board deck for the third quarter in a row and the story hasn’t changed. Medical cost trend is up. Coding intensity is up, as providers lean harder on AI-assisted coding and documentation tools that produce cleaner, faster, better-supported claims, harder to challenge and quicker to submit. The mandate from the CFO is the same one it was last quarter: catch more improper payments, with the same headcount, on the same budget.

So, the VP does what a lot of PI leaders have done for the last two years. She greenlights an AI pilot. A vendor demo promised faster triage, sharper prioritization, and a model that reads a thousand-page medical record in the time it takes a reviewer to open the PDF.

Six months later, the pilot is still a pilot. The model is only as sharp as what it’s given, and what it’s given is claims data sitting in one system, authorization data in another, and clinical documentation in a format nobody thought to make machine-readable. The reviewers are still assembling context by hand. The model is summarizing the same fragmented picture, just faster.

We’ve seen this pattern repeatedly across payment integrity organizations evaluating AI. Not a model problem. A data problem.

The Pressure is Real, and It’s Compounding

Providers are investing heavily in AI-assisted coding and documentation tools. Analysis from the Blue Cross Blue Shield Association shows that increased use of these tools is contributing to rising coding intensity, in some cases without a corresponding increase in care delivered. The claims coming in are cleaner and faster, which sounds good until you’re the one trying to question them.

Recovery economics aren’t helping. Identifying an overpayment doesn’t guarantee recovery, and industry experience suggests only a portion of identified dollars are ultimately collected once operational cost and provider friction are accounted for. That’s why prepay identification is gaining momentum: catching an improper payment before it’s made avoids the recovery problem entirely. But shifting left requires something most PI programs don’t have yet, which is a data foundation that supports timely, high-confidence decisions before payment runs.

Layer AI on top of a fragmented foundation and you don’t get faster decisions. You get the wrong answers faster, and five reviewers looking at the same case still land in five different places.

AI-ready Data Isn’t the Same As Aggregated Data

Most payers aren’t data-poor. They have enormous volumes of claims, clinical, authorization, provider, and contract data. The problem is usability. A lot of vendors will happily get that data into one place for you. That’s necessary. It is nowhere close to sufficient. Gartner draws the same distinction, noting that high-quality data by traditional standards does not automatically qualify as AI-ready, because readiness depends on the specific use case the data will serve.

For a data foundation to support AI in a payment integrity workflow, and not just a dashboard, it has to do four things a simple data lake never will:

• Structure across data types. Structured claims and authorization data has to be explicitly linked to unstructured content like medical records and clinical notes, parsed and transformed into queryable assets, not just stored next to each other and called “integrated.”
• Shared context and meaning. Diagnoses, procedures, authorization decisions, member history, and provider behavior need consistent definitions so a model can understand how they relate across an episode of care and over time.
• Consistency at scale. AI is sensitive to incomplete or ambiguous inputs in a way human reviewers can often work around. Without normalization and validation, models produce probabilistic summaries that shift from case to case instead of holding up under scrutiny.
• Governance and traceability. Every transformation needs lineage. In payment integrity, insights have to be explainable and traceable back to source data, and defensible when a provider or regulator asks why.

This is the layer Abacus builds: a unified, payer-grade data foundation that connects claims, clinical, provider, authorization, and contract data once, and keeps it usable for analytics, AI, and the operational workflows payment integrity teams run every day.

What Changes Once the Foundation is Right

Imagine that same VP six months later, but the data foundation was fixed first. Claims are connected to clinical documentation and authorizations. Member activity can be viewed longitudinally across episodes of care, services, and points of interaction. Provider behavior is visible longitudinally across episodes of care, not reconstructed claim by claim. Policy and contract context sits at the point of decision instead of in a binder someone has to go find.

Now AI and analytics can do what they’re good at, and the questions the program can ask get a lot sharper:

• Which complex claims represent the highest risk, and how do you prioritize them first?
• Where does prepay editing lack the clinical, authorization, or policy context needed for a confident decision?
• How much time is spent on postpay audits assembling information instead of reviewing it?
• Which vendors, edits, or workflows are performing across the full payment integrity lifecycle?

Those four questions are the difference between a program that can show the CFO measurable ROI on AI, and one still explaining why the pilot needs another quarter.

Start With One Workflow, Not A Rebuild

Fixing data usability and ensuring it’s ready for AI doesn’t require ripping out existing vendors or pausing the program for a year while IT rebuilds the pipeline. Most organizations that get this right start with a single high-friction workflow, often complex claim audits or targeted prepay editing, where fragmented data is visibly and measurably slowing reviewers down.

Fix the foundation there first. Context gets assembled once instead of repeatedly. Reviewers stop starting from scratch on every case. Performance can finally be measured consistently, over time, across prepay and postpay. And every investment built on top of that foundation, analytics, automation, or AI, finally has a real path to scale instead of a permanent pilot. That’s what turns a data investment into medical cost containment, better payment accuracy, and an AI budget the CFO can see a return on.

 
If any of those gave you pause, the gap probably isn’t the model you’re evaluating next — it’s the foundation underneath it. Every quarter that foundation stays fragmented is another quarter of pilots that don’t convert.

Contact our team today to walk through where your data is costing you recovery dollars.

Frequently Asked Questions

What does AI-ready data mean for a health plan?

AI-ready data is engineered so a model can interpret it reliably, not just collected in one place. In payment integrity, that means structured and unstructured data are explicitly linked, definitions are consistent across an episode of care, values are normalized, and every transformation has traceable lineage. Gartner notes that high-quality data by traditional standards doesn’t automatically mean AI-ready.

Why do payment integrity AI pilots fail to reach production?

Usually because the underlying data is fragmented so the model summarizes an incomplete picture faster instead of producing a better decision. Claims, authorization, and clinical data typically sit in separate systems, leaving reviewers to assemble context by hand regardless of what the model does.

What is the difference between prepay and postpay payment integrity?

Prepay stops an improper payment before the claim is paid, relying on edits, coding logic, and clinical or authorization review during adjudication. Postpay identifies and recovers improper payments after the fact through audits and data mining, the pay-and-chase model. Health plans favor prepay because collecting an identified overpayment isn’t guaranteed.

Is AI-assisted coding by providers driving up healthcare costs?

Analysis from the Blue Cross Blue Shield Association suggests it may be contributing. A Blue Health Intelligence study of commercial inpatient claims across roughly 62 million members found per-member costs rose about 9% between 2023 and 2024, with coding intensity accounting for around 20% of that increase, concentrated in a small share of hospitals.

How can a health plan shift from post-pay recovery to prepay prevention?

It requires clinical, authorization, and policy context available before the payment run, which fragmented data environments often struggle to deliver consistently and in time for payment decisions. Most organizations start with one high-friction workflow, like complex claim audits or prepay editing, rather than a full data rebuild.