By Sue Riddell, Director, Payer Strategy

Every conference has a question underneath it. At the Healthcare Payment and Revenue Integrity Congress in Chicago this September, part of the Medical Cost Containment Series, the question was how a health plan stops paying for the wrong thing without straining its relationships with the providers who get it right. 

It came up at the executive summit on self-insured employers, in the fraud and abuse sessions, on the payer panels, in a data operations workshop, and in a session on federal program integrity. It also came up in our own session, One Version of the Truth, where our founder and CEO, Minal Patel, MD, MPH made the case that prevention depends less on new detection logic than on having the evidence assembled before the decision gets made. 

Here are the five themes we heard most, and what each one asks of a plan’s data. 

1. The market is shifting from recovery to prevention 

The clearest signal of the week was directional. Speakers described moving away from recovery-only programs toward risk mitigation: catching an error before the payment goes out rather than chasing it afterward. 

That shift puts the current payment integrity business model to the test. Contingency fees are designed to reward recoveries, but that approach may no longer make sense when the whole goal is to prevent issues from arising in the first place. A few sessions echoed this: no single vendor covers the entire problem, and health plans make more progress by tackling the underlying issues, like confusing contract terms or policies that can be interpreted in different ways, rather than just adding more reviews. Now, with price transparency rules, it’s even easier for everyone to spot those inconsistencies. 

What it means for plans: prevention is measured differently than recovery, so the operating model, the vendor contracts, and the reporting all have to move together. Of the payment integrity trends shaping 2026, the move away from pay-and-chase recovery was the one every session returned to. 

2. Accuracy is the gate, and provider relationships are the reason

In the executive summit, speakers noted that as many as 90% of identified issues can turn out to be false positives. That figure costs more than review time. Every flag that does not hold up becomes a request sent to a provider’s office, and the plans on the panel talked about that cost openly and without defensiveness. 

Provider abrasion isn’t just a soft concern anymore; it’s a metric that gets measured. Similarly, plans are tracking the number of subject matter experts a problem consumes, and how much of that work can be automated responsibly. On the financial side, plans described measuring identified savings against realized savings, gross against net of fees, and their performance pass by pass and concept by concept. They then traced any gap back to its cause, which is often a data quality issue rather than a flaw in the logic.

What it means for plans: the shift to accuracy protects the provider relationship and the savings number at the same time. But it can only be demonstrated with measurements the plan owns. 

3. Clinical evidence is moving into payment integrity 

The most concrete ideas of the week came from teams pairing clinical data with claims. Plans described using admit, discharge, and transfer (ADT) messages and Continuity of Care Documents (CCD) through health information exchange connections to confirm when and where care happened and what clinical support exists for it. 

One payer panel described where that leads. Build an expected version of the encounter from the clinical record, a “shadow claim,” and compare it with the claim that was submitted. Where the two differ, a reviewer has a specific question to ask rather than a statistical outlier to defend. The same panel pointed to CDEX, the FHIR-based clinical data exchange, as the practical path to automating itemized bill review, which most plans still work line by line. Prepay review keeps expanding because this is where the savings sit. 

What it means for plans: the evidence usually already arrives somewhere in the organization. The hard part is connecting it to the claim, the member, and the timeline so a reviewer can use it in time. Prepay payment integrity depends on that evidence arriving early enough to inform the decision rather than explain it afterward. 

4. AI is welcome, with a human at the helm

AI came up in nearly every session, and the framing was consistent: AI should make a coder or an auditor faster and better rather than take their place. One panelist sharpened the idea usefully, “human at the helm” means directing AI, not simply reviewing its output at the end. 

Two governance themes ran alongside that. The first is transparency. Presenters encouraged plans and providers alike to be open about where AI is applied, and with litigation moving through the courts, documented AI procedures look like a shared expectation rather than a competitive risk. Payers and providers alike are adopting these tools quickly, which is why sessions on documentation integrity emphasized clinical judgment. Where a record needs a closer look, clinicians on staff read it rather than leaving it to a model. 

The second is discipline. Testing has to be segmented by line of business, because the same logic should produce the same result wherever it runs. Model cost has to be weighed against the savings it produces. And the industry now has a cautionary number to work with: MIT’s Project NANDA found that 95% of enterprise AI pilots deliver no measurable return. 

What it means for plans: explainability, segmented testing, and cost discipline are adoption criteria now, not things to sort out after go-live. 

5. Governance and data operations decide whether any of it works

Two sessions grounded the week. The first was about edit drift. Policies change, procedure codes change, technology changes, and off-cycle updates arrive, and edits gradually fall out of alignment across lines of business. The remedy described was operational rather than technical: generate edits from the policy itself, cross-reference policies across lines of business to find the gaps, and govern configuration, testing, and release as ongoing work. 

The second was about data operations, and it was the most familiar conversation of the three days. Utilization management, encounter, claims, lab, and EHR data arrive in different naming conventions and formats. Vendor reports have to be standardized before they can be compared. Analysts spend their time assembling data rather than reviewing claims. What made the session hopeful was the prescription: one governed location, documented data contracts, shared definitions, and reports built so a business user can drill down to an action. A related session made the same point from another angle, noting that 60% of one program’s fraud findings in a quarter came from sources other than claims. 

What it means for plans: this is the unglamorous layer, but it is the difference between a program that can prove its results and one that spends its meetings debating them. 

What it adds up to 

• Put the five themes together and a sequence appears. Prevention requires evidence earlier in the process. Evidence earlier requires clinical, contract, policy, and vendor data connected to the claim. Connecting it requires definitions every team shares and governance someone owns. And every AI ambition layered on top depends on all of it being true underneath. Taken together, that is the agenda the payment integrity trends of 2026 leave health plans to act on. 
• That was the argument of our session during the conference, and it is the work we do at Abacus Insights. We bring claims, clinical, pharmacy, eligibility, provider, and contract data together once, governed and traceable, with the context a reviewer needs to act on it, alongside the editing and audit platforms a plan already runs rather than in place of them. 
• If your team is working on any of this, there are two easy next steps. Read our white paper, From AI Promise to Payment Integrity Reality, for the longer version of the argument. Or reach out for a working session on your current vendor stack, where we will map out where your evidence gaps actually sit. 
• The plans furthest along were not looking for a better algorithm. They were making sure that when they pick up the phone to a provider, the record is already in hand. 

Frequently Asked Questions

What is the difference between prepay and postpay payment integrity? 

Postpay payment integrity works after the claim has been paid: the plan finds an overpayment and recovers it, usually through audits and contingency-fee vendors. Prepay payment integrity works before the money moves, using claim edits, contract terms, and clinical evidence to catch the error while the claim is still in flight. The distinction is also how the two are measured, since recovery counts dollars returned and prevention counts dollars never paid out. The shift described at HPRI MCC 2026 is from the first model toward the second.

Why do false positives matter so much in payment integrity?

Because they cost twice. Speakers at HPRI MCC 2026 noted that as many as 90% of identified issues can turn out to be false positives. Each one consumes reviewer capacity that could go to a claim that would hold up, and each one also becomes a records request or a phone call to a provider who did nothing wrong. That is why plans now track provider abrasion alongside savings.

How is clinical data used in payment integrity?

Plans are using admit, discharge, and transfer (ADT) messages, Continuity of Care Documents (CCD), and health information exchange connections to confirm when and where care was delivered and what supports it clinically. One approach discussed in Chicago is the “shadow claim”: an expected view of the encounter built from the clinical record and compared with the submitted claim, so reviewers start from a specific question rather than a statistical outlier. 

What does AI governance look like in payment integrity? 

In practice it means four things: a human directing the model rather than only reviewing its output, disclosure of where AI is applied, testing segmented by line of business so the same logic produces the same result everywhere, and a clear comparison of model cost against savings. The context matters because MIT’s Project NANDA found that 95% of enterprise AI pilots deliver no measurable return.