AI in Claims Processing: Where Payers See Real ROI (and Where They Don’t)

Earlier this year, I joined ABBYY and Fierce Healthcare for a panel discussion covering a question every payer is wrestling with right now: Where does AI in claims processing deliver a return, and where does it just burn money? In our conversation, we took a candid look at what’s working, what isn’t, and what it takes to get past the pilot stage. As Sr. VP of Insurance at Naviant, I’ve spent the past 20+ years helping payers manage the pressure to do more with less using technology and strategic process design. But I was also honored to be joined by experts who brought a variety of perspectives and expertise:

• Sushma Akunuru, Senior Vice President and CIO at Independence Blue Cross, Philadelphia’s largest health insurer.
• Bruce Orcutt, Chief Marketing Officer at ABBYY, where he’s spent fifteen years.
• Kelly Hogan of Fierce Healthcare moderated.

If you missed it, feel free to view it on demand here, and read on to get the top takeaways worth your time.

AI in Claims Processing: What Payers Are Learning from the Front Lines

Why a Claim Is a Token of Trust in Claims Processing

Sushma opened with a reframe that set the tone for the whole session: A claim is the single unit that measures the interaction as well as trust between the members, providers, and payers. This framing is important because it fundamentally changes what counts as success. We are moving away from viewing claims as a purely administrative function and now see them as a way to build trust. In this new model, a successful claim is one that eventually becomes a message to the member that they’re in good hands. But to get there, payers must move away from manual processes that leave members stuck in a frustrating loop and instead prioritize a journey that is clean, fast, and predictable. Sushma put it from the member’s point of view: when you go through a health event, you want the administrative side to move through pipes that aren’t clogged. As a patient, you deserve transparency, certainty, and fast service you can trust, and a denied claim that can’t be explained works against this. From what I’m seeing right now, the thing that’s making the case for change unavoidable is competition. The reality is that members can choose another plan at any time, so the experience a payer delivers with claims plays a pivotal role in whether they stay or go.

Why General-Purpose AI Fails in Claims Processing

Many payers have already played with generative AI and walked away unimpressed. Bruce spoke to this, explaining that the bar is 100% accuracy in claims, so when payers turn to general-purpose GenAI tools that have a habit of guessing, it’s just not acceptable. He’s even seen hallucinations pull information from one patient’s record into another document because the model decided it fit. Bruce said, “It may give you ten different answers, which now brings a level of distrust from our token of trust.” But the other problem is cost. Bruce explained that the unpredictability of GenAI spend is the first thing customers flag, often before they even get to accuracy. His alternative is what ABBYY calls purpose-built AI: smaller, focused models trained on a narrow task, in this case, getting claim data correct predictably and accurately. GenAI does have a place in claims, but it just can’t be any genAI tool out there. Automating claims adjudication today calls for a focused model you can trust. This lines up with findings from a 2025 MIT report on the state of AI in business, which found that global enterprise investment in generative AI surged to somewhere between $30 and $40 billion, and that 95% of those investments weren’t producing a return. The explanation for this was misalignment with real value. In other words, organizations were deploying AI with the fear of falling behind, not because they’d proven it would help solve their problems.

Where Claims Automation ROI Really Shows Up

The biggest claims automation ROI shows up in three places, all still painfully manual and all easy to measure:

• Ingesting documents
• Extracting data
• Adjudicating the claim

On the side of ingestion, I’m currently working with a payer that has forty people manually sorting paper claims so they can be scanned into the system. And that’s not as rare as you’d hope. Letting technology handle that intake step will cut that payer’s costs by more than $1.5 million a year. On extraction, a single claim carries hundreds of pieces of information, and keying and indexing all of it by hand is slow and error-prone. Purpose-built AI accelerators automate 95% or more of that data entry out of the box.

“What would taking 95% of your data-entry effort off the table do to your bottom line?”

The third area is adjudication itself, moving that captured data and the EDI transaction into the downstream claim systems faster. Sushma added a fourth use case that’s emerging fast: explaining claims. Teams spend hours reconstructing why a claim went the way it did, from why it was denied to why a procedure that should have cost X came back as Y. AI can tap the various source systems and produce a plain explanation along with the next best action in minutes. She walked through what that looks like in practice. You ask the AI system one question: “For this member and for this claim, tell me what happened,” and you get the full story in return. Maybe the AI will tell you that part of the claim arrived on paper, and the other part came through another channel, and the member’s benefits changed midterm, which explains the amount. Then, the AI could explain how to walk the member through it.

Where to Start for Optimal Claims Automation ROI

When Kelly asked which parts of the claims life cycle are ready for automation today, Sushma pointed to intake first. Claims intake still hasn’t fully moved past paper: members track receipts, providers submit paper forms, and she’s even been in conversations where people question whether the EDI standard itself will survive the next decade. She named three areas with near-term payoff:

• Paper claims management, where unstructured, sometimes handwritten submissions still get converted to EDI 837s and mapped to the right member and claim by hand.
• Claims inquiries, the “tell me what happened to this claim and why” work that eats so much staff time.
• Prior authorization, where decisions used to depend on specialists who’d spent decades learning why one client or provider was set up differently from another. AI can now crawl thousands of records and millions of prior claims, recognize the patterns, and even anticipate trends before a claim arrives.

In my work with payers, it’s remained clear that humans still add the most value when it comes to accuracy. When vendors promise 99% accuracy, they often mean character-level accuracy on a document, which essentially every OCR technology can hit. What you really care about is field-level capture, because that’s what gets you to straight-through processing where no human has to touch the claim. Until that field-level accuracy is consistently trusted, humans stay in the loop. The model verifies member data against the database, validates and normalizes provider NPI data, and only routes the genuine exceptions to a person. Staff stop doing manual lookups and start handling the cases that need judgment.

Claims AI and Compliance: Why It Can’t Be an Afterthought

Audience questions kept circling back to the thing that keeps payers up at night: how do you use AI without compromising the privacy, security, and auditability of patient data? Bruce’s answer started with the people who own compliance inside a health plan, the ones who won’t let anything reach production without vetting, testing, auditing, and checking it against compliance standards. You can’t deploy an AI system that makes decisions affecting confidentiality, patient outcomes, revenue, and brand without the right gates built into the platform.

Change Management is the Hardest Part

One change management reality I speak to often is that technology is the easy part, and people are the hard part. We’re creatures of routine, and AI threatens routine. The most successful rollouts I’ve seen use a formal change management process that starts at the top, with executive leadership building the story for why the change is necessary and then telling that story repeatedly. I’ve watched leaders tell “the why” a hundred times before it sticks, because often, that’s what it takes. The fear that surfaces most often is “AI is going to take my job,” and that requires a reframe: AI takes the mundane parts of the job so people can spend their time on the work that needs a human brain, which makes the role more valuable, not less. When you put it that way, I’ve seen some of my customers’ loudest critics of change turn into its biggest champions once they were brought into the process.

What “Best-in-Class” Looks Like in 2-3 Years

When he was asked to look ahead, Bruce kept coming back to two words: real time and explainability. Patients increasingly expect to know, right after a visit, what they’ll owe, and the technology to meet that expectation already exists, so Bruce sees the ability to meet patients where they are is what will drive the next wave of innovation. Explainability is the other half. With so many sources of information now feeding the claim system, the goal isn’t only to adjudicate faster but to tell the member exactly why a decision was made, in language they can understand. Bruce expects claims to get there, and he made a point about pace that’s worth sitting with: AI is moving so fast that we may see two or three full versions of this life cycle before we even reach the two-to-three-year mark.

Ready to Find the ROI in Your Own Claims Operation?

Naviant works with payers at every stage of the AI journey, from mapping the use cases worth pursuing to deploying purpose-built automation that holds up in production. Whether you’re sorting through proof-of-concepts or ready to scale what’s already working, we’d welcome the conversation.