How AI-Powered Automation Can Solve the Payment-Posting Bottleneck

The healthcare industry has spent decades moving its payments off paper. Payers now send their payment and denial detail electronically through a standard file, the Electronic Data Interchange (EDI) 835, which cut out the paper and gave everyone a common format to work from. But while standardizing the data removed some of the manual effort, it didn’t make the work automatic. A file arriving electronically isn’t the same as a payment posting itself. The EDI 835 file lands, and people still open it, read each payer’s codes, post the payments line by line, and work the denials by hand.

That manual work is expensive, and the cost keeps climbing. In Experian Health’s 2025 State of Claims survey, 41 percent of providers reported that at least one in ten of their claims is denied, and Kodiak Solutions estimated that providers lost more than $48 billion to denials and bad debt in 2025, roughly a quarter more than the year before. Staff have to post, reconcile, and work every payment and denial before an account is settled, and most of that still happens by hand.

Why Posting Stays Manual

The reason posting stays manual has less to do with technology than with the rules themselves. Every payer handles their files differently, and the rules shift with each situation and exception so that the edge cases pile up faster than anyone can write them down. Many of those rules aren’t written down at all. They live in the heads of a few experienced billers, which means the organization can’t apply them consistently and can’t easily hand them to a system. When the rules aren’t captured, claims take longer to clear, avoidable denials slip through, and the first submission often collects less than it should have.

What Large Language Models Actually Do Here

Traditional automation handles the clean, predictable claims and stalls on everything else. A rules engine only does what someone programmed into it ahead of time. The first time a payer sends an adjustment code it hasn’t seen, the claim moves to a person for review. Payers vary enough that a large share of claims end up in that queue.

A large language model closes that gap because it reads a claim the way an experienced biller does. It takes in the context; works out what the claim is, which payer rule applies, and how to handle exceptions; and hands it to the system that posts the payment. The model supplies the judgment, and the system posts the payment the same way every time.

What We Built

Marlabs built a system that does exactly that for a healthcare billing company that handles payments for emergency medical services providers across North America. The company processes about 85,000 payments a month, and both the volume and the payer complexity were outgrowing what its team could keep pace with through hiring.

The most important part was how the system handled the rules the company hadn’t written down. Instead of loading it with a payer rulebook the company didn’t actually have, we had the AI learn from how the billers worked and from what the payers paid. It proposed rules, tested them against real outcomes, and tuned toward the result the business cares about most: collecting more of what insurance will pay on the first submission rather than after rounds of appeals. Knowledge that had lived in a few people’s heads became rules the company could finally see and improve.

A Person Reviews Every Claim

Every claim still passes through a person before it’s final. The system posts the claim and shows its work so that a reviewer can see what the AI did on a clean claim and approve it in a few seconds. When the AI isn’t sure, or when a biller had to step in, the claim carries a plain explanation of what looked wrong and what the system recommends. In these cases, AI makes the review faster while accountability for every claim stays where it belongs: with the people and the company.

The Results

Those design choices, more than the technology on its own, produced the results. The company now posts about 85 percent of its payments automatically, which frees its team to spend time on the claims that genuinely need judgment. Posting errors fell by 95 percent, and reconciliation across systems is cleaner. The team resolves payer-specific denials faster, and for the first time it can watch its exceptions and their resolutions as they happen. Because we designed the framework to scale within each payer’s compliance requirements, more volume no longer forces more headcount.

What This Means If You Run a Revenue Cycle

• Start with high-volume, rule-bound work like payment posting. If your rules are already documented, use them. If they aren’t, treat that as the opportunity, and let the AI find and test them against what payers actually pay.
• Keep a person on every claim, not only the exceptions, and have the AI show its work so that a clean claim takes only seconds to approve.
• Require the AI to explain every exception. A flag with no reason just moves the bottleneck from posting to review.
• Build real-time visibility in from day one so that leaders can see how the system performs and act on the numbers with confidence.

Payment posting is one of the clearest places to start, because the sums involved are large and the return shows up directly in cash collected. By mapping the process honestly, teaching a system the rules an organization never wrote down, and keeping people in control of every claim, you can see measurable, tangible results. Handled that way, payment posting turns from a stubborn manual bottleneck into a process the business can scale.