Why Agentic Voice AI Can’t Fix RCM’s Coding Problem

·

A voice agent that can argue with a payer

On July 16, 2026, Raintree Systems — an EHR and practice-management platform built for rehabilitation and physical therapy organizations — announced it had acquired Spike Technologies, a San Francisco- and Vilnius-based startup building agentic voice AI for healthcare. The deal is a useful marker of where revenue cycle automation is heading, and a good prompt to ask a question coding and compliance teams should be asking anyway: automation of the claims process doesn’t automate correctness into it.

Spike’s agents are built to do more than read a script. According to Pulse2’s coverage of the announcement, the voice agents are designed to understand context, make decisions, and complete multi-step tasks on calls with payers — following up on claims, checking eligibility, and working prior authorizations — rather than routing a caller through a static phone tree. Raintree plans to fold that capability into its EMR, with availability and roadmap details expected at TherapyCon in Austin, Texas, August 19–21, 2026.

The economics behind the push

The numbers cited around the deal explain the urgency. Physical therapy practices face an average claim denial rate near 13%, and nearly three-quarters of those denials have to be appealed — work that, per the companies, eats up ten minutes or more of front-desk staff time per patient just on prior authorization. Raintree’s CEO, Nick Hedges, described billing as the hardest, most manual part of running a therapy practice and said that’s why the company chose to tackle it first; he has projected that RCM expense could eventually fall to roughly 1% of revenue under full automation — a company estimate, not an independently verified figure. Industry analyses the companies point to put the addressable opportunity from AI-driven RCM automation as high as $360 billion annually across healthcare.

Automating the call doesn’t automate the claim

None of that is a knock on what Spike built. Faster, more capable payer calls are a genuine improvement over hold-music purgatory. But the acquisition announcement is explicitly about the back half of the revenue cycle — appeals, follow-ups, eligibility checks — not the coding decisions that generate many of the denials in the first place. A voice agent that can out-argue a payer rep on a technical denial is still arguing from whatever the claim says. If the claim is wrong, a faster, more persistent appeal just gets you a faster, more persistent “no,” or worse, a successful appeal on a claim that shouldn’t have been billed that way.

That distinction matters because a large share of denials trace back to decisions made well before a claim ever reaches a payer’s queue:

  • Specificity gaps. A diagnosis code that’s technically valid but not specific enough to support medical necessity for the billed service.
  • Documentation that doesn’t match the code. The clinical note doesn’t contain the level of detail the ICD-10-CM or CPT code implies — a CDI problem showing up as an RCM problem.
  • Modifier and bundling errors. Missed or mismatched modifiers, or services bundled incorrectly, that trigger edits before a human ever reviews the claim.
  • Stale payer-policy mapping. Coding logic that hasn’t caught up with a payer’s current local coverage determination or medical policy update.
  • Inconsistent initial coding. The same encounter type coded differently by different coders or AI tools, creating denial patterns that look random but aren’t.

A voice agent can call about any of these. It can’t fix any of them. Fixing them means getting the code right against the documentation at the point of coding — which is a different problem than winning the appeal afterward.

What this means for coding and CDI teams

The practical takeaway isn’t that RCM automation is misguided — it’s that it’s incomplete on its own, and coding and CDI teams should treat that as leverage rather than a threat. As voice agents and other agentic tools absorb the mechanical work of chasing payers, the remaining differentiator for a practice or health system is whether the claims going out the door are accurate in the first place. That shifts the value of a CDI program from “catch problems before claims go out” to “catch problems before an automated appeal engine spends effort defending a claim that was never going to hold up.”

It also raises a quieter risk worth watching: as appeal automation gets cheaper and faster, there’s a temptation to treat a high appeal-success rate as proof the original coding was fine. It isn’t. An appeal can succeed because the clinical documentation, when someone finally reads it closely during the appeal, actually supported the code — meaning the problem was in how thoroughly the claim was reviewed before submission, not in the code itself. Automation that wins the argument after the fact doesn’t tell you whether the argument should have been necessary.

Where the real leverage is

For teams evaluating where to invest as agentic AI spreads across the revenue cycle, the Raintree-Spike deal is a reasonable signal that payer-facing automation is maturing fast, and that more of it is coming to specialties well beyond physical therapy. The coding-accuracy half of that equation — getting the right codes assigned against documentation the first time — is the half that determines whether all that automated persistence is defending good claims or just working harder to excuse bad ones.

That’s the problem Medikode’s automated medical coding platform is built to solve directly, by getting coding accuracy right before a claim ever needs an agent to argue for it.