We Didn’t Design RCM. We Accumulated It

Every dental RCM vendor is talking about AI right now. But there’s a bigger question underneath the conversation: what kind of AI really makes sense for the way RCM works?

We sat down with Zentist Co-Founder and CEO Ato Kasymov to explore why Zentist took an agentic approach with Caviar and Remit AI—and what the industry’s history has to do with it.

His starting point is simple: RCM wasn’t designed. It was accumulated.

Over time, new problems brought new reports, portals, spreadsheets, vendors, and software. Each solved a real problem. But together, they created a fragmented environment where people often became the connective tissue between systems.

So what happens when you introduce AI into that environment? Is the answer to automate more of the existing workflow? Or do you need something that can operate when the workflow stops being predictable?

We talked to Ato about where scripted automation fits, where it falls short, why context matters before an agent acts, and how much autonomy an AI system should actually have.

Why do you say dental RCM wasn't designed—it was accumulated?

RCM wasn't designed as one system. It accumulated over time as new problems required new tools, workflows, and workarounds.

Nobody sat down and architected the systems DSOs run their revenue cycle on today. We accumulated them.

Every time a new problem came up, the industry added a report. Then a portal. Then a spreadsheet. Then another vendor, another piece of software. I call this the "Urgency Trap"; revenue has to be collected today, so every fix has to ship today.

Each decision made sense in isolation. Together, they built an architecture nobody actually designed.

And because none of these systems talk to each other well enough, people became the connective tissue. Someone downloads the file. Someone checks the bank. Someone remembers what the last contractor did.

Humans are still the integration layer holding RCM together.

That's not a knock on the people; quite frankly, they're the only reason any of this works today.

What is agentic AI in dental RCM?

Agentic AI can interpret what is happening in a workflow, determine what needs to happen next, and take action within defined boundaries—rather than simply following a fixed sequence of instructions.

That's important in RCM because the systems people work with aren't always predictable.

Payer portals and clearinghouses don't give you an API to integrate with. There's no clean data pipe to build against—just a login screen and a set of buttons, the same ones a person uses.

So the industry's answer has been scripted automation: a bot that clicks through the portal the way a person would, following the same steps every time.

Why isn't scripted automation enough for dental RCM?

Scripted automation works well when a workflow is predictable. It becomes brittle when something unexpected happens.

A script follows instructions, not judgment. The moment something unexpected happens—say, for instance, a session times out, a portal layout changes, or an error code you've never seen—it just fails, and a human has to pick up the pieces anyway.

An agent works that same system a portal gives any human user, but it reads the situation instead of following a fixed script. It decides what's actually being asked, and acts, or tells you clearly why it can't.

We learned this the hard way with document recovery for Remit AI.

We could successfully find and match 90–95% of remittances automatically—a great result. We thought we were basically done.

We underestimated the remaining 5–10%.

That's where credentials expire, portals change, and sessions break—the exact place scripted automation falls over.

Once you automate the majority of a workflow, the exception becomes the product.

If a predictable workflow works, let it work. Agents belong at the failure boundary, and increasingly, that's where Caviar's Claims Review lives.

Why does an AI agent need context to work a claim?

An agent needs the full history of a claim before it can make a useful decision about what to do next.

Claims, honestly, were the moment this became clear.

Our design partners told us: help us with the hard 10–20% of claims that eat all our time. We agreed, but we also knew an agent can't work a difficult claim properly without the full history.

Was there already a remittance? Had it been resubmitted? What did the last person—internal team, offshore team, contractor—already try?

Imagine a brand-new biller sitting down at a difficult claim who doesn't read the notes, doesn't check the PMS, ignores the clearinghouse, has no idea someone already worked this two months ago.

None of us would let a person do that.

So why would we let an agent do it?

Before an agent touches a claim, it needs the whole story.

Without that context, you haven't built intelligence; you've built another contractor starting from scratch, just faster.

How much autonomy should an AI agent have in RCM?

Autonomy has to be earned, workflow by workflow.

When we pushed hard on autoposting, we automated adjustments, notes, denial decisions, claim closure—real depth.

And our customers came back and said: great, now show me exactly what the system did, and why, and let me change the rule.

That's when it clicked for us: the more we automate, the more configuration and accountability have to grow alongside it.

"The system did it" is never an acceptable answer to a patient asking about their balance.

So every agent we ship carries that same requirement: traceable actions, a rule catalog that's configuration rather than a code release, and a clear owner for every decision.

The more autonomy we hand an agent, the less fragmentation we can tolerate around it.

What can Caviar's Agentic Claims Review do?

Caviar's Agentic Claims Review can investigate rejected claims, access payer portals and clearinghouses, identify what went wrong, and resubmit the claim after team approval.

It works the claims that used to mean digging through a payer portal by hand.

The agent logs in using the practice's own credentials, scoped to that account, diagnoses the rejection, and resubmits once your team gives the go-ahead, with every action written back to the claim's activity trail in Caviar.

That covers rejections across:

• Subscriber and eligibility data
• Patient information mismatches
• Claim and charge line errors
• Missing documentation

These are the categories that make up much of what a biller sees day to day.

A handful of cases still need a human touch: a payer call, an attachment, a manual match.

For those, the agent flags exactly what's needed and the RCM team takes it from there.

Anything unfamiliar still surfaces with the clearinghouse's exact message.

What happens when an agent can't resolve a claim?

The claim isn't simply passed back to the RCM team without context. The agent flags what it couldn't complete and why human attention is needed.

That distinction matters.

The goal isn't to pretend every claim can be handled autonomously. It's to give the agent responsibility for the work it can handle while making the remaining human work more specific and actionable.

That's the model going forward.

Caviar's claims foundation is live. Agentic Claims Review is generally available today, and Agentic Claims Statusing is next.

‍