How to evaluate an AI RCM vendor: questions that separate autonomy from hype

Walk any RCM trade-show floor and every booth says "AI-powered." The phrase has been sanded down to almost nothing. It now covers a system that submits claims end to end, a dashboard that flags risky ones, and a chatbot bolted onto an old workflow tool. Those are three different products at three different price points solving three different problems, and the website copy usually will not tell them apart. You have to ask. Below are ten questions that do the separating, and the answers that should make you lean in or back away. The frame to keep in mind throughout: who does the work — the software, or your team with software open in another tab?

Autonomy vs. assistance — the line that matters

"AI" is a feature word. "Autonomous" is a who-does-the-work word, and that is the one with money attached. An assistive tool makes a biller faster. Pull the person out and nothing ships. An autonomous agent completes the task and escalates the exceptions. Pull the person out and you lose oversight, not output. Both are legitimate purchases. But you should know which one you are buying, because a copilot makes a ten-person team a little quicker, while an agent changes how many people you need on the keystrokes at all.

The ten questions below are designed to force that distinction into the open.

Ten questions to ask an AI RCM vendor

1. What work does the system complete end to end, with no human touch? Not "assists with." Completes. If every answer sounds like "helps your team do X faster," you are looking at a copilot. That may be fine, if that is what you want.

2. What is your touchless rate? This is the single most revealing number: the share of claims or tasks processed start to finish with zero human touches. A vague answer here usually means the rate is low, and the "AI" is mostly assistive.

3. You quote accuracy, but accuracy on what fraction of the work? A copilot and an agent can both claim high accuracy. The difference is how much work the system is actually carrying. High accuracy on a low touchless rate only means the software is rarely wrong on the rare occasions it acts. Pair the two numbers. Accuracy alone does not tell you enough.

4. What rules govern the system, and when does it escalate? Real autonomy still needs clear boundaries: what qualifies for autonomous processing, what gets held for review, and how exceptions are handled. If a vendor cannot explain those rules clearly, it becomes difficult to evaluate how the system will behave in production.

5. What happens when the system is not sure? The honest answer is: it escalates to a person with the reasoning attached. A system that never says "I don't know" is not safer. It is harder to supervise.

6. Can I see why a decision was made? For a code assigned, a claim held, or a payment posted, can the system show its work: the rule, policy, evidence, or payer requirement behind the call? Transparency is the difference between an agent you can supervise and a black box you have to trust.

7. What exactly is in scope — and what is not? A vendor who says "we autonomously do everything across the whole revenue cycle" is telling you something, just not what they think. Ask them to draw the boundary. Front of cycle? Eligibility? Coding? Claim scrubbing? Submission? Posting? Denials? Appeals? A/R follow-up? The ones who can name what they don't do are usually the ones doing the rest for real.

8. How does this integrate with my EHR, PM, and clearinghouse? Ask about real interfaces: EDI/X12 transactions such as 837 and 835 for claims and remits, and EHR/PM connections and FHIR where appropriate for clinical and demographic data. "We export a CSV you re-import" is not integration. It is a copilot wearing a trench coat.

9. How is this priced, and against what outcome? Pricing tells you what the vendor believes the product actually does. Per-seat pricing signals an assistive tool sized to your headcount. Pricing tied to volume or work completed signals a vendor betting that the software actually carries load. Neither is automatically wrong. But the model reveals whether they are selling assistance or actual workload transfer.

10. Do you train AI on my data? In healthcare, this belongs in the contract. A genuine no protects your PHI and tells you the vendor does not need your claims to make the product work. A yes — or a dodge — is a real consideration for a healthcare buyer. The FTC has put AI vendors on notice that privacy and confidentiality commitments about customer data bind, however they were made.

What good answers sound like

Good answers are specific, paired, and written down. A vendor who can say, "Here is our touchless rate, here is our first-pass acceptance rate, here is how we measure both, and here is how work moves through the system," is operating in the autonomous category and is not afraid of the follow-up. Treat any figures in a sales conversation as illustrative until they are substantiated for your book of business. Your claim mix, payer mix, specialties, documentation quality, and system setup matter.

Bad answers share a texture: they substitute adjectives for numbers — powerful, intelligent, seamless. They quote accuracy without showing how much work the system actually handles. They cannot name what is out of scope. They get cagey about controls, integrations, or data use. The tell is not always dishonesty. Often, it is imprecision. And in RCM, imprecision in the sales cycle usually predicts imprecision in the claims.

One more move worth making: ask the same question two ways, an hour apart. Consistent, concrete answers hold up. Marketing answers drift.

How AcuraClaim answers them

We will take the ten questions in our own voice. AcuraClaim is autonomous, front-of-cycle RCM by Emergent Connect. We complete eligibility and VOB, autonomous coding, claim scrubbing and submission through 837, and ERA/EOB posting through 835, including Medicare Part A and Part B. We measure work by what runs without human touch, not by feature count. The platform autonomously completes the work within its defined operating scope.

Our scope is deliberately bounded. We do not do denials management, appeals, payer phone calls, A/R follow-up, or underpayment recovery. Those are real RCM disciplines, and we would rather name the line than blur it. We integrate through real interfaces, including EDI for claims and remits, standard EHR/PM connections, clearinghouse connections, and FHIR where appropriate.

If you want to run these questions at us live, book a demo.