How to make a medical billing company AI-native: a playbook

In 2024, Medicare Advantage plans made nearly 53 million prior authorization decisions and denied 4.1 million of them. Providers and patients appealed 11.5% of those denials, and 80.7% of the appeals overturned the denial in whole or part[31]. In HealthCare.gov marketplace plans the gap is wider: insurers denied 20% of in-network claims in 2023, and fewer than 1% of the denials were appealed[30].

Few of those denials turn on medicine. In the marketplace data, 6% of in-network denials cited medical necessity, against 21% for administrative reasons such as missing information, duplicates or late filing[30]. When the HHS Inspector General sampled 2019 Medicare Advantage denials, 18% of payment denials met Medicare coverage and billing rules. Most of those were caused on the plan's side by "human error during manual claims processing reviews" and "system processing errors"[6].

Many of these denials go unappealed for lack of staff hours. Someone has to read the 835, find the authorization, pull the right two documents and file the payer's form before the window closes. Teams work the largest first, and the rest age out.

The other side already runs software on your claims: the PXDX figure above is one insurer's, and CMS now uses AI in traditional Medicare prior authorization in six states[5]. The question for a billing operation is whether its own agents can do the dull work at the same speed without sending a single claim nobody on staff would defend.

An AI-native operation lets agents take the first pass on every queue, keeps people on every decision that changes what a claim says about care or moves money, and can show for any claim who approved what. Success is measured in cash collected per staff hour and appeals filed against appealable, with no claim sent that a person did not stand behind.

Agents now come with the PM system, the clearinghouse, the coding vendor and the ambient note tool, and each governs only its own actions. None governs an agent that reads an 835 in the clearinghouse, finds an authorization in a client's PM system, pulls two documents from the EHR and files an appeal in one task. That task is where a billing company's exposure sits: which client's data it read, whether it touched a code, and who approved what left.

OrchKernel sits in that gap. Agents ask it before they act; it checks the rules, holds what needs a person, and records the result against the named person the agent works for. It governs only the actions agents send through it. It is not a practice management system, a clearinghouse, an encoder or a claim scrubber, and it holds no charges, claims or payments. Details are in the OrchKernel blueprint.

US health spending reached $5.3 trillion in 2024, 18.0% of GDP[1], and a 2019 JAMA study put waste from administrative complexity at $265.6 billion a year[32]. Several figures below come from hospital and physician groups arguing against payer practices, and are labelled.

What this means for agents. The money is in appeals never filed, rework done before deadlines and denials prevented at registration: clerical tasks with clear inputs, such as an 835 line, a reason code, a payer policy and a date. The risk is elsewhere: anything that changes what a claim says about care can become a false claim.

A claim runs on standard transactions: the 837 claim, the 277CA acknowledgement or rejection, 270/271 eligibility, 276/277 status, 278 authorization, and the 835 remittance, the regulated payment and remittance advice[17]. Each 835 line carries reason codes (CARC and RARC) that say why a claim was paid, reduced or denied. Those codes are what agents sort on.

Categories, not recommendations. Every performance figure below is the vendor's own claim, directional and not an industry benchmark.

In a 2023 survey sponsored by AKASA, an AI vendor, with HFMA, 46% of hospitals used AI in revenue cycle work; the AHA's examples of results are self-reported by health systems and their vendors[36]. In an MGMA quick poll, 34% of medical groups had built a workflow tool with AI[38], and 61% of physicians in an AMA survey think payers' AI is raising denials[37]. We found no independent survey of AI use in medical billing companies.

Into roll-ups of services labor. In May 2025 New Mountain Capital combined SmarterDx, Thoughtful.ai and Access Healthcare, a revenue cycle services firm, into Smarter Technologies, which by its own account handles 400 million transactions a year and manages $200 billion in revenue[47]. Commure calls itself an "AI-native healthcare operations platform" whose aim is to "turn labor into software" (company claims)[46]. R1, one of the largest outsourcers, was taken private for $8.9 billion in November 2024[51].

Into the clearinghouses. Waystar reported $319.7 million of revenue in Q2 2026, up 18%[40], and in August announced agents that resubmit eligible denied claims on their own[39].

We found no comparable investment aimed at mid-size billing companies or in-house teams. They will get AI features from their PM system and clearinghouse while the roll-ups bid for their clients. What they can still decide for themselves is which agents act on their clients' claims, and with whose approval.

Six stages, ordered by risk: read before write, draft before submit, clerical before coding, and autonomy earned one class of work at a time. The order also follows the money, since few denials turn on medical necessity. A billing company can run different clients at different stages; an in-house team can do the same by payer group.

Stage 0 across the operation, then Stages 1 and 2 on one slice: one client or one payer group, and one denial family. We do this with design partners.

Agents can prepare every one of these: find the record, draft the letter, compute the balance. A named person makes the call, and the log shows who.

Inferred from the work, not from a survey. In each role the agent assembles the worklist, the packet or the posting, and the person keeps the judgment.

Across these rules the same requirements recur: a person accountable for what each claim says; access to PHI limited by role and need; records that show who did what, kept for years; clocks that start when someone identifies a problem; and, for patient-facing AI, notice and a way to reach a person. Each one is an operational control you can build.

Real cases first. Two are not AI failures and two are on the payer side; each still shows a control a billing operation needs.

Scenarios A to H are illustrations, not reported events. Each maps to the control points that stop it.

Here is what the controls look like on one ordinary claim. The agent does most of the work; it is stopped once, and a person approves once.

OrchKernel sits between AI agents and the systems a revenue cycle operation connects, as drawn in the missing layer. Agents ask it before they act; it checks the rules, holds what needs a person, and records what happened.

What it is not. OrchKernel is not a PM system, an EHR, a clearinghouse, an encoder or a claim scrubber. Those stay your systems of record, and the controls below sit around them.

Where a revenue cycle operation needs a control whatever tools it uses, who owns it, and how OrchKernel enforces it.

• Charges, claims and payments stay in the PM system, EHR and clearinghouse. OrchKernel does not scrub claims or replace the encoder.
• Coding accuracy belongs to the coding and audit program. OrchKernel can stop agents changing codes; it cannot tell whether a code is right.
• BAAs, the compliance program and the HIPAA risk analysis are contracts and processes. OrchStack makes no HIPAA or SOC 2 claim for itself.
• MFA and user provisioning live in the identity provider and the EHR. Payer portal terms are yours to check.
• Whether something is a breach or an overpayment is a legal and compliance call. OrchKernel supplies the evidence.

Limits. An agent inside a vendor's product that does not go through OrchKernel follows that vendor's controls, so ask each vendor what its agents change without a person. OrchKernel is source-available under the Business Source License and runs on your own servers, so you can read the code that enforces these controls. If you plan to offer it to client practices as a service of your own, talk to us about licensing first.

Record a baseline in Stage 0 and track the same numbers at each stage. Public figures exist for denial and appeal rates in some payer types, and they differ a lot by payer and setting. For almost every workflow measure there is no public benchmark, and HFMA and MGMA benchmark values are member-only, so compare each stage with your own Stage 0 numbers.

Last reviewed October 2026. Sources were read in October 2026; dates are publication or data dates.