A patient-access team does not get to choose its caseload. An oncology therapy priced at one hundred thousand dollars a year or more arrives with a patient who is under-covered, a payer who has already said no once, and a clock that is counting in days the patient does not have. Behind that one case sits a benefits investigation, a prior authorization, a denial, an appeal, and a funding stack stitched together from a manufacturer copay program and one or more independent charitable foundations. It is high-volume, judgement-heavy paperwork, and the queue already outruns the specialists available to work it. That is exactly why AI agents are already inside hub and access queues. The question is not whether they are there. It is what they are allowed to do once they are.
The honest answer starts with a distinction the technology is happy to ignore: the difference between investigating a case and attesting to it. An agent can do an enormous amount of the first. It must not be the actor of record for the second.
What an agent should do: investigate, draft, assemble
Most of an access case is assembly, not judgement, and a model handles assembly well. Given a denied claim, an agent can:
- Verify benefits across the patient's plan and read out coverage, tier, and the specific reason the therapy was denied.
- Draft the medical-necessity appeal, grounded in the chart, the guideline, and the payer's own stated criteria, so a clean letter is ready the same day.
- Assemble a proposed funding stack across manufacturer copay programs and independent charitable foundations such as PAN, HealthWell, and LLS, with eligibility windows and fund-open status checked.
- Flag the conflicts that matter, including the one that ends programs: a patient with Medicare or another federal benefit who must be routed away from manufacturer copay support toward an eligible charitable foundation instead.
This is fast, consistent, and tireless work, and none of it requires the agent to decide anything that carries liability. The agent prepares the case. It proposes the funding. It does not conclude that the patient is eligible, and it does not submit.
What stays human: attesting eligibility and submitting
Two actions in patient access are not clerical, and they are where the legal exposure concentrates.
Attesting medical necessity and submitting the prior authorization or appeal is the consequential step. A submitted attestation is a representation a named human stands behind, and a mis-attested medical necessity becomes a false claim downstream. The agent can draft the letter. It must not be the actor who signs that the patient meets the criteria and pushes it to the payer.
Enrolling a patient into a copay program or a charitable foundation is the other. Routing a Medicare patient into manufacturer copay support is not a clerical slip. It is Anti-Kickback Statute and False Claims Act exposure, the kind that ends programs rather than generating a memo. The liability lives in the gap between the benefits investigation and the enrollment decision, and that gap is precisely where an unsigned eligibility attestation slips through when the team is under pressure to clear the queue.
The line is clean. The agent owns the investigation. A named access specialist or pharmacist owns the attestation and the submission. The system has to make that line structural, not a matter of good intentions, which is the job of an AI agent control plane.
The control that makes the split real
Saying "a specialist signs off" is not a control. Anyone can paste an agent's draft into a portal and click submit. The control has to guarantee that the same identity that ran the investigation provably cannot be the one that attests eligibility or submits the case, and that the attestation, with its reasoning, is captured in a form an auditor can verify later.
This is the maker-checker principle, the four-eye separation that quality and safety functions have run on for decades, applied to a machine. The agent is barred in code from attesting or submitting on its own; enrolling a patient is a high-risk skill, and a high-risk skill requires a gate. The decision routes to a named human, and the meaning of that signature is recorded word for word.
| Step | Actor | Control |
|---|---|---|
| Benefits investigation and denial reason | Agent | Deny-by-default, versioned skill grant |
| Appeal drafting and funding stack proposal | Agent | Recorded, reversible, not a final call |
| Medical-necessity attestation | Access specialist | Approval gate; requester cannot self-approve |
| Copay or foundation enrollment | Access specialist | Approval gate with reason captured verbatim |
| Prior authorization or appeal submission | Access specialist | Signed, hash-chained audit entry |
The structural part matters. It is not enough for policy to say the agent should stop before submission. The same agent must provably be unable to act as both the maker of the investigation and the checker that attests eligibility on a single run. When it tries, the attempt is refused, and the refusal itself lands in the log, which is often the exact evidence an auditor most wants to see.
A case the way it actually runs
A patient is denied an oncology therapy priced at one hundred eighty thousand dollars a year. The agent verifies benefits, identifies the denial reason, drafts the medical-necessity appeal, and proposes a funding stack of a manufacturer copay card plus an independent charitable foundation drawn from PAN, HealthWell, or LLS. Then it stops. A named access specialist signs the eligibility attestation and authorizes the submission, and only then is anything submitted.
In a second case the patient has Medicare. The agent routes the patient away from manufacturer copay support toward an eligible charitable foundation, with the reason recorded word for word, and the access specialist signs before any enrollment. The agent did the work. The named human owns the decision the agent is designed against making alone, and the record carries both.
Why the record is the point
An access case is, in the end, a chain of decisions someone will be asked to defend. Who attested medical necessity on this patient? On what date, under which version of the eligibility rules? Who enrolled the patient, into which fund, and on what stated reason? Was the record altered after the fact?
A control plane produces exactly this evidence as a by-product of doing the work. Every model call, every grant state, every gate, and every signed decision lands in an append-only, hash-chained, cryptographically signed ledger. Change one entry and the chain visibly breaks. Each case exports as an Ed25519-signed evidence pack that an OIG auditor or a payer verifies offline, against a published spec, with no access to your systems. That is the form of proof an auditor trusts, and the obligations it produces evidence for, including the audit trail and signature-meaning rules, are the same ones we work through in Part 11 and AI agents.
The honest version of the pitch
AI agents will not replace the access specialist, and any vendor who implies otherwise is selling the part of the job that carries the Anti-Kickback and False Claims Act liability. What agents replace is the backlog, the hours spent verifying benefits, decoding denials, and stitching funding before a named human ever has to attest anything.
That trade is worth taking only if the split between machine throughput and human judgement is enforced and recorded, not promised. Get it right and the agent comes out of the pilot you cannot defend into a workflow you can put in front of an auditor, because every attestation, every enrollment, and every submission has a name on it, and a record that holds. The same logic governs the adjacent trial work, see how it applies to clinical trial data with AI agents.
See how it works, or book a demo to watch an agent get blocked from approving its own work, live.