Prior Authorization Automation: Where Voice Fits in the Payer Call Loop

Prior authorization automation uses software to assemble, submit, and track the approvals a health plan requires before certain care is delivered. Every vendor in the category describes the same thing: pulling documentation from the chart, matching it to payer rules, submitting electronically, and watching for a decision.

That covers the form. It does not cover the phone, and the phone is where a surprising share of the work still sits.

This post maps the calls that surround an authorization, which of them an AI voice agent can take, and which should stay with a person.

The thread running through it is ownership. Payer menus, documentation rules and denial patterns move constantly, so the practice that can change the agent the same week is the practice that keeps the savings.

TL;DR

• Prior authorization automation replaces manual form assembly and fax or portal submission with electronic, trackable workflows.
• CMS-0057-F took effect on January 1, 2026. Affected payers now have to decide expedited prior authorization requests within 72 hours and standard requests within 7 calendar days, and to give a specific reason for a denial. The standardized Prior Authorization API follows, generally from January 1, 2027. Drug authorizations are excluded from the rule.
• Even with all of that, four call types remain: status checks to payers, payer calls asking for more documentation, peer-to-peer scheduling, and patients calling to ask whether they are approved.
• Status checks and patient status calls are the two automatable ones, because both are lookups with a script.
• Clinical justification and appeals belong to a person, so the handoff path matters more than the deflection rate.
• Anything touching patient information needs a business associate agreement and a real audit trail per call.

What is prior authorization automation?

Prior authorization automation is the use of software to run the approval a payer requires before a service, procedure, or medication is provided.

An automated flow detects that a service needs authorization at the point of ordering, pulls the diagnosis codes and clinical notes from the record, checks them against the payer's rules, submits the request, and tracks it to a decision.

Newer systems add a predictive layer, flagging requests likely to be denied for missing documentation before they are sent. That is the part most vendors mean when they say AI.

What the category does not include is the conversation. When a submission stalls, when a payer needs one more detail, or when a patient wants to know whether their surgery is approved, someone picks up a phone.

What today's automation actually covers

It covers the structured, electronic path well and the unstructured path barely at all.

The regulatory direction reinforces the structured path. The CMS Interoperability and Prior Authorization Final Rule has required affected plans since January 1, 2026 to decide expedited requests within 72 hours and standard requests within 7 calendar days, to state the specific reason for a denial, and to publicly report authorization metrics each year, with the first set due by March 31, 2026. The standardized Prior Authorization API follows, generally from January 1, 2027.

CMS put the effect at approximately $15 billion of estimated savings over ten years, most of it from taking friction out of prior authorization. The AMA's explainer is worth reading for what the rule does and does not reach.

Two limits are worth holding onto when you plan around it.

• Drugs are excluded. Medication authorizations are out of scope, so pharmacy prior authorization keeps running on its existing rails.
• Payer adoption is uneven. An API requirement for some plan types does not make every payer's portal, fax line, and phone tree disappear. Practices will be working both paths at once for years.

So the realistic picture is a mostly automated submission process with a phone loop wrapped around it.

The part still done on the phone

Four call types survive almost every automation project.

1. Status checks. Staff call the payer to ask where a submitted request stands, usually because the portal shows pending and the procedure is scheduled.
2. Documentation requests. The payer calls or the reviewer asks for one missing item, and the request sits until someone answers.
3. Peer-to-peer scheduling. A denial routes to a clinical review conversation, which has to be booked around a clinician's day.
4. Patient status calls. The patient calls the practice, not the payer, to ask whether they are approved and whether their appointment is going ahead.

Only the first and fourth are high volume, repetitive, and script-shaped. Those two are the automation candidates.

The other two involve clinical judgment or a clinician's calendar, so the goal there is to get the right human onto the call faster, not to remove the human.

What a payer status call looks like turn by turn

It is a lookup wrapped in obstacles, which is why it eats so much staff time.

Nothing in that sequence needs judgment. It needs patience, accurate identifiers, and a note written into the right record.

It is also the call staff complain about, because steps 1 and 2 consume the time while steps 3 to 6 take under two minutes.

Where voice fits, and where it should not

An AI voice agent fits the two ends of the loop: calling out to payers for status, and answering patients who call in to ask about it.

On the outbound side, the phone tree and the hold queue are the whole cost, and neither is a human problem. Retell agents can press and speak their way through another organization's phone menu using Navigate IVR, then handle the authentication and identifier exchange when an agent picks up. A batch call run can sweep every pending case on a list rather than working them one at a time.

On the inbound side, patients asking about authorization status are asking a lookup question. An agent grounded in a knowledge base and connected to the case data answers it, and books or rebooks the appointment when the answer changes the schedule.

This is not theoretical. Medical Data Systems runs three agents across inbound support, outbound collections and insurer coordination, at roughly 30,000 calls a month, with around 70% of inbound calls completed end to end. On the patient side of the same loop, Pine Park Health reports a 55.7% booking rate with senior patients and scheduling NPS +38%.

Where it should not go is anywhere the call is a clinical conversation.

• Peer-to-peer review: a clinician-to-clinician discussion of medical necessity. Automate the scheduling of it, not the call.
• Appeals and medical-necessity argument: this is written and clinical work with real consequences for the patient.
• Anything where the patient is distressed: a denial for a scheduled procedure is bad news. It should be delivered by a person, and the agent's job is to recognize that and transfer the call with the context already in hand.

What breaks on these calls

The failure modes are specific, and they are the reason a voice agent that demos well can still fail on payer lines.

• Deep phone trees that change. Payer menus get reorganized without notice, so an agent that follows a memorized key sequence breaks. It has to read the menu it is given.
• Long holds and hold music. The agent has to stay on the line through several minutes of audio and then recognize a human answering mid-sentence.
• Interruptions and overlap. Payer representatives talk over prompts and read identifiers fast. An agent that cannot handle being interrupted mid-sentence will lose the authorization number.
• Identifier accuracy. Member IDs and CPT codes are alphanumeric strings read aloud. A single mistaken character sends the case down the wrong path, so read-back and confirmation matter more than speed.
• Transfers within the payer. Status calls frequently get moved to another department, and the agent has to survive the transfer and re-authenticate.
• Cold transfers to your own staff. If the agent hands off without passing what it already gathered, the caller repeats everything and the automation has saved nothing.

Test all six before you trust a pilot. A scripted demo call will not surface any of them.

HIPAA, the audit trail, and the human handoff

Any agent handling authorization calls is touching protected health information, so treat it as a covered workflow from day one.

That means a vendor that offers a business associate agreement before any PHI is transmitted, access controls on who can hear recordings, and a retention policy you can defend. Retell offers a self-service BAA and configurable data storage settings, from full retention down to basic attributes only, so the retention question has a real answer rather than a vendor promise. The HHS guidance for HIPAA-covered entities is the reference, and no vendor claim replaces your own risk assessment. Retell publishes its security posture in a Trust Center, which is the right place to check current attestations rather than a line in a blog post.

For the research side, clinical trial patient recruitment covers the pre-screening call most sites do not staff.

The audit trail matters for a second reason, which is that authorization disputes hinge on what was said and when. Post-call analysis gives you a per-call record of the identifiers exchanged, the status given, and the reference number, attached to the case. AI quality assurance reviews whether the agent read identifiers back correctly, which is the error that costs the most downstream.

On handoff, define the rules before launch: a denial of a scheduled procedure, a distressed patient, a clinical question, and any request the agent cannot verify all go to a person, warm, with the context attached.

What to automate first

Start with the outbound status sweep. It is high volume, it is script-shaped, and it carries no clinical content.

1. Pick one payer and one procedure family. Payer phone trees differ enough that generalizing too early hides whether the flow works.
2. Automate the status check only. Dial, navigate, authenticate, capture status, write it to the case. No clinical content.
3. Instrument it against the staff baseline: minutes per case, cases per day, and identifier error rate.
4. Add the inbound patient status line next, with a firm transfer rule for denials and distress.
5. Only then expand to more payers, and expect the second payer to need changes rather than a copy.

The teams that get value from this are the ones where the person who hears the failed call can change the agent that afternoon. If every change is a vendor ticket, the flow decays as payer menus and documentation rules move, which they do constantly.

That is the structural argument for owning the improvement loop rather than renting an outcome. Unlike managed AI vendors and BPOs, a change does not become a ticket, a queue, or another SOW.

Retell, a Customer Experience AI Platform for Autonomous Customer Relations, gives revenue-cycle and access teams direct access to the flow, the prompts and the transfer rules, so the person who hears a payer menu change can fix it that afternoon. The healthcare pages cover how the same setup extends to intake, reminders and eligibility calls once the authorization loop works.

Frequently asked questions

What is prior authorization automation?

It is software that detects when a service needs payer approval, assembles the required clinical documentation, submits the request electronically, and tracks it to a decision, replacing manual forms, faxes, and portal work.

Can AI call insurance companies for prior authorization status?

Yes. An AI voice agent can dial the payer line, work through the phone menu, authenticate with provider and member identifiers, capture the status and authorization number, and write the result back to the case. Clinical discussions such as peer-to-peer review should route to a clinician.

Does CMS-0057-F eliminate prior authorization phone calls?

No. Since January 1, 2026 it has required affected payers to decide expedited requests within 72 hours and standard requests within 7 calendar days and to give a specific reason for a denial, and it standardizes an electronic authorization path from January 1, 2027. Drug authorizations are excluded, payer adoption is uneven, and status and documentation calls continue alongside the API.

Is an AI voice agent HIPAA compliant?

Compliance is a property of the deployment, not the software alone. You need a business associate agreement, access controls, a retention policy, and a defined human handoff. Ask any vendor for its current attestations and read them against your own risk assessment.

How long does a prior authorization status call take?

The information exchange usually takes under two minutes. The phone menu and hold time in front of it are what make the call expensive, which is why it is a good automation candidate.

What should not be automated in prior authorization?

Peer-to-peer clinical review, medical-necessity appeals, and any conversation delivering a denial for scheduled care. Automate the scheduling and the paperwork around them instead.

Take the hold time off your staff.

Build a Retell voice agent that calls payers for authorization status, captures the reference number, and hands the clinical calls to your team. Start with one payer and one procedure family, run it against your staff baseline, and decide from your own call recordings before you sign anything. Start a pilot on your own calls.