AI agents for prior authorization: what they do, and where a person must decide

An AI agent for prior authorization does the paperwork side of the request. It checks whether the payer needs one, builds the request from the order, the coverage and the chart, works the portal and the status calls, and drafts the appeal when the answer is no. It does not decide what is medically necessary. It does not submit on its own. A coordinator you name does that, with the evidence in front of them.
In one sentence: let the agent assemble and chase, let the chart supply the facts, and keep the submission with a person.
Why prior authorization is the queue people pick first
The volume is the reason. In the AMA's 2024 survey of 1,000 physicians, practices completed 39 prior authorizations per physician per week, and physicians and their staff spent 13 hours a week on them. 40% of physicians have staff who work only on prior authorization1.
The cost is not only time. 93% of physicians in the same survey report care delays from the process, and 89% say it adds to burnout1. A queue that large, that repetitive and that slow is exactly where an agent looks useful. It is also a queue where a wrong fact goes straight to a payer with a clinician's name on it.
What a prior authorization agent should do
This page is about the provider side: the practice or health system asking for approval. A good agent does five things.
- /01
Checks whether a request is needed. Service, payer and plan, against the payer's current list. If no request is needed, it says so and shows the rule it used.
- /02
Builds the request from the record. The order, the coverage, the clinical note and the imaging or labs the payer asks for. Every line points back to where it came from in the chart.
- /03
Fills the payer's form or portal, and saves the draft unsent.
- /04
Tracks the status. It checks portals and makes routine status calls. It logs the reference number, the date and what the payer said.
- /05
Drafts the appeal when a request is denied, using the denial reason and what the chart already contains.
The payer side is moving too. Under the CMS Interoperability and Prior Authorization final rule, Medicare Advantage, Medicaid, CHIP and some other payers must give a specific reason for every denied prior authorization since 2026. Most of them must also decide urgent requests within 72 hours and standard ones within seven calendar days2. From January 2027 they must run a Prior Authorization API. It lists covered items and services, names documentation requirements, and supports the request and the response2. An agent built for fax and portals today should be ready to use that API when payers turn it on.
Where it must stop
Five things always go to a person. None of them should depend on how confident the model feels.
- The submission. It states clinical facts to a payer on the clinician's behalf. A named coordinator sends it, and their name goes on the record.
- Anything clinical the chart does not say. If the payer wants a finding the note does not contain, the agent asks. It never writes the finding itself.
- A change to the order. If the payer would approve a different code, site or quantity, that is a clinical decision. The agent can explain the option. It cannot accept it.
- A peer-to-peer review. That is a clinician talking to a clinician. The agent prepares the file and stays out of the call.
- Any message to the patient. The agent can draft it. A person sends it.
The status call is the one place an agent talks to a payer directly. Keep that conversation narrow: reference number in, status out, notes written down. If the payer's representative asks a clinical question or suggests changing the request, the agent ends that part of the call and hands it to a person with the notes attached.
Payers use AI too. Measure your agent differently
Payers are adding AI to their own review. CMS's WISeR model, running in six states from January 2026 to 2031, uses AI and machine learning "along with human clinical review" for selected services in traditional Medicare3. Physicians are wary of the trend. 61% of physicians in the AMA survey are concerned that AI increases or will increase denial rates1.
That shapes how you judge a provider-side agent. Approval rate alone is the wrong score, because the payer controls half of it. Measure what the agent controls: requests sent complete on the first try, days from order to submission, and how many denials came back for missing documents. Run it in shadow mode on your own queue before it drafts anything a coordinator will send.
Text in a fax is data, not instructions
Prior authorization runs on faxes, scanned letters and portal messages. An agent that reads them reads everything on them, including text that tries to steer it. A line in a letter can ask it to resubmit somewhere else or to add a code. The fix is not a better filter. It is a rule the document cannot change: the agent may read, assemble and draft, and nothing it reads can widen that. Indirect prompt injection through documents covers the test.
What to ask a vendor before you buy one
- /01
Show me a request before it is sent. Does every clinical line point to the place in the chart it came from?
- /02
Who submits? The answer you want is a named person, every time.
- /03
What happens when the chart is missing something the payer wants? It should ask, not fill the gap.
- /04
What does the agent say on a status call, and when does it hand off? Ask for the script and the hand-off rule.
- /05
Where does PHI go? Ask where the model runs and what leaves your environment. Data residency and private deployment lists the questions.
- /06
Can I export every request, call note, hold and submission, with who decided? That export is your audit trail.
Gatehouse fit
Gatehouse, Surehand's control plane, sits in front of a prior authorization agent. Its scope is a signed manifest: read the order, coverage and chart, fill the payer form, check status. Submission, order changes and patient messages are held for the coordinator you name, and every request, call and hold lands in a sealed record. The healthcare page shows a prior authorization run end to end.
At a glance
| Category | Operations |
|---|---|
| Also asked as | AI agent for prior auth, prior authorization AI agent, AI for prior auth status calls |
| What it does | Checks the requirement, builds the request from the chart, tracks status, drafts appeals |
| Always held for a person | Submission, clinical facts not in the chart, order changes, peer-to-peer, patient messages |
| Never the agent's call | Medical necessity, accepting a changed order |
| Payer rules to build for | Specific denial reasons (2026), 72-hour and 7-day decisions, Prior Authorization API (2027) |
| The one test | Can anything written on a fax change what gets submitted? |
Common questions
What do AI agents for prior authorization do?
On the provider side, they check whether a service needs prior authorization with that payer, gather the order, coverage and clinical notes from the record, fill the payer's form or portal, track the status, and draft an appeal when a request is denied. A coordinator reviews and submits. The agent does not decide medical necessity.
Can an AI agent submit a prior authorization on its own?
It should not. Let it assemble the request and show the evidence it used. A named coordinator submits, because the submission states clinical facts to a payer on the clinician's behalf. Every submission should carry that person's name in the record.
Can an AI agent make prior authorization status calls to payers?
It can handle routine status calls: give the reference number, ask for the status, and write down what it was told. It should not agree to change the request, answer clinical questions or take part in a peer-to-peer review. Those go to a person, with the call notes attached.
Sources
- [1]American Medical Association, 2024 AMA prior authorization physician survey (PDF)fixpriorauth.org In text
- [2]CMS, CMS Interoperability and Prior Authorization Final Rule CMS-0057-F, fact sheetcms.gov In text
- [3]CMS Innovation Center, WISeR (Wasteful and Inappropriate Service Reduction) Modelcms.gov In text
Read next

What is shadow mode for AI agents?
Shadow mode runs the agent on live work while your people still do the job. The agent decides, nothing happens, and you compare. It is the cheapest go-live test there is.

Agent autonomy levels: where the human hold belongs
Autonomy isn't a setting for the whole agent. It's a choice you make per action. Put the hold where a mistake stops being cheap to undo.

What is an AI audit trail?
Seven fields. And why logs don't count. What your risk team will ask for, and what most systems can't produce.