AI Agents for Prior Authorization Workflows

Everyone remembers the group project where one person quietly did all the coordinating. Not the smartest person, not the loudest one. The one who texted everybody the night before, tracked who owed what, and stapled the thing together at 11pm.

Prior authorization is that group project, except the group members are separate organizations that never agreed to work together. The ordering clinician's office wants the MRI done. The payer wants documentation before it pays. The imaging center wants a valid auth number before it scans. The patient wants to know why a test the doctor ordered on Tuesday still hasn't happened two weeks later. Every one of them holds a piece of the status, and none of them can see the whole board.

The person stapling it together is the prior auth coordinator. Here is what her week actually looks like.

Monday through Friday, reconstructed

Take a coordinator at a multi-site orthopedic group. She's carrying roughly forty open authorizations at any given time. Two of them will define this week: a shoulder MRI for a patient with pain that hasn't responded to six weeks of physical therapy, and a specialty medication for a patient starting a new biologic.

Monday morning is portal rounds. She logs into six payer portals, one at a time, checking every pending case for movement. The MRI, submitted last Tuesday, still says "in review." Nine days now. She calls the payer. Forty minutes on hold, and the representative tells her the clinical documentation was never received. She is looking at the fax confirmation sheet from last Tuesday while he says this. She writes down his reference number, re-faxes the packet, and asks him to note the account. He says he will.

Tuesday, the imaging center calls. The patient is on their schedule for Thursday and they won't hold the slot without an auth number. Twenty minutes later the patient calls with the same question, more anxious. The portal still says "in review." She calls the payer again. This time a different representative finds the fax and says the request is pended: they need the note documenting conservative therapy before advanced imaging. That note exists. It's in the PT clinic's system, which is not her system, so she calls the PT clinic, waits for the records person to get back from lunch, gets the visit notes, and faxes them with the case reference number on the cover sheet.

Wednesday belongs to the other case. The specialty pharmacy rejects the biologic claim: there's an approved auth on file, but it was issued under the medical benefit and the drug is being dispensed under the pharmacy benefit. Same payer, same patient, same drug, wrong channel. Untangling it takes three calls, one of which ends when the payer's phone tree routes her to a department that handles neither benefit. In the middle of this, a peer-to-peer request lands for a third case, which means finding fifteen open minutes on a surgeon's calendar that match fifteen open minutes on a payer medical director's calendar, this week.

Thursday, the MRI is still pending. The imaging center releases the slot. The patient, understandably, is upset, and the coordinator spends part of the afternoon on the phone explaining a process she didn't design and can't speed up.

Friday morning, the approval appears in the portal. The decision is dated Wednesday. No one called, nothing pushed a notification, the status simply changed in a system she happened to check. The auth is valid. The imaging slot is gone. She calls the imaging center to rebook, calls the patient with the mixed news, and starts Monday's portal rounds early because the queue didn't shrink while she was doing all this.

Count the touches

Run the tape back and count what actually happened on the MRI case alone: six or seven portal checks, four phone calls to the payer, two calls from the imaging center, three from the patient, one records chase to an outside clinic, two faxes, one lost scheduling slot, one rebooking. Somewhere north of fifteen touches.

How many required her judgment? Two. Recognizing which PT note satisfied the conservative therapy requirement, and deciding how to sequence the resubmission so it didn't restart the review clock. Everything else was checking, chasing, re-sending, and relaying. Necessary, all of it, because if she stops chasing, cases go quiet and patients wait. But none of it used the thing she's actually good at, which is knowing this payer wants the therapy note attached up front and that payer will approve the same request in two days if you route it through their fax line instead of the portal.

She is not behind because she's slow. She's behind because the work multiplies. Every pending case generates its own orbit of status checks and callbacks, and the orbits stack.

This is a coordination problem, not an effort problem

Here's the uncomfortable part: every organization in this story is behaving reasonably by its own lights. The payer has a documented review process. The imaging center has a sensible policy about unauthorized scans. The clinic submitted a complete request. The PT clinic answered the phone and sent the notes same day.

The failure lives in the seams. The auth doesn't exist in any one system; it exists across five, and the versions don't talk. The payer's "we never got the fax" and the clinic's transmission confirmation are both true statements from inside two systems that share no memory. An approval that posts Wednesday and gets discovered Friday isn't anyone's mistake. It's what happens when status only moves if a human goes and looks.

So the coordinator becomes the integration layer. She's the shared memory, the notification service, and the retry logic, implemented as a person with a headset and a spreadsheet. Clinics respond by hiring more coordinators, and the new ones inherit the same orbits. You can't out-staff a structure problem.

She's the shared memory, the notification service, and the retry logic, implemented as a person with a headset and a spreadsheet.

The same week, replayed

Now change one thing. The team works in a shared workspace, and an AI agent works in it with them. Every authorization is a thread. The request packet, the fax confirmation, the payer's responses, the reference numbers from every phone call: all of it accumulates in one place anyone on the team can open.

The agent owns the chasing. It checks every payer portal on a schedule, several times a day, and posts to the thread only when something changes. When the MRI case pends for missing documentation, that surfaces Tuesday morning within the hour, not after a forty-minute hold. The agent reads the pend reason, finds the PT visit notes in the connected record, and drafts the resubmission packet with the case reference number on the cover sheet. It does not send anything. The draft sits in the thread waiting for the coordinator, who glances at it, confirms it's the right note for this payer's conservative therapy requirement, and approves it. Ninety seconds of judgment instead of ninety minutes of hunting.

When the approval posts Wednesday, it's in the thread Wednesday. The imaging center's scheduler, looped into that thread, sees the auth number the same hour. The slot survives. The patient gets a call with good news instead of making a third call asking for any news.

The benefit-channel mess on the biologic still takes a human. Phone trees don't negotiate with software any better than they do with people. But the agent had already flagged the mismatch between the auth's benefit type and the pharmacy's channel before the rejection came through, so the coordinator started that fight a day early with the history in front of her instead of reconstructing it from sticky notes.

Her week has maybe a dozen touches now instead of a hundred and something. The dozen are the ones that matter: the peer-to-peer prep, the channel dispute, the calls to worried patients. The thirty routine status checks happen without her.

What the coordinator becomes

Nothing about this makes the coordinator less necessary. It makes her more visible. When the chasing is automated, what's left is the work that was always the real job: exception handling and patient advocacy.

She preps the surgeon for the peer-to-peer with the payer's own criteria language pulled into the thread. She writes the appeal for the denial that's genuinely wrong, with every prior touchpoint documented and timestamped without anyone having assembled it. She calls the patient whose case is stuck and tells them what's happening before they have to ask. She notices that one payer has pended four shoulder MRIs in a row for the same reason and fixes the intake template so the fifth one goes out clean.

That last one is the interesting change. Someone drowning in status checks never gets enough altitude to see the pattern. Someone reviewing a board that maintains itself sees it in a week.

Where the line is

None of this works without hard boundaries, and they need to be stated plainly, not assumed.

The agent never makes a clinical judgment. It doesn't decide whether conservative therapy was adequate; it finds the note that documents what happened and a human confirms it's the right evidence. It never selects or adjusts codes to improve the odds of approval. It never edits clinical documentation, summarizes it into something a clinician didn't write, or files an appeal argument on its own. It assembles, tracks, chases, and drafts. Anything that leaves the building carrying clinical content gets a human sign-off, every time, and the workspace should make that sign-off a visible step in the thread, not a checkbox someone can script past.

The thread structure is itself a safeguard. When every action the agent takes is posted where the whole team can see it, review isn't a separate audit exercise. It's just reading. If the agent chased the wrong case or attached the wrong note, someone catches it Tuesday, not at the quarterly compliance review.

Prior authorization will stay a group project. Too many parties have legitimate stakes for it to collapse into one system anyone controls. But group projects don't fail because they have many members. They fail when nobody can see the whole board and one exhausted person carries the follow-through alone. Give the follow-through to something that never forgets to check, put the whole record where every member can see it, and the person who used to staple it together at 11pm gets to do the job she was hired for.