14 MRIs on Thursday's Schedule Are Still Unauthorized. Claude Cowork Hands Back the Finished Packets.

4:15 on a Wednesday, and Thursday's Magnet Is Half Unauthorized

The prior-authorization coordinator at a three-modality outpatient center has the RadMD portal open in one tab, the Carelon portal in another, and a pile of referral faxes on the desk that came in during the lunch hour. Thursday's MRI schedule shows twenty-one slots. Fourteen of them still have no authorization number in the RIS. Two of the fourteen are lumbar spine studies where the referring family practice sent a one-line order and nothing else — no note showing six weeks of physical therapy, no imaging history, nothing a utilization reviewer will accept for CPT 72148.

She knows how the next three hours go. Call the referring office. Ask the medical assistant to fax the office notes. Wait. Re-key the demographics into a payer portal that logs her out every fifteen minutes. Find the ICD-10 that matches what the ordering physician documented rather than what the fax says. Attach. Submit. Repeat.

Nobody calls this a problem, because it has always been the job. It shows up later, in the aging report, when a study got scanned on a verbal from the referring office, came back denied CO-197 — no authorization — and the appeal window closed while everyone was busy.

What the Authorization Desk Actually Costs a Center Your Size

Advanced imaging is where the money and the friction both live. Screening mammography and plain film move through the front desk with a benefit check and go. MRI, CT, PET/CT and nuclear cardiology mostly do not. They pass through a radiology benefit manager — Evolent's RadMD, eviCore, Carelon — and each one wants a slightly different assembly of the same clinical facts.

The work is not hard. It is fragmented. Demographics sit in the RIS. The order is a fax in a shared folder. The clinical justification lives in the referring practice's chart, which you cannot see. The prior study that answers the "has conservative therapy been tried" question is in your own PACS from fourteen months ago, and nobody thought to look. One case means touching five places, none of which talk to each other.

Here is the clean version of what changed this year. A work-handoff agent is software you hand a goal to — "get every advanced-imaging case on Thursday's schedule authorized" — that then works across your scheduling system, your fax folder, your PACS history and the payer portal on its own, for hours, and gives back a finished submission packet for a human to approve. It is not a chat box. It returns work.

January 12 and July 9: the Year the Assistant Stopped Answering and Started Finishing

Two releases moved this from demo to desk. Anthropic shipped Claude Cowork on 12 January 2026, and OpenAI shipped ChatGPT Work on GPT-5.6 on 9 July 2026. Both were built for people who do not write software — your authorization coordinator, your billing lead, your lab manager. Both accept a goal rather than a step-by-step instruction, connect to the applications and files you already use, break the job into steps themselves, work unattended for hours, and hand back a finished artifact: a filled form set, a completed spreadsheet, a packet.

The difference from the 2024 assistants matters to you specifically. The 2024 version could draft a letter of medical necessity if you pasted in the clinical facts. It could not go find the clinical facts. It could not open fourteen cases, notice that four of them are missing the same document from the same referring office, and produce one consolidated request to that office instead of four. It could not work while the coordinator was at lunch.

flowchart TD
  A["Thursday schedule pulled from RIS"] --> B["Agent groups cases by payer and CPT"]
  B --> C{"Clinical justification on file?"}
  C -->|Yes| D["Packet built: order, notes, prior study, ICD-10 match"]
  C -->|No| E["One consolidated fax request per referring office"]
  E --> D
  D --> F["Coordinator reviews queue, approves or edits"]
  F --> G["Submitted to RadMD / eviCore / Carelon"]
  G --> H["Auth number written back to RIS, patient confirmed"]

The Same Wednesday, Run the New Way

At 7:30am the coordinator opens one screen. Overnight, the agent pulled Thursday's and Friday's advanced-imaging schedule out of the RIS, sorted the cases by payer and by procedure family, and checked each one against the prior-authorization requirement for that plan. For each case it opened the referral fax, matched the diagnosis narrative to a defensible ICD-10, searched your own PACS for a prior study of the same body part, and drafted the clinical summary a utilization reviewer expects to see.

Twenty-two cases come back in three buckets. Sixteen are packets ready to submit — she reads the clinical summary, glances at the diagnosis pairing, and approves. Four are flagged: the referring office never sent notes, and the agent has already drafted a single request to that practice listing all four patients by name and date of service, which the coordinator sends with one click instead of making four phone calls. Two are flagged as genuinely ambiguous, and those go to the lead technologist and, if needed, the radiologist.

By 9:40am the work that used to fill the afternoon is submitted. The coordinator spends the afternoon on the thing that actually needs her — chasing the two pending peer-to-peer reviews and calling the patients whose studies moved.

Running the Numbers on 62 Cases a Week

Assume a center that books about 62 advanced-imaging cases a week requiring authorization across MRI, CT and PET/CT. These figures are an illustration; run yours against your own timekeeping.

That is roughly $334 a week, about $17,400 a year, and it is the smaller half of the case. The bigger half is the studies that get performed without a valid authorization because the desk ran out of afternoon. If two cases a week end up denied for no authorization and written off after the appeal fails, and your average collected amount on an advanced-imaging study is, say, $340, that is $680 a week walking out the door. Recovering even half of it doubles the return on the whole exercise.

What You, the Owner, Have to Change About Assigning Work

This is the part most centers get wrong. If you hand this tool to your coordinator and she uses it the way she used the old assistant — one case at a time, one question at a time — you have bought nothing. The tool is priced and built for goals, not tasks.

Three concrete changes. First, write down what "done" means for an authorization packet at your center, in one page: which documents, which diagnosis pairings you will and will not submit, what the clinical summary must contain. That page becomes the instruction. Second, decide the approval gate and put it in writing — nothing gets submitted to a payer portal without a named human approving it, and that human's name is in the log. Third, change the coordinator's day: her measure stops being "cases worked" and becomes "cases authorized before the day of service" and "denials for no authorization."

Decide too which systems the agent may read and which it may write to. Read the RIS schedule, yes. Write an authorization number back after approval, yes. Touch the billing system directly — not in month one.

Where This Still Needs a Person

The peer-to-peer review is not going anywhere. When a reviewer wants to talk to the ordering physician or your radiologist about whether the PET/CT is indicated, that is a clinical conversation between two licensed people and it stays that way.

Urgent add-ons are also a poor fit. When an oncologist calls at 2pm needing a restaging CT tomorrow, the value is a person who can call the plan, get a supervisor and push. An agent working a queue does not push.

And payers change their rules quietly. A portal adds a required field, a plan moves a CPT code into or out of the review list. The agent follows the old path until somebody notices, so build a monthly fifteen-minute review of the last month's rejections. Finally, keep this away from the Advance Beneficiary Notice conversation — telling a patient they may owe out of pocket belongs at your front desk.

Frequently asked questions

Will this work with our RIS, or do we need to replace it?

These tools connect to the applications and files your staff already open, including web portals and shared folders, so eRAD, RamSoft, Merge or an Epic Radiant build all work. Start by letting it read the schedule export and the fax folder — that alone covers most of the work — before anyone talks about interfaces.

Is patient information safe in a tool like this?

It depends entirely on the agreement you sign, not on the technology. Before a single chart note goes in, get a signed business associate agreement covering the specific product, confirm your data is not used to train anything, and turn on the audit log. If a vendor will not sign, the answer is no, regardless of how good the demo was.

What happens the first time it submits something wrong?

It should not be able to. Keep the approval gate — the agent assembles, a named human presses submit. In the first ninety days, have the coordinator spot-check every packet, then move to a sample once you have seen the error pattern. Log every case where the human changed something; that log is how you improve the instruction.

Can it handle the October screening surge?

The October surge mostly lands on screening mammography, which usually does not need authorization. What hits your desk is the Q4 deductible rush — patients who have met their deductible and want the MRI they postponed in March. That is exactly the case for an agent: the work jumps for eight weeks and you cannot hire a coordinator for eight weeks.

One last thing, because it sits right next to the authorization desk: every one of those flagged cases generates a phone call, and the phone at an imaging center rings hardest between 8am and 10am when your front desk is checking in the first three MRI patients of the day. CallSphere builds AI voice and chat agents that answer the line around the clock, book and reschedule appointments, and capture the referral details so they land somewhere your coordinator can see them — which is a different job from the packet work above, but the same afternoon it saves.