Use case · Healthcare revenue cycle
Your coordinators review the prior authorization. The agent does the gathering.
When an order is signed, an AI agent checks whether the payer needs prior authorization, pulls supporting notes and results from the EHR, lines them up against the payer's criteria and pre-fills the request. PA staff and clinicians review, correct and submit it; the agent watches for the decision.
For revenue cycle leaders and practice administrators at US health systems, specialty groups and imaging centres submitting prior authorizations to many different payers.
One imaging order
A lumbar spine MRI, traced hour by hour.
An illustrative request. Payer, order and clinical facts are identical in both versions; only the legwork changes.
Today
Physician signs the MRI order. It joins a shared work queue.
A coordinator checks the payer's policy PDF and confirms the CPT code needs authorization.
She searches the chart for conservative treatment. Physiotherapy notes sit in a scanned referral letter.
The portal asks how long treatment was tried. The note is unclear, so she messages the physician.
Answer received; request submitted. The reference number goes into a spreadsheet.
Portal shows a request for more information. Nobody checked on Thursday.
Records faxed, decision pending. The scan moves to next week.
With the agent
Physician signs the order. The agent checks the payer's rules and flags that authorization is required.
Notes, imaging history and the scanned physiotherapy letter are retrieved and linked to each criterion.
Request pre-filled. One criterion lacks clear evidence, so the agent drafts the question to ask.
A coordinator reviews the packet; the physician answers the question in the EHR inbox.
The coordinator submits. The reference number is written back to the order.
The agent spots a request for more information and routes the documents for review that morning.
What the agent handles
Order in, packet out, status watched.
The agent does the searching and paperwork. People make every call involving clinical judgement or submission.
Conceptual flow, not a live system or measured result.
Decide whether authorization is needed
When an order is signed, the agent reads the procedure or drug code, diagnosis codes, site of service and active coverage. It checks the payer's published requirements, or its electronic rules service where one exists, and records whether prior authorization is required, not required or unclear. Unclear cases reach a coordinator with the policy excerpt attached. Orders needing nothing are released untouched, with the rule and payer policy version logged against the order for later audit.
Gather evidence against the criteria
The agent splits the payer's medical policy into individual criteria, such as treatments already tried, symptom duration or prior imaging. It searches the chart for each: progress notes, results, medication history, scanned letters. Every criterion shows a linked source passage or a plain statement that nothing was found. Access is limited to what that request needs. Where evidence is missing, the agent drafts a specific question for the ordering clinician instead of filling the gap.
Pre-fill, hand over and follow up
The payer's form, portal fields or electronic request are pre-filled with demographics, codes and linked evidence, and the packet lands in a coordinator's queue. The coordinator checks it, a clinician approves any clinical narrative, and staff submit. Afterwards the agent tracks status through the portal or API, prepares documents when the payer asks for more, and warns the team when a decision is running close to the scheduled date, so the appointment can be protected or moved early.
Clinical and privacy guardrails
Paperwork moves faster; clinical judgement stays put.
Clinicians sign off clinical content
Any statement about a patient's condition or need for a service is approved by a licensed clinician before it leaves the organisation. The agent drafts and cites; it does not attest.
No automatic medical-necessity submissions
Letters of medical necessity are drafts only. Submitting any request is a deliberate action by authorised staff, never a background job.
Minimum necessary access to PHI
The agent reads only chart sections relevant to the request, under a scoped service identity, and keeps data no longer than your agreed policy.
An audit trail for every step
Rule checks, retrieved documents, pre-filled fields, human edits and submissions are logged with user, time and source.
Staff time model
Estimate the coordinator hours at stake
Use figures from your own prior authorization log. Defaults are placeholders that show the arithmetic, not benchmarks or ColdAI results.
Hours today = requests × minutes ÷ 60. Hours after = requests × (pre-assembled share × review minutes + remaining share × today's minutes) ÷ 60. Value released = the difference × hourly cost. Platform and integration costs are excluded, and effects on scheduling or denials are not modelled.
Illustrative model, not a quote or a measured result.
Options side by side
Three ways to get a request in front of a payer.
Each has a place, and many organisations use all three for different payers.
| Criterion | Phone, fax & portals | Payer portal + RPA | Agent-assembled requests |
|---|---|---|---|
| Is PA required? | Staff read each payer's policy | Scripted lookups where portals allow | Checked per order; unclear cases go to staff |
| Clinical evidence | Manual chart search | Manual; bots move data, not judgement | Located and linked to each criterion |
| Notes and scans | Read by staff | Largely out of reach | Read and cited; uncertain items flagged |
| Form completion | Typed by hand | Fast for fixed-position fields | Pre-filled from the reviewed packet |
| Portal changes | Staff adapt | Scripts can break | Still needs access; APIs preferred |
| Status tracking | Spreadsheets and reminders | Scheduled status checks | Continuous, with information requests routed |
| Clinical sign-off | Informal | Unchanged | Required and logged |
Where we would hold back
Reasons to choose something else.
- Volume is low and concentrated on one or two payers. A good portal workflow or your EHR vendor's own tools may suffice.
- Much clinical documentation is on paper or outside the EHR. The agent can only cite evidence it can reach.
- You want requests submitted without staff review. We will not remove human submission or clinical sign-off.
- You are a payer automating authorization decisions. This page covers provider-side preparation, not utilisation review.
Systems it connects to
EHRs, payer channels and standards
- Epic
- Oracle Health (Cerner)
- athenahealth
- Availity and other payer portals
- CoverMyMeds for pharmacy requests
- HL7 FHIR R4 APIs
- Da Vinci CRD, DTR and PAS implementation guides
- X12 278 through your clearinghouse
Questions from compliance and revenue cycle
What leaders ask about PHI, payers and liability.
Will you sign a business associate agreement?
Yes. The agent processes protected health information on your behalf, so a BAA is signed before any PHI is accessed, covering subprocessors such as model hosting. Deployment can sit in your own cloud tenancy, with encryption, access logging and retention set to your HIPAA policies and reviewed by your privacy officer.
Does the agent decide medical necessity?
No. It checks administrative requirements, finds documentation and shows how it maps to the payer's published criteria. Whether a service is appropriate, and anything told to the payer about the patient, remains the clinician's decision. Drafted narratives stay marked as drafts until a clinician approves them.
How does the CMS prior authorization rule affect this?
The CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F), finalised in January 2024, requires certain payers, including Medicare Advantage, Medicaid and some Marketplace plans, to phase in FHIR-based prior authorization APIs over several years. Where a payer offers one, the agent uses it. Plans outside the rule may still need portals or X12.
Is automated use of payer portals permitted?
Portal terms vary. We review each payer's terms with your team and automate only where permitted, using credentials issued to your organisation. Otherwise the agent prepares the packet and a coordinator enters it. APIs and clearinghouse transactions are preferred wherever they exist.
What happens when the agent gets something wrong?
Every field and criterion links to its source, so reviewers see why a value was chosen and can correct it quickly. Corrections are logged and used to evaluate the agent. Nothing is submitted unchecked, and unclear policy matches go to staff rather than being resolved automatically.
Where should we start?
Services with criteria-based payer policies and documentation held in the EHR, such as advanced imaging, are a natural first step. We usually pilot one service line with a few payers, agree acceptance criteria with your team, and widen scope only when review data supports it.
Go deeper
The pieces behind this workflow.
What would you like to move forward?
Start with the outcome that feels closest. We'll help give the next step a useful shape.
One conversation. A clearer direction.
Tell us what you want to improve. A few sentences are enough to start.
shayan@coldai.orgLast reviewed 25 September 2026.