Skip to main content
DoneThat

AI Adoption GuideHealthcareBill

Autonomous prior authorization

Agentic system reads payer requirements, assembles clinical evidence, submits prior authorization requests, and follows up on decisions without staff intervention, using tools like AKASA or Cohere Health.

Healthcare processAccessIntakeAssessDiagnoseTreatDischargeBillFollowup

By Don, DoneThat’s AI coach · updated

A packet is ready only when both cites are present

A prior-authorization packet is ready when it cites a specific payer coverage rule and specific chart evidence that meets that rule. If either cite is missing, the packet is not ready. The coordinator still submits or holds. Nothing in this workflow auto-submits.

Quality means a reviewer can open the rule cite, open the note cite, and see the same criterion in both places without guessing. A paragraph that asserts medical necessity without a note is not a packet. A folder of unmarked PDFs is not a packet.

Staff remain in the loop. Platforms in this class, including AKASA, Cohere Health, Waystar, and Epic, can retrieve payer questionnaires and pull chart artifacts into a draft. They do not file the request. Language about running without staff intervention is the wrong control. A human still submits. A human still holds.

Do not spend assembly time on a service the plan does not cover. Confirm member status with real-time eligibility verification. When the benefit text is unclear, use the insurance benefits explainer so you are not building necessity for a non-covered code.

Load the payer rule and the chart, then map each criterion

Load the current payer requirement for this code, this diagnosis family, and this setting. Use the policy, LCD, or payer prior-auth guideline that is in force for this member and this date of service. Do not use a remembered list, a shared spreadsheet of usual criteria, or a rule copied from a different payer.

Capture the rule identifier, the version or effective language you used, and each criterion as a separate line. Typical lines include duration of symptoms, conservative therapy, prior imaging, specialist evaluation, and any lab or functional threshold the policy names. Then load the chart: clinic notes, therapy notes, imaging and lab reports, medication history, problem list, and the order itself. Map each criterion to a document, a date, an author, and a finding. Start from the rule, not from the order comment.

If the map has a hole, the hole is the work. Fetch the missing note or leave the field blank. Do not invent a coverage rule because this is what they always want. If the portal or policy does not state the criterion, the rule cite stays empty and the coordinator holds.

Here is one illustrative path, not a measured case. A lumbar MRI is ordered. The payer guideline on file requires a period of documented conservative therapy and a focused neurologic exam. The chart contains dated physical-therapy progress notes and a neurology visit that records motor and sensory findings. Those two artifacts, lined up with the two guideline clauses, are the dual cites. If the chart contains only an order that says failed conservative care, there is no therapy cite. The conservative-therapy field stays blank.

Dual citation: payer language beside the note

Every completed field needs two pointers. The payer pointer is policy ID, section, and the criterion text. The clinical pointer is note type, author, date, and the passage or finding. Write the packet as a scan table: criterion, rule cite, chart cite, and status. Status is met, blank, or conflicting.

Conflicting evidence stays on the table. Therapy notes that show improvement while the order claims failure are not resolved by hiding a document. The coordinator sees the conflict and holds, or asks the ordering clinician to reconcile the record before anyone submits.

Denial prediction before submission is a second pass on an already cited packet. It does not replace dual cites. A score without a missing-cite list does not tell you what to retrieve.

Treat vendor pre-fill as a draft. If a questionnaire line reads as if therapy duration was documented, and the notes never state duration or visit dates, that line is a necessity claim with no note cite. Clear it to blank. The same rule applies to diagnosis justification, failed-medication lists, and functional scores: if the number or phrase did not come from a named note, it does not belong in a filled field.

Leave blanks when the chart does not support the ask

Empty stays empty. If the chart does not contain the finding the rule asks for, leave the field blank and name the gap: no physical-therapy notes in the record, specialist letter not filed, imaging report unsigned or missing. Do not invent medical necessity. Do not turn an order comment into a history of present illness.

A blank is an honest packet. A completed sentence that cannot be opened to a note is a false packet. Compliance audit risk scoring will treat uncited necessity language as a documentation risk. Catch that risk here by refusing to fill the field.

When a blank blocks the request, hold and route a precise ask to the ordering clinician or health information management: which note, which date, which finding. The draft does not complete the form by inference. Inference is invented medical necessity.

Coordinator review: submit or hold

The assembled packet is a recommendation. The coordinator checks that every filled field has dual cites, that blanks are real gaps rather than missed attachments, and that the rule text matches this payer and this code. Then the coordinator submits through the payer channel or holds.

Do not treat the packet as submitted because a worklist status says ready, a vendor dashboard says assembled, or an agent queued a portal task. Until a person confirms send, it is a draft. Auto-submit is not part of this process. After a human submit, decision follow-up can be tracked: pend, approve, or deny, using only statuses the payer returned. Do not invent a decision the portal did not show.

If the hold is benefits, stop necessity assembly and return the case to eligibility work. If the hold is a missing note, do not submit a partial packet that implies the criterion was met. Partial submission is how a blank becomes a false met.

Three ways a clean packet is still wrong

A necessity claim with no note cite looks finished. The form says conservative therapy failed, the attachments are the order and the imaging request, and no therapy note is linked. Stop at assembly and either fetch the note or leave the field blank and hold.

Treating the packet as submitted looks finished. The internal queue shows complete and the vendor workspace shows assembled, but nothing was sent on the payer path. The patient is not authorized. The coordinator owns the send.

Inventing a coverage rule looks finished. The packet argues a step-therapy sequence the policy does not state, or skips a criterion the policy does state, because another payer's checklist was reused. Load this payer's text. If you cannot cite it, hold.

The quality bar is a packet that cites the payer rule and the clinical evidence, or that shows empty fields where evidence is missing. Staff submit or hold. This workflow does not auto-submit.

Is this worth automating for you?

Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other. This one is rated high effort to implement, so the baseline matters more than usual.

DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.

Measure the baseline first