The denial is prevented in the room

The transaction cannot carry the argument

We have already covered the burden: 40 authorisations a week, 13 hours of physician and staff time, 40% of practices with someone working prior auth exclusively, the telephone still the most common method for medical services. (AMA, 2025 survey) This piece is about the design that actually changes those numbers, rather than the regulations that might.

An X12 278 — the HIPAA transaction for a services review — carries demographics, codes, units and dates. It does not carry the note. On 24 March 2026 HHS finalized the claims-attachment rule and explicitly declined to adopt an attachment standard for prior authorization, citing conflict with the FHIR work still landing in 2027. (91 FR 14350) So the transaction a vendor would build has no field to put the lumbar MRI in, or the two diagnostic blocks, or the failed conservative care. A 278 for an interventional procedure is a ticket-opener. The normal response is pended: send documents out of band.

Pain management makes the gap visible

Take a lumbar radiofrequency ablation. Palmetto GBA’s LCD for paravertebral facet joint denervation (L38765, Jurisdiction JJ — Tennessee is in it) does not ask whether the clinician thought the procedure was a good idea. It asks whether the chart shows two concordant diagnostic medial branch blocks, at the same level, at least two weeks apart, with at least 80% relief each. Fail any one of those and the authorization — and later the claim — fails for a reason that was knowable the moment the order was written.

That is not an exotic rule. It is the ordinary shape of interventional-pain coverage: conservative care of a stated duration, an outcome measure, correlating imaging, a frequency cap, sometimes a coverage exclusion (TENS for chronic low back pain is nationally non-covered under NCD 160.27 — a fact that should stop the order, not decorate the denial). The work is matching a planned service to a short list of chart artifacts. Most EHRs notice the mismatch weeks later, when a coordinator is assembling a packet from a visit nobody can quite remember.

KFF’s Medicare Advantage numbers make the cost of noticing late unusually clear. Of 4.1 million MA prior-authorization requests denied in 2024, only 11.5% were appealed — and 80.7% of appeals were partially or fully overturned. The overturn rate has exceeded 80% every year since 2019. (KFF, 2026) Nine in ten denials are never appealed. Of the one in ten that are, about four in five are reversed. The binding constraint is not the payer’s mind. It is staff time per packet, and whether the packet had the evidence in it when it left.

What AI is for, in this workflow

Not deciding coverage. Coverage conclusions are the payer’s to draw, and a model that asserts “meets criteria” is both wrong and, in a False Claims Act sense, unwise. The useful jobs are narrower, and they are the ones a model is actually good at.

Anticipate the requirement at the order, not after it

The moment an order or a prescription is signed, the question is already answerable: does this patient’s plan require an authorization for this CPT, HCPCS or drug? A payer connectivity layer can ask. A practice-owned list can answer when the payer will not — interventional-pain CPTs, advanced imaging, long-acting opioids, CGRP antagonists, Botox, genetic labs. Either way, the determination belongs on the order, in the chart, with a step-by-step record of what was checked. Waiting until a coordinator notices is how a two-week block gap becomes a six-week delay.

Bind every requirement to a chart artifact

A requirement that is a boolean the software asserts is a requirement nobody can audit. A requirement that is a pointer — this LCD clause, satisfied by this procedure-outcome row, dated this visit — is a requirement the packet can dereference. Two concordant blocks is not a sentence the model writes. It is two rows in a relief ledger, or it is missing. Missing is the useful answer, because missing is still fixable while the patient is in the room: capture the ODI, order the MRI, schedule the second block, or override with a reason.

Assemble the packet from the pointers, and do not editorialize

Once the bindings exist, packet assembly is not writing. It is listing what the chart documents, with dates and sources, and refusing to conclude medical necessity. Diagnoses, the relief-ledger paragraph, coverage, the relevant note excerpts. A coordinator who used to spend forty minutes hunting through a chart spends a few minutes confirming that the list is the list. That is the time arithmetic the AMA survey is actually describing, even though the survey does not say so.

Submit the administrative request, track it, and draft the appeal from the same record

When coverage and a diagnosis are on file, sending the 278 is an administrative act. Status polling is an administrative act. Recording a portal or fax certification by hand still has to exist, because a great many payers still live there. What should not exist is a second database of “PA facts” that are not in the chart. The appeal letter, if it comes to that, starts from the same bindings. We have no independently audited measurement of provider-side AI for prior authorization. We build this because the time arithmetic is obvious, and we will link the study when there is one.

What ChartVoyant does with an order

Every order and every prescription fires the prior-authorization agent in the background. It reads the patient’s active coverage, asks the payer whether authorization is required when we have a code, and falls back to the practice-owned list when the payer has no answer. The determination is stored on the request with a summary of what was checked and a step-by-step decision trail.

If a prior auth is required and we have coverage plus a diagnosis, the agent assembles the packet from the chart — diagnoses, relief-ledger narrative, coverage — and sends the coverage request. The packet reports what the chart documents. It never concludes medical necessity. Therapy and DME stay tracking-only: the rendering therapist and the DME supplier are the legal submitters, and a referring pain practice structurally cannot produce an initial therapy evaluation or a supplier’s Unique Tracking Number. What we owe on those two is documentation (plan-of-care certification, the Standard Written Order, the face-to-face encounter), which is a different job, and we treat it as one.

The Prior auth page is the live tracker: in-flight, certified, denied, not-required; what the agent did; the packet; the payer status; the certification number. Staff can still record a portal, fax or phone certification by hand. A request that is missing coverage or a diagnosis sits as a draft the page can finish. A clinician can always proceed over an unmet requirement with a reason; the override is audited. The engine informs. It does not practice medicine.

The question to ask a vendor

Not “do you do electronic prior auth?” Only 24% of physicians say their EHR offers it even for medications, so the answer is usually a portal screenshot. The better question is: when I order a lumbar RFA, does the system tell me, before the patient leaves, that the second diagnostic block is missing — and can I see the chart row it is looking at? If the answer is a worklist that appears next Tuesday, you are looking at a submission pipeline. If the answer is a requirement on the order, bound to a real artifact, with the packet already pointed at it, you are looking at the thing that actually reduces denials.

The 2027 federal API will make the ticket cheaper to send. It will not write the note. That is still the visit’s job, and it is the one AI is in a position to help with — not by deciding, by making sure the decision has the evidence under it before anyone hangs up the phone.