Wael Khouli wrote here on July 26 that AI will not replace physician advisors, volume will. He is right, and his argument applies elsewhere in the clinic to the people who build prior authorizations. I build this kind of software and have seen automation rescue teams as much as I have seen careless implementations bury them. The difference comes down to the quality of existing data, workflows, and knowing when automation should say “I’m really not sure about this one.”
How clinical automation breaks
Dr. Khouli’s numbers apply well beyond utilization review. AMA’s latest survey puts prior authorization at 13 hours a week of combined physician and staff time for each physician. Behind that average there’s work surveys don’t capture. A reauthorization fails because one payer accepts PHQ-9 and another insists on a different depression scale. An appeal never gets filed because the person who would file it is still assembling tomorrow’s submissions. Dr. Khouli calls this “the work that never gets done at all.” Every practice I have looked inside has a pile that can often grow until a lapsed authorization becomes a clinical event. It becomes intractable and is often made worse by isolated attempts at automation like AI voice agents calling a payer and getting more false positives than not. Volume outruns staffing, and new service lines and silent payer policy changes result in a game of endlessly falling behind.
Two kinds of automation
The obvious answer is to automate, and many organizations already have. In a 2025 HFMA and FinThrive survey, 63 percent of health care organizations reported using AI in their revenue cycle. Only 15 percent of those adopters reported a positive return though.
That gap is not a model problem, it’s arithmetic again. Chain three AI vendors together, each 95 percent accurate, and one case in seven is sure to come out wrong at the end. The problem compounds when point solutions don’t talk to each other and staff cannot make sense of different results from different vendors. These silos mean staff cannot tell where the error or confident assumption comes from, so they must review everything with equal suspicion. Their work turns into proofreading and reconciling confident machines, which is slow, demoralizing, and hardly auditable. I’ve personally never seen automation work effectively without standardization of clinical flows and data, simple auditable human review steps, and wiring systems into distinct, independently auditable and guard-railed steps.
Real physician voices, twice a week
Free, and one click to unsubscribe.
What the good kind looks like
Good automation starts unglamorously: Map where your clinical facts live and whether they contain structured, standardized data like dates for a prior failed therapy or previous medication history. Automation thrives on high-quality data together with custom workflows to account for edge cases. If a patient’s labs from another clinic are transcribed into the encounter notes rather than a distinct field with values and dates, the first step is acknowledging this reality to cleanly extract that data repeatedly. In most practices the evidence payers demand is scattered. Rating scales sit in a note, and medication trial history is a paragraph of prose in a psych eval. Part of the record is a scanned fax. A system built on the assumption that the data sits where the vendor’s demo had it is destined to fail.
What follows are boring checks that require no advanced AI whatsoever. An unsigned note should block a submission. A claim should not leave for a payer the clinician is not enrolled with. A filing deadline should be a visible, black-and-white clock. The chart should then be scrubbed for contradictions; a medication list that implies a diagnosis nobody coded is a preventable denial waiting to happen. AI is already capable enough to flag these contradictions for review at scale, as long as it’s not allowed to manufacture clinical determinations it does not have the context or authority to make when exceptions are flagged. Again, the key is subsequent steps with narrow scopes, clear guardrails, using data which has been cleaned and normalized upstream to account for the clinic’s reality.
AI earns its place after this point, and the technology has crossed a threshold only in the past year. A model can now read a payer’s written medical-necessity policy and assemble a prior authorization against it, citing the payer’s own criteria. What matters is what it does when it is unsure. A useful system gives one of three answers: this is right, this is wrong, or I’m not sure. When it says I do not know, or the data is stale, it should hand over to a person together with the reason. The A1c on file is 14 months old, and this payer requires one within 12. A 40-minute hunt through the chart and the payer’s website becomes a 5-minute task. A tool that instead announces it is “75 percent confident” tells you little about where to look or why it’s 75 percent and not 74 percent.
The last step is the one that compounds most. Treat every denial and automation failure as a chance to re-examine the previous steps. If data tells you that comorbidity data is missing in 26 percent of your authorizations, it’s time to learn which payers warrant an extra step in your intake flows. Denial patterns tracked by payer and state show where a well-built automation flow is reaching its limits and why. Each insight becomes a new rule only after a person approves it so that judgment stays human, which is where Dr. Khouli says it belongs.
If you are evaluating one of these systems, ask three questions. Firstly, what are you giving the AI to check before it runs? If your MAs are confused, the model will also be confused, but a vendor’s AI might proceed with more unwarranted confidence than your staff would. Secondly, does the AI name the specific gap when it’s not sure about its work? How does it decide whether it’s sure or not? Finally, who approves what it learns from a denial? Large vendors might be reluctant to adapt their automation to your specific situation, while others might introduce logic silently that hurts more than it helps. Demos all look alike, so ask vendors how they deal with these questions in your specific specialty, state, and payer mix.
The arms race nobody needed
There is another reason to get this right: The reviewer on the other side is increasingly a machine. Medicare’s WISeR model has applied AI-assisted review to selected services in six states since January, and ProPublica documented a commercial insurer whose system let its doctors deny claims in seconds without opening the chart. KFF’s analysis of 2024 Medicare Advantage data shows the result of that speed imbalance: 80.7 percent of appealed denials were overturned, yet only 11.5 percent of denials were ever appealed. Unappealed denials stand but not necessarily because they were correct. Tackling them takes hours few teams have.
Dr. Khouli called the volume problem simple arithmetic. Disciplined automation can and will change the arithmetic. Rules you can inspect resolve black-and-white work. A machine that knows when to stop creates high-value tasks instead of noise. Your team should spend their recovered hours where they’re needed: clinical judgment, applied by someone who knows the patient.
This essay is cited in the KevinMD record on prior authorization.
JJ Martin Quesada is a health care executive.
