Serious AI with Chris Duff
September 29, 2026, updated October 23, 2026
Yes, for a small practice’s repeatable prior authorization steps, and we’d set two conditions: a person in the practice keeps the medical necessity call and signs off on every appeal, and the practice times its own authorizations before it buys anything. The Peterson Health Technology Institute (PHTI), which publishes independent evaluations of healthcare technology, lists drafting medical necessity justifications, completing submission forms and generating appeals among what AI solutions can do on the provider side.
The money half of the answer is less comfortable. PHTI’s industry insight on administrative AI, last updated October 15, 2026, says AI may lower what an individual organization pays to execute prior authorizations but has not lowered the cost of the process across the system. It adds that while insurers and vendors report AI can speed up steps of the process, there is no evidence yet that this lowers the average cost per claim once the cost of the AI solution is counted. Plan on getting staff hours back, and measure them. Do not plan on approvals getting cheaper or more generous.
That answer comes from Serious AI Partners, an enterprise AI implementation firm run by Chris Duff, M.D., who trained as a physician at the University of Illinois College of Medicine and spent 2017 through 2019 in healthcare venture capital. He first pointed AI at his own land investment firm, years before offering the work to anyone else. The work below is the operating side of a practice, meaning the portals and the paperwork around them.
An American Medical Association (AMA) survey of 1,000 physicians in late 2024, reported in an AMA article dated April 24, 2025, found that practices complete an average of 39 prior authorization requests per physician per week, and that physicians and their staff spend about 13 hours a week completing them. 40% of physicians have staff who work exclusively on prior authorizations. The AMA’s 2025 survey, of 1,000 practicing physicians and released May 13, 2026, put it at 40 prior authorizations a week and the same 13 hours of physician and staff time.
A separate poll counts the payer portals staff log into. The Medical Group Management Association (MGMA), the oldest and largest membership organization representing group practice administration, polled medical practice leaders on March 31, 2026 and received 252 applicable responses. 61% reported their staff access seven or more payer portals in a week, 35% between seven and ten and 26% eleven or more. MGMA Stat sends a polling question each Tuesday, by text message, to a panel of healthcare leaders who sign up to receive it, so we read that as what those leaders report about their own operations.
MGMA’s write-up of the poll notes that a single prior authorization request might require confirming eligibility in one portal, submitting the auth in another and checking status in a third, each with its own login cadence and navigation logic, and with status terminology that varies between them. In our read, that is why adding staff does little to the portal work itself. Another person logging into eleven portals is eleven portals of work, at a second salary.
The table runs the AMA’s per-physician averages, 39 requests and about 13 hours a week, for practices of 3, 8 and 15 physicians. It prices every hour at the median annual wage the US Bureau of Labor Statistics (BLS) reports for medical secretaries and administrative assistants, $45,930 in May 2025, treating 40 hours a week as one full-time job.
| Practice size | Requests a week | Physician and staff hours a week | Those hours at a staff wage, per year |
|---|---|---|---|
| 3 physicians | 117 requests | 39 hours | about $44,800 |
| 8 physicians | 312 requests | 104 hours | about $119,400 |
| 15 physicians | 585 requests | 195 hours | about $223,900 |
Read the last column as a floor. The AMA’s 13 hours include physician time, so pricing every one of them at a staff wage understates what they cost the practice.
That gives a plain payback test. A tool pays for itself when the staff hours it verifiably removes, priced the same way, add up to more than it costs each year. At this wage, every 10 hours a week that come off the work are worth about $11,500 a year, so ask the vendor for its total annual cost, divide, and you know how many hours it has to return before it breaks even. Run that against your own count, since an average across 1,000 physicians says little about any one practice’s payer mix.
A federal rule changes the plumbing for part of the payer mix. The CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F) requires Medicare Advantage organizations, state Medicaid and CHIP fee-for-service programs, Medicaid managed care plans, CHIP managed care entities and Qualified Health Plan issuers on the federally facilitated exchanges to implement and maintain a Prior Authorization API (application programming interface), beginning January 1, 2027. CMS says that API has to list the payer’s covered items and services, identify the documentation an approval needs, support the request and response, and come back with an approval, a denial with a specific reason, or a request for more information.
Part of the rule already applies. Since January 1, 2026, those payers, Qualified Health Plan issuers on the federal exchanges excepted, have had to send prior authorization decisions within 72 hours for expedited requests and seven calendar days for standard ones, and impacted payers must give a specific reason when they deny one.
On the practice side, MIPS eligible clinicians report a new Electronic Prior Authorization measure beginning with the 2027 performance period. It is a yes-or-no attestation: at least one prior authorization requested electronically through a Prior Authorization API using data from certified EHR technology, for a medical item or service other than drugs, or an applicable exclusion.
So the useful question for a small practice is a short one for its EHR vendor, asked in writing: will it support that API for those payers by January 2027, and what will it take to switch on? We’d also budget for portals staying in the picture. In our read, payers outside that list will keep their portals as the main door for a while, so plan for both.
Some practices are already pointing automation at it. In a February 10, 2026 MGMA Stat poll of practice leaders, with 177 applicable responses, 16% named prior authorization as where they are turning to AI and automation to improve front-office access, behind scheduling at 31%, calls at 27% and registration and eligibility at 23%. Those focused on prior authorization said they are applying payer-specific rules and improving status tracking, including rapid denial notifications and case flagging.
Serious Land Capital is the land investment firm where this method was built, buying and reselling raw land with its own capital at title, which means it absorbs the loss when it misprices a parcel. Underwriting, working out what a parcel is worth and whether the deal clears at that price, was the highest-value hour in that business and the one where being wrong cost the most. The Deal Engine is the software that now runs it every week.
The first fully measured automated run, on August 18, 2026, took three and a half hours and burned 10 to 15% of a week’s model allowance. Correct answer, unsustainable economics. The same run now finishes in under 30 minutes unattended at roughly 1% of that allowance. The procedure became software, 40,190 lines of it with 10,036 of those being tests, so the deterministic steps stopped being model calls at all, and each remaining call was matched to the model that step needs, with the top model held back for judgment.
One piece of that maps straight onto payer portals. The Deal Engine’s website navigation costs zero model tokens, because Claude (Anthropic’s AI model) watched a screen recording of the process one time and wrote the browser work into code. Clicking through a payer portal, pasting a policy number and reading back a status is the same kind of work, and in our view paying a language model to reason through it on every request is paying a premium for a repeatable answer. The model is held back for the two or three steps that actually need a read. Full before-and-after numbers are on the Serious AI proof page.
Browser automation is where the Deal Engine ran into reliability problems. Bot detection fired when parallel instances ran, and pages failed to load their graphics. The Proof page also records a layer that failed intermittently and had to be rebuilt instead of patched.
Payer portals add authentication, session timeouts and multi-factor prompts on top of all that. So ask any vendor quoting you prior authorization automation what happens when a payer redesigns its portal on a Tuesday, who finds out, and how fast. If the answer is that somebody notices when the queue looks wrong, the queue has already been wrong for a while.
The same applies to output. When the Deal Engine disagreed with the manual valuation by roughly $19K on a test deal, the cause surfaced the same day and the re-run landed at $87,992 against a hand-built $88,000. A disagreement like that only surfaces when something compares the system’s answer against a known one. A submission tool with no equivalent check is a tool that will file a confidently wrong request, and the first person to notice will be a payer.
Count before you buy, and count the same way after. Before any tool goes in, record the authorizations each physician generates in a week, the portals your staff touch, and how long a sample of authorizations takes from eligibility check to decision. After go-live, run the same count over the same number of weeks. Add two numbers tied to the payer rules: how many requests come back asking for more information, and how many days a standard decision takes against the seven calendar days CMS sets for standard requests.
Then judge the tool by whether those hours actually leave the staff schedule, and whether staff are still using it a few weeks after launch. The same test, applied to any AI pilot, is the subject of How Do You Tell Whether an AI Pilot Actually Saved Money?.
The Peterson Health Technology Institute published an industry insight, last updated October 15, 2026, drawn from a January 2026 convening of senior leaders across health systems, health plans, technology developers, investment firms and federal agencies. Its finding on this exact question: “AI may reduce the cost for individual organizations to execute prior authorizations, but it has not reduced overall system-level costs.”
Two further points from that report matter for a small practice. Health plans are starting to answer AI-driven increases in billing intensity with across-the-board downcoding and other reimbursement reductions, but the impact of those cuts is not yet known, and PHTI notes those responses may disproportionately harm providers that have not adopted AI documentation and coding tools, who are disproportionately smaller, more rural, critical access and independent. And where AI is applied on top of flawed workflows, it makes the underlying issues worse. Our read is that a tool bought on a revenue case can see that case eroded after purchase by the payer’s response to it.
Measure first, as above. Without that baseline, you have nothing to judge a purchase against later.
Then split the work before automating it. The repeatable navigation is the candidate: logging in, pulling the documentation, populating known fields, submitting, and polling for status. What we would keep with your staff, even where a tool drafts the paperwork, is the medical necessity determination and the sign-off on any clinical narrative or appeal. Those are judgment, and the whole value of the split is that you stop paying model rates for clicking.
Then decide who owns the output. Someone in your practice has to be accountable for what gets filed under your tax ID, with a sample checked against what a person would have submitted. Federal rules (45 CFR 164.502(e)) require a written contract or other written agreement before a business associate creates, receives, maintains or transmits protected health information on a covered entity’s behalf, so ask whether any tool touching your authorizations comes with one. Settle that, and any payer contract questions, with your counsel and compliance lead rather than from an article. The questions to put to the vendor itself are in How Should a Medical Practice Evaluate an AI Vendor Before It Buys?.
Yes, for the repeatable steps. PHTI, an independent evaluator of healthcare technology, lists drafting medical necessity justifications, completing submission forms and generating appeals among what AI solutions can do for providers, and we would keep the medical necessity call and the sign-off on any appeal with a person in the practice.
An AMA survey of 1,000 physicians in late 2024 found practices complete 39 prior authorization requests per physician per week, and physicians and their staff spend about 13 hours a week on them. The AMA’s 2025 survey, released May 13, 2026, reported 40 a week and the same 13 hours.
For one practice, it may. PHTI’s industry insight, last updated October 15, 2026, says AI may reduce the cost for individual organizations to execute prior authorizations but has not reduced overall system-level costs, and finds no evidence yet of a lower average cost per claim once the cost of the AI solution is counted.
Count staff hours on prior authorization before go-live, count them the same way after, and price the difference at a staff wage. At the BLS median for medical secretaries and administrative assistants, $45,930 in May 2025, each 10 staff hours a week returned is worth about $11,500 a year. If that is less than the tool’s annual cost, it has not paid back.
Under CMS-0057-F, Medicare Advantage organizations, state Medicaid and CHIP fee-for-service programs, Medicaid managed care plans, CHIP managed care entities and Qualified Health Plan issuers on the federally facilitated exchanges must implement a Prior Authorization API beginning January 1, 2027.
Since January 1, 2026, payers covered by CMS-0057-F, other than Qualified Health Plan issuers on the federal exchanges, must send decisions within 72 hours for expedited requests and seven calendar days for standard ones, and covered payers must give a specific reason when they deny a request.
If its clinicians are MIPS eligible, yes. Beginning with the 2027 performance period they attest yes to at least one prior authorization requested electronically through a Prior Authorization API using certified EHR data, for a medical item or service other than drugs, or claim an applicable exclusion. Confirm your reporting status with whoever handles MIPS for the practice.
Under 45 CFR 164.502(e), a covered entity’s assurances from a business associate that creates, receives, maintains or transmits protected health information on its behalf have to be documented in a written contract or other written agreement, so ask any vendor touching your authorizations for one. Have your counsel review its terms.
Chris wrote up why healthcare goes first, and how an MGMA poll on practice AI policies turned a guess into a number, in the September 28, 2026 issue of Serious AI News, his weekly newsletter for operators: My Plan To Turn A $2,400 Ticket Into A $2M Pipeline.
AI can produce the analysis. It cannot be wrong, because being wrong requires having something at stake, and that stays with a person. Serious means underwritten.
=====
If your practice is losing staff hours to payer portals and you want to know what can actually come off that work, the engagement starts with diagnosing the constraint that is gating you, then building with verification sitting where a person still owns the outcome. You leave the first call knowing what your constraint is whether or not you ever hire us.
Fill out a short form, and we read every answer ourselves. If there is a fit we will reach out and find a time that works for you, and if there is not, we will say so and point you toward someone better suited. Either way, the moat question is worth answering now, in a market AI has already reset.