An algorithm denied your fusion 🚫, AI's real risk is never-skilling 🧠, The care with no billing code 🕯️
An algorithm started denying spine fusions — and the bug was hiding in a 15-year-old code
Edward DelSole, MD, a fellowship-trained orthopaedic spine surgeon in Reading, PA, documented a sudden wave of denied fusion authorizations and traced all of them to one culprit: CPT 20930, the morselized-allograft add-on.
The timeline is the tell. Effective May 1, his payer moved musculoskeletal prior auth in-house and switched to an automated review engine; the denials clustered about three weeks after cutover.
Here’s the trap. A 2011 CPT revision folded several different bone-graft materials into that single code, so 20930 is materially blind — you cannot tell from the code which material was used, and some uses are flagged experimental.
A human reviewer used to read the operative narrative to resolve that ambiguity. The automated engine can’t, so it resolves ambiguity the only way an unvalidated classifier does: it denies.
This isn’t a new coverage policy. It’s a latent flaw that’s been sitting in every payer’s policy library since 2011, and automation just activated it.
For builders, this is the eval thesis with a patient on the table. The fix was never a better model — it was the boring step that got skipped: validate the engine against the human reviewer’s historical approvals, and route the discordant cases to a person. The most durable product on this seam isn’t a denial-predictor. It’s a tool that watches the rules — the policy bulletins and code edits — and flags which of your in-flight cases just became “ambiguous” to a machine.
😤 “This is just a billing glitch, not an AI story.” It’s both, and that’s the point. A glitch that lives in one ambiguous code becomes a systematic denial pattern the moment you put a literal-minded reviewer in front of it at scale. The codebook was always flawed; we just hired something that reads it exactly as written.
😤 “So humans were better. Great, hire more humans.” No — humans were validated, by 2011-to-2026 of reading narratives. You can absolutely automate this; you just have to benchmark the engine against the baseline it’s replacing before it touches a live case. Skipping that step isn’t an AI problem. It’s a deployment-discipline problem.
😤 “Spine codes are a niche. Doesn’t generalize.” Every specialty has its materially-blind codes — the bundled add-on, the unlisted-procedure catch-all, the modifier that means three different things. Spine just got there first because it’s full of them.
❓ If the most defensible product here is “monitor the rule changes, not the claims,” who’s actually positioned to build it — the surgeon who feels the denials, or the RCM vendor who’ll never see the operative note? I keep thinking the edge belongs to whoever has read both the bulletin and the chart, and almost nobody has read both.
📡 Builder’s Radar
The real risk of clinical AI isn’t a wrong answer — it’s “never-skilling”
Jan Beger reframes why AI deployments stall: it’s not the technology, it’s the unfunded psychological work of a clinician renegotiating their identity once a machine offers an opinion first.
The data underneath it is striking. A spring AMA survey of ~1,700 physicians found 81% now use AI (more than double 2023) — but 88% worry it’s eroding their skills, with the worry highest among those ten years or less into practice.
Stanford’s Carl Preiksaitis names the failure mode: not de-skilling but never-skilling — a clinician who never builds the judgment to catch the AI, because the AI did the reasoning before they got the reps.
If you build clinical tools, you’re not just shipping a feature — you’re shaping whether the next generation can still tell when you’re wrong.
😤 “That’s a training problem, not my problem as a builder.” It’s a design problem, and it’s yours. Deep workflow integration lowers identity threat; forcing a per-decision sign-off raises it. How you ask for the human’s attention determines whether they stay sharp or check out.
The most important thing a clinician does often has no billing code
The most awesome John Lee, MD, an emergency physician and longtime Epic physician-builder, tells the story of canceling an ambulance transfer so a dying child could spend her last hours at home — an act with no RVU, no quality measure, nothing for the dashboard to see.
His argument cuts at most health-tech pitches: deploy AI on a system that mistook the score for the game and you don’t fix the foundation, you industrialize the dysfunction.
The design question he leaves you with isn’t “how much friction does this remove” — it’s “which friction.” Strip the faxes and the redundant alerts; protect the eye contact and the quiet room.
😤 “Nice sentiment, but sentiment doesn’t ship.” It ships a better filter. “Which friction” is a concrete product spec: it tells you what to automate aggressively and what to leave gloriously manual. That’s more useful than another throughput metric.
💡 80/20: Before you remove a step, ask whether it’s wrong friction (faxes, prior auth, duplicate alerts) or right friction (the human moment). Build a tool that kills the first and defends the second.
Malpractice insurance wasn’t built for AI or for the care models eating the edges
Nikhil Krishnan lands on an uncomfortable truth about how malpractice actually works: “standard of care” is whatever an expert witness can convince twelve non-clinicians of — and AI plus new-model practices are fracturing it fast.
The vivid example: a whole-body-MRI screening case where an AI-assisted scan allegedly missed a 60% stenosis. Screen more healthy people continuously and you expand the surface area for missed-finding suits — even as there’s no settled standard to be measured against.
If your tool surfaces a finding, you’ve also created a record of what a “reasonable” system should have caught. Liability is a product surface now, not a footnote.
😤 “Carriers will just write an AI exclusion and move on.” Maybe — but an exclusion is a confession that nobody’s priced the risk yet, which is the opening. The builder who can document a tool’s miss-rate hands the underwriter the one thing they’re missing: an actual number.
💡 80/20: For any tool that flags or screens, write down — before launch — what happens to the finding it doesn’t surface. The malpractice question isn’t “is it accurate,” it’s “who owns the miss.”
Quick hits
Code review just became the most leveraged skill in software. A widely-read essay from a Google Chrome engineer argues that as coding agents get good, the bottleneck moves from writing code to deciding whether to trust it — and understanding stays as expensive as ever. Read “code” as “clinical output” and it’s the same job: generation is cheap, judgment isn’t. The reviewer is the product.
A man with ALS is communicating again through a brain-computer interface — at home, unsupervised. A UC Davis–led team (with Brown and Mass General Brigham) reports in Nature Medicine that a person with severe paralysis decodes neural signals into text and cursor control to run a standard computer independently. The decoding algorithm — not new hardware — did the heavy lifting. The frontier is moving from “restore function” to “restore independence.”
🎙️ From the Pods
🎙️ Radio Advisory — “The hard truths behind the fight for commercial volumes“
The framework worth stealing: stop running every ambulatory site on its own P&L and sort them into three jobs — margin drivers (imaging, surgery, infusion: ~25% of volume, ~100% of the profit), access engines (primary/urgent care: ~60% of volume, you celebrate breaking even), and system stabilizers (outpatient behavioral health that prevents costlier admissions). The network should be intentionally imbalanced.
💡 Builder take: If you build for a service line, know which of the three jobs it’s hired to do — a stabilizer evaluated like a margin driver gets cut, and so does the tool attached to it.
🔇 Speaker Blindspot: Composition fallacy — they note that 90% of health systems are chasing the same shrinking commercial pie, then prescribe a portfolio model as if every system can win it. The strategy that works for one system mathematically fails when the whole market runs it.
🎙️ Lifers with Christina Farr — “Ric Sinclair, CEO of Cotiviti, on why payer-provider friction is finally fixable“
The reframe: healthcare’s ~$1T of administrative waste isn’t a data problem or a technology problem — it’s a coordination problem and an infrastructure problem. The data exists; the rails to move it cleanly between payer and provider don’t.
💡 Builder take: The fundable opportunity is the seam, not the silo — the boring plumbing that lets a clean claim and its clinical justification travel together.
🔇 Speaker Blindspot: Motivated reasoning — the CEO of a payment-integrity company calling payer-provider friction “finally fixable” (and AI “a jobs creator in healthcare”) has a product that only works if both are true. Worth taking seriously; worth pricing in the incentive.
🎙️ Health Tech Nerds Radio — “How Alignment Health gets health systems to come to them“
The operational pearl: Alignment’s edge isn’t the contract type, it’s the admissions-management infrastructure — the data and care-model visibility to tell an inpatient stay from an observation stay (a ~$23K vs. ~$3.5K difference) and to surplus the acute risk pool back to provider partners.
💡 Builder take: “Never bet against Medicare Advantage” only pays off if you own the back-end that turns risk into a surplus. The infrastructure is the moat; the contract is just paperwork.
🔇 Speaker Blindspot: Dismissing the consensus — when a dozen health-system CFOs independently report being paid “86 cents on the dollar,” waving it away as “a brilliant PR campaign” explains the data away instead of engaging it. A signal that consistent is usually measuring something real.
📅 Upcoming Health AI Events
Free virtual events for clinician-builders — attend live or catch the recording later. [These are not sponsored, just interesting.]
Thu Jun 18 — Interoperability Imperative: Connected Data for AI-Ready Operations (AHIP + SAS)
2:00 PM ET · Virtual · Free
The data plumbing question every builder eventually hits: getting health data clean and connected enough to actually run AI on it. Payer, digital-health, and policy panelists on what “AI-ready” really takes.
Thu Jun 25 — Adoption of AI in Clinical Care: Updates from the HHS RFI (ONC / HHS)
3:00 PM ET · Virtual · Free
HHS leadership shares what came back from the national AI-in-clinical-care RFI. If you build clinical AI, these are the policy signals you’ll be designing around.
Tue Jun 30 — Clinician in the Loop: AI Investment to Real-World Impact (CHIME, LinkedIn Live)
12:00 PM ET · LinkedIn Live · Free · Replay available
For CMIOs and clinical leaders: how to tell whether your AI is actually working post-deployment, and the gap between vendor promises and clinical reality.
Notable further out: Becker’s AI + Digital Health Virtual Event — Tue Jul 21, 1 PM CST, free to register.
💡 BTW: Ric Sinclair, the CEO steering one of healthcare’s biggest payment-integrity companies, started out as a drummer in Nashville — and says the best lesson for leading a team came from the drum seat: keep the beat, make everyone else sound better, and when you mess up, everyone hears it.
What are you building this week? Email and tell me (kevin@clinicians.build) — I read every one.
— Kevin