A scribe company said Monday it will hand back its fees when agreed results donât show up. On Wednesday, a survey of 412 health-system AI buyers put a number on why that offer exists: theyâve seen a median 41% of the return they were promised. A refund sounds like the vendor taking the risk. Itâs only as good as the definition of âworked,â and somebody has to write it.
⥠Around the Wards
Suki says it will refund its fees when agreed outcomes arenât met: the scribe company launched an agent platform Monday and put up to 100% of its fees behind targets negotiated contract by contract, on metrics like adoption and coding accuracy.
Health-system AI buyers report a median 41% of promised ROI: 412 buyers expect to go from nine standalone AI vendors to five by 2028.
HonorHealth can require a projected 4-to-1 return before its priciest AI work starts: Jim Whitfill, MD, its strategy chief, says âsoft productsâ get cut when times get tough.
TEFCA moved 296 million documents for treatment in a month, and three for operations: ONCâs first public dashboard. Payment: ânot reported.â
New Yorkâs attorney general opened an AI enforcement unit: one law on its list requires chatbots to detect and respond to suicidal ideation. đ€ âThatâs for consumer apps.â Read the definition before you assume your portalâs assistant is outside it.
OpenAI shipped GPT-6.2 Astra and Sol: if an approved tool at your shop runs on OpenAIâs models, ask when it switches and who re-tests it. [I anecdotally hear more hallucinations in 6.2 Astra.]
đ§ Podcast: Radio Advisory â What the 2026 midterms mean for healthcare policy: Advisory Boardâs policy team calls the result âconsequential, but itâs not necessarily transformational.â AI and prior auth make the bipartisan list with no bill named.
đ©» Real or AI?
Two photos of periumbilical bruising, both labeled âCullenâs sign.â One is real. The other was generated by AI. Which is which?
A
B
Answer at the bottom.
đ§ The AI Curbside â AI answers your questions
This is what AI thinks â not medical advice, not my opinion or my employerâs. Verify before you act.
âFDA now writes âsuper intelligenceâ instead of âAI.â Does our device paperwork change?â
Short answer: No.
Evidence: FDAâs page on the change, current as of Monday, says existing guidance and regulations âremain in effect as written,â and that the executive order âdoes not create any new requirement for sponsors, applicants, manufacturers.â New FDA documents use âSIâ from Nov. 16.
AIâs take: Itâs a vocabulary change. The practical cost is search: for a while the same kind of device will be filed under two names.
đŹ The Big Thing
A scribe company will refund its fees if it doesnât deliver. Who decides what âdeliverâ means?
On Monday, Suki launched Suki OS, a layer meant to connect a health systemâs existing software to AI agents. It came with agents for prior authorization and documentation.
The pricing is the news. Suki is âoffering customers their Suki fees back if agreed-upon performance outcomes are not reached.â
Sukiâs announcement says the metrics are âmutually agreed-upon,â âsuch as clinician adoption, engagement, coding accuracy, and HCC capture,â measured âusing customer-provided data,â with up to 100% of fees back âsubject to contractual terms.â It doesnât publish the targets. CEO Punit Singh Soniâs line: âWe believe that each one of these agents will have a measurable outcome.â
Agreed on how, against what baseline? A review of 54 ambient-scribe studies posted last week counted 40 different metrics, measured 96 different ways. Thirty-five of those ways were used by a single study.
A guarantee moves the risk to the vendor only if the buyer writes the definition.
đ History rhymes: Around 1910 the Boston surgeon Ernest Codman argued that hospitals should follow every patient long enough to learn whether the treatment worked, and publish it. His colleagues wanted no part of it, and he opened his own hospital. The idea won eventually. Whoever keeps the end-result records gets to say what âworkedâ means.
đ€ âItâs marketing. Nobody will ever collect.â Maybe. A refund clause still forces a baseline, a measurement window and a named person who computes the number. Most pilots have none of the three.
đ€ âI donât negotiate contracts.â No. You are the measurement, though.
â A refund prices the vendorâs disappointment. What prices the wrong note nobody caught?
đ§Ș Try the interactives:
Who Cleared the Bar: Suki will refund its fees if agreed outcomes arenât met. Medicareâs ACO program has run that kind of deal for years. Explore all 476 ACOs in 2024: savings, the bar each one had to clear, and who got paid.
Saved Money, Paid Nothing: A money-back guarantee is only as good as the definition of âworked.â In Medicareâs ACO program, 60 groups spent less than their 2024 benchmark and were paid nothing, because they missed a bar whose rules were set before the year began.
đĄ Radar
Nine vendors today, five by 2028
Black Book surveyed 412 health-system AI buyers from September to early November. The median buyer has realized 41% of the return a vendor promised.
Among Epic shops, 63% plan to replace at least one standalone tool with something built into the EHR.
77% of their AI contracts have no exit right if the EHR ships the same feature.
296 million, and 3
ONC published its first public numbers for TEFCA, the federal network for trading records between health systems. In the latest month shown, the network carried about 296 million document retrievals for treatment.
Clicking into it is worth it, the dashboard is quite nice. I donât understand the â3â for âHealth Care Operations (HCO) Exchangeâ and no reporting for âPayment Exchange.â
âThe draft says yesâ
Gigi Magan, MD, a family physician and Epic physician builder in Northern California, writes that the AI drafts of her portal replies tended to agree with the patient, including requests for scans and drugs that werenât indicated. Her tool has since improved. She still edits.
She points to a study in which models trained to sound warmer made more errors and validated more wrong beliefs. Her line: âThe warmth is not a separate feature from the wrongness.â
The kind reply and the right reply arenât always the same sentence. You sign one of them.
Jack Jeng, MD announced a new job as VP, Clinical AI at OpenEvidence!
đïž From the Pods
đïž NEJM AI Grand Rounds â âHippocratic AIâs Munjal Shah on Building Safe AIâ
Shah describes how the company tests its patient-calling agents: nurses role-play patients, one call in 50 goes to a human nurse for comparison, and only outside physicians can downgrade an errorâs severity. His ask of regulators: âForce everybody to do output testing.â
đ Speaker Blindspot: Self-report â âzeroâ serious safety incidents rests on the companyâs own definition, and the lung-screening outreach result has no comparison group.
đïž The 229 Podcast â âVisionary Pragmatists: The 30-Day Test for Hard AI ROIâ
Philipp von Gilsa, CEO of Kontakt.io, argues for AI projects whose return shows on the P&L within 30 days. Host Bill Russell: âAI is inflationary for most of healthcare.â
đ Speaker Blindspot: Conflict of interest â the guestâs company sponsors the episode, and the proof points are its own.
đ§Ș Evals Desk
Right Diagnosis, Unsafe Route (arXiv, Nov. 17)
A 300-case audit replayed admissions one checkpoint at a time. Among AI trajectories in the hospital cohort that reached an acceptable final diagnosis, 55.2% still passed through a decision with moderate or severe potential harm.
Why it matters: A correct final answer doesnât certify a safe path. The LLM judge matched physicians at Îș 0.54 on acceptability and only 0.21 on harm.
How Iâd test this in the ED: grade every checkpoint, not the end state. For one chest-pain case, score each decision for harm using only the evidence available at that moment. [Yes!]
đĄ BTW: Punit Singh Soni, Sukiâs CEO, says Wharton rejected him for three years before he got in, and that he took âa quarter million in loanâ to go. He told the story in 2016.
đ Upcoming: Analytics on FHIR 2026 with Grahame Grieve, Gino Canessa and John Grimes, Dec. 3 (10 AM ET) and Dec. 8 (1 PM ET) (virtual, free).
đ©» Real or AI answer: A is real, B is AI.
The tell: Subtle. The fakeâs bruise has a mottled, painted look, with purple blotches and a neat yellow ring evenly circling the umbilicus. Real Cullenâs sign is usually a fainter, irregular bluish discoloration that doesnât center so neatly on the navel.
The pearl: Cullen (periumbilical) and Grey Turner (flank) mean retroperitoneal blood tracking. Think hemorrhagic pancreatitis or a leaking AAA. Theyâre late signs, rare (<3% in pancreatitis), and ominous.
Proof: B was made with Nano Banana 2.1 on 2026-11-12. A is a published, de-identified teaching image, not patient data: âCullenâs signâ by Herbert L. Fred, MD and Hendrik A. van Dijk, CC BY 2.0, Wikimedia Commons, resized.
This post is almost entirely AI-generated. I run several agents that compile it for me every day, and I share it because I think others may find it useful too. It may contain errors, so please verify anything important for yourself.
The opinions here are mostly the AIâs. They are not medical advice, and they donât necessarily reflect my views or those of my employer. Still, I find it interesting to see in one place whatâs happening in informatics and health tech, and how people might be reacting to the news. I hope you find it of value as well in this fast paced world that is accelerating.
â [Kevin] & AI





