The week medical AI had to reconcile its own books: vendors say $862–1,004 per clinician per month, peer review says $167, and the only new randomised trial — 249 physicians across three countries — lands between +7.2 and +18 points. The gap is not error; it is two sets of accounts.

EFF obtained roughly 1,000 pages of records on the WISeR programme through FOIA litigation. The programme launched in January 2026 across six states. The records show Innovaccer telling CMS about a month before launch that it would ship partially functional, incompletely tested software, writing that "given CMS's decision not to delay the model start date, auto-affirming is the only path available"; a second vendor, Virtix, denied more requests than it approved in the first three months. One prior-authorisation request sat unanswered for 83 days, against a stated CMS target of 72 hours. The penalty for a poor quality score is a payment reduction of just 5–10% (EFF Deeplinks; see also STAT, Sep 15 and Medscape).

This is the week's only gap documented by the government's own paperwork. Every other argument about whether AI works is still vendors and academics quoting numbers at each other; the WISeR records spell out the incentive structure directly — the launch date was immovable and the quality penalty caps at 5–10%, so a rational vendor ships first and fixes later. For any country contemplating AI in coverage review, this file is the cautionary text.

Coverage disagrees on the denominator and window behind the denial counts (EFF's own text puts it at 5,944 requests denied by two vendors in the first three months). Cite the released records rather than the aggregated figures in secondary coverage.

Electronic Frontier Foundation, 2026-09

FDA's TEMPO digital health devices pilot lets generative-AI devices reach patients and a reimbursement pathway before marketing authorisation (STAT, Sep 3). The participant list grew to four this week, with Limbic the first "AI-led" mental health company selected (Businesswire, Aug 19; MedTech Dive). Meanwhile FDA's discussion paper on generative-AI-enabled devices is still open for comment, closing October 19.

Set this beside the previous item and the same federal government is doing two opposite things at once: using untested AI to deny payment while letting unauthorised AI collect it. Both rest on the same justification — the schedule cannot slip. The October 19 comment deadline is one of the few remaining points this year where outside input can still change the outcome.

TEMPO is a pilot, not a marketing authorisation; selection is not an FDA finding on safety or effectiveness. Separately, five states have made AI-delivered psychotherapy unlawful, and the federal–state gap has yet to be tested in court (Becker's).

U.S. FDA — TEMPO participants, 2026-09

Published in npj Digital Medicine on September 9: a parallel-group randomised controlled trial in which 249 physicians were randomised, with the intervention arm given access to GPT-4o. The result was +18.0 points in Kenya (95% CI 12.7–23.2), +10.7 in Indonesia (5.7–15.7) and +7.2 in the Netherlands (3.7–10.7), all at p < 0.001. The authors describe it as the first experimental evidence that LLM assistance lifts physician performance across diverse economic settings.

It corroborates Monday's EU I3LUNG study: with real physicians in the loop, explainable AI lifted immunotherapy-response judgement from 57% to 65% (this daily, Sep 14) — eight points. Two independent studies, different diseases, different countries, same order of magnitude. That magnitude is the benchmark a procurement contract should use, not uncontrolled descriptive claims like "influenced 92.6% of surgical decisions." The largest gains landed in the more resource-constrained settings, which carries an implication for Taiwan's regional and rural hospitals too.

The authors flag two limits plainly: the results come from controlled conditions rather than real practice, and harms were not assessed. The control arm was barred from the internet and clinical guidelines, which inflates the intervention's relative gain — a physician in a real clinic can already open a guideline database. Treating +18 points as a deployment expectation would repeat exactly the error in the week's first two stories.

npj Digital Medicine, 2026-09-09

Since Epic shipped native AI Charting on February 4, turning ambient documentation into a built-in EHR feature (HIT Consultant's analysis), the pricing premise under third-party scribes has been gone. Two answers surfaced this week. Ambience launched The Ambience Standard on August 19, shifting the software fee from usage-based to outcome-based. Abridge went where Epic will not, into pre-bill DRG and coding review and the nursing workflow. And the benefit figure exists in two versions: vendor and consultant case studies land at $862–1,004 per clinician per month, while the peer-reviewed study across five academic centres and 8,500 clinicians computes $167 (fully unpacked in this daily, Sep 18).

A sixfold gap is not estimation error, it is two sets of accounts: the vendor case study measures best-case deployment with willing physicians plus improved charge capture, while the peer-reviewed figure averages across an entire academic centre. A hospital that puts the first number in the numerator and the second's real implementation cost in the denominator can make any ROI work. The week's real product news is not who announced what, but that Ambience volunteered to carry the burden of proof — the first time a seller in this market has done so.

The $862–1,004 range comes from vendor and third-party consultant case studies (including KLAS's ROI validation) and is not independently audited; and nothing public states who measures or audits the "outcome" in The Ambience Standard.

HIT Consultant, 2026-08-19

Published in Scientific Reports on September 14, covering 6,166 US hospitals across 3,143 counties (DOI 10.1038/s41598-026-70027-1). Between 2022 and 2024 the number of AI-enabled hospitals rose an estimated 56% and population coverage went from 66.2% to 75.2%. Yet in 2023 roughly 114.6 million people still lived more than a 30-minute drive from the nearest AI-enabled hospital, with only 65.8% (220.1 million) inside that radius; the 90th-percentile drive was 61.3 miles against 1.9 miles at the 10th. On outcomes, staff-scheduling AI was associated with a 2.24-percentage-point increase in sepsis bundle completion (3.9% relative) and routine-task automation AI with a 0.87-point reduction in 30-day pneumonia mortality (5.4% relative) (News-Medical summary).

This is the week's only study that puts "does AI work" and "who can reach it" on the same page. Note that even the optimistic outcome figures are 2.24 and 0.87 percentage points — single digits again. When a technology's real effect is a few percentage points and the access gap runs to nine figures of people, the allocation question stops being "which is the best model to buy" and becomes "where do we lay it down first."

This is an association study, not causal inference: hospitals that adopt AI already skew larger, better resourced and more process-mature, and this design cannot tell how much of the sepsis and pneumonia difference comes from the AI itself.

Scientific Reports, 2026-09-14

Published in Nature Medicine on September 7, a commentary by Kristina Lång, who led the mammography randomised trial. She writes that "the first generation of medical artificial intelligence was judged on whether algorithms could match clinicians." The next phase sets a far higher bar — systems must demonstrate improved patient outcomes — and what matters is "carefully designed human–AI systems," not standalone algorithms. Her conclusion: the challenge is not deciding which is superior but designing systems in which each compensates for the other's limitations.

This is precisely why almost every accuracy story this week reads as insufficient. Apply her standard to the week just past: npj's three-country trial measures decision accuracy, not patient outcomes; Scientific Reports reaches mortality but can only claim association; the scorecards from Oracle and the scribe vendors are denominated in hours. Evidence that meets the next-generation bar appeared exactly zero times this week.

Lång declares competing interests with AI healthcare companies and holds breast-screening roles; this is a position piece, not a systematic review. The article is paywalled, and this item is written from the publicly visible passages and abstract.

Nature Medicine, 2026-09-07

Epsilon Health came out of stealth with nearly $27.6M, not to sell reading software but to open its own AI-native radiology group. Nara Health raised $14M to become an AI-native third-party administrator. Forus raised $150M at a $3B valuation, taking the whole stretch between prescription and dispensing. Put those back in the denominator: Rock Health's H1 2026 report shows $7.4B raised and brisk M&A, with IPOs at zero (Fierce Healthcare, Healthcare Dive).

Connect this to the $167 item and the strategy is legible: if selling software only earns you that small verifiable number, stop selling software. Become the provider, bill for the whole service, and never argue with a hospital about the ROI numerator again. It is the week's most explanatory piece of business logic, and why a valuation can triple in a year.

Round sizes and valuations are company-announced or reported on investors' accounts, with no audited filings behind them; Epsilon's $27.6M is early-stage and not on the same risk tier as Forus's $150M.

Rock Health H1 2026, 2026-07

An Imprivata survey released September 15 of 250 healthcare executives found that 72% of organisations have AI tools or agents deployed without formal IT approval, with only 17% believing their identity and access controls are adequate (see also Healthcare IT News, Sep 15). At the model layer the same week: TechCrunch confirmed on September 15 that OpenAI, Anthropic and Google have been discussing a self-standards body since July, with Cohere's chief executive calling it a cartel; California's governor signed an executive order on September 18 directing work on a frontier-model kill switch and faster independent oversight; and on September 17 Anthropic disclosed internal metrics for the first time, saying Claude now leads 26% of its own R&D.

Only together do the two layers show the shape of the problem: the safety floor at the top is being negotiated by the sellers, and the deployments at the bottom are invisible to hospital IT. The layer in between — the one FDA, CMS and national regulators are meant to fill — is exactly the layer the WISeR records showed cannot keep pace. The question a hospital should be asking is not "is this model safe" but "which agents are running in my institution that I never approved."

Imprivata sells identity and access management, so the survey carries an obvious commercial motive, and the sampling method behind its 250 respondents is not detailed. Anthropic's 26% is a company-defined internal metric with no third-party verification, and the same disclosure puts fully autonomous work at 0%. The California order so far directs study, not rules.

GlobeNewswire / Imprivata, 2026-09-15

At Confluence Health in Washington state, primary care physicians cut EHR documentation time by 34 minutes a day after adopting ambient AI, while MyChart message handling rose from 48% to 59%. In Florida, Tampa General handles two million calls a year with agentic AI, reporting a 58% reduction in wait time and a 21% increase in appointment capacity, plus a halving of sepsis mortality through predictive analytics (HIStalk roundup, Sep 16).

These two deserve separate billing because they demonstrate the week's converse: on the administrative and process side, the numbers are far better than on the clinical side and far easier to check. Thirty-four minutes and 58% are auditable operating metrics that need no randomised trial. That also answers a practical question — if the clinical evidence bar is rising towards patient outcomes, the most certain near-term returns sit in scheduling, call centres, rostering and coding, where nobody argues about the endpoint.

Both are single-institution self-reports relayed through a secondary roundup, with no control group and no independent audit. Tampa General's halving of sepsis mortality warrants particular caution — that is a clinical outcome claim with no verifiable study design behind it, and it does not sit at the same evidentiary tier as the other clinical figures in this issue.

HIStalk, 2026-09-16