State of AI for Healthcare

Two directions of the same risk, and both cut against incumbents. (1) THE GATE IS CHEAP. A 510(k) costs $26,067 at the FY2026 standard MDUFA fee ($6,517 for a small business) and produced 96.2% of the 1,524 authorisations on FDA's AI-Enabled Medical Device List, so anything whose defensibility rests on a clearance count is being commoditised by the pathway that grants it. A De Novo is worse than neutral for the winner: the product code it creates becomes every follower's predicate, and the largest such code now carries 83 devices from 29 filers. (2) THE RATE IS DISCRETIONARY. Every New Technology APC rate carrying an AI service is held up by CMS's equitable-adjustment authority under section 1833(t)(2)(E), which the CY2027 proposed rule calls 'an interim policy', with CY2027 described as 'a transitional period' — the phrasing recurs six times. CPT 0721T/0722T would have dropped to APC 1502 ($51–100) on its own claims data, in CMS's words 'the resulting approximate 90 percent reduction in payment', but for a manual override. The CY2027 OPPS/ASC final rule, expected on public display around 1 Nov 2026 for a 1 Jan 2027 effective date, can re-rate the whole SaMS set at once, and CMS explicitly solicited comment on a status-'T' specification that would add multiple-procedure discounting when the software is billed alongside the underlying scan. Hits HTFL most directly — 75580 is already proposed at $850.50, −3.1% after −13.8% the prior year — and hits every private reimbursed-diagnostic mark on the watchlist.

The most under-priced risk on this board, and it gates adoption at every layer rather than creating a pool anyone can invest in. Two mechanisms landed in the same window and they compound. All-party-consent wiretapping and medical-confidentiality statutes carry statutory damages with no actual-harm prerequisite and count per recording, so an ambient deployment converts routine consultations into a countable exposure with a limitations tail measured in months rather than years; and general-liability and errors-and-omissions towers are being endorsed away from generative AI at renewal, with the exclusions filed and approved rather than merely floated. The party left holding it is the deploying health system, not the AI vendor: the vendor sits inside the recording statutes' party exception, and that same exception is exactly what leaves the buyer exposed. The named defendants in the filed ambient-consent class actions are health systems. That asymmetry is what makes this cross-cutting rather than name-specific — vendor revenue slows while vendor loss ratios are never touched, and the buyer being asked to fund an unreimbursed subscription is the same buyer absorbing the tail, which raises the hurdle rate on the least creditworthy layer in the stack at the worst possible moment. There is nothing to own on the other side either: the defendant systems are nonprofit corporations with no equity, the listed hospital operators carry the affected states as a minority of each footprint, and the carriers shedding the exposure book it as a loss-elimination effect that will never be separable in a reported combined ratio. Falsifier: a court holding that a general notice of privacy practices satisfies the all-party-consent requirement, together with no ambient-consent class settling above a material threshold — either alone weakens it, both together kill it.

Three separable exposures the market tends to net into one. (a) CONSENT — the all-party-consent exposure above is a data-privacy failure before it is a liability failure, and it lands inside a single fiscal year rather than amortising across the contract. (b) DATA GRAVITY — no AI vendor accumulates a comparable clinical corpus by integrating with a record owner; it accumulates a COPY under a business-associate agreement the record owner can decline to renew. The distinction between a corpus and a copy is the one most 'proprietary clinical data' theses elide, and it is the single largest risk to any data-differentiation claim on the board, including the listed data-licensing lines. (c) REGULATORY DIRECTION — information-blocking enforcement has produced exactly zero actions: OIG's database shows 11,097 enforcement actions with entries as recent as 30 Jul 2026 and contains no information-blocking action, and the realised transfer to date is $0 against a $1,327,209-per-violation ceiling. But the proposed doctrine turned pro-agent. HTI-5 would amend 45 CFR 171.102 so that access and use of electronic health information expressly include 'automation technologies such as robotic process automation and autonomous artificial intelligence systems', which would make refusing an AI agent programmatic access information blocking per se at that penalty. It is not final — HTI-5 remains a proposed rule as of 2 Aug 2026 — and its own impact analysis books $0 implementation cost and $1.53B of present-value savings onto certified health IT developers, the moat-holders, with no dollar figure at all on the requestor side. Net: it erodes the DATA moat while leaving the WORKFLOW moat intact, which is the wrong half for anyone underwriting proprietary clinical data and the right half for the record owner.

The most direct threat to any thesis that ambient documentation or AI coding monetises through higher realised reimbursement — and the honest version is narrower than the loud version. The provider-side gain is measured but its mechanism is not settled: the JAMA authors leave open whether the $167.37 per clinician per month came from extra visits booked in available or nontemplated time, or from a shift in the composition of level-4 and level-5 E/M coding. Only the second is clawable, and only the second is what payers organise against. So far the clawback is being beaten back rather than enforced. Cigna's R49 auto-downcoding policy, which would have cut one level off 99204-05, 99214-15 and 99244-45, was paused after specialty-society advocacy; on 13 Mar 2026 the Maryland Insurance Administration fined Cigna $80,000 and ordered it to stop, then extended the prohibition to every insurer in the state through Bulletin 26-9 on 7 Apr 2026. Two things survive that pause and both matter: the underlying audit right is untouched, and CMS's statutory Medicare Advantage coding-intensity adjustment is unaffected, so the risk-score channel is still clawed back mechanically even while the commercial channel is protected. The rate-side vector is real but small — extending the −2.5% Medicare physician-fee-schedule efficiency adjustment to E/M would take $167.37 to roughly $163.19, a 2.5% haircut and not an evaporation; a documentation-intensity budget-neutrality offset is a separate and plausibly larger mechanism, and it is not on the table today. The structural carve-out outweighs both: the E/M channel is worth about zero wherever the clinician is not paid fee-for-service, which covers staff-model plans, the VA and full-risk MA and ACO arrangements. Hits the documentation-to-billing attach thesis, the CDI leg of any RCM long, and every private coding-automation mark.

The category's evidence base is thin and its benchmarks disagree, which means there is no shared standard to fall back on when something goes wrong. On the safety set with the most physician labour behind it — 1,100 physician-derived scenarios across ten specialties, more than 4,000 annotated actions, roughly 13,000 annotations from 29 board-certified physicians — the top-performing models still recommend severely harmful actions in 11.8 to 14.6 cases per 100 and the worst exceed 40 per 100, with 76.6% of harmful errors being omissions rather than unsafe recommendations. The composite leaderboard's top four sit within three points of each other, and purpose-built clinical benches remain unsaturated with a 0.310 spread in mean win rate across ten ranked models; a field with four near-tied leaders has no referee. Reformulating static multiple-choice medical questions as sequential decision-making with incomplete information drops diagnostic accuracy for some models to below a tenth of the original — an inference about agentic capability, not a measurement of what shipped products do, because no public benchmark yet measures whether an agent drafts a correct order set or assembles a valid prior-authorisation packet. Procurement has already priced the uncertainty: health-system leaders report benchmarks have minimal impact on purchasing and use them only as a preliminary screen before internal validation against local patient populations. A single well-documented patient harm attributed to an ambient note, a CDS recommendation or an autonomous coding agent would hit adoption velocity, procurement cycle length and insurance terms simultaneously, across public and private names alike, and faster than any of them could differentiate out of it. The asymmetry is the same one that makes liability cross-cutting: the deploying health system carries the clinical exposure, so vendor revenue falls without vendor loss ratios ever being touched.

Every layer above the provider is selling into a balance sheet that cannot fund an unreimbursed subscription at scale. Kaufman Hall's April 2026 metrics, across more than 1,300 hospitals, put the median calendar-year-to-date operating margin at 2.5% including corporate allocations and 8.3% excluding them. Two things qualify the obvious reading, and both cut against a lazy bear case. First, the direction of travel is UP, not down: the CYTD median runs 1.9% → 2.1% → 2.3% → 2.5% across January to April 2026, though it remains below the 3.6% at which 2025 closed — so 'the worst credit in the stack' survives on level and has to be argued on level, not on trend. Second, the pain is not concentrated in small hospitals: year over year the 26–99 bed cohort is worst at −11.5%, but 500+ beds is second worst at −9.0%, which contradicts any argument that large academic systems are the cohort with spare budget to fund AI. Kaufman Hall also restated its corporate-allocation adjustment in 2026 to represent the sample more accurately, so 2025-to-2026 comparisons are not strictly like-for-like and any cross-year claim has to say so. The consequence for this board is the whole argument in one line: a vendor whose invoice lands on provider opex is underwriting a 2.5%-margin buyer's discretionary budget, which is why payer identity is the screen and not a preference.