Super or Scam: What is AI?

In this issue

Market value and 2026 return for the companies the sector trades on. Payers and the recovery names are carrying the year; the big hospital operators are split.

Market cap and year-to-date price return per Yahoo Finance at the most recent close; YTD measured from the Dec 31, 2025 close. Selected companies, not exhaustive. Not investment advice.

Healthcare AI's ROI is real. It is just lopsided.

A new Bessemer and Bain survey of 226 healthcare executives is the clearest answer yet to the super-or-scam question. Returns arrived in about 12 months, not the 24 buyers budgeted, averaging 3.5x. So, not a scam. But the money is almost entirely in the back office: revenue cycle returns 4.0x with two-thirds of those deployments already semi or fully autonomous, while clinical AI sits at 2.9x and only 4% semi or fully autonomous, held back by trust, liability and reimbursement, not by model capability. So, not super either.

The discipline gap is the story. Across the wider economy, a16z finds that while roughly 30% of S&P 500 companies now claim quantifiable impact from AI, only about 2% disclose a metric they track over time. Meanwhile a Mercor study clocked Anthropic’s Claude Opus 5 at 100% on bounded month-end accounting tasks, against a 37% average for the licensed CPAs doing the same work. Capability is not the constraint. Measurement is.

The executive question: for every AI dollar you are spending, what is the one number that tells you it worked, and are you tracking it? Source: Bessemer + Bain

A Sept 29 executive order tells federal agencies to replace “artificial intelligence” with “super intelligence” in official use. The legal definitions do not change, and the six firms that signed the voluntary White House safety accord, Anthropic and OpenAI among them, have not renamed a single product. For health leaders the practical effect is narrow for now: agency guidance, procurement language and non-statutory documents will start carrying “SI,” while the rules you actually comply with stay put. Watch the 60-day deadline for a proposed federal definition, where any real teeth would be. Washington Examiner

At the MAHA Summit, HHS Secretary Robert F. Kennedy Jr. called AI a second opinion better informed than any physician, and relayed Sam Altman’s line that diagnosing without it would be malpractice. Six physician groups pushed back, and health system leaders told Becker’s that the underlying claim, that AI can already perform a fifth of outpatient care, is far too high for their settings. The signal is not the soundbite. It is that federal health leadership is now anchoring to AI’s most optimistic case, while the clinicians who carry the liability are not convinced. Fierce Healthcare

Cigna is embedding OpenAI’s frontier models into clinical workflows, starting with oncology through its Accredo specialty pharmacy, under clinician oversight. At its Sept 30 investor day it put numbers behind the AI it already runs: call volume down 20%, clinician notetaking time at its MDLIVE virtual care down by up to 90%, and specialty benefit reviews running 43 times faster. The pattern worth noting is who is moving: a payer, not a provider, building the intelligent entry point into complex care, which is where the patient relationship and the margin both sit. Healthcare Finance News

Everlywell is now selling Clairity Breast nationwide, the first FDA-authorized AI that estimates a woman’s five-year breast cancer risk from a routine mammogram, with results reviewed by a licensed provider. NCCN’s 2026 guidelines now build AI mammogram-based risk assessment into screening from age 35, treating a five-year risk of 1.7% or higher as increased risk. They name no vendor, and Clairity is for now the only FDA-authorized tool that produces that number. Two things matter here. Regulated clinical AI moved from pilot to guideline to market far faster than the usual decade, and the distribution is running through a consumer platform rather than a health system, another front door opening outside the walls you control. Business Wire

A study in PLOS Digital Health ran a census of every AI device the FDA has cleared for patient care. Of 1,357 devices, only three were tested on whether patients actually lived longer or better. Most cleared through a path that asks only whether the tool resembles something already on the market. The lesson for anyone deploying imaging or risk tools: FDA clearance is a floor, not evidence of benefit, so demand outcome data before you trust the output. Earth.com

A new Science Advances study built a speech clock from 2,928 people across five Latin American countries, using AI on voice and language to estimate biological age. Its speech age gaps separated healthy adults from mild cognitive impairment, Alzheimer’s and frontotemporal dementia. It is early research, not a product, but two features make it worth watching: a voice sample is about the cheapest, most scalable screen imaginable, and training on an underrepresented population is exactly the generalizability evidence most clinical AI still lacks. Science Advances

EliseAI raised $350 million at a $4 billion valuation, double a year ago, co-led by a16z and Bessemer. The company automates administrative work in two industries, property management and healthcare, and in healthcare it serves specialty physician groups: inbound calls, referrals, scheduling, insurance verification and intake. It has crossed $200 million in annual recurring revenue. It is the clearest funding signal yet of where investors think healthcare AI pays: not the bedside, but the paperwork. That lines up exactly with the ROI data up top, where administrative automation is returning multiples clinical AI has not reached. TechCrunch

Google’s Gemini 4 Argon leads or ties on 13 of 18 benchmarks Google disclosed, ahead of OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5, and priced below both while introductory pricing holds. It is in limited release, going to cyber defenders first, and the leaderboard is Google’s own disclosure, so weigh it accordingly. The detail that should hold a health executive’s attention is not the benchmark table: in early access, security firm Wiz used the model to find a critical vulnerability exposing personal data in healthcare software used by hospitals worldwide. The same capability that defends your systems is now available to whoever points it the other way. VentureBeat