EHR vendors built AI into documentation, coding and prior authorization

For health systems, physician practices and payers, AI now arrives inside the EHR and revenue-cycle platforms they already run. The quarter’s strongest study found that an ambient scribe raised the volume and intensity of physicians’ billed work, and payers, states and Medicare are now contesting where that gain belongs.

Editorial snapshot: 31 October 2026

The central finding

With the same platforms now drafting the note, the charge and the authorization request, documentation AI’s gains are appearing as higher billed intensity that payers and regulators contest code by code, rather than as lower administrative cost.

01Key findings

What changed in health care AI in the third quarter of 2026

Epic and Oracle put AI into documentation, coding and prior authorization within five weeks, and the quarter’s strongest scribe study found its largest effects in what physicians billed. Payers automated coding review and appeals, two states limited algorithmic downcoding, and Medicare’s own AI prior-authorization model came under scrutiny.

Capability

Epic and Oracle built AI into the note, the charge and the authorization

At its users group meeting on Aug. 18, Epic showed native ambient charting live at more than 60 organizations and real-time prior authorization running with UnitedHealthcare, Aetna and Network Health. On Aug. 19, Oracle extended its agent to suggest professional-fee charge codes from the visit conversation.

Capital

Digital-health venture held at $3.7 billion

Rock Health counted $3.7 billion across 114 US digital-health deals in the third quarter, against $3.5 billion and 107 deals a year earlier. Acquirers bought prior-authorization and patient-access automation, including R1’s agreement to buy Humata Health.

Adoption and economics

AI scribe eligibility raised physicians’ work RVUs by about 20%

An NBER study of 1,868 physicians at one academic network estimated that eligibility for Abridge’s ambient scribe added five to six appointments and about three level 4–5 visits per physician-month. In fee-for-service medicine, that gain is revenue for the provider and cost for the payer.

Accountability

Medicare’s AI prior-authorization records showed denial-linked vendor pay

Records obtained by the Electronic Frontier Foundation show that WISeR vendors are paid for denials not reversed on appeal and that two vendors denied 5,944 requests in three months. On Sept. 29, Bloomberg Government reported internal CMS communications showing plans to extend the model nationally.

Operating question

When the same AI drafts the note, the charge and the authorization request, which of those outputs does a named person review before it leaves the organization, and what record shows that review?

02Capability

EHR platforms absorbed documentation, coding and authorization agents

Frontier models kept improving and became cheaper at the standard tier, but the change that mattered most for care delivery was where AI runs. Documentation, coding and authorization functions moved into the platforms that already hold the record, the order and the claim, so the right to write into those systems became the main point of difference between vendors.

Agents · Early adopters; broad release planned for 2027

Epic opened agent building to its customers

Epic’s Agent Factory, shown Aug. 18, lets customers build, run and monitor agents without code, starting from 120 out-of-the-box AI features. Epic also moved from 18-month to quarterly software versions, with weekly updates to its Nebula cloud. Derek De Young of Epic R&D expects a few core agents sharing thousands of skills rather than many bespoke agents.

Coding

The visit conversation now drafts the charge

Oracle’s Clinical AI Agent suggests professional-fee charge codes from the recorded visit inside the ordering workflow, for the clinician to confirm. Epic’s Penny is a revenue-cycle copilot, and Optum’s Integrity One automates claims coding for providers. The transcript that drafts the note now also feeds the claim, which is why coding intensity became the quarter’s economic question.

Prior authorization · Live at four health systems

Real-time prior authorization runs over a standard API

Four health systems, including Froedtert ThedaCare, Denver Health and Summit Health, run Epic’s real-time prior authorization with UnitedHealthcare, Aetna and Network Health through the Coverage Requirements Discovery (CRD) API; Epic is testing with 16 more payers. Federal rule CMS-0057-F requires payers to offer a FHIR-based prior-authorization API, with compliance dates beginning in 2027.

Prediction · In validation

A generative model trained on pooled patient records

Epic’s Curiosity is trained on Cosmos, Epic’s pooled data set of 320 million patients and 23 billion encounters. It is being validated at 20 organizations and is planned as an EHR component in March 2027. Its supporting Yale study, of about 3,000 emergency department patients with abdominal pain, is a medRxiv preprint that has not been peer reviewed.

Frontier models

Self-serve HIPAA access and diverging task costs

Anthropic made HIPAA configuration self-serve for its Enterprise and API customers on July 14. Artificial Analysis measured output tokens per benchmark task from about 27,000 for GPT-6 Astra to about 119,000 for Claude Opus 5.5, so a task’s cost does not follow token price. OpenAI’s GPT-6 Astra, launched Sept. 3–4, operates applications without APIs and is off by default for Enterprise customers.

Security

Agent tooling became an attack surface

Epic disclosed that it joined Anthropic’s Project Glasswing cyberdefense program and said it will patch hosted customer environments; its chief security officer, Stirling Martin, stressed that rapid patching matters more than ever. In August, attackers targeted agent tooling directly: a malicious Model Context Protocol (MCP) server, the kind that connects agents to tools, circulated through GitHub pull requests, and Langflow flaws were exploited.

Cost and consequence

Model prices matter less to providers than integration rights: who can write into the chart, the charge and the payer’s interface. Every Epic and Oracle customer now weighs building agents against waiting for the platform’s own, and carries the cost of monitoring them either way.

03Capital

Digital-health venture held steady as acquirers bought authorization and patient-access tools

US digital-health venture ran between $3.2 billion and $4.2 billion a quarter in 2026, close to 2025 in dollars and deals, while a growing share went to rounds of $100 million or more. Global venture reached $159 billion in the quarter, 64% of it to AI, by Crunchbase’s count. The quarter’s health care acquisitions bought automation at the points where providers, patients and payers exchange information.

US digital-health venture by period

Rock Health’s counts of US digital-health venture funding, deals and acquisitions. H1 2026 comprised $4.2B in Q1 and $3.2B in Q2.

US digital-health venture by period
Period FundingDealsAcquisitions
H1 2025 Funding $6.4B Deals 245 Acquisitions — No comparable figure
H1 2026 Funding $7.4B Deals 244 Acquisitions 115
Q3 2025 Funding $3.5B Deals 107 Acquisitions — No comparable figure
Q3 2026 Funding $3.7B Deals 114 Acquisitions 45
January–September 2025 Funding — No comparable figure Deals — No comparable figure Acquisitions 172
January–September 2026 Funding $11.1B Deals 358 Acquisitions 164

The average deal through September 2026 was $31.0M, against $29.5M for 2025 and $20.6M for 2024.

45%

Share of H1 2026 digital-health dollars in rounds of $100M or more

20 rounds at 19 companies; 22% in 2024 and 42% in 2025.

$14M

Median digital-health deal, H1 2026

Highest since 2022; $12M in 2025.

Acquisition · Announced Aug. 18

R1 agreed to buy Humata Health

R1, a revenue-cycle outsourcer for providers, agreed to acquire the AI prior-authorization company for its Phare OS platform at an undisclosed price and said it expected to close by the end of the third quarter. Humata is also the WISeR vendor reviewing traditional Medicare requests in Oklahoma; whether that contract continues under R1 has not been disclosed.

Acquisitions · September

Patient-access agents changed hands

Hello Patient, which automates outpatient practices’ phone work, bought Converse Health on Sept. 15 for its back-office agents that read and write to the EHR. DexCare, whose health system clients include Kaiser Permanente and Tampa General, acquired Mila Health on Sept. 21; Mila’s phone, text and chat agents handle scheduling and post-discharge follow-up. Terms were not disclosed.

Venture · Reported Sept. 24

OpenEvidence reportedly raised $250 million at a $15 billion valuation

AI Weekly, citing Business Insider, reported that the physician-facing AI company quietly closed the round, up from a $12 billion valuation in January, and that its founders were fielding acquisition interest. Sacra estimates its annualized revenue at about $300 million.

Exits

Health care listings came from outside digital health

Rock Health recorded no US digital-health IPO since HeartFlow’s in August 2025. PitchBook and NVCA counted 12 of 18 US venture-backed listings in the quarter as health care, a broader category than digital health.

Federal R&D · Awarded Sept. 9

ARPA-H funded agentic cardiovascular AI

The ADVOCATE program awarded up to $62.7 million over four years to teams building agentic AI for cardiovascular care, including Atman Health, Tempus AI, UpDoc, Kaiser Permanente and teams led by Duke and Stanford.

What it signals

The quarter’s acquisitions favored the exchanges around a visit, such as authorization requests, patient calls and back-office forms, over standalone documentation tools, which EHR vendors now bundle. Buyers of authorization and patient-access tools should expect a vendor to change owner during a contract term; three agreed to sales in the quarter.

04Products

What Epic, Oracle and others shipped, and what is still scheduled

Epic at its Aug. 18 users group meeting, and Oracle between Aug. 19 and its Sept. 22–24 summit, shipped or scheduled AI in every administrative layer of a visit within five weeks. Most items are at early-adopter or announcement stage, so availability has to be read function by function.

EHR platforms

Ambient documentation became an EHR feature

Epic’s native ambient charting, Chart with Art, arrived by the early-2026 date customers were given in August 2025 and now covers more than 70 specialties. Advisory Board, an Optum subsidiary, wrote that ambient documentation, patient assistants and coding support may soon be expected functionality in health care software.

60+Organizations using Chart with Art

300+Organizations using Art visit summaries

Point solutions

Abridge is repositioning toward decision support

As EHRs absorb documentation, Abridge says clinicians at more than 300 health systems have adopted its decision-support tools since April and that more than half of eligible clinicians use them monthly.

Payer-owned vendors

A payer’s subsidiary sells coding to providers

Optum Integrity One, UnitedHealth’s AI claims-coding product for providers, is live at 15 clients with 24 more implementing, the company said on its third-quarter call. UnitedHealth, whose insurance arm pays claims, also joined Anthropic’s partner network.

Clinician tools

Free and low-cost tools reach physicians directly

Doximity, which offers the Ask assistant and a scribe, reported AI search queries up more than 25% from the prior quarter in results released Aug. 6. OpenEvidence and ChatGPT for Clinicians, free to verified US clinicians since April, reach practices that enterprise scribe contracts do not.

Regulated pilots

AI takes on licensed tasks under regulatory agreements

Utah’s Office of AI Policy signed sandbox agreements in the quarter with August AI, for refills from a fixed list of noncontrolled drugs; Nolla Health, for first-time topical acne prescriptions; and Expect Fitness, for pelvic floor physical therapy plans reviewed by a therapist. The FDA’s TEMPO pilot added Cadense, for hypertension medication management, and Limbic, an AI voice agent delivering cognitive behavioral therapy.

Status of EHR-vendor AI presented in the quarter

Dates and status as the companies described them at Epic’s Aug. 18 users group meeting, in Oracle’s releases and at Oracle’s Sept. 22–24 summit.

Status of EHR-vendor AI presented in the quarter
Product What it doesPresentedStatus
Epic Agent Factory What it does No-code building and monitoring of agents Presented Aug. 18 Status Early adopters; broad availability planned for 2027
Epic Chart with Art What it does Ambient charting inside Epic Presented Aug. 18 Status Live at 60+ organizations, 70+ specialties
Epic real-time prior authorization What it does Payer decisions through the CRD API Presented Aug. 18 Status Live at four systems with three payers; 16 testing
Epic Ergo Visit What it does AI-redesigned office visit Presented Aug. 18 Status Live at Ochsner and a few community physicians
Epic Emmie What it does Patient assistant Presented Aug. 18 Status Live at Sutter and Ochsner; scheduling due in October
Epic Curiosity What it does Generative prediction trained on Cosmos Presented Aug. 18 Status Validation at 20 organizations; EHR component March 2027
Oracle Clinical AI Agent What it does Pro-fee coding, dictation and chart review Presented Aug. 19 Status US availability; release stage not stated
Oracle Clinical AI Agent for nurses What it does Inpatient nursing documentation Presented Sept. 14 Status US availability
Oracle next-generation ambulatory EHR What it does Outpatient record Presented August Status Live for early adopters
Oracle revenue-cycle agents What it does Five revenue-cycle functions Prior authorization, clinical documentation integrity, charge capture, pro-fee coding and appeals Presented Sept. 22–24 Status Announced; no dates given
Oracle Oncology EHR What it does Oncology record with embedded agents Presented Sept. 23 Status Planned features

Buying

Epic’s own scorecard for the 167 projects it announced in 2025 lists 84 delivered on time, seven complete or nearly so and 76 in progress. Treat a keynote as a roadmap, and contract for functions that already run at named reference sites.

05Adoption

Adoption by segment and where the gains land

Clinician use of AI is broad, and near-universal where a large system has licensed an enterprise scribe. No survey in the quarter measured provider organizations by segment, and the figures in circulation measure different things. The differences that are visible concern staff: who can govern agents, and who can contest a payer’s coding decision.

What the common adoption figures measure

Each figure uses its own definition and population. The AMA and Doximity surveys were fielded before the quarter; the others are single-organization or vendor counts.

What the common adoption figures measure
Source What is measuredFigurePopulation and period
AMA physician survey What is measured Physicians using AI in their practices Figure 81% (38% in 2023) Population and period Released March 2026; sample not on release
Doximity member survey What is measured Any AI use; voice-based documentation Figure 63%; 29% Population and period Latest wave, November 2025–January 2026 3,151 member physicians across two waves
Six-hospital academic network (NBER) What is measured First AI-scribed note within a month of eligibility Figure 56%; near-universal within 14 months Population and period 1,868 physicians, May 2024–November 2025
Optum What is measured Employed clinicians using AI chart summarization Figure More than 40% Population and period Third quarter 2026; company statement
Epic What is measured Clinicians assisted by Epic AI each month Figure More than 1.4 million Population and period August 2026; vendor figure
Microsoft (Dragon Copilot) What is measured Patient encounters automated Figure 28 million, double a year earlier Population and period April–June 2026; vendor figure
OpenEvidence What is measured US physicians using the service Figure About 40% Population and period September 2026; company claim, as reported

Large systems and academic medical centers

Large systems produce nearly all the evidence

Inova runs 70 AI features in production, about 20% of them agentic. Mass General Brigham is rolling out Microsoft’s ambient documentation to more than 4,000 providers, and Ochsner, ECU Health and Sutter report results from Epic’s agents. These systems have the governance committees and multirole teams that agentic workflows require.

Community and regional systems

Community systems get AI through their EHR vendor

BayCare and Beacon Health appear as early adopters of Oracle’s agents, and Epic’s sales leadership said its rural customer base keeps growing. EHR-native AI reduces the number of vendors these systems manage but deepens their dependence on one vendor’s releases and pricing, and they are less likely than large systems to have the data scientists that agent monitoring needs.

Independent and small practices

Small practices rely on free tools and face payer automation

No practice-size breakdown of AI adoption was published in the quarter. Small practices reach AI mainly through free or low-cost clinician tools and through vendors such as Hello Patient, which automates outpatient phone work and bought Converse Health for back-office agents. They meet payer downcoding with the fewest staff to contest it, and Illinois’s limit takes effect only in 2028.

Payers

Payers automate coding review, appeals and authorization

UnitedHealth reported that AI information gathering shortened appeal cycle times by 26% and that an agentic project lets a Medicaid plan adjudicate complex claims in seconds rather than days. Commercial plans are connecting real-time authorization to Epic through the CRD API ahead of the 2027 prior-authorization API compliance dates.

NHS England

The NHS is deploying a general assistant at scale

Microsoft said NHS England is rolling Copilot out to 505,000 clinicians and staff after a trial that saved 43 minutes a day. These are Microsoft’s figures for its general-purpose assistant, not for ambient documentation.

Spend and price

Prices and spend remain undisclosed

Epic does not publish AI prices, and no survey of provider AI spend by segment appeared in the quarter. The figures in circulation are vendor claims, such as about $13,000 a year in recovered revenue per clinician from ambient AI. In April, the Peterson Health Technology Institute cited estimates that each prior-authorization submission costs plans $40 to $50 and providers $20 to $30.

Economics

In fee-for-service, a scribe that adds visits and captures more complexity raises provider revenue and payer cost together. Payers are answering with their own coding and appeals automation, so the near-term result is faster, more automated exchange on both sides, with the price of each visit contested code by code.

06Outcomes

Scribe studies point to more visits and higher billed intensity

The quarter’s most rigorous result came from a staggered rollout at one academic network, released as an NBER conference paper on July 22. Other analyses point the same way on coding with incompatible measures, and most implementation results are vendor or customer reports without a stated method.

NBER paper · July 22

Ambient scribe rollout raised physician output

Erkmen Aslim, Manuel Hoffmann and Haonan Yin compared physicians newly eligible for Abridge’s scribe with those not yet eligible at a six-hospital academic network in two states, May 2024 to November 2025. They estimate that eligibility raised monthly work relative value units (wRVUs) by about 20%, and physicians accepted 80% to 90% of drafted text unmodified.

Scope: The paper’s absolute estimate, about 35 wRVUs on a pre-eligibility mean of 226.8, implies nearer 15%. It does not measure patient outcomes or coding accuracy, and Abridge provided administrative support.

About 20%More wRVUs per physician-month, authors’ estimate

5–6More appointments per physician-month

1,868Physicians studied

Background · JAMA, April 1

An earlier multisite study found smaller effects

Rotenstein and colleagues followed 1,809 scribe adopters among 8,581 clinicians at five academic health systems, with access from June 2023 to August 2025. Designs, vendors and measures differ from the NBER study, which shifts attention from time saved to what is billed.

16 minutesLess documentation per eight scheduled hours

0.49More visits a week

$167.37More E/M revenue per adopter-month

Coding intensity after ambient-scribe adoption

Each analysis uses a different unit: a count per physician, a percentage-point share, a percentage change and a dollar total. They point in the same direction and cannot be averaged.

Coding intensity after ambient-scribe adoption
Analysis DesignPopulationReported change
NBER (Aslim, Hoffmann, Yin), July 2026 Design Staggered rollout; not-yet-eligible physicians as comparison Population 1,868 physicians, one academic network Reported change Level 4–5 visits up about 3 per physician-month
Trilliant Health, as relayed Aug. 11 Design Before and after, claims 2018–2024; no control Population Outpatient visits at six large systems Reported change New-patient 99204–99205 share up 12–20 points Established-patient share up 7–12 points
PHTI workshop example, April 2026 Design Before and after Population One multihospital system Reported change Level 5 up 5%, level 4 up 7% About $1,004 more revenue per provider-month
Blue Cross Blue Shield Association, March 2026 Design Claims trend, Q2 2022–Q1 2025 Population Hospital maternity admissions Reported change $2.3 billion in added spending Postpartum anemia coding rose from 4% to 12.3% at high-growth hospitals

Only the NBER study has a comparison group, and none tests codes against the clinical record. BCBSA attributes its rise to AI-enabled coding; the Healthcare Financial Management Association (HFMA) disputes its method.

Results reported from deployments

Organization and vendor reports from the quarter, without published comparison groups. Community Health Network went live in July.

Results reported from deployments
Organization AI useReported resultReported by
Ochsner Health AI use Epic Emmie patient assistant Reported result 1,000+ call-center staff hours saved in 10 months Reported by Customer, at Epic’s meeting
Community Health Network AI use Epic Emmie for billing Reported result 73% fewer billing InBasket messages since go-live Reported by Customer, at Epic’s meeting
ECU Health AI use Agent summarizing transfer-center requests Reported result About 20 hours a week saved Reported by Epic estimate
Inova Health AI use 70 AI features in production Reported result Chart review 15x faster Measles response in about 6 hours vs. 72 Reported by Health system
Optum AI use Chart summarization; appeals research Reported result Chart review time down 33%; appeal cycles down 26% Reported by Company
Oracle customers AI use Clinical AI Agent notes Reported result 400,000+ physician hours saved over about two years Reported by Vendor; method not stated
Humata Health customers AI use AI prior authorization Reported result Up to 96% first-pass approval; 30% fewer write-offs Reported by Vendor
Mass General Brigham AI use Ambient documentation Reported result 21% less burnout Reported by Customer study cited by Microsoft’s CEO
07Corrections

Four claims that need a distinction

Each of these statements circulates widely. The distinction behind each changes what a provider, payer or investor should conclude.

Litigation · Allegation, not finding

The 90% figure is an allegation in the Lokken complaint, not a court finding

Claim under review: UnitedHealth’s nH Predict tool has a 90% error rate.

A discovery response filed with plaintiffs’ Aug. 17 brief in Estate of Lokken v. UnitedHealth Group repeats the figure as an allegation from their amended complaint. The plaintiffs’ surviving claims rest on contract and good-faith theories about how UnitedHealth used naviHealth’s tool.

Cost · Not supported

The comparative evidence shows more visits and higher billed intensity, not lower system cost

Claim under review: AI scribes are lowering health care costs.

The NBER rollout study found higher physician output and more high-level visits, which in fee-for-service raise payer spending. In April, the Peterson Health Technology Institute found that AI may cut each organization’s cost of handling prior authorizations but “has not reduced overall system-level costs.”

Capital · No current basis

Rock Health stopped labeling startups as AI-enabled, so its data give no 2026 AI share

Claim under review: A stated share of 2026 digital-health funding went to AI startups.

Rock Health dropped the label in its H1 2026 report. The 64% AI share of global venture in the third quarter is Crunchbase’s cross-sector figure and does not describe digital health.

UK adoption · Excludes the NHS

The ONS business survey behind the UK figure does not cover the NHS

Claim under review: About 30% of UK organizations, including the NHS, use AI.

The Office for National Statistics’ business insights survey reported 30% of businesses using AI in September, but it samples only the private sector and excludes public provision of health. It says nothing about NHS adoption.

08Accountability

Failures in prior authorization, note quality and agent upkeep

Most of the quarter’s failures came from payment incentives, review habits and maintenance load; the one direct test of output quality rated AI-scribe notes below clinicians’. Outside health care, a September VB Pulse survey of 103 self-selected teams found that 52% of those that test before release had seen an AI feature pass its tests and then fail customers.

Medicare prior authorization · Payment design

WISeR vendors are paid on denials that are not reversed

Since Jan. 15, the WISeR model has used AI vendors to run traditional Medicare prior authorization for 13 services in six states. Records the Electronic Frontier Foundation (EFF) obtained through a FOIA lawsuit, published Sept. 8, show that vendors are paid for denials not reversed on appeal and that low quality scores reduce that pay by only 5% to 10%. CMS says a clinician reviews every denial.

Medicare prior authorization · Operations

The records show delays and uneven vendor conduct

Two vendors denied 5,944 requests in the first three months. Virtix denied more requests than it approved and was put on a corrective action plan, while Innovaccer told CMS a month before launch that it would auto-affirm all requests until its software had full functionality. One request went unanswered for 83 days against a public 72-hour target, a span that counts the full request pathway. On Sept. 29, Bloomberg Government reported internal CMS plans to extend WISeR nationally.

Documentation

Graders rated AI-scribe notes below clinicians’ notes

A 2026 Annals of Internal Medicine study, reported by Healthcare Dive on Sept. 3, rated primary care notes from 11 commercial AI scribes significantly lower than notes from 18 clinicians in all 10 quality domains. The Texas Medical Liability Trust warned that clinicians may review notes less as reliance grows. Healthcare Dive reported no public malpractice case tied to an AI scribe so far.

Coding

Coding exposure runs in both directions

Billed intensity rises after scribe adoption, and agents now suggest the codes. Providers that let codes pass without review face False Claims Act exposure, a risk a physician commentator raised in August, while payer algorithms that lower codes now face state limits.

Operations

Running agents takes staff many systems lack

Parkview Health’s associate CMIO, Hasan Ahmad, said running an agent inside a health system differs from building one: monitoring for drift needs data scientists and informaticists, which some organizations cannot supply. Epic says it is building an Evaluation Suite to test and monitor agents, but its August general session left executives with little detail on how agents would be run.

Litigation

Lokken v. UnitedHealth heads toward class certification

In the suit over UnitedHealth’s use of naviHealth’s nH Predict, UnitedHealth moved on Aug. 10 to add a filed-rate-doctrine defense, and plaintiffs opposed on Aug. 17. Expert disclosures were due Sept. 14, and class-certification motions are due Dec. 31.

Who is accountable

Responsibility stays with identifiable parties: the clinician who signs the note, the provider who submits the code, and the payer or contractor whose denial stands. Each needs a record of human review it can produce on request.

09Regulation

Federal proposals, state limits on payer algorithms, and EU and UK timetables

Federal action in the quarter built frameworks rather than binding duties: two payment proposals, an FDA discussion paper and a new FDA leadership post. Binding change came from the states, where Manatt counts more than 280 health-AI bills introduced and 40 enacted across 24 states in 2026. The EU pushed device-grade obligations to 2028, and the UK set out a lifecycle model it has not yet legislated.

CMS · Outpatient payment · Proposed; comments closed Aug. 31

Medicare proposed a payment category for algorithmic services

The 2027 outpatient payment proposal, released July 2, renames what Medicare called software as a service “Software as a Medical Service” (SaMS) and designates 36 HCPCS codes, including AI retinal-image analysis and CT-derived fractional flow reserve. It moves 21 codes to New Technology APCs at roughly 2026 rates and 10 algorithmic lab-analysis codes off the lab fee schedule. CMS calls 2027 transitional. The category covers device-like diagnostics, not documentation or administrative AI.

CMS · Physician payment · Proposed July 14; comments closed Sept. 14

The physician fee proposal adds AI quality credit and lowers the conversion factor

The 2027 physician fee schedule proposal adds MIPS improvement activities for clinician AI use and interoperable decision support, and asks how Medicare should pay for technology-enabled primary care. The conversion factor falls 1.19% for qualifying APM participants and 1.68% for other clinicians as a one-year 2.5% increase expires.

FDA · Discussion paper; comments due Oct. 19

The FDA floated competency-based review for generative-AI devices

The Aug. 18 discussion paper, which is not guidance, scores risk on two axes, how much and how independently a device acts and how severe an incorrect output could be, and pairs device benchmarking with risk-scaled clinical confirmation, periodic postmarket benchmarking and voluntary foundation-model master files. It lists clinical documentation among agentic uses that may not be a focus of device oversight. On Sept. 8, HHS named Jared Seehafer the FDA’s first deputy commissioner for technology and AI.

States · Payer algorithms

Illinois and Indiana limit algorithmic downcoding

Illinois, in a law signed July 10 and effective Jan. 1, 2028, bars payers from using algorithms to override submitted billing information and requires downcoding to be made or reviewed by a person under CPT guidelines, not solely on diagnosis codes. Indiana’s HB 1271, in effect since July 1, bars AI-only downcoding by insurers and requires providers to have a person review AI-generated claims.

California · Effective Jan. 1, 2027

California set duties for clinical decision support and health chatbots

SB 503, signed Sept. 30, requires developers of AI clinical decision support to document intended uses, training data and bias risks, and deployers to monitor and mitigate them. A second law treats certain consumer health chatbot deployers as health care providers under the state’s medical privacy law and bars AI from directing unlicensed staff in licensed functions. Gov. Gavin Newsom vetoed a bill restricting AI recording of psychotherapy without consent as overly broad.

EU and UK

Europe deferred device duties as the UK proposed lifecycle oversight

The EU’s AI omnibus, in force since July 27, applies high-risk obligations to AI medical devices and IVDs from Aug. 2, 2028, and to Annex III systems from Dec. 2, 2027. In the UK, the National Commission into the Regulation of AI in Healthcare published 44 recommendations on Sept. 10, including lifecycle oversight, contractual allocation of each risk control and telling patients when AI is used. The next King’s Speech, where a comprehensive AI bill could be announced, is not expected before late 2027.

Compliance and decision dates known at quarter end

Dates as published by the end of September 2026. CMS’s 2027 outpatient and physician payment rules were still proposals at quarter end.

Compliance and decision dates known at quarter end
Date Obligation or eventApplies to
July 1, 2026 Obligation or event Indiana HB 1271 takes effect No AI-only insurer downcoding; human review of AI-generated claims Applies to Insurers and providers in Indiana
Aug. 2, 2026 Obligation or event EU chatbot disclosure duties apply Applies to Chatbot deployers in the EU
Oct. 19, 2026 Obligation or event Comments due on FDA generative-AI discussion paper Applies to Device makers and providers
Dec. 31, 2026 Obligation or event Class-certification motions due in Lokken v. UnitedHealth Applies to Parties to the case
Jan. 1, 2027 Obligation or event California SB 503 and health chatbot law take effect Applies to Decision-support developers and deployers; chatbot deployers
2027, from January Obligation or event CMS-0057-F prior-authorization API dates begin AHIP standard fields also start Applies to Impacted payers
July 1, 2027 Obligation or event California SB 1119 companion-chatbot provisions Applies to Companion-chatbot operators
Oct. 1, 2027 (proposed) Obligation or event CMS-0062-P drug electronic prior authorization Applies to Payers
Dec. 2, 2027 Obligation or event EU AI Act Annex III high-risk obligations Applies to Providers and deployers in the EU
Jan. 1, 2028 Obligation or event Illinois payer downcoding law Applies to Payers in Illinois
Aug. 2, 2028 Obligation or event EU AI Act obligations for AI medical devices and IVDs Applies to Manufacturers
10Voices

Where regulators, clinicians, operators and vendors agree and differ

The quarter’s positions divide on three questions: what the coding rise means, how much human review AI needs, and whether automated authorization speeds care or adds a barrier.

Three views converge across constituencies. Ambient documentation is becoming a platform feature, a view Epic, Oracle and Advisory Board share and Abridge’s repositioning reflects. The constraint has moved from licenses to the staff who govern and maintain agents. Regulators, the AMA and the UK commission all want stronger evidence behind generative AI: the FDA’s Abramson expects evidentiary standards to change, and the AMA’s Whyte wants AI held to the standards of other interventions. The disagreements are sharper. On coding, the NBER authors read rising intensity as fuller capture of care delivered. Before the quarter, in March, BCBSA attributed $2.3 billion in added hospital spending to AI-enabled coding, and HFMA said the study conflated documentation accuracy with aggressive coding; Dashevsky says the claims data cannot yet tell the two apart. On human review, Whyte and Ehrenfeld stress clinical context while Adams warns that mandatory review can delay care, partly because they describe different tasks. On authorization, vendors and UnitedHealth describe faster approvals through commercial integrations, while EFF’s records describe delays under Medicare’s denial-linked model.

Regulators

Stephanie Carlton

Deputy administrator and chief clinical AI officer, CMS

Reimbursement is the unsettled part of CMS’s four-part AI strategy: when Medicare should pay for an AI tool, and when that makes sense. She also said CMS does not want a large reporting apparatus to grow up around AI quality.

CMS’s July outpatient proposal would set up a payment category for algorithmic diagnostics, not for documentation tools.

Regulator · reported remarks, Sept. 9

Rick Abramson, MD

Digital Health Center of Excellence, FDA

The current device framework fits generative AI poorly, and he expects evidentiary standards to change.

The FDA’s Aug. 18 discussion paper sets out one possible approach: competency-based review scaled to risk.

Regulator · reported remarks

Alastair Denniston and Henrietta Hughes

Co-chairs, UK National Commission into the Regulation of AI in Healthcare

Regulate AI across its lifecycle, allocate each risk control between manufacturer and provider by contract, and tell patients when AI is used, with an opt-out where appropriate.

Hughes is also the Patient Safety Commissioner. The 44 recommendations await a cross-government response.

Commission recommendations, Sept. 10

Clinical leaders

John Whyte, MD

Chief executive, American Medical Association

Cautioned against overstating AI’s capabilities and asked for the evidence base that safety decisions require, with AI held to the same standards as other interventions.

The quarter’s strongest study measured output and billing, not patient outcomes.

Professional body · reported remarks

Jesse Ehrenfeld, MD

Global chief medical officer, Aidoc; former AMA president

Some of the context clinicians rely on cannot be automated.

His example concerned tasks such as refills that depend on changing clinical context.

Vendor executive · reported remarks

Laura Adams

National Academy of Medicine

Challenged requiring a clinician in the loop for every AI decision, arguing that mandatory review can delay care where AI performs well.

Her example concerned tasks such as image analysis, where AI performs well.

Reported remarks

Health-system operators

Jon McManus

Chief data and AI officer, Inova Health

Each agentic workflow needs its own multirole team, and for some problems AI can be the riskiest option available.

Large systems such as Inova can staff those teams; Parkview’s concern below is about systems that cannot.

Buyer · reported interview

Hasan Ahmad

Associate CMIO, Parkview Health

Building an agent is different from running it inside a health system, and he wants Epic to be clear about how customers are expected to operate and maintain its agents.

His concern is the staffing that drift monitoring demands.

Buyer · reported interview

Jacob Parrish

Vice president of clinical operations, ECU Health

The operational value of agents is real, but he does not yet know how many agents is too many.

Epic plans broad availability of customer-built agents in 2027.

Buyer · reported interview

Payers

Stephen Hemsley

Chief executive, UnitedHealth Group

Reported that UnitedHealth has extended AI chart summarization to more than 40% of its employed clinicians, cutting chart review time by 33%, as part of building what he called modern agentic infrastructure.

UnitedHealth both insures claims and, through Optum, sells coding tools to providers.

Payer · third-quarter earnings call, Oct. 13

Vendors

Judy Faulkner

Founder, Epic

Told customers that what Epic announces at its users group meeting is very likely to happen.

Most of Epic’s 2026 announcements were at early-adopter or announcement stage.

Vendor · users group meeting, Aug. 18

Shiv Rao, MD

Chief executive, Abridge

Abridge is moving in a different direction from Epic while still partnering with it, and change management is the hardest part of deployment.

Point-solution scribes face absorption into the EHR platforms they integrate with.

Vendor · reported interview

Researchers and advocates

Erkmen Aslim, Manuel Hoffmann and Haonan Yin

Authors, NBER study of AI adoption and productivity in health care

Eligibility for an AI scribe raised physician output and coding intensity; they attribute the intensity rise to fuller documentation of care delivered, which the study does not test.

Their reading sides with providers in the coding dispute, though the study cannot settle it.

Academic study · NBER conference paper, July 22

Jared Dashevsky, MD

Physician commentator, Healthcare Huddle

“The False Claims Act does not distinguish between intentional fraud and negligent overcoding.”

He sees a case that AI corrects past undercoding but says the data cannot separate accurate coding from upcoding, and that physicians often never see the final codes.

Commentator · published article, Aug. 11

Lena Cohen

Electronic Frontier Foundation

CMS’s safeguards are weaker than it says: low quality scores cut vendor pay by only 5% to 10%, many requests took far longer than the 72-hour target, and AI recommendations can sway the clinicians who must review denials.

EFF obtained the records through a FOIA lawsuit against CMS.

Advocacy group and litigant · Sept. 8

11Decisions

Three decisions for provider and payer leaders

Chief financial officer and revenue-cycle leader

Decide who confirms AI-suggested codes before they reach a claim

Agents from Oracle, Epic and Optum now suggest codes from the visit record, and billed intensity rises after scribe rollout. Payers are automating their review in response, and Indiana already requires human review of AI-generated claims.

Which codes does a clinician or certified coder confirm, and can we show that review to an auditor?

Chief information officer and CMIO

Decide whether to build agents, wait for the EHR’s own or buy from a specialist

Epic plans broad availability of Agent Factory in 2027, and Oracle has announced five revenue-cycle agents without dates. Parkview and Inova describe the staff each agent needs for monitoring, so the choice depends on staffing as much as on features.

For each agent we turn on, who monitors it, and is that person in place before go-live?

Payer chief medical officer

Decide how a person reviews algorithmic denials and downcoding

Illinois will require human review of downcoding from 2028, Indiana bars AI-only downcoding now, and Lokken v. UnitedHealth is moving toward class certification. The WISeR records show how payment tied to denials draws scrutiny of a vendor’s incentives.

Where a model recommends a denial or a lower code, which person decides, and how is any vendor paid for that work?

12Actions

Four actions before the end of 2026

  1. 1

    Baseline E/M levels before each scribe rollout

    Record each physician’s level 4–5 share and wRVUs before go-live and track them monthly afterward. The result is a documented explanation of coding changes when a payer or auditor asks.

    Revenue-cycle leader

  2. 2

    Inventory the AI features in production

    List each AI feature with its vendor status (early adopter, announced or generally available), its owner and how it is evaluated. The result is a staffing gap you can size before more EHR-native agents arrive in 2027.

    CMIO

  3. 3

    Map authorization volume to payer readiness

    Group prior-authorization volume by payer route: real-time CRD integration, a planned CMS-0057-F API or manual handling, including traditional Medicare in the WISeR states of Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington. The result shows where automation will shorten turnaround and where staff will still be needed.

    Patient access and payer contracting

  4. 4

    Assign risk controls and patient notice in AI contracts

    Write into each AI contract who monitors performance and bias, who patches and who tells patients when AI is used. California’s SB 503 puts monitoring duties on deployers from Jan. 1, 2027, and the UK commission recommends allocating each risk control by contract.

    Compliance and legal

13References

References

Covers care delivery and its financing in the US, EU and UK from July 1 through Sept. 30, 2026; pharma and biotech research are excluded. Earlier sources are cited for background, and October publications for what they report about the quarter.

Research and outcomes

Regulation and courts

Market and practice

Capability and cost

Thoughtful engineering for what’s next.

Have a question about what this means for your team? Let’s talk it through.

Start a conversation

Have a project in mind, or just exploring an idea? Let’s talk.
Fill out the form or give us a call.

413-570-3299

What can we help you with?