AI in care delivery moved into core platforms as payers and providers disputed its effect on claims
Between January and March, Epic made ambient documentation a built-in EHR feature; OpenAI began selling products built for HIPAA-covered work to health systems, and Anthropic to providers and payers; and traditional Medicare began AI-assisted prior authorization in six states. The tools now sit inside the systems that produce notes, codes and authorization requests, so their effects reach payers as well as clinicians.
Editorial snapshot: 30 April 2026
The central finding
Ambient scribes saved clinicians minutes a day in controlled trials, while a payer-affiliated claims analysis attributed about a fifth of a 9% rise in commercial inpatient cost per member between 2023 and 2024 to coding intensity. AI’s financial effect in care delivery is being measured, and disputed, at the level of the claim.
What changed in care-delivery AI in the first quarter of 2026
The quarter’s main developments meet at the claim. Tools that draft notes, look up coverage rules and assemble authorization requests became features of the EHR and of enterprise AI products sold directly to providers and payers, and both sides began measuring what those tools do to coding and approvals.
Capability
Frontier labs and Epic packaged AI for clinical work
OpenAI (Jan. 8) and Anthropic (Jan. 11) launched products for HIPAA-covered health care work: OpenAI with business associate agreements, Anthropic with connectors to Medicare coverage rules and code sets. The week of Feb. 4, Epic released AI Charting, which drafts notes and queues orders inside its EHR.
Capital
Digital health funding concentrated in 12 large rounds
Rock Health counted $4.0 billion across 110 digital health deals, 59% of it in 12 rounds of $100 million or more, and OpenEvidence raised $250 million at a $12 billion valuation. The largest disclosed technology budgets were payers’ own, led by UnitedHealth Group’s planned investment of nearly $1.5 billion in 2026.
Adoption and economics
Physician AI use reached 81% as payers measured rising coding intensity
The American Medical Association found 81% of physicians used AI in early 2026, up from 38% in 2023. Blue Health Intelligence, the Blue Cross Blue Shield Association’s analytics partner, attributed about 20% of a 9% rise in commercial inpatient cost per member between 2023 and 2024 to coding intensity and linked its timing to AI documentation tools; the Healthcare Financial Management Association said the analysis may mistake more accurate documentation for aggressive coding.
Regulation
Federal oversight loosened for clinical software and tightened for prior authorization
FDA widened enforcement discretion for clinical decision support in final guidance, and comments closed Feb. 27 on a proposal by HHS’s health IT office (ASTP/ONC) to drop AI transparency criteria from certification. CMS decision deadlines for payers took effect Jan. 1, public prior authorization metrics were due March 31, and traditional Medicare began AI-assisted review in six states.
Operating question
For each AI-assisted note, code or authorization decision, can your organization show what the tool produced, who reviewed it and which clinical facts support the result?
AI arrived packaged for HIPAA-covered work and inside the EHR
General models advanced: Anthropic released Claude Opus 4.6 on Feb. 5, and OpenAI released GPT-5.4 on March 5 as its first general-purpose model with native computer use, reporting 75.0% on the OSWorld-Verified benchmark against a 72.4% human baseline; that capability is relevant to payer portals that offer no API. For providers and payers, the larger change was packaging: compliance terms, health data connectors and native EHR features that turn earlier custom builds into products a health system can switch on.
Compliance terms
Frontier models under a business associate agreement
OpenAI for Healthcare offers ChatGPT for Healthcare with a business associate agreement (BAA), data residency options, audit logs and customer-managed encryption keys, plus an API with BAAs for eligible customers, both on GPT-5.2 models. OpenAI bases its performance claims on HealthBench, which it says does not measure percentage accuracy, and on GDPval.
Payer rules
Coverage rules and code sets as standard connectors
Claude for Healthcare connects to the CMS Coverage Database of local and national coverage determinations, ICD-10, the NPI Registry and PubMed, with skills for FHIR development and a sample prior authorization review. That turns coverage checks and appeal drafting into configurable workflows. Anthropic cites model benchmarks, not customer results.
Documentation
Ambient notes and order queuing inside the EHR
Epic’s AI Charting, built with Microsoft on Azure-hosted models, drafts the note and queues orders from the visit conversation. Epic’s chief medical officer, Jackie Gerhart, said it will extend to orders and diagnoses, which would bring it closer to the line where FDA regulates decision-support software.
Patient records
Patients can load their records into general assistants
ChatGPT Health connects U.S. medical records through b.well, plus Apple Health and wellness apps; it launched Jan. 7 with a waitlist and is not offered in the European Economic Area, Switzerland or the U.K. Claude’s consumer health integrations are in a U.S. beta. Consumer health apps are not covered by HIPAA, and OpenAI does not describe ChatGPT Health as HIPAA compliant.
Cost and consequence
When ambient notes, coverage lookups and appeal drafts come built into the EHR or a BAA-covered model, less of the cost is in building data pipelines and more is in configuration, evaluation, alignment with coding policy and security review. Agents with file or connector access also inherit known attack types, including the prompt-injection exfiltration found in Anthropic’s Cowork and the remote-code-execution flaws fixed in its Git MCP server, both disclosed in January.
Venture funding concentrated in a few large rounds while payers funded their own technology programs
Digital health funding followed the wider venture market on a much smaller base: more money in fewer, larger rounds. AI took about 80% of global venture funding in the quarter, by Crunchbase’s count. Rock Health no longer tracks AI deals separately because AI is now standard across digital health, so it reports no AI share of the sector total.
$4.0B
Digital health funding across 110 deals
First quarter of 2025: $3.0 billion across 122 deals. The average deal, $36.7 million, was the highest since the fourth quarter of 2021.
59%
Share of capital in 12 rounds of $100 million or more
Rock Health counts these as megadeals.
Disclosed financing
Digital health rounds of $100 million or more, Q1 2026
The 12 megadeals in Rock Health’s first-quarter count, in USD millions. Several went to consumer, wearable or virtual-care companies rather than provider-facing AI.
Whoop
Funding round — $575M
Verily
Funding round — $300M
OpenEvidence
Series D — $250M
$12 billion valuation; led by Thrive Capital and DST Global.
Talkiatry
Funding round — $210M
eMed
Funding round — $200M
Grow Therapy
Funding round — $150M
Honest Health
Funding round — $140M
Solace
Funding round — $130M
Qualified Health
Funding round — $125M
Named as an Anthropic partner.
Garner Health
Funding round — $118M
Cognito Therapeutics
Funding round — $105M
Midi Health
Funding round — $100M
Rock Health, Q1 2026 funding overview (April 6, 2026); OpenEvidence announcement (Jan. 21, 2026).
Clinical reference
An ad-supported clinical reference reached a $12 billion valuation
OpenEvidence’s $250 million Series D, announced Jan. 21, brought its total raised to nearly $700 million. The service is free to doctors and supported by advertising; the company reports about 18 million consultations in December 2025, up from about 3 million a month a year earlier.
Incumbent spending
Payers disclosed the largest technology budgets
UnitedHealth Group plans to invest nearly $1.5 billion in modern intelligent technologies in 2026 and at least as much in 2027, and UnitedHealthcare expects nearly $1 billion in operating cost reductions in 2026, many AI-enabled. Elevance Health said it is investing more than $1 billion in digital and AI-enabled capabilities. HCA Healthcare named technology investment among the reasons its other operating expenses rose 90 basis points as a share of revenue.
Consolidation
Acquisitions rose as the largest planned combination collapsed
Rock Health counted 43 digital health M&A deals, up from 30 in the previous quarter, including OpenAI’s acquisition of Torch. New Mountain Capital’s planned $32 billion combination of five portfolio companies into Thoreau fell apart in March, and Rock Health said the public-listing window for digital health remains narrow.
What it signals
The venture money reaching care delivery went to a few large rounds, several outside provider-facing AI. The largest disclosed technology budgets are payers’ own, and payers describe their AI gains as administrative cost reduction and payment integrity; providers are answering with revenue-cycle staff, appeals and automation of their own, as HCA described.
Epic built the scribe into the EHR as frontier labs sold to health systems directly
Three ways of supplying AI to care delivery now meet in the same accounts: features native to the EHR, frontier-lab products bought directly under a BAA, and point solutions such as ambient scribes, some of which build on the same frontier models.
EHR platform
Epic made ambient documentation an EHR feature
Epic released AI Charting to customers the week of Feb. 4, with several systems piloting it; STAT also put Epic at 55% of hospital beds. In Epic accounts, third-party scribes now compete on what the native tool lacks, such as revenue-cycle integration or coverage across several EHRs.
42%Epic’s share of the acute-care hospital EHR market, per STAT
85%Epic customers live with its generative AI, per Epic
Frontier labs
OpenAI and Anthropic sell to health systems and to scribe vendors
OpenAI named eight health systems rolling out ChatGPT for Healthcare, among them AdventHealth, Cedars-Sinai, HCA and UCSF, and lists Abridge and Ambience among its API customers. Anthropic’s named partners include Banner Health, Elation Health, Commure and Premier. The labs now sell to scribe vendors and to those vendors’ customers at once.
Revenue cycle
Revenue-cycle agents arrived inside the platforms
Epic says more than 200 organizations use Penny, its revenue-cycle AI, and that Insights chart summaries run more than 16 million times a month, nearly triple November’s level. Optum plans to integrate Optum Real’s AI-driven revenue-cycle products with payment and financing services for providers.
Availability of the quarter’s main launches
Dates and status as announced by each vendor or reported in trade press at launch.
| Product | Launched | Status | Built for |
|---|---|---|---|
| ChatGPT Health (OpenAI) | Launched Jan. 7 | Status Waitlist; not offered in the EEA, Switzerland or U.K. | Built for Patients; U.S. record connections |
| ChatGPT for Healthcare and API (OpenAI) | Launched Jan. 8 | Status Rolling out to institutions; BAA offered | Built for Health systems and health tech developers |
| Claude for Healthcare (Anthropic) | Launched Jan. 11 | Status Connectors available; consumer features in U.S. beta | Built for Providers, payers and vendors |
| AI Charting (Epic) | Launched Week of Feb. 4 | Status Released to customers; several systems piloting | Built for Epic customers |
| Penny, Emmie and Insights updates (Epic) | Launched Week of Feb. 4 | Status Updates to existing agents | Built for Revenue cycle, scheduling and bills, chart summaries |
OpenAI did not state when ChatGPT for Healthcare would be generally available, and Epic did not state how widely AI Charting is available.
Physician AI use reached 81%, and the scribe’s financial case runs through coding
The quarter’s adoption figures come in units that cannot be combined: physicians’ self-reported use, vendors’ customer counts and hospital surveys from 2024. A physician who checks a question in a free app counts in the AMA’s 81% whether or not the employer licenses any AI, and what employers license and integrate is measured mainly in older hospital data. The economics are clearer at the level of the claim.
How AI use was measured in the quarter
Each row is a separate measure from a different publisher and population; the rows cannot be compared or combined.
| Measure | Reported value | Population | Period |
|---|---|---|---|
| Physicians using AI in practice (AMA) | Reported value 81% (38% in 2023) | Population U.S. physicians; 1,692 surveyed | Period Jan. 15 to Feb. 2, 2026 |
| AI use cases per physician (AMA) | Reported value 2.3 (1.1 in 2023) | Population Same AMA survey | Period Jan. 15 to Feb. 2, 2026 |
| U.S. physicians using OpenEvidence daily | Reported value More than 40% | Population Vendor figure; base not stated | Period January 2026 |
| Epic customers live with Epic generative AI | Reported value 85% | Population Epic customers (vendor figure) | Period February 2026 |
| Hospitals using generative AI | Reported value About one-third | Population U.S. hospitals (ONC data) | Period 2024 |
| Eligible visits where the assigned scribe was used | Reported value 29.5% to 33.5% | Population 238 UCLA physicians in a trial | Period November 2024 to January 2025 |
| UnitedHealthcare member calls using AI tools | Reported value More than 80% | Population UnitedHealthcare member calls | Period Stated Jan. 27, 2026 |
| Elevance commercial members using its AI assistant regularly | Reported value 22 million | Population Elevance commercial members | Period Stated on first-quarter 2026 earnings call |
| People asking ChatGPT health questions each week | Reported value More than 230 million | Population ChatGPT users worldwide (OpenAI) | Period January 2026 |
In the AMA survey, 1,342 of the 1,692 respondents answered the use question.
Independent and small practices
Small practices face per-seat prices and downcoding
Training is nearly as common in solo practice as in hospitals, by the AMA’s count. Nong and Neprash estimate scribes at $300 to $500 per clinician per month, and practices billing office visits face payers’ algorithmic downcoding of level-4 and level-5 claims, adopted by Cigna and Aetna in 2025, without large systems’ contracting leverage.
74%Solo physicians with AI training
79%Hospital-based physicians with AI training
Community and regional hospitals
Native EHR features lower the integration bar
Hospitals without data-science teams can adopt Epic’s AI Charting and Penny as EHR features instead of integrating separate tools. The American Hospital Association told ASTP/ONC that hospitals use the certified decision-support transparency data that HTI-5 would remove to evaluate the AI tools they buy. No survey in the quarter separated community hospitals from academic centers.
Large systems and academic centers
Large systems test products and contract with labs directly
Duke and UCLA tested scribes in randomized trials on their own clinicians, and most of the health systems OpenAI named are large systems or academic centers. HCA’s chief executive credited ambient listening with physician productivity without quantifying it, and contract terms and returns were not disclosed.
Payers
Payers tied cost targets to AI in their operations
UnitedHealth Group’s chief financial officer, Wayne DeVeydt, attributed part of the company’s 12.8% operating cost ratio outlook to early AI benefits. Elevance, which said more than 60,000 associates have access to AI tools, reported that its adjusted operating expense ratio improved 20 basis points to 10.5%.
Breakeven
Four extra level-4 visits a month cover a scribe’s fee
In a Jan. 9 JAMA Health Forum viewpoint, Paige Nong and Hannah Neprash of the University of Minnesota estimate that a $500 monthly scribe fee is recovered by about four additional level-4 established-patient Medicare visits a month. With trials measuring minutes saved per day, the larger financial lever for fee-for-service practices is the level of service coded.
Inpatient claims
Coded severity rose fastest at a tenth of hospitals
Blue Health Intelligence, the Blue Cross Blue Shield Association’s analytics partner, measured the shift in admissions coded with complications or comorbidities between the second quarter of 2022 and the first quarter of 2025. It attributes about 20% of a 9% rise in commercial inpatient cost per member from 2023 to 2024 to coding without matching care, across plans covering about 62 million members.
46.8% to 59.8%Complex-coded share of admissions, top 10% of hospitals
+4.2 pointsSame measure, other 90% of hospitals
Office visits
Visit levels shifted at six systems that adopted scribes
Trilliant Health, an analytics company, told TechTarget that all-payer claims from 2018 to 2024 at six large systems that announced scribe adoption show visits shifting to the two highest levels for new and established patients, with high-intensity new-patient visits up 12 to 20 percentage points.
Economics
For a fee-for-service provider, a scribe’s measurable return comes mainly through the level of service coded, which is what payers are now measuring. Expect those gains to be contested through audits, downcoding and payment adjustments; Nong and Neprash cite Medicare Advantage’s mandatory 5.9% coding-intensity adjustment as a precedent and argue that higher codes also raise patients’ cost-sharing.
Scribe trials found small time savings and no burnout difference between products
Duke’s crossover trial, published Feb. 23, compared two ambient scribes against each other rather than against usual care. Read with UCLA’s three-arm trial from November 2025, the controlled evidence shows a real but small effect on documentation time and modest gains in well-being. Results reported by vendors and payers in the quarter were their own process measures, without comparators.
Documentation time and well-being in scribe studies
UCLA’s trial ran Nov. 4, 2024, to Jan. 3, 2025, and was published in NEJM AI in November 2025; Duke’s ran October to December 2024. Mini-Z is scored from 10 to 50.
| Study and tool | Design | Documentation time | Burnout or well-being |
|---|---|---|---|
| UCLA Health: Nabla | Design Randomized vs. usual care; 238 physicians | Documentation time 9.5% less time in note (significant) | Burnout or well-being Mini-Z well-being +2.69 points |
| UCLA Health: Microsoft DAX | Design Same trial | Documentation time 1.7% less time in note (not significant) | Burnout or well-being Mini-Z well-being +2.83 points |
| Duke Health: scribe B | Design Randomized crossover; 160 clinicians enrolled | Documentation time 9.07 fewer minutes a day in notes vs. baseline | Burnout or well-being Burnout down from baseline; no difference between tools |
| Duke Health: scribe A | Design Same trial | Documentation time 5.88 fewer minutes a day in notes vs. baseline | Burnout or well-being Burnout down from baseline |
| The Permanente Medical Group | Design Organization’s own tally; no comparator | Documentation time About 15,000 hours over 2.5 million uses in a year | Burnout or well-being — No comparable figure |
In UCLA’s release, Nabla users’ time per note fell from 4:30 to 3:49, against 4:22 to 4:04 with usual care. The Permanente Medical Group’s figure is its own one-year tally, reported before the quarter.
Head-to-head trial
At Duke, satisfaction differed more than outcomes
Physicians preferred scribe B, which also cut more time in notes. Differences per appointment, in after-hours EHR work and in burnout were not significant. The trial had no external funding, and the products were not named.
3.19 minutesDaily time-in-notes difference between the two scribes
+2.51 vs. +1.91Satisfaction change on a 1-7 scale
Revenue cycle · Vendor-reported
Epic reports fewer coding-related denials for Penny users
Epic says many Penny customers report coding-related denials down more than 20% and medical-necessity appeal letters produced 23% faster. It gave no comparator or baseline.
Prior authorization · Payer-reported
Elevance reports fewer missing-information denials
Elevance Chief Executive Gail Boudreaux said AI and the company’s HealthOS data-sharing platform cut prior authorization denials caused by missing information by close to 70%. The company gave no baseline or period.
What holds
Controlled trials support a modest documentation-time benefit that varies by product and by how often clinicians use the tool. Health systems can measure the same effects with EHR audit-log data they already collect, as Duke did with minutes in notes, after-hours time and note closure.
Four claims from the quarter that need narrowing or correcting
Several widely repeated claims overstated a finding or misdated a rule. Each correction changes a decision: how to value a scribe, how to read coding data, how Medicare’s new model works and when EU device obligations begin.
Documentation time · Overstated
Controlled trials measure minutes a day, though volume can make aggregate totals large
Claim under review: Ambient scribes give clinicians back hours of documentation time; Commure, on Anthropic’s launch page, describes “saving clinicians millions of hours annually.”
Duke’s trial measured reductions of 5.88 and 9.07 minutes a day in notes from baseline, and UCLA found a 9.5% reduction in time in note for one product and no significant change for the other. Aggregate tallies can still be large: The Permanente Medical Group counted about 15,000 hours over 2.5 million uses in a year.
Coding costs · Overstated
The research tied part of a cost rise to coding intensity and inferred the AI link from timing
Claim under review: Hospitals’ AI systems inflate health care costs, as a March 10 trade headline summarized the Blue Cross Blue Shield research.
Blue Health Intelligence attributes about 20% of a 9% rise in commercial inpatient cost per member from 2023 to 2024 to coding intensity without matching care. It links that rise to AI through timing, vendor announcements and hospital surveys rather than by measuring AI use at each hospital, and says only that the shift aligns with AI deployment. The Healthcare Financial Management Association said the study appears to conflate more accurate documentation with aggressive coding.
Medicare prior authorization · Mischaracterized
In Medicare’s WISeR model, AI assists review and a clinician reviews every request that is not approved
Claim under review: Traditional Medicare now lets AI deny care.
Under WISeR, which began Jan. 1 in six states, technology vendors use AI and machine learning to review prior authorization requests, and a human clinician reviews every request that is not affirmed. Vendors are paid a share of the spending averted and penalized for inappropriate denials, a design that KFF notes has raised concern about an incentive to maximize denials. A House bill introduced Dec. 2, 2025, H.R. 6361, would prohibit the model.
EU AI Act · Wrong date for devices
AI in medical devices falls under the act’s 2027 date, and EU institutions have proposed 2028
Claim under review: EU high-risk AI obligations reach medical AI on Aug. 2, 2026.
The Aug. 2, 2026, date applies to stand-alone high-risk systems listed in Annex III. AI that is, or is a safety component of, a medical device or in vitro diagnostic device requiring third-party (notified-body) conformity assessment falls under Article 6(1), whose obligations apply from Aug. 2, 2027. The Council of the EU (March 13) and the European Parliament (March 26) backed moving that date to Aug. 2, 2028, but the change had not been agreed by March 31.
Chatbot errors, scribe inaccuracies and coding audits showed where responsibility sits
The quarter’s failure evidence concerns three things: chatbots used outside clinical validation, documentation accuracy at the margin and the integrity of what gets billed. Two further questions, patient privacy and liability, apply across all three.
Patient safety
ECRI ranked chatbot misuse the top technology hazard for 2026
ECRI’s experts said chatbots such as ChatGPT, Claude, Copilot, Gemini and Grok have suggested incorrect diagnoses, recommended unnecessary tests, promoted subpar supplies and invented anatomy; when ECRI asked one chatbot, it endorsed an unsafe placement for an electrosurgical return electrode. ECRI published the list Jan. 21, the same month OpenAI and Anthropic launched consumer health features.
Documentation accuracy
Trial scribes made omission and misattribution errors
In Duke’s trial, five of 112 free-text responses (4%) raised safety concerns such as misattributing speakers and over-summarizing, with no serious safety events reported. UCLA’s earlier trial reported occasional clinically significant inaccuracies, mostly omissions and pronoun errors, and one mild adverse event.
Coding integrity
A plan audit found most postpartum anemia codes unsupported at one system
At high-growth hospitals in Blue Health Intelligence’s analysis, postpartum anemia coding rose from 4.0% to 12.3% of maternity admissions while transfusion rates stayed roughly flat, an estimated $22 million in additional spending. One Blue Cross Blue Shield plan audited maternity cases at one outlier hospital system and found fewer than 20% met established clinical criteria for postpartum anemia.
Privacy
Records patients load into chat assistants sit outside HIPAA
Records patients connect to ChatGPT Health are held by a consumer service that HIPAA does not cover, and OpenAI does not describe the product as HIPAA compliant. Privacy advocates cited by The Register questioned where health data shared with Claude could end up.
Liability
Liability for AI-assisted care is unsettled
Physicians ranked clear liability frameworks first among their regulatory needs in the AMA survey. Verisk’s generative AI exclusion endorsements for general liability policies became available in January. FDA’s new decision-support guidance moves more clinician-facing software outside device review, which leaves more of the risk with the clinicians and organizations that deploy it.
Who is accountable
In each case, responsibility stays with the deploying organization and the signing clinician: a scribe’s draft becomes the physician’s note, a code generated from it becomes the hospital’s claim, and a consumer chatbot’s answer reaches patients without clinical validation. ECRI also lists sudden loss of access to electronic systems and implementations that create unsafe workflows among 2026’s top hazards, both relevant as documentation and coding run through AI features inside the EHR.
Federal policy loosened software oversight and standardized prior authorization
U.S. federal policy moved in two directions in the quarter. FDA reduced oversight of clinical software, and comments closed on an ASTP/ONC proposal to reduce it further, while CMS put deadlines, published metrics and AI-assisted review into prior authorization. State law remained the main source of AI-specific duties, and EU device obligations stayed on a 2027 date in law.
CMS-0057-F · In force Jan. 1, 2026
Payers face decision deadlines and public denial data
Under CMS’s interoperability and prior authorization rule, Medicare Advantage plans and Medicaid and CHIP programs must decide expedited requests within 72 hours and standard requests within seven days, and give specific reasons for denials. Their first public prior authorization metrics, covering 2025, were due March 31. Prior authorization and data-exchange APIs follow on Jan. 1, 2027.
WISeR · Operating from Jan. 1, 2026
Traditional Medicare adopted AI-assisted prior authorization
In Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington, the Wasteful and Inappropriate Service Reduction model has six technology vendors use AI and machine learning to review prior authorization requests for selected services, with clinician review of every nonaffirmation. KFF found these services were 5.3% ($12.3 billion) of traditional Medicare Part B spending in 2024, 83% of it skin substitutes, and expects a modest first-year effect.
FDA · Final guidance, January 2026
FDA widened enforcement discretion for decision support
The revised guidance, issued without a prior comment period, applies enforcement discretion to decision support that gives a single recommendation when only one option is clinically appropriate, a phrase it does not define. Software that predicts time-critical events or analyzes medical images or signals remains regulated. A separate wellness guidance brings more noninvasive wearables, including wrist-worn blood pressure features, within its scope.
ASTP/ONC · Proposed; comments closed Feb. 27
HTI-5 would drop AI transparency criteria from certification
The proposal would remove 34 of 60 health IT certification criteria and revise seven, including the decision-support source-attribute requirements that work as model cards, and would treat automated access, including by autonomous AI, as covered by information-blocking rules. HHS estimates $1.53 billion in savings. No final rule had been issued by March 31.
States · Effective Jan. 1, 2026
State laws carry most AI-specific duties
California’s AB 489 bars AI from implying health care licensure, its SB 243 sets safeguards for companion chatbots, and Texas’ HB 149 (TRAIGA) requires telling patients when AI is used in their care. Laws enacted in 2025 in Arizona, Nebraska and Texas bar AI-only medical-necessity denials. Colorado’s AI Act (SB 24-205) takes effect June 30 unless replaced; a replacement draft agreed March 17 was not enacted.
EU and U.K.
EU guidance slipped and a U.K. commission took evidence
The European Commission missed its Feb. 2 deadline for guidance on classifying high-risk AI systems, and the Council and Parliament adopted positions in March that would delay obligations for AI in medical devices from Aug. 2, 2027, to Aug. 2, 2028. In the U.K., the National Commission into the Regulation of AI in Healthcare, convened by the Medicines and Healthcare products Regulatory Agency and chaired by Alastair Denniston, closed its call for evidence in February and met three times in the quarter.
Compliance dates and their status at March 31, 2026
Dates as published by the end of the quarter. Proposed changes had not been adopted by March 31.
| Obligation | Date | Applies to | Status at March 31 |
|---|---|---|---|
| CMS-0057-F decision deadlines and denial reasons | Date Jan. 1, 2026 | Applies to Medicare Advantage, Medicaid and CHIP | Status at March 31 In force |
| WISeR prior authorization model | Date Jan. 1, 2026, to Dec. 31, 2031 | Applies to Providers in six states | Status at March 31 Operating |
| California AB 489 and SB 243; Texas TRAIGA | Date Jan. 1, 2026 | Applies to Deployers in those states | Status at March 31 In force |
| HTI-5 comment period | Date Closed Feb. 27, 2026 | Applies to Certified health IT developers | Status at March 31 Final rule pending |
| First public prior authorization metrics | Date March 31, 2026 | Applies to Payers under CMS-0057-F | Status at March 31 Due |
| Colorado AI Act (SB 24-205) | Date June 30, 2026 | Applies to Deployers of high-risk AI | Status at March 31 Scheduled; replacement drafted |
| EU AI Act obligations for stand-alone high-risk systems (Annex III) | Date Aug. 2, 2026 | Applies to EU providers and deployers | Status at March 31 Scheduled; Dec. 2, 2027, proposed |
| CMS-0057-F prior authorization and data-exchange APIs | Date Jan. 1, 2027 | Applies to Payers under CMS-0057-F | Status at March 31 Scheduled |
| EU AI Act obligations for AI medical devices and IVDs needing third-party assessment | Date Aug. 2, 2027 | Applies to Manufacturers and deployers in the EU | Status at March 31 In law; Aug. 2, 2028, proposed |
What it changes
Less clinical software will pass through FDA review or carry certified transparency data, so health systems will need to obtain that information from vendors themselves. The authorization process, meanwhile, is becoming faster, more documented and more public, which gives providers data to contest denials and gives payers deadlines they must meet.
Where payers, providers, regulators and researchers agree and differ
The positions converge on direction. Payer-affiliated analysts, independent researchers and the Healthcare Financial Management Association all accept that coded intensity is rising; they dispute whether it reflects fuller documentation or inflation, a counterfactual no one has measured directly. Payers, CMS and providers are all automating prior authorization. Several apparent contradictions describe different things: Elevance and HCA measure denials at different stages of the claim, and the American Hospital Association and ASTP/ONC describe different parts of the market — hospitals without in-house evaluation capacity and developers carrying certification costs. The quarter produced few on-record positions from named patient advocates, small-practice physicians or community-hospital leaders.
Coding intensity
Payer-affiliated analysts, provider finance staff and clinicians describe the same claims from opposite positions.
Blue Health Intelligence
Analytics partner of the Blue Cross Blue Shield Association
“Coding intensity is increasing commercial inpatient spending”
Its brief attributes about 20% of a 9% rise in commercial inpatient cost per member from 2023 to 2024 to coding without matching care and links the timing to AI tools.
Payer-affiliated analysis · issue brief, March 2026
Shawn Stack
Director of Perspectives and Analysis, Healthcare Financial Management Association
Said Blue Health Intelligence’s study appears to conflate more accurate documentation with aggressive coding.
HFMA represents health care finance staff, mostly on the provider side.
Provider-finance association · HFMA report, March 17
Dr. Luann Racher
University of Arkansas for Medical Sciences
Said scribes let her engage with patients, that clinicians review notes before billing and that the aim is accurate documentation rather than upcoding.
Her account describes the clinician review step that payer audits will test.
Clinician · interview with TechTarget, March 17
Paige Nong and Hannah Neprash
University of Minnesota
Argue that using scribes for billing could raise spending through higher codes, more recorded diagnoses and care-gap revenue, and that productivity pressure could absorb the time saved; they propose a national AI-use registry, clinician-in-the-loop claims audits and downward payment adjustments.
Neprash said at a March 11 conference that it is unclear whether the increases reflect more complex care.
Independent researchers · JAMA Health Forum viewpoint, Jan. 9
Prior authorization and denials
Payers report fewer avoidable denials while a large hospital operator reports denials still rising; they are measuring different stages of the claim.
Tim Noel
CEO, UnitedHealthcare
“We anticipate operating cost reductions of nearly $1 billion in 2026 — many AI-enabled.”
UnitedHealth Group plans nearly $1.5 billion of technology investment in 2026.
Payer · earnings remarks, Jan. 27
Gail Boudreaux
CEO, Elevance Health
Described AI as part of Elevance’s operations rather than a separate experiment, and said AI with its HealthOS platform cut prior authorization denials caused by missing information by close to 70%.
The figure covers pre-service authorization denials at one payer, with no baseline given.
Payer · first-quarter earnings call, April 22
Mike Marks
Chief Financial Officer, HCA Healthcare
Said denials and underpayments rose across payers, led by Medicare Advantage, and remain high; revenue-cycle staff, technology, appeals and dispute resolution kept the earnings effect small, and digital partnerships with payers are at an early stage.
He describes post-service claim denials across a hospital operator’s payers, a different stage from Elevance’s figure.
Provider · first-quarter earnings call, April 24
Alex Cottrill, Jeannie Fuglesten Biniek, Juliette Cubanski, Tricia Neuman and Misha Segal
KFF
Expect WISeR’s first-year effect to be modest, partly because CMS separately cut skin-substitute payment rates, and note concerns that paying vendors a share of averted spending could reward denials.
Independent policy research · analysis, Feb. 10
Oversight of clinical software
Regulators presented the quarter’s changes as removing burden; hospitals argued they would lose information they use to buy AI safely.
Marty Makary
Commissioner, U.S. Food and Drug Administration
Said the January guidances cut unnecessary regulation and promote innovation with AI and medical devices.
Regulator · Jan. 6 statement, as reported by Arnold & Porter
Dr. Tom Keane
Assistant Secretary for Technology Policy and National Coordinator for Health IT
“These proposals reflect a commonsense approach that removes redundant requirements on health IT developers”
The HTI-5 proposal he introduced drew public comments through Feb. 27.
Regulator · HHS release, Dec. 22, 2025
Ashley Thompson
Senior Vice President, Public Policy Analysis and Development, American Hospital Association
“These data are critical to foster trust in AI tools and ensure patient safety.”
AHA opposed removing the certified decision-support transparency criteria and asked ASTP/ONC to issue broader guidance on AI and information blocking first.
Hospital association · HTI-5 comment letter, Feb. 27
AI in clinicians’ and patients’ hands
Physicians report everyday use and want a say in adoption; ECRI warned about unvalidated chatbots in the month frontier labs launched consumer health features.
Dr. John Whyte
CEO, American Medical Association
“AI has quickly become part of everyday medical practice.”
In the same survey, 85% of physicians wanted to be consulted on adoption decisions and 88% were concerned about losing skills.
Physician association · survey release, March 12
Dr. Marcus Schabacker
President and CEO, ECRI
“Medicine is a fundamentally human endeavor.”
He said it in presenting chatbot misuse as the top health technology hazard for 2026.
Patient-safety organization · hazard list, Jan. 21
Fidji Simo
CEO of Applications, OpenAI
Presented ChatGPT Health as a way to organize patients’ scattered health records.
OpenAI says the product is not intended for diagnosis or treatment.
Vendor · reported by Fortune, Jan. 7
Platforms and buyers
Vendors describe the quarter in terms of distribution and integration; the available buyer statements explain why a system chose a supplier rather than what it achieved.
Eric Kauderer-Abrams
Head of biology and life sciences, Anthropic
Called health care and life sciences one of Anthropic’s largest bets and described the launch as connectors to tools professionals already use rather than new models.
Vendor · interview with Fierce Healthcare, January
Corey Miller and Jackie Gerhart
Vice President of R&D (Miller) and Chief Medical Officer (Gerhart), Epic
Plan to make AI Charting an active assistant that extends from notes to orders and diagnoses.
That extension would bring the product closer to FDA’s line for regulated decision support.
EHR vendor · reported by Healthcare Dive and STAT, February
Mike Reagin
Chief Technology Officer, Banner Health
Said Banner chose Anthropic for its focus on AI safety.
A purchasing rationale; Banner reported no results.
Health-system buyer · interview with Fierce Healthcare, January
Three decisions for clinical, financial and technology leaders
Chief medical information officer
Native EHR scribe or a separate product
Epic’s AI Charting is now an option inside the EHR for Epic customers, and Duke’s head-to-head trial found satisfaction differed between products more than time or burnout outcomes did. A separate scribe now has to justify its cost against a feature already in the record system.
Which measured result, beyond clinician preference, would justify paying for a separate scribe where the native tool is available?
Chief financial officer and revenue-cycle leader
How to defend coded severity after scribe deployment
Blue Health Intelligence’s analysis, Trilliant’s visit-level data and payers’ downcoding policies all focus on the codes that rise when documentation becomes fuller. Whether those codes are accurate will be settled claim by claim, against the clinical indicators and treatment in the record.
For the diagnosis and visit codes that rose most after scribes went live, can we show the clinical evidence and treatment behind each one?
Chief information officer
What to require from AI vendors when certification no longer does
HTI-5 would remove the certified source-attribute data that hospitals use to evaluate decision-support tools, and FDA’s guidance moves more of those tools outside device review. Frontier labs now sell BAA-covered products and connectors directly, and agents with file or connector access face the prompt-injection attacks disclosed against Anthropic’s Cowork in January.
If certification and FDA review no longer supply model information, what will we require vendors to disclose and test before deployment?
Four actions for the next quarter
1
Baseline coding mix before and after scribe deployment
Compare complex-coded admissions and office-visit levels by service line and clinician for the periods before and after scribes went live, using Blue Health Intelligence’s top-decile pattern as a reference point. The result is a record of where coded intensity rose and which notes support it, ready before payers ask.
Revenue-cycle director
2
Evaluate scribes with existing EHR audit data
Track minutes in notes, after-hours EHR time, note closure and the share of eligible visits where the scribe is used, as Duke and UCLA did. The result is a product decision based on the organization’s own clinicians and a usage rate that shows how much of the licensed capacity clinicians use.
Chief medical information officer
3
Align authorization workflows with the new deadlines
Map CMS-0057-F’s 72-hour and seven-day decision limits and denial-reason requirements into payer or provider workflows, and in the six WISeR states track nonaffirmations and appeal outcomes by service. The result is evidence for disputes and a plan for the Jan. 1, 2027, API requirements.
Prior authorization lead
4
Review AI tools that connect to clinical and claims data
For BAA-covered frontier products and agents connected to EHR, coverage or claims data, confirm audit logs, data residency, encryption-key control and testing against prompt injection through files and connectors. The result is an approved configuration that security and compliance teams can sign off on.
Chief information security officer
References
Covers Jan. 1 through March 31, 2026. Earnings calls and reports published in April are cited for what they report about the first quarter.
Products and platforms
- Introducing ChatGPT Health — OpenAI, 2026-01-07
- OpenAI for Healthcare — OpenAI, 2026-01-08
- Claude for Healthcare and Life Sciences announcement (Anthropic news: healthcare-life-sciences) — Anthropic, 2026-01-11
- Claude Opus 4.6 — Anthropic, 2026-02-05
- Introducing GPT-5.4 — OpenAI, 2026-03-05
- Epic AI charting ambient scribe (Epic details AI Charting) — STAT, 2026-02-04
- Epic's AI scribe goes live as EHR giant touts strong adoption of built-in AI features — Fierce Healthcare
- Epic rolls out AI Charting (Art) notetaking documentation tool — Healthcare Dive, 2026-02-05
- JPM26: Anthropic launches Claude for Healthcare to turbocharge AI efficiency at health systems, payers — Fierce Healthcare
- OpenAI launches ChatGPT Health in a push to become a hub for personal health data — Fortune, 2026-01-07
- Claude joins the ward as Anthropic eyes US healthcare data — The Register, 2026-01-12
Capital and markets
- Q1 2026 funding overview: Capital continues concentrating and four other market signals — Rock Health, 2026-04-06
- OpenEvidence Raises $250 Million to Build Medical Superintelligence for Doctors — OpenEvidence via Business Wire, 2026-01-21
- Record-breaking funding: AI global Q1 2026 — Crunchbase News, 2026-04-01
Payer and provider disclosures
- UnitedHealth Group Q4 2025 earnings prepared remarks — UnitedHealth Group, 2026-01-27
- Elevance Health (ELV) Q1 2026 Earnings Transcript — The Motley Fool, 2026-04-22
- HCA (HCA) Q1 2026 Earnings Call Transcript — The Motley Fool, 2026-04-24
Adoption and practice
- AMA: AI usage among doctors doubles as confidence in technology grows — American Medical Association, 2026-03-12
- Physician sentiments around the use of AI in health care (2026 report) — American Medical Association, Center for Digital Health and AI
- Important news: Aetna and Cigna announce automatic downcoding policies — American Society of Interventional Pain Physicians, 2025-09-11
Research and outcomes
- Comparing ambient scribes: a randomized crossover clinical trial addressing ambient scribe technologies' impact on physician burnout — Journal of the American Medical Informatics Association (Chowdhury et al., Duke), 2026-02-23
- UCLA study finds AI scribes may reduce documentation time and improve physician well-being — UCLA Health, 2025-11-26
- Ambient AI Scribes in Clinical Practice: A Randomized Trial (PubMed record, NEJM AI 2025;2(12)) — PubMed / National Library of Medicine, 2025-11-26
- Unintended Consequences of Using Ambient Artificial Intelligence Scribes for Billing — JAMA Health Forum (Nong and Neprash, viewpoint), 2026-01-09
- Rising Coding Intensity and Its Impact on Health Care Affordability (Issue Brief) — Blue Health Intelligence / Blue Cross Blue Shield Association
- BCBS research shows hospitals' AI systems inflate health care costs — BenefitsPRO, 2026-03-10
- AI hospital coding costs study (Fast Finance) — HFMA, 2026-03-17
- How AI scribes are shifting coding intensity, reimbursement — TechTarget RevCycle Management, 2026-03-17
- Examining the Potential Impact of Medicare's New WISeR Model — KFF, 2026-02-10
Safety, liability and security
- Misuse of AI chatbots tops annual list of health technology hazards — ECRI, 2026-01-21
- Verisk to Roll Out New General Liability Exclusions for Generative AI Exposures — Independent Insurance Agents & Brokers of America (Independent Agent), 2025-10-21
- Contagious Claude Code bug Anthropic ignored promptly spreads to Cowork — The Register, 2026-01-15
- Anthropic quietly fixed flaws in its Git MCP server that allowed for remote code execution — The Register, 2026-01-20
Regulation and policy
- FDA Cuts Red Tape on Clinical Decision Support Software — Arnold & Porter, 2026-01-21
- Key Updates in FDA's 2026 General Wellness and Clinical Decision Support Software Guidance — Faegre Drinker, 2026-01-09
- Clinical Decision Support Software (guidance document page) — U.S. Food and Drug Administration, 2026-01-29
- CMS WISeR Model — DLA Piper, 2026-01-14
- CMMI ushers in new era with WISeR model — Norton Rose Fulbright, 2025-07-14
- CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F) fact sheet — Centers for Medicare & Medicaid Services, 2024-01-17
- Payers' prior authorization denial rates go public: 5 notes — Becker's Payer Issues, 2026-03-30
- State Legislative Activity: AI in health care (issue brief) — American Medical Association
- ASTP/ONC's Year-End Moves Mark a Strategic Pivot — Holland & Knight, 2026-01-07
- HHS Proposes HTI-5 Rule — U.S. Department of Health and Human Services, 2025-12-22
- AHA comment letter on HTI-5 proposed rule (ASTP/ONC Deregulatory Actions to Unleash Prosperity) — American Hospital Association, 2026-02-27
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal text — EUR-Lex / Publications Office of the European Union, 2024-07-12
- EU Digital Omnibus on AI update: Council and Parliament agreed positions — Addleshaw Goddard, 2026-03-27
- European Commission misses deadline for AI Act guidance on high-risk systems — IAPP, 2026-02-03
- New state AI laws are effective on January 1, 2026, but a new executive order signals disruption — King & Spalding, 2025-12-29
- Artificial intelligence working group agrees on framework to replace Colorado law — Colorado Politics, 2026-03-17
- National Commission into the Regulation of AI in Healthcare — Medicines and Healthcare products Regulatory Agency (gov.uk)
Thoughtful engineering for what’s next.
Have a question about what this means for your team? Let’s talk it through.
Start a conversation