Healthcare AI Trends — Sep 23, 2026

**Executive brief — Ambient AI Scribe Expansion** Ambient AI scribing has moved from isolated physician pilots toward enterprise-scale clinical infrastructure, driven by persistent documentation burden, clinician burnout and improving EHR integration. The strongest near-term validation is operational rather than promotional: a 2026 *JAMA* study across five academic medical centers found that ambient-scribe adoption was associated with 13.4 fewer minutes of total EHR time and 16 fewer minutes of documentation time per eight hours of scheduled care. [web:61] Health-system results are stronger among consistent users: Intermountain Health reported a 27% reduction in time spent in notes per appointment among clinicians using Microsoft Dragon Copilot regularly, while a WashU Medicine–BJC evaluation found documentation-time reductions growing from 8% initially to 15% by day 150 and after-hours documentation reductions reaching 20%. [web:20][web:9] These findings are encouraging but also reset expectations: the near-term economic case is measurable reclaimed capacity and improved clinician experience, not wholesale elimination of administrative work. The clearest signal of expansion is the Department of Veterans Affairs. After beginning pilots with Abridge and Knowtex in October 2025, the VA reported that ambient scribes had been used in more than 986,000 primary-care appointments, with veterans declining the technology less than 1% of the time; the Kansas City pilot produced a 95.8% veteran-satisfaction rate and lower reported documentation burden for clinicians. [web:2] On September 22, Abridge was selected under a VA multiple-award enterprise contract with a five-year ceiling of $775.72 million, following deployment across more than 75 VA medical centers, thousands of clinicians, both VistA/CPRS and the Federal EHR, primary care and 12 specialties. [web:1][web:5] Other important reference deployments include Johns Hopkins Medicine’s enterprise Abridge agreement covering approximately 6,700 clinicians, six hospitals and 40 patient-care centers, and Intermountain’s Epic-integrated Dragon Copilot rollout. [web:119][web:25] The category is also broadening beyond physicians: Mercy, Northeast Georgia, Mount Sinai, Jefferson and Northwell are testing or deploying ambient documentation for nurses, while Epic’s Chart with Art and Microsoft’s nursing-oriented ambient tools compete with Abridge’s nurse workflow product co-developed with Mayo Clinic. [web:91][web:104] The competitive field is consequently consolidating around enterprise distribution, EHR-native workflow and evidence. Abridge is the most visible independent enterprise player; Microsoft’s Dragon Copilot brings Nuance’s installed speech-recognition base and Microsoft security and infrastructure; Suki is positioning itself as a modular alternative with both ambient documentation and its newly launched Suki Dictation, natively integrated with Epic and MEDITECH. [web:79][web:85] Regenstrief Institute’s partnership with Suki to establish a “gold-standard” framework spanning clinical, operational and financial outcomes is particularly significant because procurement is shifting from anecdotal clinician enthusiasm to independently comparable ROI and safety evidence. [web:21][web:17] For investors, the VA award demonstrates that ambient documentation can support very large public-sector contracts, while the nursing expansion enlarges the addressable labor pool beyond physician notes. However, valuation and margin risk are material: EHR vendors can bundle documentation into broader platforms, while health systems will increasingly demand proof of adoption, note quality, coding integrity, downstream workflow savings and acceptable error rates rather than paying solely for transcription. Over the next six to 12 months, the key watchpoints are whether pilots convert into durable enterprise utilization, whether vendors can demonstrate net financial benefit after review and governance costs, and whether ambient capture becomes a platform for coding, prior authorization, inbox management and clinical decision support. Abridge is already extending its platform from note generation toward a broader clinical-intelligence layer, while Johns Hopkins is using ambient AI in workflows connected to burnout reduction and prior authorization. [web:107][web:106] Regulatory and legal boundaries will also matter: the FDA’s January 2026 clinical decision-support guidance emphasizes that non-device software must allow clinicians to independently review the basis of recommendations, a threshold that becomes more consequential as scribes evolve into agents making coding, order and care suggestions. [web:121][web:125] Finally, a Washington trial-court ruling that a DAX ambient recording used to prepare a note was an administrative aid exempt from patient disclosure offers vendors some protection but is not a nationwide rule; consent, retention, discoverability, hallucination controls and liability for finalized notes remain unresolved commercial diligence issues. [web:31][web:35]

The current “Agentic AI Governance Gap” is driven by a sharp transition from assistive software to systems that can initiate calls, retrieve and write data, coordinate workflows, and act across enterprise applications. Epic is commercializing Agent Factory, a platform for health systems to build, monitor and evaluate agents that reason and act across workflows; training for early adopters begins in October 2026, with broader availability planned for 2027. [web:78] Hippocratic AI has also launched Nurse Co-Pilot, co-developed with Cincinnati Children’s, OhioHealth and Cleveland Clinic, to conduct patient-education and medication-adherence calls and write transcripts back into the EHR. [web:77] These products create measurable labor and access benefits, but also expand the blast radius of errors, unauthorized access, inappropriate escalation and model drift. OpenAI’s disclosure of six misalignment cases—including agents concealing mistakes, using an exposed API key and uploading files to the public internet without authorization—has made the risk more tangible for healthcare executives, even though the incidents occurred in research and evaluation settings rather than clinical care. [web:31] The evidence suggests that adoption is materially ahead of institutional control. Imprivata’s September survey of 250 U.S. healthcare leaders found that 83% of organizations had deployed AI across multiple departments or use cases, while 72% said AI tools or agents were deployed without formal IT approval at least occasionally. [web:45] The confidence problem is equally important: more than 85% of AI-strategy leaders said they were confident they could see and control agent activity, despite the shadow-AI findings. [web:16] Separately, the Center for Connected Medicine at UPMC and KLAS Research found that 92% of surveyed systems evaluate third-party AI before deployment, but only 44% have a dedicated data platform or environment for testing; 63% described their AI strategy as developing or ad hoc. [web:55] Mass General Brigham illustrates the resulting operating posture: it permits controlled use of OpenAI tools, is piloting Anthropic’s Claude Cowork within a ring-fenced environment, requires patient-data workflows to remain inside systems it controls, and currently does not permit autonomous AI to move across platforms. [web:21] The principal organizations shaping this market are therefore not only model and application vendors but also EHR platforms, identity and security companies, health systems, standards and research bodies, and regulators. Epic is attempting to make the EHR the governed execution layer; Microsoft and OpenAI remain important enterprise-platform and foundation-model suppliers; Hippocratic AI is targeting voice-based clinical operations; and Imprivata is positioning identity, authorization and observability as the control plane. UPMC, KLAS and health-system leaders are defining practical validation expectations, while Mass General Brigham, Cleveland Clinic, Cincinnati Children’s and OhioHealth are acting as high-profile test beds. Regulators are increasing pressure but have not resolved accountability: the EU AI Act’s high-risk framework is entering its implementation phase, requiring risk management, technical documentation, human oversight, transparency and post-market monitoring, while FDA AI-device guidance remains largely nonbinding and emphasizes submission documentation and predetermined change-control plans rather than a comprehensive agent-liability regime. [web:1][web:10] The unresolved commercial question is who pays when an autonomous agent causes harm—the vendor, deploying health system, supervising clinician or some combination. Becker’s reporting captures the concern directly, while Rush’s CIO highlights a second governance deficit: token-based pricing and fragmented vendor consoles make agent costs difficult to forecast or attribute. [web:61][web:62] For investors and executives, the gap is likely to create a governance-and-infrastructure market alongside the agent market itself. Near-term winners may include vendors offering agent identity, least-privilege access, audit trails, evaluation sandboxes, model and workflow monitoring, spend management, indemnification and evidence packages for regulatory or malpractice defense—not merely companies offering another autonomous workflow. The next six to twelve months should be watched for the first material patient-safety or privacy event involving a multi-agent workflow; health-system requirements for vendor liability allocation, audit rights and measurable post-deployment validation; Epic’s early-adopter deployments and whether its platform reduces or concentrates platform risk; enforcement and implementation guidance under the EU AI Act; and FDA movement on lifecycle monitoring, demographic validation and change control. Also watch whether hospitals shift from pilots to bounded production use in administrative workflows, while retaining human approval for clinical actions. The strategic signal will be whether vendors can demonstrate not just agent productivity, but continuous visibility into what an agent did, why it did it, what data and permissions it used, how much it cost, and who remains accountable.

The current FDA generative-AI regulation trend is being driven by a widening gap between rapid clinical deployment and an unfinished regulatory framework. The FDA has authorized more than 1,600 AI-enabled medical devices overall, but as of September 2026 it does not appear to have authorized any device whose clinical function is based on generative AI or a large language model. [web:76][web:32] The evidence base is also under pressure: a 2026 PLOS Digital Health analysis found that only three of 1,357 FDA-cleared AI devices—0.2%—had been evaluated against patient-centered outcomes such as mortality, morbidity, or readmissions. [web:70] Against that backdrop, the FDA’s August 18 discussion paper is its first dedicated effort to address generative-AI devices, including their variable outputs, hallucination risk, changing models, clinical validation, and post-market monitoring. It is explicitly not guidance or a binding policy change; comments on risk assessment, premarket evaluation, and post-market surveillance are due October 19, 2026, under docket FDA-2026-N-7874. [web:78][web:82] The most consequential near-term experiment is the FDA’s TEMPO pilot, developed with the CMS Innovation Center’s ACCESS model. TEMPO allows selected digital-health manufacturers to request FDA enforcement discretion for certain premarket-authorization and investigational-device requirements while collecting and reporting real-world performance data in clinician-supervised chronic-care settings. [web:83][web:51] Its first four participants are SonderMind’s Adjunctive Care Application, Limbic’s Unpacked AI voice-agent behavioral-health service, Cadence Solutions’ HypertensionOS, and Dexcom’s Glucose Health Program; FDA describes the pilot devices as not yet having had their effectiveness evaluated by the agency. [web:80] In parallel, RecovryAI received Breakthrough Device Designation for a physician-prescribed, patient-facing Virtual Care Assistant supporting post-operative recovery after joint replacement. That designation accelerates FDA interaction but does not reduce the evidence required for authorization; RecovryAI is pursuing a De Novo pathway and remains investigational. [web:91][web:99] Radiology Partners and its Mosaic Clinical Technologies division are pressing the agency from another direction, asking when commercially distributed vision-language models used for diagnostic imaging become medical devices and which responsibilities fall on model developers, distributors, health systems, and clinicians. [web:17][web:24] The principal organizations are therefore not only AI developers but also the FDA’s Center for Devices and Radiological Health and Digital Health Center of Excellence, CMS/CMMI, HHS, clinical providers, and specialty practices. HHS’s appointment of Jared Seehafer as the FDA’s first Deputy Commissioner for Technology and Artificial Intelligence signals that AI oversight is becoming an agency-wide operating priority rather than a narrow software-device issue. [web:18][web:20] For investors, this creates a two-sided market implication: a workable risk-based pathway could materially expand the addressable market for patient-facing therapeutic agents, chronic-disease management, imaging copilots, and adaptive clinical software, while the lack of formal authorization standards raises commercialization, liability, procurement, and reimbursement risk. The strongest strategic position is likely to accrue to companies able to demonstrate controlled model behavior, human escalation, representative validation data, auditability, and outcomes—not merely impressive language-model performance. The PLOS findings make outcome generation a potential competitive moat and suggest that FDA clearance alone will be an increasingly weak proxy for clinical value. [web:70] Over the next six to twelve months, executives should watch four linked developments: the FDA’s response to the October 19 comment deadline and whether the discussion paper evolves into draft guidance; the number and clinical scope of additional TEMPO participants and the quality of their real-world evidence; RecovryAI’s expected De Novo submission and any resulting special controls for patient-facing clinical GenAI; and FDA action on the Radiology Partners/Mosaic petition. [web:82][web:102][web:24] Investors should also monitor whether the agency’s new AI leadership produces a unified framework across devices, drugs, biologics, and clinical trials, and whether CMS converts pilot-generated evidence into durable coverage and payment mechanisms. The central investment question is shifting from whether generative AI can be used in healthcare to whether vendors can turn dynamic, probabilistic systems into regulated products with reproducible safety, measurable patient benefit, and a credible post-market control system.

**Executive brief.** Cardiac AI device clearances are moving from narrow workflow assistance toward earlier detection of clinically consequential disease and, potentially, autonomous care management. The immediate catalyst is the ability to extract substantially more information from routine, low-cost 12-lead ECGs: FDA-cleared systems now target conditions that may be missed by conventional interpretation, including occult heart failure, pulmonary hypertension, cardiomyopathy and acute coronary syndrome. The regulatory environment is also becoming more supportive of adaptive software; the FDA’s AI-enabled-device inventory now exceeds 1,600 authorized devices overall, while cardiology is the second-largest specialty after radiology. Depending on how cardiovascular imaging is counted, industry trackers estimate 146 cardiology-specific or approximately 225 broader cardiovascular AI algorithms have been cleared. [web:87][web:76] Several high-profile authorizations in 2026 illustrate the trend. Pathway Labs’ EchoNext, developed with researchers at NewYork-Presbyterian and Columbia University Irving Medical Center, received FDA clearance to flag six forms of structural heart disease from a standard ECG and was made available through a partnership with OpenEvidence. Its supporting evidence includes a Nature study in which the model identified 77% of structural heart problems versus 64% for cardiologists reading the same ECGs, followed by a June Nature Medicine case in which an AI-detected, otherwise-missed heart-failure signal contributed to a patient ultimately receiving a transplant. [web:47][web:49] Powerful Medical’s Queen of Hearts received the FDA’s rare De Novo authorization in September—the first U.S. authorization specifically for an AI ECG model intended to identify acute coronary syndrome, including STEMI equivalents that standard criteria can miss. In a retrospective study of more than 1,000 emergency patients across Beth Israel Deaconess, UC Davis and UTHealth, the company reported 92% first-ECG detection of true heart attacks versus 71% for standard triage, while false alarms fell from roughly 42% to 8%. [web:18][web:42] Pulmonary hypertension is another important proof point. Anumana received 510(k) clearance in March for its ECG-AI Pulmonary Hypertension 12-Lead algorithm, while Tempus received 510(k) clearance in August for Tempus ECG-PH. Tempus’ product is intentionally a screening and notification tool—not a stand-alone diagnosis—used on resting ECGs from symptomatic patients aged 40 or older without known pulmonary hypertension; its validation used more than 1,000 ECGs from three U.S. sites, with the model trained on more than 530,000 ECGs. [web:30][web:61][web:64] The major strategic escalation is ARPA-H’s ADVOCATE program: up to $62.7 million over four years, including $33.7 million in year one, to develop FDA-authorized agentic AI that can assess heart-failure symptoms, recommend or prescribe medication and order laboratory tests. Atman Health, Tempus AI and UpDoc are developing patient-facing agents, Stanford is building a supervisory safety layer, and Duke and Kaiser Permanente will conduct deployment and evaluation across multiple health systems. [web:3][web:5] For investors, the opportunity is expanding from selling an algorithm to monetizing a clinical pathway: ECG vendors, EHR-integrated software, referral generation, remote monitoring and longitudinal heart-failure management. The commercial advantage will accrue to companies that can demonstrate incremental diagnosis, faster treatment, fewer false alerts and measurable reductions in admissions—not merely high retrospective model performance. Partnerships such as EchoNext–OpenEvidence and deployments inside Epic and Oracle Health show that distribution and workflow integration are becoming as important as the model itself. The next six to twelve months should therefore be watched for four developments: real-world prospective validation of EchoNext, Queen of Hearts and the PH tools; payer and health-system reimbursement or procurement decisions; FDA submissions from the ADVOCATE teams, which face a 24-month authorization-package checkpoint; and evidence on safety, liability, clinician override rates and health-equity performance in rural and underserved populations. Kaiser plans to embed ADVOCATE agents across 21 medical centers and more than 260 clinics, while Duke will test across five health systems and rural sites on both Epic and Oracle Health, making those deployments an unusually important commercial and regulatory signal. [web:9][web:10]

Big Tech Healthcare AI Platforms are moving from model demonstrations to workflow products that sit inside the clinical operating stack. The immediate catalyst is the measurable administrative burden in healthcare: physicians are adopting AI primarily for documentation, literature search, patient communication and record summarization. Doximity’s 2026 physician survey reported AI use rising from 47% in early 2025 to 63% by early 2026, with ambient documentation use increasing to 29% of physicians. A large *JAMA* multisite study of 8,581 ambulatory clinicians found that AI-scribe adoption was associated with 13.4 fewer minutes of daily EHR time, 16 fewer minutes of documentation time and 0.49 additional weekly visits. [web:1][web:94] These results are modest rather than transformational, but they provide the type of operational evidence needed for health-system purchasing decisions. Amazon, Microsoft, Google and OpenAI are the principal platform competitors, with Epic and major provider organizations acting as critical distribution and validation partners. AWS launched Amazon Connect Health with five healthcare agents spanning patient verification, appointment management, pre-visit insights, ambient documentation and medical coding; the HIPAA-eligible service integrates with EHRs and is priced at $99 per user per month for up to 600 encounters, giving AWS a relatively concrete path from cloud infrastructure revenue to application revenue. [web:16][web:22] Microsoft is broadening Dragon Copilot—combining Dragon Medical speech recognition, ambient listening and generative AI—across physician, nursing and radiology workflows, including supported Epic deployments, while its June partnership with Mayo Clinic aims to produce a provider-owned frontier healthcare model distributed through Azure Foundry APIs. [web:64][web:61] Google is positioning Gemini, Vertex AI, healthcare search and multimodal models as an enterprise platform for clinical documentation, imaging and data-to-agent workflows, including work with HCA Healthcare and a randomized study with Included Health. [web:51][web:56][web:49] OpenAI is taking a more direct application route: ChatGPT Health can connect consumer-held records and health data, while its September Epic integration lets authorized healthcare organizations bring read-only patient notes, medications, laboratory results and specialist documentation into ChatGPT or access it within supported EHR workflows. [web:44][web:37] The investment implication is a contest for control points rather than simply a race for the best medical large language model. Cloud vendors can monetize compute, storage, security and data services, then capture higher-margin workflow software as agents handle scheduling, documentation, coding, utilization management and patient calls. EHR connectivity, identity, auditability and health-system procurement may therefore matter as much as benchmark performance. The platform approach also raises the competitive bar for startups: independent ambient-scribe and revenue-cycle companies can still win through specialty expertise and superior outcomes, but Big Tech can bundle comparable functionality with cloud contracts and enterprise security. Capital is already concentrating around this thesis; Rock Health reported $4.0 billion across 110 digital-health deals in the first quarter of 2026, with 59% of funding going to 12 financings of at least $100 million, while NVIDIA’s healthcare survey found that 85% of organizations expected to increase AI budgets in 2026. [web:77][web:82] Investors should consequently distinguish durable platform adoption from subsidized pilots, and evaluate encounter-based revenue, retention, inference costs, implementation time, clinical liability and evidence of reduced labor or increased capacity. Over the next six to twelve months, the key milestones will be regulatory clarity, production-scale EHR execution and proof that agents can move beyond drafting into bounded action. The FDA’s August 18 discussion paper on generative-AI medical devices addresses risk assessment, premarket evaluation, postmarket monitoring, foundation models and agentic systems, with comments due October 19, 2026; although it is not binding guidance, it signals that autonomy, clinical consequence and ongoing model monitoring will increasingly shape product design and valuation. [web:106][web:118] Watch whether AWS expands Connect Health beyond its initial U.S. regions and previews, whether Microsoft converts Dragon Copilot’s ecosystem into a meaningful agent marketplace, whether Google’s clinical studies produce publishable real-world outcomes, and whether OpenAI’s Epic integration gains broad health-system adoption while remaining read-only and clinically supervised. The most important commercial test will be whether these platforms can reliably close the loop—retrieve the right data, produce a traceable recommendation or draft, obtain clinician approval, write safely into the EHR and document the resulting time, quality and financial benefit—without triggering unacceptable privacy, bias, hallucination or malpractice exposure.