For two years, healthcare AI meant one thing: documentation. Ambient scribes listened to patient visits, drafted clinical notes, and gave doctors their eyes back. The U.S. Department of Veterans Affairs is scaling that exact tool to all 130+ medical centers this year. Mount Sinai became the first academic medical center to deploy Microsoft’s Dragon Copilot system-wide. Epic shipped over 150 AI features. athenahealth launched a free ambient scribe. Documentation AI won.
But while the industry celebrated AI that types, a different category of AI started doing things.
Agentic AI, the kind of system that doesn’t wait for a prompt but instead reasons, plans, and executes across workflows, entered healthcare in force this year. Amazon launched a five-agent suite for patient engagement. The NHS launched a national initiative to test agentic AI across care settings. Anthropic and OpenAI both released healthcare-specific products within days of each other in January. And a wave of smaller companies deployed AI agents that interact directly with patients between visits: texting them, delivering assessments, and logging clinical data without a human initiating the workflow.
Documentation AI listens and transcribes. Agentic AI reads a patient’s chart, notices a missed follow-up, sends the reminder, and logs the interaction. Nobody asked it to. That’s the difference.
The documentation era earned its win
Credit where it’s due. The VA’s ambient scribe program is the largest government healthcare AI deployment in the United States. It launched as a pilot in October 2025, and by early 2026, Rise8 and Thoughtworks had a contract to scale from 10 sites to more than 130 VA medical centers. The early numbers are hard to argue with: pilot users saved an estimated 15,700 hours of documentation time in one year. That’s roughly 1,794 working days. A 2026 UCLA study published in NEJM AI found that using AI scribes for just one month significantly reduced physician burnout.
Veterans noticed, too. VA providers reported that patients said they felt more connected during visits. The doctor was actually looking at them instead of typing into a screen.
Mount Sinai’s deployment of Dragon Copilot followed a similar logic: ambient, voice-activated documentation across clinical departments, with plans to expand system-wide through 2026. Becker’s Hospital Review tracked at least 15 health systems that signed enterprise AI deals this year, spanning clinical documentation, decision support, virtual care, and workforce management.
Documentation was the safe bet. It sat between the clinician and the computer, not between the clinician and the patient. What’s happening now is different.
The agents arrive
At the 2026 JPM Healthcare Conference in January, Anthropic launched Claude for Healthcare, a set of HIPAA-ready AI tools with native connections to the CMS Coverage Database, ICD-10 codes, and PubMed. The product supports prior authorization, claims appeals, and care coordination. Patients can sync health records from Apple Health and Android Health Connect. OpenAI had announced ChatGPT Health less than a week earlier. Both companies let users upload medical records and get plain-language explanations of their results. Both emphasized that health data stays private and isn’t used for model training.
NVIDIA’s VP of healthcare, Kimberly Powell, described 2025 as an absolute breakout year for agentic AI at the same conference and pointed to workforce shortages as the driver pushing health systems toward autonomous AI tools.
Two months later, Amazon made its move. Amazon Connect Health went generally available in March with five AI agents designed to reduce administrative burden across the care continuum: patient engagement, intake, scheduling, and point-of-care workflows. Not a pilot. Not a waitlist. Generally available, positioned explicitly as agentic AI for healthcare.
On the engagement side, IKS Health debuted MyCareHub at the American Medical Group Association conference in April. The platform uses behavioral modeling and what the company calls “longitudinal patient memory” to personalize outreach. It scores patients on awareness, ability, and willingness to comply. It then adjusts the channel, cadence, and tone of reminders accordingly. IKS Health deployed MyCareHub at OrthoNY, an orthopedic and pain management practice in New York, and reported improvements in key financial metrics since its 2025 rollout.
The UK moved on governance first. TrustX Health, launched in December 2025, is a partnership between Health Innovation Kent Surrey Sussex, the University of Cambridge’s Trustworthy AI Lab, the Responsible AI Institute, and The King’s Fund. The goal: build a verification and safety framework for deploying agentic AI across NHS and social care before scaling it. Early testing is already underway at Sussex Partnership NHS Foundation Trust. Where the U.S. deployed first and governed later, the UK built the governance layer first.
Then came the proof that the operational model works at institutional scale. In May, NHS Shared Business Services partnered with Salesforce to launch SBS One, an AI-powered finance and procurement platform. An AI agent named “Agent Murphy” handles incoming queries, resolves most within 24 hours, and has cut handling times by 20%. NHS SBS processes approximately £395 billion of NHS funding annually and expects AI agents could eventually manage up to half of service demand. It’s not clinical. But it proves that agentic AI can operate inside a healthcare system at real volume, with real money, on real timelines.
Where agents meet patients directly
The platforms above handle scheduling, documentation, triage, and administrative workflows. A different set of tools focuses on what happens after a patient enrolls in a monitoring program — keeping them engaged day to day.
This is the less glamorous end of healthcare AI. No billion-dollar launches. No JPM keynotes. But it’s where patient outcomes and program revenue are actually determined, because a monitoring program that patients stop using is a monitoring program that stops billing.
The financial context matters here. CMS expanded remote monitoring reimbursement significantly in the 2026 Physician Fee Schedule Final Rule. New CPT codes now allow billing for as few as 2 days of patient data transmission in a 30-day period, down from the previous 16-day minimum. A new code covers management interactions as short as 10 minutes. Reimbursement rates across remote monitoring, chronic care management, and advanced primary care management increased by roughly 7–21%. CMS created financial pathways that didn’t exist a year ago. But those pathways only generate revenue if patients participate consistently enough to meet billing thresholds. When engagement drops, data transmission stops, thresholds go unmet, and the revenue disappears.
That’s the problem a growing number of AI-powered tools are now designed to solve.
Some approach it through behavioral nudges. The American Telemedicine Association has reported that AI-enabled remote monitoring programs can see up to 36% higher patient adherence when automated nudges adapt their timing and tone based on patient response patterns. Prevounce has documented similar results, with AI-driven prompts adjusting to each patient’s history. The system might send a reminder at the time a patient is most likely to respond, or change the message entirely when previous attempts went ignored.
Others go further than nudges. Some remote therapeutic monitoring platforms have deployed AI agents that contact patients via text message or WhatsApp when they stop completing assessments. The agent initiates a conversation, reminds the patient to log in, and if the patient still doesn’t engage through the app, delivers the clinical assessment directly via text. Responses are logged into the patient’s chart automatically. Actuvi, a leading digital care platform with an award-winning RTM product, uses this approach and reports a 91% medication compliance rate across its active patient population.
A 2025 systematic review published in PMC, covering 26 studies on AI tools for medication adherence, found that conversational agents, mobile AI tools, and smart device prompts all showed improvements in adherence outcomes. The methods for measuring adherence varied, which limits direct comparison across studies, but the direction was consistent: AI-driven patient contact works better than leaving patients to remember on their own.
What’s different this time
Readers who sat through the chatbot hype of 2023 and 2024 have reason to be skeptical. Healthcare has a long history of adopting technology in pilot programs and then stalling before production deployment. Three things make this cycle different.
First, these are deployed, not demoed. The VA didn’t run a hackathon. They signed a contract with Rise8 and Thoughtworks to scale to 130+ sites. NHS SBS is processing real invoices through Agent Murphy. OrthoNY measured real financial results from MyCareHub. Every example above is a production deployment with measurable outputs, not a conference slide.
Second, reimbursement now pays for the outcomes that these agents enable. CMS didn’t just adjust rates. It created entirely new CPT code families for shorter monitoring durations and shorter clinical interactions. Post-surgical patients who need 7-14 days of monitoring are now billable. Brief 12-minute check-ins to review an alert and adjust a care plan are now billable. The unit economics of AI-enabled remote care work for the first time across a much wider range of patient types.
Third, governance is forming alongside the technology instead of years behind it. TrustX Health in the UK built a verification framework before scaling deployments. The FDA issued draft guidance on clinical decision support software incorporating generative AI in January 2026. Both Anthropic and OpenAI have emphasized that their healthcare tools require professional review before any clinical decisions are made.
The honest caveat: most health systems are still in the documentation phase. Healthcare Dive reported in January 2026 that the majority of providers have focused their AI efforts on administrative and back-office work. The patient-facing, decision-making, autonomously-acting agents described here are the leading edge, not the norm.
But the gap between “AI that helps you type” and “AI that acts on your behalf” closed faster than anyone expected. In January, the biggest healthcare AI story was an ambient scribe. By June, the story is AI agents managing patient outreach, processing NHS invoices, and coordinating care across health systems.
The organizations moving first aren’t running experiments. They’re signing contracts, deploying at scale, and measuring results. The question for the rest of the industry is whether they’ll get there before the early movers lock in the operational advantages that come with going first.
