Capability was the easy part. Cost, Liability, and Payment caught up.

Epic’s AI agents can now act inside other vendors’ software  AI Agents

At Open@Epic, Epic announced Connectors: links that let its AI agents pull information from third-party apps and carry out tasks in them. That covers Emmie for patients, Art for clinicians, Penny for revenue cycle, and the custom agents health systems build in Agent Factory. Launch examples include an ambulance app sending arrival times to the emergency department and a dietary app taking a meal order. Epic will use Model Context Protocol at launch, with developer-built Connectors reaching its Showroom marketplace in March 2027. Each organization can review what a Connector is able to read and change before enabling it. Availity, the first developer named in Epic’s release, is targeting coverage, prior authorization and claims management. healthsystemCIO

Zocdoc invites AI agents in while everyone else locks them out  AI Agents

Registered agents can search Zocdoc’s provider directory, verify insurance, check real-time availability and book a confirmed slot on a provider’s calendar, with explicit patient consent required before booking, changing or cancelling. Google Gemini and Amazon Health AI are live; a separate partnership with the personal assistant Instinct was announced as rolling out in the coming weeks. Agents connect through an MCP server or REST API and are barred from training on, selling or republishing the data. Zocdoc says agent traffic has grown nearly 30-fold since July, though it has published no baseline or measurement method. Amazon, Delta and several major US banks have blocked personal AI agents from their platforms. Healthcare Dive

Ro ships a patient-facing agent with no human review  AI Agents

Telehealth company Ro launched an AI Agent for its members that answers questions, checks orders, billing and insurance, navigates its services and drafts messages to the care team. Ro states that it is not a clinician, does not diagnose or prescribe, and that its answers are not reviewed by a person before users see them. Business Insider

NEJM AI: AI-recommended care costs 3.2x physicians’  Clinical AI

Researchers ran 24 AI systems against 200 real outpatient and emergency department notes and priced every resulting order at 2026 Medicare rates. Across the 20 general-purpose systems, recommended care came to $314 a visit against $97 for the treating physicians. The four medical-specific systems were no cheaper, at 2.7 to 3.5 times. In two-thirds of visits the physician ordered no diagnostic tests at all, and in every one of those cases the AI recommended testing anyway, often a panel such as a complete blood count and metabolic panel. AI also recommended roughly twice as many new medications, 1.14 per visit against 0.59. A cost-aware prompt cut the ratio to 2.5 times and a parsimonious-but-safe prompt to 1.4 times. About 20 percent of the additional workup was judged guideline-concordant. The authors write that the pattern “parallels human defensive medicine,” and note the study measures cost rather than outcomes or harm. Published Oct 15. NEJM AI

AMA: liability follows control, and you are the deployer  Governance

In an Oct 29 issue brief, the AMA’s Center for Digital Health and AI set out who should answer when clinical AI harms a patient, aligning responsibility with each party’s ability to prevent harm. Developers must demonstrate and maintain safety and effectiveness. Deployers, meaning health systems and practices, are accountable for how AI is selected, implemented, governed and monitored. Physicians are responsible for the judgment they can meaningfully exercise, should have strong liability protection when relying on well-validated tools used as intended, and should be free to decline unproven AI without penalty. Patients should retain avenues for remedy. The brief warns that indemnification clauses, liability caps and disclaimers can place responsibility on parties with little power to prevent harm. More than a third of respondents to the AMA’s 2026 physician survey named clear liability frameworks as the regulatory step most likely to increase their trust in AI. The AMA describes this as an initial contribution rather than adopted policy, and named six priority areas for further work, including liability for increasingly autonomous AI and automation bias. Becker’s Hospital Review

A Senate draft would have Medicare test paying for autonomous AI  Policy

Senate HELP chair Bill Cassidy released a legislative discussion draft on Oct 29 called the PATIENT AI Act. It would specify whether a given AI-based technology falls under FDA or ONC oversight, with applications that are not FDA-regulated devices moving into ONC’s certified health IT program. It would give FDA a risk-based review process for software-based devices, including authority to review changes on a rolling basis. It directs the CMS Innovation Center to run a five-year payment model for applications “capable of performing a clinical function autonomously,” with patient notice and appeal rights written in, and restates federal preemption of state product liability law for these devices. The draft has not been introduced or voted on. Comments are due Dec 1. Becker’s Hospital Review

An agent ordered 7,000 cancer screenings, and that was the point  In Practice

Beacon Health System in northern Indiana was behind on colorectal cancer screening heading into fiscal year end. The quality team’s plan was to mass-sign paper orders, which would have returned results as PDFs. Chief Information and Digital Experience Officer Stacey Johnston, MD instead pushed for an agent that placed electronic orders so results came back as discrete data. It was built in about three weeks. Roughly 7,000 screenings were ordered and about 6,500 kits delivered, with more than 2,600 returned, around 200 positive, and one early colon cancer detected in time for surgery. The agent was then converted from a year-end catch-up into ongoing monitoring: when a patient turns 50 it notifies them, places the order, ships the kit and returns discrete results. The Digital Patient

A new benchmark scores sycophancy, not accuracy  Model Behavior

Artificial Analysis launched PersonalityBench, which scores models on communication style, warmth, transparency, sycophancy and willingness to admit uncertainty rather than on accuracy. It found wide variance in default behavior, from highly transparent to markedly sycophantic, with some models rarely conceding uncertainty. Artificial Analysis

$90M into agents, and all three rounds lead with the humans  Funding

Rock Health counted $7.4 billion flowing into digital health in the first half of 2026, with 20 rounds taking 45 percent of the total. Three agent-focused Series A rounds have closed recently.

Baba, $40M Series A led by FirstMark Capital with CVS Health Ventures and Valtruis joining returning investors General Catalyst, Genius Ventures and Ground Up Ventures, for Medicare and Medicaid navigation. Every member gets a dedicated team of care coordinators and licensed therapists, with AI agents handling paperwork, insurer calls and billing follow-up. Selected for the CMS ACCESS Model. The company also released BabaBench, a benchmark for long-horizon patient advocacy tasks. Fierce Healthcare

Avante, $30M Series A led by Frist Cressey Ventures, building AI agents for corporate benefits teams to analyze programs, evaluate vendors and give employees personalized plan guidance. Its Ava Agents beta launched Oct 20. GeekWire

Midstream Health, $20M Series A for an autonomous platform that flags revenue leakage and supply-chain savings for hospitals, already deployed at CommonSpirit, Mount Sinai and Houston Methodist. MobiHealthNews

Mayo Clinic Platform pays startups in data, not dollars  Accelerator

Mayo Clinic Platform welcomed its largest Accelerate cohort yet, 31 digital health companies, split between a 30-week data-access accelerator for 24 companies and a multiyear pathway for seven building AI tools for clinicians. Participants receive de-identified Mayo clinical data and mentorship rather than cash. Mayo Clinic

Forrester: 5 big healthcare predictions for 2027  What Comes Next

Forrester’s Predictions 2027: Healthcare forecasts a consumer-driven ‘identity crisis’ as patients bypass traditional routes:
• Over half of US consumers turn to public AI tools first for health guidance
• AI agents handle 20% of routine care check-ins
• Five of the 10 largest insurers launch alternative health plans
• One in five Americans takes a GLP-1, and direct-to-consumer care spreads
• One in four regularly uses at-home diagnostics, opening a new digital divide
Becker’s Hospital Review