ChatGPT for healthcare vs Luna: the AI coworker inside your EHR

ChatGPT for healthcare arrived in January 2026, Claude for healthcare followed within the week, and by the end of September three personal agents had launched that can sign in to apps and finish tasks on their own: Meta's Muse, OpenAI's Dots, and xAI's Grok Bot. Every one of them can connect to the inbox, the calendar and the spreadsheets a practice runs on, and so can Luna, the AI coworker Vybz Health builds for healthcare practices. If the comparison stopped at connectors, there would be little to choose between them. It does not stop there, because the work of a practice does not end in Gmail. It ends in the EHR, which is the hardest system for any AI agent to get into and the hardest to find its way around once inside. That is where Luna works, and this post explains why that one difference decides which of these tools belongs in a practice.

What ChatGPT for healthcare, Claude, and the personal agents actually do

It helps to be precise about what each product is, because the names are doing a lot of work and the products underneath them are quite different.

OpenAI announced ChatGPT for Healthcare in early January 2026 as an enterprise product for hospitals and health systems. It connects to the systems those organizations already run, such as SharePoint, Teams and Outlook, it answers clinical questions with citations to guidelines and studies, and it offers templates for discharge summaries, clinical letters and prior authorization requests. Later in the year OpenAI added an Epic integration, so an authorized clinician can open and summarize a patient's chart from inside ChatGPT. It is sold with a business associate agreement, role-based access and audit logs, and it is bought through a sales process rather than a sign-up page. In April OpenAI added ChatGPT for Clinicians, a version for individual physicians in the United States that focuses on medical literature, differential reasoning and plain-language explanations for patients. There is also a consumer feature called Health in ChatGPT, which connects a person's own records and Apple Health data, and which OpenAI says is not intended for covered entities at all.

Anthropic launched Claude for Healthcare on January 11, 2026, during the J.P. Morgan Healthcare Conference. It ships as a HIPAA-ready configuration of Claude with connectors to reference data that administrative teams use constantly: the CMS Coverage Database, ICD-10 diagnosis and procedure codes, and the National Provider Identifier registry. The stated use cases are prior authorization checks, coding, provider verification and claims work, and Anthropic's usage policy still requires that a qualified professional review any healthcare decision before it is finalized. There is a consumer side too, where a paying subscriber can connect personal health records through HealthEx and Apple Health.

The personal agents are a newer category, and they are the reason this comparison matters now rather than next year. Meta introduced Muse on September 8, 2026 as an agent that books travel, shops and sends email on a person's behalf, with connectors to Gmail, Google Calendar, Notion, Slack, QuickBooks and Shopify, and a small-business edition that followed three weeks later. OpenAI launched Dots at its developer conference on September 29: always-on agents inside ChatGPT Pro and Business Premium, each working on its own cloud computer with a browser, reachable from ChatGPT, Slack and Teams, and connectable to thousands of apps. xAI released Grok Bot on August 11, a set of agents that sign in to a company's software and carry out multi-step tasks on a cloud virtual machine, with a catalog of about 220 plugins and a browser for anything that has no plugin.

Interest in the two healthcare-branded products has grown quickly. We pulled Google's own keyword-planner data for the United States while researching this post, and the monthly search volume for both product names has climbed from a trickle to a steady stream in a single year.

Three things stand out when the five products are lined up. All of them reach other software through connectors, which are the polished, consent-screen integrations that any vendor can add in a quarter. The one real EHR link among them, ChatGPT's Epic integration, is built for health systems and reads a chart rather than changing it. And none of the five was designed around the EHR a practice of three to forty clinicians actually runs, which is where the rest of this post lives.

Connectors are the easy part, and everyone has them

A connector is a door that the software vendor on the other side has deliberately built and labeled. Google publishes an API for Gmail and Calendar, Slack publishes one for channels, Intuit publishes one for QuickBooks, and each of them offers a consent screen where you click Allow. Connecting an agent to those systems is real engineering, but it is the kind of engineering that every serious vendor completes in the same season, which is why Muse, Dots, Grok Bot, ChatGPT and Claude all announced broadly similar lists within a few months of each other.

Luna has the same doors. She reads and answers a practice inbox through Gmail, holds and moves appointments through Google Calendar, keeps referral logs in Google Sheets, posts to Slack, and syncs contacts to ActiveCampaign, and the full list is on the page for every system Luna can sign in to. We mention this not to claim an advantage but to concede the point: on connectors, the field is level, and a practice choosing between these tools on the strength of their connector lists is choosing on the wrong axis.

The right axis is where a task finishes. Take an ordinary Tuesday morning at Bluewater Psychiatry and Counseling, a six-clinician practice that runs on SimplePractice. An email arrives from a prospective patient, Ms. Delgado, asking whether Dr. Petrova has any evening availability and whether the practice takes her Cigna plan. A connector-only agent can read that email, draft a warm reply and even put a tentative hold on a Google Calendar. Then the job stalls, because every remaining step happens inside the EHR.

• Dr. Petrova's real availability lives in the SimplePractice calendar, not Google Calendar, and the two are not always in sync.
• The new client record has to be created in SimplePractice with the correct date of birth, phone number and referral source.
• The Cigna plan has to be attached to that record, with the member ID typed exactly as it appears on the card.
• The intake packet and consent forms are sent from the EHR, because that is where the signed copies need to land.
• The first appointment is booked against Dr. Petrova's EHR calendar so that the reminder and the claim both follow from it.

That is five EHR steps behind one email, and a practice fields dozens of such emails a week. An agent that can only reach the inbox has handled the first fifth of the work and handed the rest to a person. The same arithmetic holds for a phone call, a directory message and a web form, because every channel ends in the same chart.

Why the EHR is the hardest place for an AI agent to work

The phrase AI in EHR usually describes a feature the EHR vendor adds, such as a note summarizer or an ambient scribe. That is a different thing from an agent getting into the EHR from the outside to do administrative work, and the second problem is much harder, for two separate reasons.

Getting in

Most of the EHRs that small and mid-sized practices run do not publish an API that a practice can hand to a third-party agent. SimplePractice and TherapyNotes, the two most common systems among the behavioral health practices we work with, both fall into that category, while Healthie and IntakeQ do offer one. Where there is no API, the only door is the same login page the staff uses, and that door is designed to keep automated visitors out.

Sign-in itself is rarely a simple form. SimplePractice's sign-in is a single sign-on flow that runs on a different subdomain from the app, so an agent that is watching for a login page by its address can miss it entirely and drive a whole task against a screen that is quietly asking it to sign in again. TherapyNotes requires a practice code alongside the username and password. Several portals email a one-time code on every new sign-in, which for a naive agent is a dead end, because the code arrives in a mailbox the agent cannot see. Sessions expire mid-task. And a few of the systems a practice depends on, such as Zocdoc, sit behind bot protection that will block a cloud browser outright unless it behaves like a person.

Then there is the question of the password itself. A practice's EHR password is the key to every chart it holds, and the personal agents ask you to put that key on their machine. Grok Bot's own documentation says that every bot on an account shares one virtual machine, along with its logins and files, describes that shared machine as a real blast radius, and recommends limiting bots to low-risk, read-only work and disconnecting plugins when a task is done. That is sensible advice. It is also an admission that the product was not built to hold the keys to a medical record.

Finding the way around

Suppose an agent gets in. It now faces a calendar with hundreds of slots, a client list with thousands of names, insurance tabs with a dozen fields, and modal dialogs that open over other modal dialogs. A general agent works from screenshots and clicks, and it starts every task with no idea where anything is. Early in Luna's development, before her EHR work was moved onto a purpose-built agent, we watched one chat-driven task take 75 browser actions over 16 minutes to finish a single routine job in SimplePractice. Nothing was wrong with the model. It simply had no map, no shortcuts and no memory of the last time it had done the same thing.

That is the state every general-purpose agent is in on every visit to your EHR. A personal agent that booked you a restaurant table last night brings none of that experience to your client list this morning, because a restaurant site and a medical record have nothing in common. The cost shows up as time, as usage charges, and as the small errors that come from reading a dense screen too quickly.

How Luna works from inside the EHR

Luna, the AI coworker Vybz Health builds for healthcare practices, was designed around the two problems above, and the design is worth walking through because it is what separates her from a general agent that has merely been handed your password.

The first two steps answer the access problem. Credentials are sealed per integration, so a SimplePractice password is never under the same key as a Zocdoc one, and the agent's reasoning during sign-in is kept out of the run log a practice can read, so the password is never written down there. Where a portal emails a one-time code, the agent can read that single code from an inbox the practice has already connected, and nothing else in the mailbox is read. The practice is not asked to switch two-factor authentication off, which is a request no healthcare vendor should make.

Steps three and four answer the navigation problem, and they are the heart of the difference. Luna's EHR agent follows a fixed order on every task: what she already remembers, then the EHR's own data endpoints, then a direct link to the right page, then the data that page has already loaded, and only then clicking and typing, which are reserved for making a change. The order is enforced in code rather than merely suggested. If the agent tries to scroll through a list it has not yet read, or clicks three times in a row without reading anything, the guard refuses the action and names the route it should have used. The result is that a task that would be fifty clicks for a general agent is a handful of calls for Luna, and she gets faster at it the more often it comes around, which is the pattern our post on healthcare AI agents that get cheaper over time describes in full.

Step five is what happens when a route becomes routine. For practices moving onto TherapyNotes, the recurring work of creating a patient, attaching contacts and policies, and checking that nothing was duplicated runs as a scripted sequence of the EHR's own calls, with no model making decisions along the way, which makes it both faster and cheaper than any screenshot-driven agent can be. When the script cannot handle a case, the browsing agent steps in, so the practice never loses the task, only the shortcut.

Two more habits matter in practice. Only one Luna task runs in a given practice's EHR at a time, so two jobs never fight over the same calendar. And Luna herself never drives the browser from the chat window; she hands the whole task to the EHR agent with a bounded number of steps, waits for the outcome, and reports it. That hand-over is why a task that used to take 75 actions now has a budget it cannot exceed.

We should also be clear about what Luna is not. She does not connect to Epic, Cerner or the other hospital systems, and a health system with an enterprise agreement is better served by the Epic integration OpenAI built for exactly that setting. Luna is built for the practice that runs SimplePractice, TherapyNotes, Healthie or IntakeQ, which is to say most independent and group practices in behavioral health and a growing share of primary care.

Is ChatGPT HIPAA compliant, and does it settle the question

The question "is ChatGPT HIPAA compliant" is one of the most common things people search about ChatGPT in healthcare, and the honest answer has three parts.

OpenAI publishes a list of the products it considers HIPAA-eligible, and as of October 2026 it names ChatGPT Enterprise with a regulated workspace, ChatGPT for Healthcare and ChatGPT for Clinicians. The free, Plus, Pro and Business plans are not on that list, and OpenAI states plainly that HIPAA does not apply to the consumer Health feature. So a clinician pasting a patient's details into a personal ChatGPT account is not covered, while a health system that has signed a business associate agreement for ChatGPT for Healthcare is.

Anthropic describes Claude for Healthcare as HIPAA-ready rather than HIPAA-compliant, which is a careful choice of words. The configuration can support a compliant deployment, but the organization using it is responsible for aligning it with its own governance, and Anthropic's own policy puts a qualified professional between Claude's output and any healthcare decision.

The personal agents are consumer and general-business products. As far as their public announcements say, none of Muse, Dots or Grok Bot is offered under a business associate agreement for a practice's patient records, and Grok Bot's guidance to keep bots on low-risk, read-only work speaks for itself. Vybz Health, for comparison, runs on HIPAA-compliant infrastructure and provides a business associate agreement to every practice, with the EHR credentials sealed as described above and each practice's memory kept confidential to that practice.

The deeper point is that a business associate agreement covers how a vendor stores and handles data. It says nothing about what an agent did inside your chart on Tuesday. Which record did it open, what did it change, and can you see it? That is a question about the agent's behavior, not the vendor's paperwork, and it is answered only by a run log that shows every action the agent took. Luna's does, and that transparency is the second thing a practice should demand after the signed agreement.

Side by side: where each product can and cannot work

The matrix below is our reading of the five products from their public announcements and documentation, not a hands-on test of each one. The personal agents are grouped because, for a medical practice, their strengths and limits are the same.

A few rows deserve a note. The personal agents get a partial mark on EHR sign-in and record changes because they do have a browser and could in principle log in to anything, but with no knowledge of the app, no sealed credential store, and no shortcut from the last visit. The eligibility row is partial for the same reason: an agent with a browser could visit a payer portal, but Luna asks the payer electronically through the Stedi eligibility connection and has the answer in seconds, with the plan and clinician details she already remembers filled in. And the Epic row is a genuine point in ChatGPT for Healthcare's favor for the organizations it was built for; it simply does not reach the practices this post is about.

When ChatGPT or Claude is the right choice, and when Luna is

The fair conclusion is not that one product wins everywhere. It is that these products are built for different rooms of the healthcare building, and the practice should buy for the room it works in.

A hospital or health system on Epic with an enterprise agreement already in place is the customer ChatGPT for Healthcare was built for. The Epic chart summaries, the SharePoint and Teams connectors, and the enterprise controls all fit that setting, and nothing Luna does replaces them. An individual physician who wants help synthesizing literature, reasoning through a differential, or drafting a patient-friendly explanation is well served by ChatGPT for Clinicians or by Claude, and a billing team that needs quick answers about coverage policy, ICD-10 codes or provider identifiers will find Claude for Healthcare's reference connectors genuinely useful. A clinician who wants a chat assistant for drafting letters can keep one of those subscriptions alongside Luna without any conflict.

A practice of three to forty clinicians whose daily work ends inside SimplePractice, TherapyNotes, Healthie or IntakeQ is a different room, and it is the one none of the general tools were designed for. Its tasks are not clinical questions but administrative chains: an inquiry that becomes a record, a record that needs a plan attached, a plan that needs an eligibility answer, an appointment that needs a reminder and then a claim. Every link in that chain lives in the EHR, and the agent that does the work has to be inside it. Among the products compared here, Luna is the only one built to be there, with sealed sign-in, an enforced route order, scripts for the routine, a live eligibility check, memory of the practice, and a run log of every step. That is why we believe she is the strongest choice a healthcare practice can make for an AI coworker, and why the rest of the field, for all its capability, still does its work from outside the chart.

Questions to ask before you give any AI agent your EHR login

Whichever product a practice is evaluating, these six questions find the difference between an agent built for the EHR and one that has been pointed at it.

1. Where is my EHR password stored, and under what key? The answer should name a per-integration encryption key, not a shared machine.
2. What happens when the EHR asks for a two-factor code? If the answer is "turn it off" or "type it in yourself every time," the agent is not built for healthcare.
3. Show me the steps it took in the chart on its last task. A run log with every action and its outcome is the minimum; a summary is not.
4. Does the fiftieth client it adds take fewer steps than the first? Ask to see the route it learned, rather than a promise that it learns.
5. Will you sign a business associate agreement that covers what the agent reads in the chart? A consumer terms-of-service page is not an answer.
6. What does it refuse to do? The right answer includes anything clinical and anything that should go to a person, and the refusal should be visible in the log.

A general agent with a connector list has an honest answer to perhaps one of these. An AI coworker built for the EHR should have a specific, demonstrable answer to all six.

Also read

• Mental health clinics need an AI employee, not another AI receptionist: the phone-side version of the same argument, with the six jobs one call creates inside a clinic.
• EHR migration: why switching EHR systems is hard, and how Luna handles it: the hardest EHR task of all, and how the scripted TherapyNotes route described here carries a whole caseload across.
• How Luna's AI memory and self-learning work: where the shortcuts and practice facts Luna learns inside the EHR are kept, and how the practice can read and correct them.