Stanford ChatEHR reaches 1,450 clinicians inside its Epic record
Stanford ChatEHR, the chart chatbot Stanford Health Care built inside its Epic electronic health record, was used by 1,450 of roughly 7,000 eligible clinicians as of mid-January 2026, STAT reported on January 28, 2026. The tool reads a patient's full record and answers questions about it. The lab behind it says its automated reviews cut tasks such as infection surveillance chart review by 40%. For hospitals, it shows that a team inside IT can turn hours of chart reading into short lists of cases staff can act on.
Why clinicians needed a faster way through the chart
Reading a patient's record before a visit is slow because the information sits in many places. STAT described how Penn Medicine's chief health information officer, cardiologist Srinath Adusumalli, prepares the day before his appointments by working through prior visits, labs, imaging and scanned documents from other hospitals, each in a separate tab of the electronic health record.
Stanford Health Care faced the same work at scale. Its own lab gives one example: a screening task that required a chart review of 5 to 20 minutes across 150 patient charts.
What Stanford built instead of buying a tool
Stanford built ChatEHR itself, connected it to its Epic electronic health record and rolled it out across the health system in September 2025. The Shah Lab, which documents the project, calls it a vendor-agnostic, internally governed platform that places a patient's full longitudinal record, up to twenty million tokens, directly into a model's context window in real time.
The team behind it is a data science group embedded in the IT organization: software engineers, product managers and data scientists with direct access to academic researchers. The lab dates the work to a sandbox in April 2023. It became ChatEHR in November 2023, went through several prototypes by mid-2024, entered a pilot in early 2025 and reached 150 clinical users by August 2025.
How ChatEHR works during a clinician's day
A clinician opens a ChatEHR tab inside the record and first chooses which data to load for the model, by data source and time range. A chat window follows, where the user asks questions, rates answers with a thumbs up or down, or returns to the menu to start a new session.
That chat accounts for 40% of use. The other 60% runs through APIs that automate high-volume, repetitive tasks.
"Automations are fixed prompts applied to defined sets of patients." — Stanford's Shah Lab, on how ChatEHR runs its automated reviews
The lab lists seven operational workflows built this way, from placing patients in the right care setting and strengthening infection surveillance to validating billing accounts.
The time savings Stanford reports from automation
The lab gives figures for each workflow, and they are its own measurements. Patient screening turns the review of 150 charts into roughly 30 highly probable cases. Surveillance of surgical site infections cuts chart review by 40% and, the lab says, makes infection reporting more timely.
At the Contact Center, referral triage takes 17.7% less time, throughput rises 17.8% and the backlog fell from five days to two. A screen for inpatient hospice reads 600 patient charts to find 10–40 eligible cases, which the lab says avoids 18–20 hours of work each day. In orthopedics, referral triage processes more than 4,700 documents and saves at least 1.7 staff hours a day. Review of accounts held by an LoA stop bill saves 10–15 minutes per account. A prospective study by Wang and colleagues in JAMA Network Open examined ChatEHR in surgical patient triage.
How adoption grew after the rollout
STAT reported that as of mid-January 2026, ChatEHR was used by 1,450 clinicians out of roughly 7,000 eligible users. The interface pilot that ended in June 2025 had 150 users and 3,000 chat sessions. In the first 3 months after the enterprise rollout, 1,075 users completed the training video and ran more than 23,000 sessions, with more than 19 billion tokens processed. Users in that period included 180 APPs, 424 physicians, 151 residents and 60 fellows.
The lab's later count, as of July 2026, is more than 3,400 active users, more than 150,000 sessions and 140 billion tokens processed.
What hospital-built AI tools mean for other health systems
Stanford is not alone. In January 2026, Penn Medicine introduced Chart Hero, a similar tool that lets clinicians query and summarize a patient's record through a chatbot-like interface, and around 70 Penn clinicians were testing it. Teams at Duke Health and the Children's Hospital of Philadelphia are developing similar applications.
For a hospital, the case shows where the measured savings sit so far: referral queues, infection reporting, hospice screening and billing review, rather than the chat window alone. Stanford's approach also depends on a team inside IT that can build and govern the tool, and on a platform that is not tied to a single AI vendor.
Questions and answers
What is Stanford ChatEHR?
It is an internally built tool that lets Stanford Health Care clinicians query a patient's full record inside the Epic electronic health record. It also runs fixed prompts across defined groups of patients to automate chart review.
How many clinicians use ChatEHR?
STAT reported 1,450 clinicians out of roughly 7,000 eligible users as of mid-January 2026. Stanford's lab reported more than 3,400 active users as of July 2026.
Does ChatEHR save staff time?
The lab says it does in specific workflows: referral triage time fell 17.7%, chart review for surgical site infections fell 40%, and orthopedic referral triage saves at least 1.7 staff hours a day.
MOZGII view
How we would build it
Where AI helps most in a hospital today is the reading nobody has time for: chart review before a visit, screening lists, referral queues and billing holds. Stanford's numbers sit in exactly those places. If we were building a system like ChatEHR, we would start with one queue that has a clear owner and a clear end state, such as referral triage, and only then add a chat window. The automations, not the chat, are where the hours come back.
The system would need three things. First, the full patient record pulled from the EHR in real time, with the data sources and time range recorded for every answer. Second, a model that can be swapped, because the record is the asset and the model is not. Third, a named nurse, coder or physician who signs off each short list before anyone acts on it. It breaks when the record is incomplete, for example a scanned outside document the model cannot read, and when a fixed prompt drifts away from how the clinic now works.
The one metric we would watch is how many of the cases the automation flags turn out to be real when a person checks them. A falling share means the prompt or the data has changed. Next, we would add a feedback loop that sends every thumbs-down to the team that owns the prompt, so each workflow gets reviewed on a schedule rather than after a complaint.
Sources
Where the facts in this story come from.