What AI Tools Does Epic Build Into Its EHR?

Epic builds AI assistants directly into its EHR, including Art for clinicians, Emmie for patients, and Penny for revenue cycle and operations. It also offers ambient note drafting (“AI Charting”), publishes research on models trained on its Cosmos dataset, and sells a sepsis prediction model. Most performance numbers come from Epic or its customers. The independent peer-reviewed studies I found cover only a few features, and their results are mixed. This article reflects sources checked on 2026-10-10, and Epic’s product list changes often.

The named assistants (vendor descriptions)

• Art (clinicians). Epic’s AI page describes it as “Epic’s AI for clinicians.” In its February 4, 2026 announcement, Epic said Art’s Insights feature was used more than 16 million times a month.
• Emmie (patients). According to Epic, Emmie is a built-in assistant that “answers patients’ questions, surfaces information from their charts, and takes action for them” through MyChart. It covers scheduling, plain-language explanations of results, and billing help. Epic’s February 2026 post says it is also available by text message.
• Penny (revenue cycle and operations). Epic’s February 2026 post says over 200 organizations use Penny for professional billing coding. It also says some organizations report more than a 20% reduction in coding-related denials, and that appeal letters for medical necessity denials are created 23% faster. These are Epic’s figures, not independent measurements.

Ambient documentation: AI Charting and Microsoft

Epic’s own announcement says it released AI Charting, “a built-in feature that listens during patient visits, drafts the clinician’s note, and queues up orders based on the conversation.” That post does not mention Microsoft, so I could not confirm from Epic what technology powers it.

Separately, Microsoft’s Epic partnership page describes its Dragon technology as integrated with Epic and built on “conversational, ambient and generative AI.” The Microsoft page gives no performance figures, and I did not find a statement linking it to Epic’s AI Charting.

Cosmos-based models

Epic Cosmos is a dataset built with a community of health systems that use Epic. A preprint on arXiv (2508.12104) by authors including Epic staff and Microsoft Research collaborators describes generative medical event models trained on 115 billion medical events from 118 million patients. The paper describes Cosmos itself as holding more than 300 million patient records across 310 health systems, which is larger than the training sample. The authors report that the models “generally outperformed or matched task-specific supervised models” on 78 clinical tasks without fine-tuning. This is a preprint, not a peer-reviewed result, and the authors are the developers. The version cited here (v3, November 2025) calls the model family “Curiosity.”

Drafted replies to patient messages

Independent academic groups have studied EHR-integrated generative AI drafts for patient in-basket messages. A July 2024 JAMA Network Open study had 16 primary care physicians review 344 messages. They rated the AI drafts higher on communication quality (3.70 versus 3.38), but this is a rating of style, not proof that the drafts were more accurate; the accuracy difference was not statistically significant. The authors also found the drafts were more linguistically complex, which raises equity concerns for patients with limited health or English literacy.

Sepsis prediction: two independent evaluations

The Epic Sepsis Model is one feature with published independent evaluations. A 2021 JAMA Internal Medicine external validation at Michigan Medicine covered 38,455 hospitalizations. It found sensitivity of 33%, specificity of 83% (at the alert threshold studied), and an AUC of 0.63, against developer-reported performance of 0.76 to 0.83 as cited by the authors. The model alerted on 18% of all hospitalized patients and missed about two-thirds of sepsis cases.

Epic later released an updated version. A multicenter prospective study published February 27, 2026 in JAMA Network Open covered 227,091 inpatient encounters at four health systems. It found AUROC between 0.82 and 0.92 and positive predictive value between 0.13 and 0.26. The authors concluded the updated model showed “improved discrimination” but also “high institutional variability, low positive predictive value, and high alert burden,” and advised local validation.

Reported results: vendor and customer claims

Epic’s March 10, 2026 post reports that more than 85% of its customers use Epic AI. Its other figures are customer-reported on Epic’s own site: Riverside Health clinicians using Art’s inpatient insights spend up to 32% less time on documentation and communication, and Summit Health cut medication prior authorization submission time by 42%, according to Epic. Rush University Medical Center saw a 58% reduction in billing-related messages with Emmie, according to Epic. These numbers have not been independently verified. I found no peer-reviewed evaluation of Art, Emmie or Penny on the pages I checked.

How to read all this

• Treat usage counts and percent improvements on epic.com as vendor-reported, even when attributed to a customer.
• The peer-reviewed studies reviewed here support a cautious reading (this is not a systematic review of all the literature): draft replies can be good but complex, and the sepsis model needs local validation.
• Check Epic’s AI page for the current feature list.