HugoScore hugoscore.org

Full review

Epic Emmie (incl. Ask Emmie) CAIHL draft report

Evidence-linked HugoScore draft report for a health AI tool that affects patients.

Back to profile Markdown source

HugoScore CAIHL Draft Report: Epic Emmie (including Ask Emmie)

  • Status: Draft for human review
  • Last reviewed: 2026-07-21
  • Review method: Deep public-source review of current Epic product, newsroom, and privacy pages, current Sutter Health and Reid Health deployment materials, Reid's generative-AI policy, trade reporting and a product-design interview, a June 2026 peer-reviewed study of a closely related pre-release Epic MyChart assistant, and current HHS and California governance context. No authenticated MyChart access, Epic UserWeb documentation, implementation contract, production-model or data-flow inspection, transcript-retention test, patient or care-team interview, accessibility audit, security audit, or independent production validation.
  • Service: Epic Emmie, including Ask Emmie
  • Vendor: Epic Systems Corporation
  • Category: Patient portal AI assistant
  • CAIHL classification: Institutional patient-facing AI embedded in MyChart and deployed by health systems
  • Agency posture: Mixed, institution-led and directly patient-facing
  • Agency axis: 38 of 100, where the axis indicates control and agency orientation rather than overall product quality or safety

Bottom Line

Emmie is not one uniform chatbot. It is Epic's expanding family of patient-facing AI functions inside MyChart and related text workflows. Epic's current catalog includes record-grounded questions, plain-language summaries of notes and results, scheduling, billing, payments, reminders, and other actions. A patient does not receive all of those functions merely because their health system uses Epic. Each health system decides what to enable, for whom, and under what workflow.

That deployment variation is central to the CAIHL assessment. Sutter Health currently describes Ask Emmie as a signed-in assistant that can answer general health questions, summarize information already visible to the patient, and help with portal navigation. Sutter says it cannot schedule or cancel appointments, message the care team, diagnose, make care decisions, or change the medical record. Reid Health separately describes text scheduling, billing help, imaging-result explanation, and a symptom checker for eligible primary-care patients. Some Reid functions require the patient to enable Preview Features. Reid described the initial release as English-only and multilingual messaging as future work.

Emmie can expand practical agency. A patient may understand a result sooner, prepare a better question, reschedule without waiting on the phone, or make sense of a bill. Its location inside the health system's portal also gives it access to a more complete institutional record than a consumer chatbot would usually have. Yet the system remains institution-led. Epic and the deploying health system determine the model, access, configuration, guardrails, data handling, and success measures. Patient design input is documented, but patient governance and control are not.

The evidence is strongest for product scope, deployment differences, and early usability. It is weakest where the stakes are highest. Public sources do not identify the current production model provider or inference host, chat retention, deletion or export controls, training or secondary use, subgroup performance, or a patient correction and appeal process. A June 2026 peer-reviewed study found strong usability for a closely related pre-release Epic MyChart assistant, but it was not a production Emmie evaluation and did not systematically review outputs for errors or unsafe recommendations. No independent product-specific production study of accuracy, safety, equity, or patient outcomes was identified.

Who Does This AI Serve?

Emmie directly serves patients in use, but Epic and health systems govern it.

Epic markets Emmie partly through institutional outcomes. Its current materials cite a 58 percent reduction in billing-related service messages at Rush, 14,900 appointments rescheduled at Ochsner with an estimated 750 staff hours saved, and 94 percent patient satisfaction with AI reminders. These may also benefit patients. Fewer calls and faster self-service can reduce access friction. The measures still reflect organizational definitions of success more clearly than patient-defined outcomes such as improved understanding, successful correction of an error, reduced anxiety, equitable access, or better decisions.

The related design interview describes input from individual patients, patient advisory councils, patient and family advisory councils, and safety councils. That is meaningful evidence of patient input. It is not evidence that patients have governing authority over deployment, model selection, retention, safety thresholds, or future expansion of chat content into clinical workflows.

CAIHL answer: Institution-led, with direct patient utility.

What Emmie Can Do, and Where

| Public source | Functions described | Important boundary |
| --- | --- | --- |
| Epic product catalog | Scheduling and rescheduling, reminders, plain-language endoscopy, imaging, and after-visit summaries, billing questions and payments, record-grounded questions | Catalog scope does not establish that every health system enables every function |
| Sutter Health | General health questions, summaries of notes, results, visits, and messages, follow-up questions, portal navigation, authorized proxy access | No scheduling or cancellation, care-team messaging, diagnosis, personalized care advice, care decisions, or record writeback |
| Reid Health | Record questions, imaging-result explanation, opt-in text scheduling, billing help, and symptom routing for eligible primary-care patients | Some features require Preview Features, the initial release was English-only, and capabilities vary by patient eligibility |

This variation means that claims about "Emmie" should always identify the source and deployment. The Epic catalog is evidence of offered capability, not proof of a function being live for every patient. The Sutter and Reid pages are evidence about those deployments, not universal product rules.

Visibility

Sutter and Reid clearly identify Ask Emmie as generative AI and provide patient-facing explanations. That is better than an invisible algorithm acting behind a portal page.

The disclosure is not yet complete across the broader product family. Epic also describes immediate AI-generated imaging summaries, AI-rewritten after-visit summaries, reminders, and other functions. The public materials reviewed do not establish how each generated output is labeled at the moment a patient sees it, whether the label persists when content is copied or printed, or whether a patient can tell which parts came from source records and which parts were generated.

California law provides relevant context for some deployments. Assembly Bill 3030 requires specified generative-AI patient clinical communications to include an AI disclaimer and clear instructions for contacting a human provider, subject to exceptions including human review. Scheduling and billing communications are excluded from that defined patient clinical information. This report does not determine whether any Emmie deployment is or is not compliant.

CAIHL answer: Partial, feature- and deployment-dependent.

Meaningful Choice

Choice exists at the level of opening the assistant, but not at the level of selecting the platform or its governance.

Reid documents patient opt-in to text messaging and Preview Feature toggles for some functions. Sutter's current FAQ says Emmie is available to all patients after sign-in and does not document a separate Emmie opt-out. A patient can avoid opening the chat, but the health system decides whether the assistant exists, which functions appear, and what alternatives remain available.

The current sources do not answer whether patients can decline automatically presented AI summaries or reminders, whether declining changes access to ordinary portal functions, or whether a choice follows the patient across proxy, family, and linked-account contexts.

CAIHL answer: Partial and deployment-dependent.

Contestability and Correction

Sutter provides thumbs-up and thumbs-down feedback. It also tells patients to contact their provider when they have health concerns or disagree with information. Those are feedback and escalation paths, not a documented correction or appeal system.

No reviewed source explains whether a patient can edit an answer, obtain an audit trail, export or delete a transcript, require review of a disputed output, learn the result of an investigation, or correct the assistant's future behavior. Sutter says chats are not part of the legal medical record and do not update the chart. That boundary limits accidental chart writeback, but it also means a patient who spots a chart-grounded error still needs a separate process to correct the source record.

CAIHL answer: Partial feedback, weak correction.

Action Support

Emmie's most credible agency benefit is reducing the distance between a patient's question and a useful next action. The product catalog describes scheduling, bill explanation, payments, payment plans, reminders, and record-grounded explanations. These are not merely informational. They can affect access, money, preparation, and follow-through.

The clinical stakes vary. Sutter draws a relatively firm public boundary around education and summary. Reid's symptom checker may recommend self-care, an appointment, urgent care, or emergency care. Epic also promotes immediate imaging-result summaries. These functions can help patients act sooner, but they also make calibration, uncertainty, escalation, and labeling more consequential.

CAIHL answer: Yes, with variable scope and safety boundaries.

Data Governance

Emmie operates within a covered-care context rather than as a stand-alone consumer chatbot. Sutter says its Ask Emmie chat uses information already visible to the patient, cannot access non-Sutter Care Everywhere information, is not part of the legal medical record, and cannot be seen by providers. Sutter also says authorized proxies can use Emmie within their existing permissions.

Those statements answer only part of the data-governance question. Public sources reviewed for this profile do not identify:

  • the current production model or models
  • the production inference host and downstream processors
  • whether configurations differ by health system
  • how long chat prompts, responses, feedback, and safety logs are retained
  • whether patients can export or delete chat data
  • whether chat content is used for model improvement, evaluation, or another secondary purpose
  • whether future visit-agenda or intake functions would change provider visibility or medical-record status

Epic's mobile-app privacy statement says MyChart connects to servers and systems operated and maintained by Epic customers, and that the mobile app does not receive or store copies of transmitted health and fitness data. That does not resolve server-side Emmie processing or retention. It also does not support a blanket claim that Emmie is Epic-hosted.

The related peer-reviewed pre-release study used GPT-4.1-2025-04-14 and stated that study data were not used to train or validate the system. Those facts belong to the study configuration. They should not be projected onto current production Emmie without a production disclosure.

HIPAA is relevant because the assistant is offered through a healthcare organization. HHS explains that when an app is provided on behalf of a covered entity, HIPAA responsibilities can apply through the covered entity and its business-associate relationships. That context is a meaningful protection compared with many consumer tools, but it is not a substitute for product-specific disclosure, and this report makes no legal determination.

CAIHL answer: Covered-care context, important product details undisclosed.

Clinical Boundaries and Safety

Sutter warns that generative AI can make mistakes. It says Emmie provides general education, not personalized medical advice, diagnosis, or care decisions, and does not replace the care team. It directs patients to their provider for health concerns. These are clear public warnings.

The practical boundary is less simple. An answer grounded in a patient's record can feel personalized even if it is labeled educational. Immediate imaging summaries may shape a patient's interpretation before a clinician provides context. Reid's symptom routing can direct a patient among self-care, routine appointments, urgent care, and emergency care. The more the assistant recommends or triggers action, the less adequate a disclaimer becomes as the sole safety control.

The reviewed public record does not establish production performance for hallucination, omission, uncertainty calibration, conflicting records, medication or allergy errors, high-risk symptoms, pediatric use, caregiver questions, mental-health crises, or subgroup differences. Nor does it disclose production safety thresholds, incident counts, rollback criteria, or patient-notification practices after an error.

CAIHL answer: Clear warnings, expanding practical stakes.

Evidence and Evaluation

Product and deployment evidence

Current Epic, Sutter, and Reid pages provide strong evidence that Emmie exists, has expanding functions, and varies by deployment. Epic's operational figures are useful but are vendor-reported and focus on service-message reduction, scheduling volume, saved staff time, and satisfaction.

Reid says it plans to track usage, navigation, and friction during early adoption. Its general generative-AI policy describes security risk review, accountable owners, human confirmation where harm could result, labeling of AI content, restrictions on data and functions, and annual review. The policy is evidence of an institutional governance framework, not proof of how every Emmie function implements each control. Its broad language about de-identification also does not explain how an identifiable, chart-grounded patient assistant is processed.

Peer-reviewed evidence

A June 2026 JAMIA Open article reported a prospective usability study of a patient-facing AI assistant integrated into a non-production Epic MyChart environment. The paper does not name the product as Emmie. This review therefore treats it as a closely related pre-release system, not as production Emmie validation.

The study included 131 adults using their own data. It reported a mean System Usability Scale score of 81, 76.9 percent likelihood of reuse, and a Net Promoter Score of 43.9. These are encouraging usability findings.

The limitations are substantial for a safety claim. Participants came from one health system and were a selected cohort. The mean age was 60 and 82.4 percent were college educated. The study was non-production. Outputs were not logged and systematically reviewed for errors or unsafe recommendations beyond initial research review. Epic employees were among the authors. The paper reported no conflicts and acknowledged AHRQ grant support. These facts do not invalidate the study, but they constrain what it can establish.

The paper describes a broader vendor testing pipeline using simulated personas, automated grading, and periodic human audits. That is useful process evidence. Public results from that production pipeline, including error rates, safety failures, subgroup performance, and external audit findings, were not identified.

Evidence not found

The review did not identify an Emmie-named peer-reviewed or independently conducted production evaluation of answer accuracy, safety, equity, clinical outcomes, financial outcomes for patients, or access outcomes. Absence from this search is not proof that no such evaluation exists. Restricted Epic UserWeb materials, nonpublic customer evidence, and unpublished internal studies were not accessible.

CAIHL answer: Mostly vendor- and institution-defined, with patient design input and encouraging but limited pre-release usability evidence.

Equity and Access

Embedding AI in MyChart can place assistance where many patients already manage care. Text scheduling may reduce portal and phone burden. Sutter's proxy support can help parents, caregivers, and others who already have authorized access.

The same delivery model can reproduce existing portal inequities. Availability depends on the health system, the MyChart account, digital access, enabled features, and local policy. Reid described its May 2026 launch as English-only and 14-language messaging as future work. Public evidence does not establish disability access, low-literacy performance, broadband and device burden, support for people with limited portal experience, or safety and usability across racial, language, disability, age, and socioeconomic groups.

The pre-release usability study should not be treated as representative evidence because its participants were selected and highly educated. A system intended to simplify health information needs evaluation with the patients most likely to face access and comprehension barriers.

Structured HugoScore Profile

| Measure | Finding | Evidence strength |
| --- | --- | --- |
| Institutional or patient-directed | Institutional patient-facing | High |
| Primary user | Patients and authorized proxies at enabled health systems | High |
| Hosting and control | Epic-built, health-system-configured, production model and inference host undisclosed | Medium |
| Service alignment | Institution-led with direct patient utility | High |
| AI visibility | Partial and feature-dependent | Medium |
| Meaningful choice | Partial and deployment-dependent | Medium |
| Contestability | Feedback exists, correction and appeal not documented | Medium |
| Action support | Strong but variable | High |
| Data governance | Covered-care context, material details undisclosed | Medium |
| Clinical boundaries | Clear warnings, higher-stakes functions expanding | Medium |
| Equity | Portal and deployment dependence, limited subgroup evidence | Medium-low |
| Evaluation ownership | Mostly vendor and institution-defined | High |
| Independent production validation | Not identified | Medium |
| Overall confidence | Medium draft | Medium |

Unknowns for Human Review

  • Which production foundation model or models does Emmie use now, and can each health system choose among them?
  • Where does inference occur, and which entities receive identifiable prompts, record context, responses, feedback, or safety logs?
  • What are the retention, export, deletion, training, and secondary-use rules for chat content?
  • How is AI authorship disclosed for each generated summary, reminder, and action, not only for the named chat interface?
  • Can patients decline automatic AI-generated content without losing ordinary portal access?
  • What happens after a patient reports a wrong, unsafe, or misleading answer?
  • Can a patient see whether feedback led to review, correction, or system change?
  • How are urgent symptoms, mental-health crises, pediatric questions, pregnancy, medication risk, and conflicting chart data handled?
  • What production error, escalation, incident, and subgroup-performance data exist?
  • How do proxy access, adolescent confidentiality, sensitive notes, and linked records affect Emmie context and visibility?
  • Which functions are available by deployment, language, disability access need, and device type?
  • If future chat content feeds visit agendas or intake, what new consent, visibility, record, and correction rules apply?

Risks

  • Patients may over-trust a fluent answer because it is grounded in their record.
  • The institutional chart may contain errors or omissions that the assistant repeats with added authority.
  • Immediate result summaries may shape interpretation before a clinician provides context.
  • Symptom routing may have high consequences if uncertainty or escalation is wrong.
  • Patients may not recognize AI involvement outside the named chat interface.
  • Undisclosed retention or secondary use may undermine candid questioning.
  • Deployment differences can produce unequal access and inconsistent protections.
  • Patient feedback may improve a product without giving the patient a remedy for the specific error they experienced.
  • Future reuse of conversations for intake or visit agendas could change the privacy boundary.
  • MyChart integration may deepen institutional and vendor lock-in even when it improves convenience.

Patient Agency Summary

Emmie can give patients something they often lack: a usable way to interrogate the institution's own record and act without waiting for a phone call. That is a genuine agency gain. It may be especially valuable when it turns a confusing note, result, bill, or scheduling barrier into a comprehensible next step.

The deeper question is who controls the conditions of that help. Patients do not choose the platform, the model, the available functions, the retained data, or the safety thresholds. They can provide feedback, but no public correction or appeal right is documented. Patient input helped shape the interface, but patient governance is not visible. Emmie therefore sits in a mixed position: directly useful to patients, institutionally controlled, and not yet transparent enough for stronger confidence.

Publication Recommendation

Publish as a medium-confidence, AI-assisted public-source draft for human review. Do not present the profile as verified or treat the related pre-release usability study as production Emmie validation.

Human review should prioritize a live, deployment-specific walkthrough at Sutter and Reid, screenshots of disclosures and choices, direct questions about models and data retention, testing of correction and escalation, and review by patients who use proxy access, have limited English proficiency, use accessibility technology, or have lower portal familiarity.

Sources