Symptom intake, red-flag screening and acuity routing, built on your protocols and your EHR.
We build AI patient triage software that interviews patients by chat, SMS or voice, applies your licensed nurse triage protocols, escalates red flags to a person immediately, books the right visit, and writes a complete, auditable record to Epic, athenahealth or eClinicalWorks.
Workflow scoping · Evaluation design · Human oversight · Integration planning
Updated September 2026
AI patient triage software takes a patient from "here is what is wrong" to the right level of care. A language model captures the complaint in the patient's words, deterministic red-flag rules screen for emergencies, a licensed protocol such as Schmitt-Thompson sets the disposition and time frame, and the case moves to a nurse, a scheduler or emergency services. Clinicians stay in the loop for every decision that matters, and the full session is written back to the EHR over FHIR.
Every session follows the same governed path. The model handles language; the protocol and the rules decide acuity; a person owns every disposition that leaves the safe defaults.
The session starts on web chat, SMS, the patient portal or a voice line. We match the caller to an EHR Patient record with demographics plus a second factor, capture consent to proceed, and record who is answering: the patient, a parent or a caregiver.
The language model turns free text such as "my chest feels tight when I climb stairs" into structured symptoms, onset, duration and severity. It extracts; it does not decide. Every extracted value is shown back to the patient to confirm.
A deterministic rule layer checks for emergency signs on the first message and again after every answer. A match ends automated questioning and moves straight to escalation.
The complaint maps to one protocol from your licensed set, for example an adult chest pain or pediatric fever protocol. Questions are asked in protocol order, and the first positive answer sets the disposition, exactly as a nurse would work the guideline.
Output is a protocol disposition such as call 911 now, go to the ED now, see a clinician within 4 hours, book within 24 hours, routine appointment, or home care advice. Care advice text comes from the approved protocol content, not from model generation.
Urgent and uncertain cases go to the nurse queue with the full transcript. Routine dispositions go to scheduling with the visit type and time window already set. The patient can ask for a person at any step.
The completed session is written to the EHR as a triage encounter with the question-and-answer record, the reason for contact and any nurse communication, so the next clinician sees what was asked and why.
Acuity logic comes from content your clinical leaders already trust. We encode it; we do not invent it.
The adult and pediatric Schmitt-Thompson protocols are licensed clinical content that your organization brings under its own license. We load them as versioned decision tables, keep the question order and dispositions intact, and give your nurse informaticist a review screen to approve each protocol version before it goes live.
For ED front-door and waiting-room use, the Emergency Severity Index gives the five-level acuity structure (ESI 1 to 5), maintained by the Emergency Nurses Association. The software gathers the inputs for the triage nurse: vital signs from the device feed, expected resource count and danger-zone vitals. The ESI level is confirmed by the nurse.
Specialty practices often run their own pathways, such as post-operative calls, oncology symptom management or behavioral health intake. We model them in the same format with a named clinical owner, an effective date and a review date, so local content is governed the same way as licensed content.
The model interprets free text, handles follow-up phrasing, translates, and summarizes the session for the nurse. Protocol selection is checked against the extracted complaint, and dispositions are computed from the protocol table, so a change of model never changes clinical logic.
Red flags are code, not prompts. They run before the protocol, after every answer and on the final summary, and they are tested on every release.
Chest pain with shortness of breath or sweating, stroke signs (BE-FAST: balance, eyes, face, arm, speech, time), severe difficulty breathing, signs of anaphylaxis, uncontrolled bleeding, and fever of 100.4°F or higher in an infant under 90 days. A match shows the call-911 instruction and alerts the nurse line at the same time.
Statements of suicidal intent or self-harm trigger a warm handoff to your crisis workflow and display the 988 Suicide and Crisis Lifeline. The session transcript goes to the on-call clinician with the triggering statement highlighted.
Contradictory answers, low extraction confidence, a complaint with no matching protocol, pregnancy, and patients your rules mark as high risk all route to the nurse queue instead of an automated disposition. The system never downgrades acuity to fit available capacity.
Escalations land in a nurse worklist or the EHR in-basket with a callback deadline set by disposition. Missed deadlines page the next owner, and after-hours cases route to the on-call schedule.
Triage is only useful when the result lands where clinicians work. We integrate with the EHR's FHIR R4 APIs and fill write gaps with the vendor's own APIs or HL7 v2 feeds.
Search and read to match the caller, pull age, sex and preferred language, and pick the right pediatric or adult protocol set.
A triage encounter (virtual or telephone class) anchors the session, the disposition and any follow-up appointment to one record.
Reads the problem list and pregnancy status for risk rules, and records the reason for contact where the EHR accepts it.
Stores every protocol question, the patient's answer and the resulting disposition as structured data rather than a pasted note.
Carries the nurse summary and the care advice given to the patient, so the message trail sits in the chart next to the encounter.
Epic, athenahealth and eClinicalWorks each expose different FHIR write support. We map each resource to what the target supports, using SMART on FHIR backend services authorization, athenahealth's own API for gaps, and HL7 v2 or document write-back where FHIR writes are not offered.
Epic integrationathenahealth integrationeClinicalWorks integrationFHIR integration services
The disposition sets the visit type, the location type (in person, video, urgent care) and the time window. We search open slots in the EHR scheduling API and book or hold one. If nothing fits the window, the case goes to a scheduler or nurse instead of being booked later than the protocol allows.
Each session stores the protocol and version, every question and answer, each red-flag rule evaluated with its result, the disposition, the model and prompt versions, any nurse override with its reason, and timestamps for every step. A clinical reviewer can replay any session exactly as the patient saw it.
We sign BAAs, encrypt data in transit and at rest, and call language models only through endpoints covered by a BAA with zero data retention for training. Application logs carry session IDs, not symptoms. Nirmitee is ISO 27001:2022 certified.
Role-based access separates nurses, schedulers, clinical reviewers and engineers. Retention follows your medical record policy, and transcripts stored outside the EHR are deleted on the schedule you set.
FDA's Clinical Decision Support Software guidance and the four criteria in section 520(o)(1)(E) of the FD&C Act shape the architecture from day one.
Recommendations beyond the approved protocol dispositions are presented to a nurse or physician, never acted on automatically. The clinician sees the basis for each recommendation, including the protocol, the answers that triggered it and the rule that fired, so they can review it independently rather than rely on it.
What the patient reads is the disposition and care advice text your clinicians approved for that protocol. The model does not generate diagnoses or treatment advice for patients.
We write the intended-use statement, the user population and the criteria mapping into the design file, which gives your regulatory and clinical governance teams a clear record for their classification decision.
Evaluation is built into the delivery plan. These are the measures we track on every release and in production.
A clinician-authored vignette suite covers every red flag, including indirect phrasing, typos and other languages. The release gate is that every red-flag vignette escalates.
Dispositions are compared with a nurse reference disposition on the same vignettes and, in silent mode, on live contacts. Under-triage is reviewed case by case; over-triage is tracked for nurse workload.
Every override captures a reason code. Overrides by protocol show where content, extraction or rules need work.
Completion and abandonment by step, time from escalation to nurse contact, callback deadlines met, and share of routine dispositions that reach a booked appointment.
We confirm channels, protocol licenses, EHR environment access, escalation owners and the red-flag list with your nursing and medical leads.
Protocol tables, rule layer, EHR integration and nurse worklist are built in a non-production EHR environment and tested against the vignette suite.
The system runs alongside your nurses on live contacts without reaching patients. Its dispositions are compared with the nurse's decision before anything goes live.
Go-live starts with one channel and a limited protocol set, with weekly override review and a rollback switch owned by the clinical lead.
AI patient triage software collects a patient's symptoms through chat, SMS, portal or voice, screens for red flags, applies an approved triage protocol to set an acuity and time frame, and hands the case to a nurse, a scheduler or emergency services with the full record written back to the EHR.
No. It removes the repetitive intake work so nurses spend their time on the cases that need clinical judgment. Urgent, uncertain and red-flag cases route to a nurse, every automated disposition is reviewable, and nurses can override any outcome with a recorded reason.
It runs the protocol content your organization approves: licensed Schmitt-Thompson telephone triage protocols that you bring under your license, the Emergency Severity Index for ED intake, and your own local pathways. We encode each one as a versioned decision table signed off by a named clinical owner.
Red flags are deterministic rules that run before the protocol and after every answer. A match stops automated questioning, shows the call-911 or ED instruction, and alerts the nurse line. Every red flag has test vignettes, and a release ships only when each one escalates.
We integrate with Epic, athenahealth and eClinicalWorks through their FHIR R4 APIs, adding the vendor's own APIs or HL7 v2 feeds where a FHIR write is not available. The same design applies to other EHRs with FHIR or HL7 interfaces.
Patient for matching, Encounter for the triage contact, Condition for problem list context and reason for contact, QuestionnaireResponse for the structured question-and-answer record, and Communication for the nurse summary and care advice. Scheduling uses Slot and Appointment where the EHR supports them.
It depends on the intended use. We design to FDA's Clinical Decision Support Software guidance: recommendations go to clinicians who can see the basis and review it independently, and patient-facing output is limited to clinician-approved protocol content. We document the intended use and criteria mapping for your regulatory team's classification decision.
We sign BAAs, encrypt PHI in transit and at rest, use language model endpoints covered by a BAA with no training retention, keep PHI out of application logs, and apply role-based access. Nirmitee is ISO 27001:2022 certified.
We track red-flag sensitivity on a clinician-authored vignette suite, under-triage and over-triage against nurse reference dispositions, nurse override rates with reason codes, abandonment by step and time from escalation to nurse contact. A silent-mode period compares the system with live nurse decisions before patients see it.
Yes, for routine dispositions. The disposition sets the visit type and time window, the software searches open slots through the EHR scheduling API and books or holds one. If no slot fits the window, the case goes to a scheduler or nurse rather than being booked late.
More Agents
Explore complementary solutions for your practice.
Bring your protocol set, EHR and escalation model. We will map the channels, red-flag rules, FHIR resources and evaluation plan for your first go-live.
