Claude Community New YorkClaude Impact Lab for Digital Healthwith ElevenLabs · Alterra Labs and Claude Community New York

Referral Tracker

Every handoff of a specialist referral writes to one shared record. The patient sees a plain status, each office sees its work queue, and an AI agent drafts the paperwork and flags stalls while people make every decision.

Referral trackerFollow one referral through six stages as the patient, the primary care office and the specialist office. Demo scenario: type 2 diabetes to endocrinology. Referral note drafterPick a patient and a specialty, see the chart extract a specialist needs, and draft the referral note.

Live AI demo, by arrangement

This version runs the real model inside a mock record system on a teammate’s computer, so it is only up when we switch it on. To see it, coordinate with us first: message any of us on LinkedIn (links under Team 3 below) and we will set a time.

  1. Open the live link. If a notice from ngrok appears first, choose “Visit Site”.
  2. Click Sign in as Dana.
  3. Under Have Fable do it, leave the specialty as Endocrinology.
  4. Optionally type a question, such as “Next step in glucose-lowering treatment?”
  5. Press Run Fable.

Team 3

The problem

Once a referral leaves the primary care office, nobody can see where it is.

The primary care office sends the referral and loses sight of it. The specialist office receives a packet that may lack what it needs to triage. The patient waits without knowing whether anything is happening. When a referral stalls, nothing tells anyone.

34.8% of 103,737 referral scheduling attempts in one large health system ended in a documented completed appointment. Source: Patel et al., Journal of General Internal Medicine, 2018 (24 primary care sites, 2015–16). One setting; it does not measure Referral Tracker.

How it works

Six stages, one shared record

Every handoff updates the same record. The line under each stage says who decides.

  1. ReferThe agent drafts the packet from the chart and asks follow-up questions. Patient consent is recorded.Coordinator and doctor approve
  2. SendThe packet goes out with its urgency. The patient gets a QR code that holds only a referral number.Coordinator sends
  3. ReviewThe specialist office accepts with a triage priority, asks for more information, or declines.Specialist decides
  4. BookThe patient picks an offered appointment time or asks for another.Patient chooses
  5. VisitThe patient shows the QR code at the front desk and is checked in.Specialist office
  6. ReplyThe specialist writes findings and a plan. Primary care is notified and the patient sees a summary.Specialist writes

Three views

Each person sees the part that is theirs

PatientA plain-language status and timeline, an appointment picker, the QR code, and messaging. Stalls and the clinical packet are deliberately not shown.
Primary care officeA referral queue filtered to “needs you”, the packet editor with open items marked, and an activity log of every update.
Specialist officeA board from incoming to closed, where the specialist triages, asks for what is missing, and writes the reply.

AI and people

The agent drafts and flags. People decide and send.

TaskWhat the AI agent doesWhat a person decides
Referral packetDrafts it from the chart, lists what it could not confirm, asks follow-up questions.The coordinator reviews it; the doctor confirms the clinical question and urgency.
Stalled referralFlags it and drafts a nudge or a booking reminder.The coordinator reviews the draft and sends it.
Patient messagesAnswers clear status or scheduling questions, only from the tracker’s status.A nurse handles anything about health and anything unclear.
VisitSends visit preparation notes; alerts the clinic if the visit is missed.Staff check the patient in; someone calls to rebook a missed visit.
Specialist replyMatches the reply to the referral and notifies the primary care doctor.The specialist writes the findings and the plan.

Safety and trust

Design rules that limit what can go wrong

No medical adviceAny message about symptoms, medicines, doses or results goes to a nurse with no automatic answer.
A keyword rule comes before any AIHealth messages are caught by a plain rule first. Unclear messages also go to a nurse.
Unconfirmed facts are namedThe agent lists what it could not confirm. A packet with open placeholders cannot be approved.
The QR code carries no medical detailsIt holds a referral number only, on a phone or on paper.
The model sees a curated extractEach specialty gets the part of the chart that specialist needs, not the whole record.
Not yet reviewedThese are design choices in a prototype. It has not been through a clinical safety or security review.

Where it stands

A prototype, not a product

What works
  • The six-stage flow runs end to end for type 2 diabetes, primary care to endocrinology.
  • Three views in one app: patient, primary care office, specialist office.
  • With an API key, Claude drafts the packet, asks follow-up questions, revises the draft and routes patient messages.
  • Note drafting also covers nephrology, cardiology and orthopedic surgery referrals.
What is simulated or missing
  • All patients are synthetic. No real patient data has been used.
  • Office, insurer and specialist actions are simulated. There is no EHR, payer or scheduling integration.
  • Demo state lives in one browser; a record shared across sites is not built.
  • Not clinically validated. No pilot, no customers, no outcomes data.

What comes next

Five steps to a first pilot

  1. EHR integration through FHIR. Read the chart from a real record system.
  2. A shared multi-site record. One record both offices and the patient reach.
  3. Real payer checks. Live eligibility and referral approval.
  4. Clinical safety review. Independent review of the drafting and message rules before real patients.
  5. Pilot. One clinic and one specialist group, measuring time to scheduled, loop closure and coordinator workload.

No dates are committed, and the order is a proposal.

Prototype built at the Claude Impact Lab for Digital Health in New York. All patients are synthetic (Synthea); no real patient data is used. Office, insurer and specialist actions are simulated, and what you do stays in your own browser. This public demo has no AI model behind it: drafts are labelled templates. Not for clinical use. Code: github.com/nitchakorn/ClaudeDigitalHealthHack