Patient Enquiry Triage with ChatGPT for Clinic Websites

Patient enquiry triage software built on ChatGPT solves a problem every private clinic recognises: the contact form, the shared enquiries inbox and the WhatsApp Business number fill up faster than two receptionists can empty them. A mid-size London clinic with a GP, two physiotherapists and a dermatology list receives 60 to 120 written messages a day: new-patient requests worth hundreds of pounds, repeat prescription questions for the nurse, invoice queries for the practice manager, and a handful describing symptoms that should never wait in a queue. Today they sit in one list, in arrival order, and whoever opens the inbox decides what happens next.

The cost shows in three numbers the practice manager already tracks: first-response time, conversion of new enquiries to booked appointments, and reception hours spent on messages rather than patients at the desk. ChatGPT changes the first step of the process, not the clinical decision. It reads the raw message, works out what kind of request it is, extracts the details a human will need and places the item in the right queue with a suggested reply. It never diagnoses, never advises on treatment and never decides a message is safe to ignore. Clinicians and trained staff make every clinical judgement; the model makes sure the right person sees the right message first.

Where Traditional Enquiry Handling Falls Short

Most clinic websites still end at a contact form that sends an email. It lands in a shared mailbox alongside supplier invoices, CQC correspondence and newsletters. Reception scans it between phone calls, and whoever has a spare minute deals with whatever is at the top. A new-patient enquiry about a private ultrasound might wait four hours; a question about a missing sick note gets answered in five minutes because it was easy. Nothing in the process weighs urgency, value or the skill required to answer.

The other weak point is context. "Same problem as last time, can I have the same tablets" is meaningless without the patient record, so reception opens the practice management system, searches by name, reads the last consultation and then asks the GP. Each hop costs minutes and risks mis-keying a name or date of birth. When the practice grows, the answer is usually another receptionist, and the backlog returns a few months later.

What ChatGPT Adds to a Clinic Enquiry Website

• ChatGPT reads free-text messages from any channel and classifies them into the categories the clinic actually uses, such as new booking, follow-up, prescription, results, billing or complaint.

• It extracts the practical details a human will need, including the service requested, preferred dates, named clinician, insurer and any red-flag wording, into a fixed JSON structure.

• It drafts a reply and a routing recommendation that a receptionist, nurse or clinician approves or edits before anything reaches the patient.

Understanding Messy, Emotional Patient Language

Patients do not write in the clinic's categories. One message can ask for a physio appointment, mention that the last invoice looked wrong and add that a child has had a rash for three days. ChatGPT reads for intent rather than keywords. It recognises that "the tablets I had in March" means a repeat prescription, that "my Bupa number is" signals an insured self-referral, and that "getting worse each day" changes the tone of the whole message.

The system prompt is written to be conservative. It tells the model that its job is administrative sorting, that it must never suggest a cause for a symptom, and that any mention of chest pain, breathing difficulty, severe bleeding, suicidal thoughts, a very young child or a pregnancy concern must be flagged as urgent regardless of how calm the wording sounds. The flag does not replace a clinical assessment; it guarantees a human sees the message within minutes rather than hours.

Producing a Structured Record Instead of a Paragraph

The valuable output is not the summary but the structure around it. Each enquiry becomes a record with a category, urgency band, confidence score, extracted fields and the exact quote that triggered any flag. Because ChatGPT returns JSON against a schema, the backend validates it, rejects anything malformed and stores the result beside the original message without anyone retyping.

That structure feeds everything downstream. The practice manager filters by category to see how many billing queries arrived this week. The clinical lead reviews every urgent flag and confirms or downgrades it. Marketing sees which services new patients ask for that the clinic does not yet list. None of this was possible while enquiries lived as loose emails.

Routing to the Right Queue With a Draft Ready

Once the record exists, routing is a set of rules the clinic controls, not something the model decides. Prescription requests go to the nurse queue with the last issue date noted. Insured new-patient requests go to the bookings team with insurer and authorisation status highlighted. Anything urgent goes to the duty clinician's phone as well as the inbox.

Each item arrives with a suggested reply in the clinic's tone: a booking link, a request for missing insurer details, or a note that the query has been passed to the nursing team. Staff read it, adjust it and send. The saving is not only in typing; the thinking about what to say has been done once, consistently, for every message.

Best Use Cases for ChatGPT Patient Enquiry Triage

• Independent GP and multi-specialty private clinics that handle bookings, prescriptions and results queries through one small reception team.

• Physiotherapy and musculoskeletal practices where most enquiries are self-referrals that need a quick decision on suitability and insurer cover.

• Diagnostic and imaging centres that receive referrals and appointment requests from GPs, consultants and patients through several channels at once.

Private GP and Multi-Specialty Clinics

A clinic offering GP consultations, minor surgery, travel vaccinations and blood tests receives enquiries that span every category on the same day. Triage matters here because the same inbox contains a request for a same-day appointment for a feverish toddler and a request for a fit-to-fly certificate. ChatGPT separates the two in seconds, puts the first in front of a clinician and gives the second a reply with the price list and booking link.

The practice manager also gains a picture of demand. When 30 per cent of new enquiries ask about a menopause clinic the site barely mentions, that is a service decision the clinic would not otherwise have noticed. A human approves each reply throughout, which keeps the clinic within its own governance rules.

Physiotherapy and Musculoskeletal Practices

Physio practices live on self-referrals, and speed of first response is the biggest single driver of conversion. A prospective patient who describes a shoulder problem and asks whether the clinic accepts AXA cover wants a yes or no and a booking link within the hour. ChatGPT extracts the body area, the duration of the problem, the insurer and any previous treatment, and drafts the reply while the therapist is still in a session.

The model is instructed not to comment on likely causes or suitability for treatment. It notes the facts, highlights anything that reads like a serious injury or post-surgical concern for the lead therapist, and hands over. That division lets a two-therapist practice respond like a larger one without anyone giving clinical opinions by email.

Diagnostic and Imaging Centres

Scanning centres and pathology services receive a different mix: GP referrals with attached forms, consultant secretaries chasing report turnaround, and patients asking whether to fast before a test. ChatGPT reads the referral text, extracts the requested modality, body region and clinical question, and checks that referrer and patient details are complete.

Incomplete referrals are the biggest source of wasted imaging slots. Flagging a missing referrer signature or pregnancy status field the moment the enquiry arrives, and drafting the request for it, saves the bookings team a phone call and the patient a cancelled appointment. A radiographer or radiologist still vets every referral before it is scheduled.

Core Features of a ChatGPT Patient Enquiry Triage Website

• A single intake layer that gathers messages from the website form, email, WhatsApp and patient-portal chat into one queue with consent recorded at the point of entry.

• An intelligence layer that turns each message into a validated JSON record with category, urgency, extracted fields and a suggested reply.

• A staff dashboard and workflow layer that routes items, tracks who owns each one and reports on response times and conversion.

Multi-Channel Intake With Consent Built In

The website form captures the minimum data needed and shows a clear privacy notice explaining that messages are sorted by automated tools and that a member of staff reads every one. Email and WhatsApp messages are pulled into the same queue through their APIs. Attachments such as referral letters or insurer authorisations stay in the clinic's own storage and are not sent to the model unless the clinic has decided that is appropriate.

Every intake record carries a timestamp, channel, consent flag and a link to any existing patient record the practice management system can match. Matching is done by the backend against the clinic's own database, so the model never needs the full record.

The Intelligence Layer and Its Guardrails

The backend sends the message text and a short context block to ChatGPT with a system prompt defining the categories, urgency bands, red-flag list and the exact JSON schema expected back. The response is validated before it is trusted. An unknown category or a confidence below the clinic's threshold sends the item to a general queue for manual sorting rather than a guess.

The prompt is versioned and reviewed by the clinical lead. When the clinic adds a service, categories and routing rules change in configuration, not code, so the practice manager adjusts behaviour without a developer while the guardrails that stop the model giving clinical advice stay fixed.

Dashboard, Ownership and Audit Trail

Staff see a board with a column per queue, colour-coded by urgency, showing the message, extracted fields and suggested reply side by side. Taking ownership, editing the reply and sending are single clicks. Every action is logged with who did what and when, which meets the audit expectations a CQC inspector will have about how urgent messages were handled.

Reporting comes from the same data: first-response time by category, conversion of new enquiries to bookings, volume by channel and urgent flags confirmed versus downgraded, with nothing exported to a spreadsheet.

Step-by-Step Integration Process

• Agree the categories, urgency rules and red-flag list with the clinical lead before any code is written.

• Build the intake, ChatGPT call and validation as a backend service that the website, email and messaging channels all feed into.

• Launch in shadow mode, where staff see the suggestions but the old process continues, and measure before switching over.

Step 1: Define Triage Goals, Categories and Business Rules

The project starts with a workshop involving the practice manager, the lead receptionist and a clinician. They list the categories enquiries actually fall into, decide which carry a time expectation, and write the red-flag list in plain language. This document becomes the core of the system prompt and routing configuration.

The same session sets the success measures. For most clinics they are median first-response time, the share of new enquiries booked within 48 hours and reception hours spent on messages. Recording the baseline now makes the later comparison honest.

Step 2: Design the User Journey for Patients and Staff

The patient-facing part is deliberately simple: a short form that asks for the essentials, explains how the message will be handled and confirms receipt with a realistic response time. It avoids asking for detailed medical history, which is neither necessary for triage nor wise from a data-minimisation perspective.

The staff-facing part is the dashboard, prototyped with real anonymised messages so reception can say what they need at a glance and what can wait behind a click.

Step 3: Connect Your Website Backend to ChatGPT

The backend, typically a Node.js service on the clinic's existing stack, receives each message, adds minimal context, calls the ChatGPT API with the system prompt and schema, and validates the JSON that comes back. Retries, timeouts and a fallback to the manual queue are built in so a slow API response never blocks a patient message.

Prompt and schema live in version control, and every change is tested against historical messages with known correct answers before it goes live.

Step 4: Save Results, Trigger Workflows and Keep Humans in Control

Validated records are saved to the clinic's database and pushed to the relevant queue. Urgent flags trigger an immediate notification to the duty clinician through the channel the clinic already uses. Suggested replies are stored as drafts, never sent automatically.

The rule that no message reaches a patient without a person approving it is enforced in code, not just in policy. That single design decision is what keeps the system firmly on the administrative side of the line and away from anything that could be read as clinical advice.

Step 5: Measure Triage Accuracy and Improve Over Time

After launch the clinic reviews a sample of classifications each week: how often staff changed the category, how often an urgent flag was downgraded, and whether any urgent message was missed. Findings feed prompt adjustments and, occasionally, changes to the categories.

Over a few months the practice manager sees whether first-response time and conversion have moved. Improvements come in small, reviewed steps, and the clinical lead signs off any change to the red-flag logic.

Security, Privacy, Cost Control and Long-Term Scalability

• Health data is special category data under UK GDPR, so the clinic needs a lawful basis, a completed DPIA and a data processing agreement with every provider involved.

• Costs are predictable because each enquiry is a single short API call, and the model is never used for open-ended patient conversations.

• The same backend scales from one site to several without changing the patient experience or the governance model.

Privacy is the first design constraint. Enquiries contain health information, which UK GDPR treats as special category data, so the clinic documents its lawful basis, completes a Data Protection Impact Assessment and makes sure its ICO registration and privacy notice describe automated sorting accurately. Only the message text and minimal context go to the model; identifiers are pseudonymised where not needed, and API settings that exclude data from training are used. Clinics working with NHS patients also map the setup against the Data Security and Protection Toolkit and the Caldicott principles, and record where data is processed.

Because the system prepares and routes rather than advises, it stays outside the scope of a medical device, but that boundary must be kept deliberately. The clinical safety lead reviews the red-flag logic, the practice keeps a written description of what the model does and does not do, and any drift towards giving patients clinical information is treated as a governance issue. Cost stays low: a clinic handling 2,000 enquiries a month typically spends less on API calls than on a single day of reception time, and a second site reuses the same service with its own routing rules.

Davydov Consulting scopes and builds this kind of triage layer on the stack a clinic already runs, whether that is a Wix Studio site with Velo, a Node backend on Google Cloud or Firebase, a WordPress site or a patient app. We start with the governance conversation, keep humans approving every reply, and connect the result to your practice management and messaging tools. If your reception team is drowning in enquiries, we would be glad to talk through what a first version could look like.