AI appointment enquiry routing.

What routing does, and the line it never crosses

AI appointment enquiry routing reads each incoming patient message and puts it in the right administrative queue: book, change or cancel; new patient enrolment; forms and ACC paperwork; billing; records requests; prescription requests for the clinical team’s queue; and general enquiries. Reception still handles each one, but starts from a sorted list with a drafted administrative reply instead of an unread inbox.

The line it never crosses is clinical judgement. Health NZ tells its own staff not to use generative AI tools for any clinical decisions, or for personalised advice to patients. This workflow adopts the same rule. The model does not grade urgency on a scale. It has one clinical output: “may contain clinical content”, which sends the message to a person straight away. When unsure, it flags.

How the workflow runs

Inputs

Messages from the practice email address, website forms and the patient portal, voicemail transcripts if used, the practice’s appointment types and policies, and staff rosters for routing.

Systems

The practice management system (for example Medtech Evolution) and portal, read-only except for creating tasks; a review queue; a model running under a contract with no retention or training on your data.

Steps

Strip signatures and quoted history; detect possible clinical content with keyword rules plus the model; if flagged, send the fixed holding reply and alert staff; otherwise classify into an admin queue; draft an administrative reply; log category, flags and sources.

Human review

A nurse or designated staff member checks flagged messages within an agreed time during opening hours. Reception confirms the category and sends or edits each draft. The practice manager reviews misroutes weekly.

Outputs

Sorted queues, drafted administrative replies, flagged messages with timestamps, and a weekly misroute report.

Measures

Time from arrival to first human read for flagged messages, clinical messages not flagged (found in weekly audit), misroutes, and reception minutes per message.

Risks

Missed clinical content, replies that read as advice, details disclosed to a third party, and data processed somewhere the practice has not approved.

The NZ rules that apply

The Health Information Privacy Code 2020 replaces the information privacy principles for health agencies such as general practices, PHOs and pharmacies. Rule 5 requires reasonable security safeguards and, where health information is given to a service provider for storing or processing, that everything reasonably within the agency’s power is done to prevent unauthorised use or disclosure. Rule 12 limits disclosure outside New Zealand. On 23 September 2026 the Privacy Commissioner issued compliance notices to Manage My Health and Health NZ over security failings, so expect scrutiny of how patient data reaches new tools.

The Medical Council’s guidance on AI in patient care says doctors must not use AI to impersonate themselves, including through chatbots. Administrative replies are signed by the practice team, never as a named doctor.

Illustrative example: a six-GP practice in Lower Hutt

Illustrative example, synthetic numbers: a practice with about 9,000 enrolled patients receives 1,100 written messages a month outside phone calls. Assumptions: reading, routing and answering the administrative part takes reception 2 minutes a message today and 0.75 minutes with sorted queues and drafts; about 4% are flagged as possibly clinical.

• Reception time: 1,100 x 1.25 minutes saved = about 23 hours a month.
• Flagged messages: about 44 a month, each read by a nurse soon after arrival instead of waiting in a general inbox.
• Model fees: at 1,500 input and 150 output tokens per message on Claude Haiku 5.5, at Anthropic’s list prices of US$0.10 and US$0.50 per million tokens (as listed on 19 October 2026), about US$0.25 a month. For processing kept inside Australia and New Zealand, AWS says Bedrock’s AU cross-Region inference keeps Claude inference within that geography; check AWS pricing separately.

What stays with people

Every clinical question, prescription decision, result discussion and urgency call stays with nurses and doctors. Reception decides how to respond to administrative requests. The privacy officer handles requests for access to or correction of health records. Anything from a third party about a patient goes to staff, who check authority before saying anything.

Where routing goes wrong

• Understatement. “Bit of chest tightness, no rush, can I see Dr Ngata next week?” reads as a booking. Keyword rules for symptom terms run before the model and can only add flags.
• After-hours silence. A flagged message at 10pm is not read until morning. The holding reply must say so and give 111 and Healthline (0800 611 116).
• Drafts that sound clinical. “That sounds like it can wait” is advice. Restrict drafts to approved administrative templates.
• Unaudited misses. Each week, a nurse reviews a sample of unflagged messages for clinical content. Read our guide on managing AI mistakes and escalation.

Queue design is covered in human review queues. Related workflows are referral administration and nonclinical policy search. Our AI governance advisory covers the privacy impact assessment, and the healthcare administration overview covers the sector.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.