Short answer
An AI referral administration workflow is where a model reads incoming referrals, extracts patient and referrer details, proposes a match to an existing patient record, checks the referral against the clinic’s completeness checklist, and drafts acknowledgement and missing-information letters. Clinicians still read and prioritise every referral. In New Zealand the Health Information Privacy Code 2020 applies, including rule 3A on telling patients about information collected from their referrer.
- The model prepares the administration; clinicians read the original referral and decide priority.
- Patient matching uses exact NHI and date of birth; anything fuzzy goes to staff.
- The referrer’s stated priority is copied exactly, never inferred by the model.
- Since 1 May 2026, HIPC rule 3A can require telling the patient about information received from a referrer.
What referral administration involves
A referral administration workflow covers everything between a referral arriving and a clinician reading it: registering it, finding or creating the patient record, checking it is complete, acknowledging the referrer and the patient, chasing anything missing, and tracking status until an appointment or outcome letter goes back. Referrals arrive through e-referral systems, secure email, fax-to-email and self-referral forms, often with attachments split across messages.
The model does the reading and drafting for the administrative part. It does not touch the clinical part. Health NZ tells its staff not to use generative AI for clinical decisions, and prioritising a referral is one.
| Step | Model prepares | Person decides |
|---|---|---|
| Register | Extracted patient, referrer and funding fields | Admin confirms |
| Match patient | Exact NHI and date of birth match, or a candidate list | Admin chooses when not exact |
| Completeness | Missing items against the clinic’s checklist | Admin sends or edits the request |
| Priority | Referrer’s stated priority, copied verbatim | Clinician sets clinical priority |
| Acknowledge | Draft letters to referrer and patient | Admin sends from approved templates |
| Triage | Nothing beyond a list of attachments | Clinician reads the original referral |
How the workflow runs
- Inputs
- Referral letters and forms, attachments (imaging reports, results, ACC or insurer documents), the clinic’s referral criteria and completeness checklist, and letter templates.
- Systems
- The practice or patient management system, the e-referral inbox and secure email, a review queue, and a model API under a contract with no retention or training on your data.
- Steps
- Group messages and attachments into one referral; extract administrative fields; attempt exact patient match; check completeness; copy the referrer’s priority wording; draft acknowledgement and information-request letters; place the complete referral in the clinician’s triage queue; log every field with its source page.
- Human review
- Administrators confirm every registration and any non-exact match before the record is created or linked. Clinicians triage from the original. The practice manager approves templates and reviews a weekly sample.
- Outputs
- Registered referrals linked to the right patient, drafted letters, a missing-information queue, and a status tracker.
- Measures
- Days from receipt to clinician review, referrals returned as incomplete, wrong-patient matches found (target zero), and admin minutes per referral.
- Risks
- Wrong-patient linkage, lost attachments, duplicated referrals, inferred priority, and clinical content summarised for triage.
The HIPC rules that apply
The Health Information Privacy Code 2020 replaces the information privacy principles for health agencies. Three rules shape this workflow. Rule 8 says a health agency must not use or disclose health information without reasonable steps to ensure it is accurate, up to date, complete, relevant and not misleading, which is why patient matching is strict. Rule 5 requires safeguards when a service provider, such as a model vendor, processes health information for you.
Rule 3A is new. Under Amendment No 2, in force from 1 May 2026, a health agency that collects health information about someone other than from them or their representative must take reasonable steps to make them aware of the collection, its purpose, the intended recipients, the agencies involved and their access and correction rights, unless they were already made aware. A referral is exactly that. Agree with referrers what they tell patients, and cover the rest in your acknowledgement letter. The Privacy Commissioner’s IPP 3A page explains the principle behind the rule.
Illustrative example: a Tauranga orthopaedic clinic
Illustrative example, synthetic numbers: a private clinic receives 260 referrals a month. Assumptions: administration takes 12 minutes a referral today and 5 minutes with extraction, matching and drafted letters; 25% arrive incomplete.
- Admin time: 260 x 7 minutes = about 30 hours a month back.
- Incomplete referrals: about 65 a month get an information request the same day, instead of being found when the clinician opens them.
- Model fees: at 8,000 input and 1,000 output tokens per referral on Claude Sonnet 5.5, at US$2 and US$10 per million tokens (Anthropic, as listed on 19 October 2026), about US$7 a month.
What stays with people
Clinicians decide priority, whether to accept or redirect a referral, and what to tell the referrer clinically. Administrators decide every non-exact patient match and send every letter. The privacy officer handles requests from patients to see or correct what the referrer sent. The model never declines a referral or contacts a patient on its own.
Where referral administration goes wrong
- Wrong patient. Same name and date of birth, or a missing NHI. Only exact matches link automatically; everything else is a person’s decision.
- Split attachments. An imaging report arrives in a second email. Hold referrals open for a set period and match follow-ups by referrer and patient.
- Duplicates. A GP resends after two weeks. Detect and link, rather than registering twice.
- Inferred urgency. The model reads “worsening” and calls it urgent. Copy the referrer’s priority wording only.
- No trail. Without a record of what was extracted, matched and sent, errors cannot be traced; see AI audit logs.
This is an AI document processing build with workflow automation around it. Related workflows are appointment enquiry routing and nonclinical policy search; the healthcare administration overview covers the sector.
Sources
- Health Information Privacy Code 2020 (incorporating Amendment No 2, in force 1 May 2026), Office of the Privacy Commissioner, 2026
- Health Information Privacy Code 2020, Office of the Privacy Commissioner, accessed 19 October 2026
- IPP3A, Office of the Privacy Commissioner, accessed 19 October 2026
- Using generative AI and large language models, Health New Zealand | Te Whatu Ora, accessed 19 October 2026
- Claude API pricing, Anthropic, accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.