Anima's three AI teammates: clinic phone waits from 50 minutes to 10

Anima is a UK healthcare platform built around the three most clerical parts of clinic life: patient requests, clinical letters and consultation notes. In the official customer story's terms, it now serves more than 14 million patients across 1,000+ UK clinics, processes over 800,000 cases a month, and is expanding into the US and Australia. Founder Shun Pang's starting point came from his clinical training: one patient's abnormal test results sat in a queue for six months, unactioned — and that patient missed potentially life-saving care. His conclusion: headcount alone would never keep up; the change had to come from software.

On model choice, Anima says it benchmarked closed and open models before settling on OpenAI; Shun Pang's stated reasons were "superior intelligence, intelligence per dollar, and steerability with in-context learning." In the product this becomes three AI teammates with distinct jobs: Andi the receptionist takes calls with GPT-Live, the document pipeline runs clinical letters on GPT-6, and Scribe writes notes with GPT 5.6 Luna. Worth noting is the split itself — voice, documents and note-writing each get a fitting model rather than one generalist doing everything, and deterministic rules stay at the critical points (the next two sections).

• Read the official customer story ↗

Andi is a voice receptionist running on GPT-Live. A patient calls and describes symptoms; Andi listens and asks. But what gets asked next is not the model improvising — it is Anima's own policy engine walking a deterministic, clinically approved intake workflow. Each call starts with GPT-5.4 mini selecting the right questionnaire from the library; once the history is complete it becomes a structured case in the clinic's dashboard, where staff review and act. The model carries the "understanding human speech" part, while every clinical step stays pinned to approved rules.

The numbers: clinics using Andi cut phone queues from 40–50 minutes to 10 — roughly a 75% reduction in the headline's terms; 86% of patient surveys say they would use Andi again, and users in their eighties and nineties rank among the happiest, with quoted praise for being "infinitely patient and very polite." These are Anima-reported figures published by OpenAI, not our trial; but the order of change — fifty minutes to ten — maps to a real difference on the patient's side of the line. The clinic phone you can't get through is among the most concrete doors AI walks into daily life.

• Next: what the GPT-6 family's Terra does at the desk →

The document line shows the capability-plus-integration combination best. The GPT-6 pipeline first extracts sender, recipient, document type and event dates from each letter, matches the letter to the right patient through a multi-stage pipeline, maps it to SNOMED clinical codes — engineering lead Michael Skupien cites a number worth keeping: SNOMED has roughly 800,000 possible codes — and then suggests the next action. Adaptation happens through in-context learning per practice, with no per-customer fine-tuning. The throughput figures: staff move from 15 items per hour to 30–45, the best clinics to 45–60, and historical backlogs above 100,000 items clear within days.

Scribe, on GPT 5.6 Luna, turns consultation recordings into structured notes. The official screenshot shows one 14-minute-52-second consultation: "cough for over a week" and chest pressure are highlighted in the transcript, SNOMED codes appear alongside (Cough 49727002, Chest Pressure 29857009), and a blue underline creates the chest X-ray task. The same context is reused for referral letters — five to fifteen minutes down to seconds — and follow-up messages. One quoted doctor's version: finishing on time instead of at eight or nine at night. Most cases resolve within 24 hours, with a roughly 3.5-hour median.

• Next: where Luna sits in our cost comparison →

Pin the basis first: every figure here comes from the Anima customer story published by OpenAI — Anima-reported, with patient surveys likewise vendor-cited; this site trialed nothing and logged into no clinic system. By design the AI teammates do not act alone — every case reaches the dashboard for staff review before action, and phone intake walks deterministic, clinically approved workflows. Expansion beyond the UK has just begun, with no US or Australia figures yet. How large the real-world difference is, and how errors get caught, will take independent material.

Placed in this site's thread, it is a full source-to-life arc: voice (GPT-Live) and document understanding (GPT-6) are the source capabilities; Anima's policy engine, per-practice in-context learning and staff review are the integration work; the life end is the unanswered phone, the piled-up letters and the late notes. On-site references: the GPT-6 family's Terra covers the same source at the office desk; Luna, which appeared in our Sol/Luna/Opus cost comparison, here takes dictation; and the personal AI companions topic's running question — agents taking over everyday chores — is told here in a clinic's voice.

• Next: the personal AI companions topic →
• Next: GPT-6.1 Terra's desk tasks and prices →