Anima on GPT-6 and GPT-Live Cuts Clinic Phone Wait Times by 75%

UK startup Anima, a platform for clinics with more than 14 million patients, 1000+ clinics and 800 thousand cases a month, has built three products on OpenAI models. They are Andi, a voice administrator running on GPT-Live; a pipeline for processing incoming documents on GPT-6; and Scribe, a clinical notes service on GPT 5.6 Luna. According to the company, phone wait times fell from 40–50 to 10 minutes, a reduction of at least 75%. Document throughput rose 2–3x, and the median case resolution time is about 3.5 hours.

Voice on a Short Leash

Andi answers patient calls and takes the medical history. A language model alone does not run the conversation. Anima's proprietary policy engine drives the dialogue through deterministic, clinically approved scripts. The model does not improvise on medical questions; it works within a fixed structure.

GPT-5.4 mini then steps in to select the appropriate questionnaire. The completed medical history lands in the clinic's dashboard as a structured case that a doctor or administrator reviews and processes. The division of roles is pragmatic. The expensive voice model handles the live conversation, the compact model does quick classification, and the safety rules live in ordinary code, where they can be tested and audited.

In Anima's surveys, 86% of patients said they would use Andi again. One unexpected detail: people aged 80–90 were among the most satisfied. CEO Shun Pang explains it this way: the voice agent is "infinitely patient and very polite."

800 000 SNOMED Codes and No Fine-Tuning

The Documents product, built on GPT-6 models, extracts the sender, recipient, document type and event dates from incoming letters. A multi-stage pipeline then matches the document to the right patient, selects clinical codes and suggests the next action to a staff member.

Coding is the hardest part. According to Michael Skupien, head of the engineering team, there are around 800 thousand possible SNOMED codes. Only a small fraction is used regularly, but different practices choose codes differently. Anima does not make clinics configure the system and does not fine-tune a separate model for each customer. Instead, behavior adapts through in-context learning: the model receives a specific practice's examples and preferences in its context.

From an economic standpoint, this means one set of models for all customers, with no zoo of weights, training pipelines and versioning. Costs shift toward context tokens, but onboarding a new clinic requires no ML work. The approach rests on one property of the models: they must follow instructions consistently. That, according to Skupien, is why the team chose OpenAI. Shun Pang adds that the team compared closed and open models and evaluated intelligence, intelligence per dollar and controllability.

The results in numbers: practices that processed about 15 documents an hour now handle 30–45, and the best clinics handle 45–60. According to Pang, backlogs of more than 100 thousand documents were cleared in a few days.

Full List: Products, Models and Metrics

• Andi, an AI administrator on GPT-Live: takes calls, collects symptoms; phone wait times fell from 40–50 to 10 minutes (a reduction of 75% or more). • GPT-5.4 mini: selects the appropriate questionnaire after the conversation.

• Documents on GPT-6: extracts sender, recipient, type and dates, matches the document to a patient, selects SNOMED codes, suggests the next step. • Scribe on GPT 5.6 Luna: notes from the consultation transcript, plus referrals and patient messages generated from the same context.

• Scale: 14 million patients, 1000+ clinics, 800 000+ cases a month, expanding into the US and Australia. • Speed: median case resolution time of about 3.5 hours; most cases are closed within 24 hours.

• Documents: 15 an hour before rollout, 30–45 after, 45–60 for the best clinics. • Notes and letters: notes are ready in seconds instead of several minutes; referrals in place of 5–15 minutes. • Satisfaction: 86% of surveyed patients would use Andi again.

What to Take Back to Your Stack

The case shows an architectural pattern that carries over beyond healthcare. Keep critical logic, meaning dialogue routes and permitted actions, in a deterministic rules engine. Bring in the model where you need language and meaning extraction. Assign different tasks to different classes of model: voice and complex document parsing go to the top-tier models, while questionnaire selection goes to the mini version.

If you have many customers with different rules, start with in-context learning rather than per-customer fine-tuning. It is cheaper to maintain, but it requires a budget for long context and tight control over instruction following. Finally, in all three Anima products a human stays in the loop: a doctor reviews the case, the note and the letter. The gain comes from faster preparation, not from dropping the review. Keep in mind that all figures in this article come from the company itself and have not been independently verified.