AI Receptionist for healthcare: 3,440 calls in 4 months, an enterprise AI lesson

For four months an AI has been answering the phone at a private outpatient clinic in Cremona. Not in a controlled trial, not in a demo room: in the real call centre, 24/7, with real patients calling to book visits or ask questions. Its name is YoDa Health, it is our vertical AI Receptionist for private healthcare, and the numbers from these four months (January-April 2026) are the most solid public case we have to talk about enterprise AI without lying.

We pull the numbers directly from the production database (container `sgz-db`, Postgres 15, server `aifabric-yoda-automation-prod`). They are public within the limits of aggregate processing and we report them here because the customer — SGZ Solutions, Poliambulatorio MedicinaPO Cremona — authorised the sharing of the case after the first quarter of operation.

The numbers, no marketing on top

• 3,440 calls handled by YoDa Health in the first 4 months (Jan-Apr 2026)
• Monthly growth curve: Jan 81 → Feb 896 → Mar 1,189 → Apr 1,274
• 22% of calls happen out of front-desk hours (evening, weekend, holidays)
• About 800 unique patients per month identified from the call
• Average call duration: 65-80 seconds end-to-end
• 99.4% of recap emails sent after the call
• 94.4% of WhatsApp messages delivered with read receipt
• 75 medical specialties handled, 171 services in the catalogue, 66 doctors in scheduling, 232 FAQs in the knowledge base

What actually works: three components few people talk about

1. The vertical knowledge base, not a generic one

YoDa Health cannot answer everything. It can answer the 232 FAQs of this specific clinic. It knows that Dr Rossi only sees patients on Tuesday afternoons. It knows that a knee MRI takes 30 minutes and goes into specific slots. It knows that electronic payment is available but the POS at the counter doesn't take Diners cards. These are operational data of the individual client, structured in YAML and loaded into a dedicated knowledge base for each installation.

The classic temptation of a first AI product is to say "oh we do RAG" and dump an entire website into embeddings. In a real clinic this doesn't work: the website doesn't know when Dr Rossi is on holiday, when the radiologist swaps shifts, when the clinic shifts hours for the Easter holiday. That information lives in the practice management software, in the shared calendar, in the heads of the front-desk team. Extracting, structuring and maintaining that data is where the real work happens — and where the product wins or loses.

2. The integration with the practice software, not the AI

The most important piece of YoDa Health is not the LLM. It is the bidirectional integration with the clinic's practice management software (calendar, patient records, service catalogue). The LLM is the front end. The value comes from the fact that, after a 70-second conversation, the appointment is actually scheduled in the doctor's calendar, the patient record is updated, the confirmation email is sent, and the service is listed in the day's billing.

Without that integration, you are a 2010s call-centre voicebot in new clothes. With that integration, you replace the receptionist for standard cases, offload the human one to complex cases (unclear reports, complaints, sensitive clinical situations), and — not minor — produce high-quality structured data on booking patterns.

3. Human fallback, always

YoDa Health is designed to recognise when it does NOT know what to answer. The situations that trigger "pass to operator" mode are coded: specific clinical question, agitated patient, legal request, any question that falls outside the knowledge base with confidence below threshold. In those cases, the call is queued for the human receptionist the next day or — if the office is open — transferred in real time.

The counterintuitive point: this fallback function is what makes the product sellable. A clinic would never buy an AI that claims to know everything. They buy an AI that handles standard cases autonomously and admits when it has to hand over. The customer's mental model is not "replace the receptionist", it is "help the receptionist not lose calls".

What doesn't work: three hard lessons

Lesson 1: the model is not the problem, the voice stack is

The first three months of YoDa Health were spent fixing the quality of the text-to-speech (TTS) voice. The model knew what to say, but the voice was robotic enough to make 18% of callers hang up within the first 10 seconds. We swapped TTS provider three times, tuned the voice for specific Italian pronunciation (Lombardy surnames that ChatGPT-voice gets regularly wrong), and invested in voice cloning of the long-standing human receptionist. Drop-off went down to 4%. Lesson: in a voicebot, voice matters more than brain.

Lesson 2: the first call from a patient is a wall

The patient segment with the most resistance is the one calling the clinic for the first time. They don't know the receptionist is an AI, have no historical reference, and — most of all — have no reason to trust it. That segment has 15% drop-off in the first 5 seconds. Recurring patients have 2% drop-off: they get used to the system, find it convenient, some prefer it because "I don't have to wait in queue".

The practical solution we adopted: in the first 3 seconds of the call, YoDa Health declares itself as an automated assistant (never hides it), explains what it can do in two sentences, and offers "press 0 to speak with a person" as an explicit option. Transparency cuts drop-off by 40% compared with the script we tried at the beginning.

Lesson 3: integration with the practice software is where you die (or live)

The clinic's practice management software is a legacy product — common in private Italian healthcare. Limited APIs, custom auth, exotic date formats. The first month of YoDa Health was spent building a bidirectional adapter layer that spoke that practice software reliably. Without that work, the LLM would have answered well over the phone and then been unable to do anything concrete. Our initial mistake was underestimating that layer: in retrospect, it is worth at least half the total development effort.

What it means for other verticals

YoDa Health for healthcare is the first case of AI Fabric, our platform of vertical AI Personas for enterprise. The pattern we distilled works — we believe — also for:

• Customer care of utility and telco companies: automatic handling of standard information and administration requests, human escalation on technical cases
• Back-office of professional firms: routing email/phone by case type, preliminary data collection, appointment setting
• Sales support for mid-market: qualifying inbound leads from website or phone, booking discovery calls only on leads validated for fit and budget
• Internal IT help desk: triage of tier-1 tickets with automatic resolution for known patterns, escalation to tier-2 when needed

In each of these verticals, the three components that matter are the same: well-structured knowledge base for the individual client, bidirectional integration with operational systems, disciplined human fallback. The LLM is a commodity. The product is what surrounds it.

For those thinking of trying

AI Fabric is today a consulting + technology programme for the mid-market. We work on one vertical at a time: today YoDa Health is the most mature case, but we are building the patterns for telco customer care and professional back-office. The typical model is: 6-8 weeks of assessment + MVP implementation, 3-6 months of guided operation to measure ROI, progressive scaling after the first quarter.

The YoDa Health numbers above are not a benchmark: they are the particular case of a specific clinic. If your context is similar in orders of magnitude (recurring operations, finite but deep knowledge base, integration needed with one or two practice systems), let's talk. The first call is an assessment, lasts 60 minutes, is free.