Standard-Hochdeutsch · medium-warm
For private GP practices, internal medicine, dermatology. Even cadence, no regional colour, audibly calm.
No rough estimate. No "depends on scope." Seven business days is the commitment, on every voice agent build, at every tier. Below is exactly what happens on each of those days.
One structured email, ~20 questions, sent as a fillable Markdown form. You complete it in 30–45 minutes whenever it suits the practice. I reply within the same working day with clarifications and the signed scope document. No scheduled phone calls in this entire process.
ElevenLabs Flash v2.5 persona selected from a curated shortlist or custom-cloned from a 60-second reference recording. The voice has to fit the practice register without inheriting AI-uncanny artefacts. The intro line is written and pressure-tested against natural-language pickup patterns from the briefing.
For private GP practices, internal medicine, dermatology. Even cadence, no regional colour, audibly calm.
For pediatric, family-medicine, dental-family practices. Slightly softer attack, gentler pitch range, parental register.
For cosmetic / aesthetics / private specialists. Crisper diction, "Sie" exclusive, no informalisms.
If you'd like callers to recognise your own front-desk person. 60-second WAV reference. Cloned voice tested against side-by-side blind test of three callers before approval.
Your existing FAQs, opening hours, services, insurer list, accepted forms of payment, location and directions, and any practice-specific protocols are ingested into the agent's retrieval store. Every source document gets chunked at sentence boundaries, embedded with a multilingual model, and indexed in Postgres with pgvector. The store is then pressure-tested against 20+ sample questions sourced from your briefing.
faq.pdf14 pages
services.docx6 pages
hours.html1 table
aufnahmebogen.pdf3 pages
~ 180 chunks
sentence-aware
deduplicated
metadata tagged
multilingual-e5
1024-dim vectors
cosine similarity
avg latency 38 ms
Postgres + pgvector
HNSW index
per-practice tenant
RLS isolated
Gaps surfaced by the retrieval test (questions with no confident answer) are reported back to you in writing the same day, with a precise list of source sentences that would close each gap.
Live read/write access wired to your existing calendar system through its documented API. Slot availability is checked in real time during the call. Bookings are written back instantly and confirmed to the caller. Cancellation and rescheduling flow tested end-to-end. Outside-of-hours rules and emergency-day overrides set.
Fifty scripted test calls run against the agent before it is allowed near a real caller. The mix is engineered: routine bookings, hard edge cases from your own briefing, deliberate safeguard-triggering scenarios, and three adversarial calls designed to provoke prompt-injection. Pass threshold: 48 of 50 correct outcomes. Below that, the issue is logged, fixed, and the whole battery is rerun the same day.
If a scenario fails twice, the build window pauses. Most of the time it doesn't. When it has, the failure has been useful: it surfaces a real misalignment between the practice's expected handling and the agent's default policy.
The agent goes live on a shadow number that you control, never on your public practice number. You can forward as much or as little real traffic to it as you choose. I monitor every call live for the first four hours from my side. Any failure that day is patched within the same business day.
The agent switches from the shadow number to your public practice number on a date you pick. Full admin dashboard is handed over: live call feed, transcripts with confidence scores, booking log, escalation log, and an editable knowledge-base front-end. I stay on standby for the first 48 hours after launch: email reply within one hour, faster on flagged issues. Two written checkpoint reviews are delivered in week 1 (edge-case transcripts plus the adjustments I made). Ongoing support is part of the monthly fee from day one.
If the agent isn't answering live calls by the end of day 7 after the briefing is signed, the setup is free. No partial billing, no "almost there" negotiation. The 7-day clock is the commitment. The only caveat: calendar credentials must be provided on time on day 1, if that's delayed, the clock pauses accordingly.
Send a briefing with your practice setup and I'll confirm whether the 7-day timeline applies to your configuration. Most practices qualify. Custom EHR integrations can take 10–12 days.