The Orkelia method.

A generic agent talks about "patient + appointment". A verticalized Clinic agent talks about "initial consultation", "aesthetic treatment", "local anesthesia", "post-op", "recovery". The difference is invisible in a demo but becomes critical in production.

• Clinic: 200+ pre-integrated medical terms (aesthetics, dental, physiotherapy, general medicine, dermatology)
• Hospitality: HR vocabulary (fixed-term contracts, extras, maximum durations, weekly rest) + service (covers, amuse-bouche, mise en place)
• Transport: logistics terminology (POD, ETA, clearing, tail lift, pallet, semi-trailer)
• Trading: financial terms (resistance, breakout, MA50/200, RSI, MACD, regime, smart money)

A generic LLM happily invents an "approximate" price based on what it has learned. For a patient who books an appointment based on that invented price, a dispute is in the making. Our guardrail: post-generation regex that detects the patterns "X€", "from X", "around X" and rejects the draft if the price is not in the agent's database.

An agent must NEVER handle certain conversations. Our method explicitly lists the automatic escalation cases:

• Life-threatening emergency keywords (severe pain, hemorrhage, severe allergy) → immediate escalation + SMS alert to the professional mobile
• Minors without a guardian → systematic escalation to the legal guardian
• Financial requests (refund, invoice dispute) → mandatory escalation
• Complaints / reports → escalation + compliance log
• Quality or GDPR claims → escalation to the DPO

Each escalation is configured before activation, not after an incident. The client sees the exhaustive list in their dashboard and can add their own rules.

A generic agent speaks in a "robotic neutral" that resembles no one. A verticalized Orkelia agent is calibrated on 10-15 examples of the client's real replies. The agent extracts the tone and applies it systematically.

• Family dental practice → "Hello Mrs. Dupont, we can offer you Tuesday at 2pm or Wednesday at 10am, which works best for you?" (warm, calm)
• Premium aesthetic clinic → "Hello, we have two information slots suited to your request: Tuesday at 2pm or Thursday at 4pm. Would you prefer one or the other?" (elegant, reserved)
• Sports osteopathy practice → "Hi! For your ankle, we can see you Tuesday at 2pm or Thursday at 11am, let me know what works." (casual, energetic)

The same system prompt does not produce these three outputs. Tone calibration is what differentiates an agent that "gets the job done" from an agent that patients find "like your receptionist".

Orkelia is not a self-serve SaaS. It's a service where we configure, calibrate, monitor and optimize for you. Initial setup done by an industry expert (24h for Clinic, 1-3 weeks for Hospitality / Transport depending on the stack). Maintenance, industry updates, adding integrations — included in the monthly fee.

For regulated industries (health, finance, legal), compliance is not an option to switch on. It is on by default in Orkelia:

• Supabase eu-west-3 (Paris) hosting — no data leaves Europe
• "Zero retention" mode on the AI provider side (Claude Anthropic) — no storage of conversations on the LLM side
• DPA (Data Processing Agreement) provided BEFORE signature, not after
• Anonymization of patient data before transmission to the LLM
• Right of access / erasure operational within 30 days
• Audit logs kept for 12 months for data protection compliance
• Native TOTP 2FA · SSO Google/Azure/Okta (Group) · SAML 2.0 (Enterprise)

Without this layer, a practice or clinic adopting an AI agent takes a real data-protection risk. Our 2026 health AI GDPR checklist details the 12 points to verify before signing with any provider.