The demo is the product: Ergofon, our AI agents that answer the phone, from 90 demo configurations to a real customer pilot

The challenge

Small and medium businesses lose revenue through missed phone calls. Nobody answers a call after hours, during lunch, or while the staff serve customers. A receptionist costs EUR 1,500-2,000 or more per month. A receptionist still leaves nights, weekends, and callers in another language without cover. We built Ergofon as our own answer: an AI voice agent that is always on. It answers every call, holds a natural conversation in the caller's language, and takes bookings.

A voice product has a second problem: trust. Nobody buys a voice agent from a slide. You must hear it. So we made the demo the product.

What we built

We built ergofon.com around a conversation with the product itself. A visitor speaks with ERGO, a sales agent that sells Ergofon as Ergofon. Its script contains a deliberate proof point: "You're talking to a voice agent right now". ERGO's whole sales conversation exists as a flowchart. It sorts callers into four types, and each type has its own script, decision points, and polite exit. We can therefore review and improve the conversation visually.

A visitor can also try a demo agent for their own industry, in their own language: ten industries, nine languages, and 90 configurations in total. The demo agents do real work. They check the available time slots and book an appointment during the call.

The third piece proves the product outside a demo. Eva is an agent in German that guides students and parents through apprenticeship opportunities. We built Eva for a real pilot customer, and Eva runs on the same system.

How we built it

One system serves every agent. We did not write 90 agents by hand. The system assembles each agent from shared building blocks at the moment a call starts. A common base defines the agent identity, the safety rules, the answer to an off-topic question, and natural speech on the phone. The system then adds one of ten industry profiles and one of nine language profiles.

A new industry or a new language needs one new profile, not a rebuild. The system checks every profile automatically at startup, so a mistake appears before the first call, not in the middle of one.

Automated tests check the behavior before every change goes live. The tests cover all ten industries. A second AI listens to scripted test conversations and confirms that the agent stayed on script. The agent for a medical clinic must refuse a diagnosis. An agent set to German must answer in German.

A phone call fails if the timing is wrong, so we engineered the conversation itself. The agent starts its reply before the caller finishes the sentence. It answers without an awkward pause, and it does not talk over the caller. A filter removes the AI's internal deliberation before the voice stage, so the agent never speaks its own deliberation aloud.

What this build demonstrates

Ergofon is past the prototype stage. Two agents are live, and we built the Eva pilot for a real customer. The next step is a separate phone number and a separate agent setup for every customer. That step is already fully planned.

Ergofon taught us where voice AI becomes hard. The hard part is not an answer from an AI. The hard part is the timing, a natural turn between the speakers, the separation of the AI deliberation from the voice, and the proof of an agent's behavior before each release. That is the experience we bring to any voice or conversational AI project.