SERVICES
AI Chatbots and Voice Agents
A useful assistant cites its sources. A confident one does not.
Short answer
An AI chatbot answers customer questions from your own documented knowledge rather than from general training data, and shows the source behind each answer. PHOENIX VERUS builds text and voice agents that can also act — booking, looking up an order, opening a ticket — with the access rules and logging that acting safely requires.
How it actually works
The difference between a useful assistant and an embarrassing one is a single design decision: what it does when it does not know. A chatbot that guesses costs you more trust than it saves in time, and it costs it in front of customers.
Ours answer from your own documents and show the source under every reply. When a question falls outside what they have been given, they say so and hand over to a person. That refusal path is designed before the interface is, because it is the part that decides whether the thing can be trusted.
From 2 August 2026 there is a second reason to get this right. Under Article 50 of the EU AI Act, an assistant that talks to customers must tell them it is an assistant, and the obligation sits with you as the deployer even when the software is a vendor’s. Every chatbot we build discloses itself at first contact, and we will tell you if the one you already have does not.
Article 50 of Regulation (EU) 2024/1689 has applied since 2 August 2026. An interactive AI system must inform people they are dealing with a machine, no later than the first interaction.
What the engagement covers
- Grounded answers
- Retrieval over your documents, with the source shown per answer so a wrong answer is traceable rather than mysterious.
- Scope control
- A defined boundary and an honest refusal outside it. A bot that guesses costs more trust than it saves time.
- Actions
- Booking, order lookup, ticket creation — each with its own permission model and audit trail.
- Languages
- Greek, English and German, each with its own evaluation set rather than one set translated.
How it runs
- Define the scope and collect the knowledge base.
- Build retrieval and answer evaluation before any interface work.
- Add actions and permissions.
- Launch behind a review period, then measure deflection and escalation.
Is it for you?
It is, when
- The same twenty questions arrive by email and phone every week, and the answers exist in writing somewhere.
- You want the assistant to complete something — a booking, a form, an order lookup — not only to answer.
- You sell in more than one language and each one needs to be tested separately.
It is not, when
- The answers are not written down anywhere. That knowledge base has to exist first, and building it is your work, not ours.
- The questions genuinely require judgement. A model will answer them confidently and sometimes wrongly.
Frequently asked
What does a chatbot cost?
Four variables move it: number of languages, size and volatility of the knowledge base, integrations, and whether it is text or voice. The pricing page sets out each of them.
How long does it take?
A scoped text assistant on a defined document set is measured in weeks. Voice, multiple languages and write access each extend that, and the knowledge base is usually the slowest part — not the software.
Will it hallucinate?
Grounded retrieval with per-answer sources makes hallucination visible and rare, not impossible. The design choice that matters is what the bot does when it does not know: ours say so and hand over.
Contact us
Send us the brief or book a short call. You get concrete next steps in reply.