AI voice agent or chatbot for customer service: how to choose

Most teams start with the wrong question: should we put a chatbot on the website or a voice agent on the phone line? Both work in production today, so ask instead where the requests that cost you time come from, and which of them a machine can close on its own without causing damage. As CTO of IGEA I built a voice agent that runs clinic reception, and I have worked on text assistants for e-commerce and on systems that route written tickets. This comparison comes from those projects.

What is the difference between a chatbot and an AI voice agent?

A chatbot answers in writing: on your site, in an app, over WhatsApp or by email. The customer can reread, paste an order number, send a photo. A voice agent answers the phone and has to understand speech in real time, with background noise, accents and half-finished sentences. The language model behind both can be the same kind. The channel changes almost everything else: the reply speed you need, how errors are caught, what it costs to run and who will use it.

For a small or mid-sized business the main difference is how you start. A chatbot can launch small, on one page, and grow from there. A voice agent takes over a phone number, so it has to work well on day one or callers hang up.

Which channel do your customers actually use?

Count before you choose. For two to four weeks, log how many requests arrive by phone, by email, through the site form or on messaging apps, and above all how many go unanswered. Missed calls are the number nobody tracks, and it is usually the one that settles the decision.

At a medical clinic, bookings come in by phone and every unanswered call is a visit that may go to a competitor. That is why IGEA is a voice agent. At Centro Medico Luni Care, a clinic with 40+ doctors, its first month brought 0 missed calls and about €3,500 of bookings handled by the AI, including on mornings when the clinic was closed and at weekends. A fashion retailer has the opposite profile: customers type, and they want to see and compare products. There a text assistant fits, like the one built for Estyl that searches a catalogue of over 2 million products.

How fast does a voice agent need to reply to feel like a conversation?

In chat, a few seconds of waiting is fine: the customer sees the reply coming. On the phone, silence is obvious at once. If the agent takes too long, the caller thinks the line dropped, repeats the question or talks over the agent, and the call falls apart.

IGEA's voice agent replies in under 1 second, measured from the end of the caller's speech to the start of the reply. Getting there takes a real-time pipeline in which speech recognition, the model and speech synthesis stream into each other instead of waiting for each step to finish. When you evaluate a vendor, ask for reply latency measured on real calls.

Which requests can an AI assistant handle alone, and which can't it?

Scope matters more than channel. An assistant that does a few well-defined things, with data it can check, holds up in production. One that is expected to answer everything gets things wrong in the worst way, which is confidently.

IGEA's scope is deliberately narrow: bookings, double bookings, changes, reschedules and cancellations. It does no medical triage. That decision kept the agent predictable, cut the sensitive data it handles, and made it possible to go fully live, with staff trained, in 2 weeks. The first month closed with a 96.7% success score and around 10 hours a week of front desk time recovered.

How does the assistant hand over to a person?

Design the handoff together with the assistant, with clear rules for when it triggers: an out-of-scope request, a customer asking for a person, data that does not match, several failed attempts to understand the request.

In chat, handoff is easy: the conversation is already written down and the agent reads it before replying. On the phone it is harder. If someone is free, the agent transfers the call. If nobody is, or the business is closed, the agent should collect name, number and reason and trigger a callback, so the customer never has to start over. For written requests at volume, the hard part is getting each one to the right team on the first try, which is the subject of the guide to AI ticket triage.

What do GDPR and the EU AI Act require for chat and voice?

If you serve customers in the EU, the same transparency duty covers both channels. Since 2 August 2026, Article 50 of the AI Act requires that people interacting with an AI system are told so, unless it is obvious from the context (Regulation (EU) 2024/1689, Art. 50, applicable from August 2026). On the phone, where a synthetic voice can pass for a human, say it in the first sentence.

Under GDPR the phone raises a few more questions. Do you record the call or keep only the transcript? For how long? Where do the models run? IGEA keeps each clinic's data separate, and staying out of triage reduces the health data the agent ever sees. Settle these points with your DPO before launch.

What drives the cost of a chatbot versus a voice agent?

I don't quote generic figures because they vary too much, but the cost drivers are well known. A chatbot pays mainly for model tokens per message and for the knowledge base it draws on. A voice agent adds per-minute costs for speech recognition, speech synthesis and telephony, plus more tuning work to get latency down. For both, the biggest line is often integration with the booking system, CRM or ERP. Without it the assistant can only give information; it cannot close a request.

When should you pick a chatbot, and when a voice agent?

What mistakes should you avoid when choosing between them?

The first is choosing the channel because it is in fashion instead of counting where requests come from. The second is giving the assistant a wide scope from day one: a few tasks it closes alone beat many it half-finishes. The third is judging a voice agent on a demo in a quiet office instead of on real calls. The fourth is skipping the handoff, or designing one that does not collect what you need to call the customer back. The last, for EU customers, is forgetting to tell people they are talking to an AI.

If you want an agent built around your own process, the AI agent development page explains how I work, from defining the scope to running it in production.