Service 02 / 06
A voice agent that answers the phone like it wants to help.
A voice AI agent picks up immediately, holds a real conversation (you can interrupt it), does the lookups and updates mid-call, and passes anything it should not handle to a person with the transcript already there. It is the same assistant as your chat agent, using its voice.
What I build
- It answers straight away, and you can interrupt it like a person
- It books, checks and updates things in your systems mid-call
- It hands over to a human with the transcript already there
- Same assistant in text and voice, so a chat can become a call
What you get
A number people can ring, and a record of what it handled and what it passed on.
Good fit ifyou get repetitive calls that follow a known pattern.
The call button on this page is one of these. Try it before you believe any of this.
Why most phone bots are worse than a menu
Because they cannot be interrupted, and because they start over. A caller who has to sit through eight seconds of options, then repeat everything to a human who has none of it, has had a worse call than if nobody had answered. The technology is not the problem; the turn-taking is.
What makes a voice agent tolerable is unglamorous: it starts speaking fast enough that the silence does not read as a dropped call, it stops the moment you talk over it, and when it gives up it gives up *to someone*, carrying everything you already said.
Can a voice agent speak Dutch, or actually Flemish?
This is the question that decides the project in Belgium, and the honest answer in 2026 is: Dutch yes, Flemish carefully. Most vendors ship nl-NL, label it "Dutch", and let you discover the difference on a real call with a real customer in Ghent or Antwerp.
The Belgian-specific state of play, from evaluating this directly: Azure ships genuine nl-BE and fr-BE for both recognition and speech, and Deepgram added nl-BE recognition to Nova-3 in 2026. Several models that top the general quality leaderboards are unusable here for exactly this reason: they are excellent at a language your customers do not speak. Any voice project for a Flemish market should be scoped on nl-BE from the first meeting, not adapted to it afterwards.
This site can call you, in English, right now. That is the fastest way to judge latency and turn-taking without taking my word for either.
What the agent handles, and what it should not
The right first scope is the five questions you already get all day: opening hours, order status, booking and rescheduling, the routine change of address, the "has it shipped" call. These follow a known pattern, and a pattern is what an agent is good at.
What it should not take is anything where being wrong is expensive or being human is the point: a complaint, a cancellation someone is upset about, anything medical or financial that needs a person accountable for the answer. Drawing that boundary narrowly at the start is what makes the first month a success instead of an incident.
One brain, two channels
The agent that answers your phone and the agent on your website should be the same system, not two builds that drift. On this site they are: a chat conversation can become a call mid-thought, and the voice picks up with the history already in hand.
That is an architectural choice with a commercial consequence: every improvement to the agent's knowledge lands on both channels at once, and there is only ever one set of instructions to keep true.
Questions people ask
- How fast does a voice agent answer?
- The number that matters is time to first word, and under about a second is where a caller stops hearing the pause as a fault. Getting there is an engineering budget spent across recognition, the model and speech synthesis. That is why the model choice for a voice agent is a different decision from the model choice for chat, where a second of thinking is invisible.
- Can callers interrupt it?
- Yes, and they must be able to. Barge-in (the agent stopping mid-sentence the moment the caller speaks) is the single feature that separates a conversation from a recorded menu, and it is table stakes rather than an upgrade.
- Does it speak Flemish (nl-BE) or only Dutch (nl-NL)?
- Both exist, but not from every vendor, and the default almost everywhere is nl-NL. Azure supports nl-BE for both speech recognition and synthesis, and Deepgram Nova-3 added nl-BE recognition in 2026. If your callers are Flemish, insist on hearing a sample in nl-BE before anything is signed.
- Does it work with our existing phone number?
- Usually yes: the agent sits behind a telephony provider, so an existing number is ported or forwarded to it rather than replaced. The practical constraint is your current provider, so that is worth checking in the first conversation.
- Are the calls recorded, and what about GDPR?
- A transcript is what makes the hand-off to a human work, so most deployments keep one, which makes it personal data with a retention period, a lawful basis and a deletion path, decided by you and built in rather than added later. Both the retention job and the erasure path are things I build as part of the system, not afterwards.