What is an AI voice assistant?

When a call comes in, the assistant reads the greeting. The caller speaks, the speech is turned into text, and an answer is built from the sources the business supplied. The answer is then read back in a synthetic voice.

One thing is said at the start of every call. The caller is told plainly that this is software, not a colleague. A caller who would rather speak to a person can say so in the first seconds. A greeting that hides the point costs trust later.

What the assistant knows is a closed list. Opening hours, the service list, the address and the common questions are loaded before it goes live. A question outside that list gets no invented answer. The call moves to a person instead.

The voice is synthetic. Tone, pace and language are picked during the build, and the same voice speaks the same way on every call. A business that works in Dutch and English needs a written greeting for each language.

The greeting belongs to the business. The first sentence carries the name, the opening hours and the topics the assistant can help with. It stays short, because a caller decides within seconds whether to keep listening.

A greeting is not written once and left alone. Incoming questions are read every month, and the three most common ones are added to the greeting or to the source list. The share of calls the assistant can finish rises with each pass.

Handover happens in three cases. The first is a question with no answer in the sources. The second is a caller who asks for a person. The third is a listed topic. Complaints, price negotiation and contracts sit on that list by default.

A summary travels with the handover. The colleague who picks up sees the summary of the conversation, the caller's record and the answers given so far. That way nobody has to tell the same story twice.

Calls outside opening hours are not lost either. The assistant takes the message, leaves the written summary in a queue, and somebody calls back the next working day. For most small teams that is the part which pays for itself.

Call recording is a separate decision. Written conversations are stored as a matter of course. A voice recording is made only when the business asks for one and the caller agrees to it. Without that agreement, nothing is recorded.

Three questions are answered in writing before it goes live: where the recording sits, how long it is kept, who opens it. Under GDPR those answers belong in the privacy notice. Where health details come up, the rules are stricter and consent is explicit.

Conversations and recordings are not used to train models. That limit is written into the contract. A caller can ask for their own record to be deleted, and the request is logged like any other.

Any business whose phone rings while the work is happening. In a salon, a clinic, a hotel, a restaurant or a field service team, the call arrives mid-job and often goes unanswered.

Busy hours are the second case. When four calls land at once, all four are answered and nobody sits listening to hold music. Out of hours is the third: a night call is taken as a message and reaches the right person in the morning.

It does not suit every call. A complex support case, a long negotiation or an upset customer needs a human voice. Those topics go on the handover list and reach a person directly. ALTAI Digital agrees that list with the client during the build.