AI call assistants: what they can handle today, and what still needs a person

Voice is the AI use case clients ask about most, and the one with the widest gap between the demo and the reality. A good demo is ten seconds long. A real call has background noise, interruptions, accents, and a caller who wants to talk about three things at once. We have built enough of these to have clear views on where they work.

What they do well

• After-hours and overflow. The calls that currently go to voicemail or ring out. An assistant that answers at 11pm, takes the details, books a slot and sends a confirmation is a plain improvement over nothing.
• Structured tasks. Booking, rescheduling, cancelling, order status, opening hours, “is this in stock”, “where is my delivery”. Anything where the caller’s intent is one of a known list and the answer comes from a system you already have.
• Qualification. Asking the five questions your team always asks first, so that when a human calls back, they are calling back a lead that is worth their time.
• Repetitive outbound. Appointment reminders, payment reminders, delivery confirmations, satisfaction follow-ups.
• Languages. Handling the first minute in the caller’s language and routing accordingly.

Where they fall short

• Emotionally loaded calls. Complaints, bereavement, anything where the caller is upset. The assistant should recognise this quickly and hand over, not try to resolve it.
• Negotiation and judgement. Discounts, exceptions to policy, “can you make an exception just this once”. These need a human with authority.
• Very noisy environments. Construction sites, busy roads, poor mobile reception. Speech recognition has improved enormously but it is not magic.
• Anything you would not put in writing. If the answer depends on nuance you would not trust to an email template, do not trust it to a voice assistant either.

The hand-off is the product

The thing that separates a useful assistant from an infuriating one is how it gives up. We design every assistant with:

1. A short path to a person. “Speak to someone” works at any point, and the assistant offers it itself when confidence drops or the caller repeats themselves.
2. A warm transfer. The human receives a two-line summary — who is calling, what about, what has been collected — so the caller does not start again.
3. A fallback when nobody is available. Take a message with a promised callback time, create the ticket, and actually create it.

Where the data lives, and what it costs

The assistant connects to your calendar, CRM or order system through their APIs, in your accounts. Call audio and transcripts are stored where you decide, with retention you set. Running costs are mostly per-minute telephony and model usage; we estimate them from your call volume before the project starts and show you the number.

How to tell if it is for you

Count the calls your team takes in a week that fall into the first list above. If it is a meaningful fraction, an assistant pays for itself quickly and your people get their day back for the calls that need them. If almost every call is in the second list, we will tell you that, and suggest something else.