Skip to content
Client login

Free Audit

Artificial intelligence

AI voice & call agents

An agent that answers, books an appointment or takes a simple order, and hands over to a person the moment the call leaves the script.

A demonstration

An assistant answering, on a real site.

An imaginary pizzeria in Blida, one page. You speak to Zitouna aloud, in derja, Arabic, French or English.

Try the demo (opens in a new tab)

The business is fictional: no order placed here is real.

An unanswered call is lost, not postponed. The caller dials the next number, and in most trades they do not ring back. That is what a voice agent replaces: not your switchboard, but the calls ringing out at midday, on Fridays, and through the three weeks somebody is on leave.

We apply one test before accepting these projects: can the typical call be handled in thirty seconds with three pieces of information? Booking an appointment, confirming an address, saying whether an item is in stock, logging a complaint for callback — yes. Understanding a technical fault, settling a dispute, negotiating — no, and no amount of tuning changes that.

Voice is harder than text, and that has to be said before selling anything. A noisy line, a call from a moving vehicle, a caller speaking darja with numbers in French, a name transcribed as its homophone: each of these degrades transcription before the model even has to understand anything. We measure that on your real calls, and sometimes the result makes us advise against the project.

Finally, a voice agent is not a phone menu with a nicer voice. A keypad menu makes the caller wait in order to send them to the wrong department; the point here is to remove the menu, not to dress it up. If you want to keep the tree, the saving will not happen.

What we usually find

  • Some of your calls ring out at peak times, and you do not know which ones.
  • Appointments are booked by phone and written in a notebook nobody else can read.
  • Your phone menu keeps callers waiting and then routes them to the wrong extension one time in three.
  • The same call — "are you open today" — occupies your team all day long.

What changes

  • Missed calls measured, then recovered

    First the real count of unanswered calls by time of day — often the most useful number in the project — then the share the agent handles.

  • A handover that does not restart the call

    When the agent passes over, the person receives the reason and what has already been said. A caller made to repeat everything would have preferred it to ring out.

  • A written trace of every call

    Transcript, decision taken, appointment created. It is also what lets you see, a month in, what the agent consistently mishears.

What you get

  • Booking into your real calendar

    The agent writes into the calendar your team already uses, under the same slot rules. A second calendar to reconcile cancels the gain.

  • Simple orders and status callbacks

    Low-variation orders taken and read back before confirmation. Anything involving an option or a discount goes to a person.

  • Replacing the keypad menu

    The caller says what they want in one sentence and lands in the right place. That is most of the felt benefit, before any automation at all.

  • Tested on your real recordings

    Your past calls replayed to measure transcription before you commit. Studio demos predict nothing about what happens on a real line.

  • Immediate switch to a human

    On explicit request, on silence, on detected frustration, or after two misunderstandings. The threshold is set with you, and it is set low.

  • Declared hours and holidays

    What the agent covers and when, said to the caller instead of letting them believe somebody is there.

How we work

  1. Counting the missed calls

    One to two weeks of measurement before any decision. Sometimes that count is the whole answer, and extra human cover at peak is the right response.

  2. Transcription test on your lines

    Your recordings run through the recognition engine, with an error rate measured per call type. That is the door to the project, and it is sometimes shut.

  3. Writing the script and its exits

    The three or four paths the agent holds, and every exit to a human. A script without an exit becomes a trap for the caller.

  4. Launch on a secondary line

    First on an overflow number at peak hours, never on the main line. Full switchover only follows measured results.

Is this the right fit for you?

This is for you if

  • You lose calls measurably, at hours that can be identified.
  • Your typical calls resolve in under a minute with little information.
  • You have a calendar or order system reachable programmatically.

This is not for you if

  • Your calls are diagnoses, complaints or negotiations.
  • Your line quality is poor and your callers are frequently on the move.
  • You want the agent to pass as human — we do not build that.

What we commit to

  • The agent says it is a machine

    At the start of the call, plainly. We turn down projects that ask for the opposite, including when the client insists: being found out is paid for in trust.

  • Nothing firm without confirmation

    An appointment or order is read back to the caller before being saved, and confirmed in writing. Voice leaves the caller no record.

  • We will tell you when the line cannot carry it

    If the transcription test comes back poor, we stop at scoping and you do not pay for a rollout that will frustrate your customers.

What changes for an Algerian business

The phone remains the decision channel for a substantial share of purchases, even when discovery happened on Instagram. Somebody who saw a product online rings to check availability and price before travelling — and that short, repetitive call is exactly what an agent handles well.

Mixed language is harder spoken than written. In writing, a darja sentence in Latin script stays readable; spoken, the switch between darja, French and Arabic within one sentence, with numbers in one language and street names in another, degrades recognition before any understanding is attempted. That is the reason for the upfront test.

Line quality is part of the project. Calls arrive overwhelmingly from mobiles, often on the move, and a voice agent amplifies that weakness rather than compensating for it. When that happens, diverting to a written message gives a better result than an agent that asks for a third repetition.

Frequently asked questions

Does the caller know they are talking to a machine?

Yes, from the first sentence. It is a condition we set rather than an option: a customer who works it out mid-call feels deceived.

What happens if the agent does not understand?

It transfers after two attempts, or immediately if the caller asks. The reason and the exchange travel with the call.

Can it work in darja?

Partly, and that is precisely what the upfront test on your recordings measures. We give you the error rate before you decide.

Do we need to change our phone system?

Not necessarily. We work with your existing installation; an overflow number is the only usual addition.

Does it replace our switchboard?

No, and we do not propose that. It takes the calls nobody currently takes, which is a more modest aim and a far more achievable one.

How long to know whether it works?

The transcription test takes a few days and decides everything. The supervised rollout that follows runs a few weeks.

How to start

Give us about thirty recordings of ordinary calls, without picking the clearest ones.

We return an error rate per call type and our view, including if that view is not to do the project.

What we have written on this subject

Let us talk about your project

A free audit, no commitment: we look at your online presence and tell you what is holding it back.

  • The agent says it is a machine
  • Nothing firm without confirmation
  • We will tell you when the line cannot carry it

We measure how this site is used with Google Analytics, to learn which pages actually help. You can stop that measurement at any time from the footer. Cookie policy