Short answer. An AI voice agent is a chain of four links: the telephony that brings it the call, speech recognition, a model that decides by consulting your tools, and the synthesis that answers. It takes the repetitive calls, hours, appointments, tracking, and hands the others to a person with a summary. It is judged on two figures: the share of calls resolved, and the response time. In this page's example, of 530 calls that were going unanswered in a week, the agent resolves 307 and hands over 159.
This page is for clinics, practices, dealerships, agencies, after-sales services and call centres whose phone overflows, in Algeria, and for diaspora founders setting up a service here. The definitions come from public sources: interactive voice response, speech recognition, word error rate, speech synthesis, code-switching, Algerian Arabic, telephony signalling protocol. The 1,400 calls, their split and the delays are a constructed example. The page covers neither the rules on call recording and personal data, nor telephony authorisations, and quotes no hardware or subscription price.
Five kinds of call
The figure sorts five kinds of call, and what the agent does with each.
Before AI. Interactive voice response is an automated telephony system that interacts with callers, gathers information and routes the call; it understands keys or a few words. A voice agent goes further: it understands a sentence, and acts. What it takes. Repetitive calls whose answer can be checked: opening hours, an appointment, the status of an order. What it does not take. A negotiated price, a complaint, an emergency: there, a person decides.
The rule. The agent handles only what its instructions allow, and says it is an automated assistant in its first sentence. Handover. At the first sign of misunderstanding, anger or urgency, it passes the call to a person, with what has already been said: the caller does not repeat. On other channels. The same requests arrive by messaging; the written agent is described in the AI agent on WhatsApp, and the general approach in the AI agent in a company.
The calls nobody was taking
The figure follows the 1,400 calls of one week.
In the example. 1,400 calls in one week on one number. The reception takes 870, that is 62%. 530 get no answer: 220 at peak hours, when every line is busy, and 310 after closing. What is given to the agent. Those 530 calls, and only those, to begin with. It resolves 307, hands 159 to a person with a summary, and 64 hang up before the end.
What that says. The agent takes no call away from the reception: it takes those that were being lost. The 64 abandoned calls. They are listened to again: a word not understood, too long a wait, a request out of scope. That is next week's work list. The call log. The switchboard already gives these figures, calls received, taken, missed, per hour, see the IP switchboard. What is measured. Calls taken by the agent; share resolved without handover; share handed over; abandoned calls; requests not understood.
«The agent takes no call away from the reception: it takes those that were being lost.»
A call, in six steps
The figure follows a call taken by the agent.
Arrival. The call arrives through your switchboard: the signalling protocol sets up the session and passes it to the agent when nobody picks up, or outside opening hours. The greeting. One short sentence, saying it is an automated assistant and offering to speak to a person. The request. Speech recognition turns speech into text; the model looks for the intent in it: an appointment, tracking, something else.
The action. That is the accented step. The agent does not answer from memory: it consults the diary, the order or the file, through tools with limited rights, see what the agent is allowed to do. An appointment taken follows the rules of booking without double booking. The answer. Speech synthesis, the artificial production of speech, reads the answer out; a confirmation message goes out in writing. The record. A summary, the next step, and the call attached to the customer's record, see the CRM.
Response time, and language
The figure breaks down one turn, from the end of the question to the start of the answer.
Four delays. In the example: 400 milliseconds to be sure the person has finished speaking, 200 of transcription, 700 for the model and the look-up in the diary, 200 before the first sound. Total: 1.5 seconds of silence. The example's markers. Under one second the exchange stays fluid; beyond two, the caller says “hello?”. Where to gain. On the model and the tool: a short answer, a diary that responds fast, the start of a sentence sent while the rest is being computed. Barge-in. Voice activity detection spots the presence or absence of speech: as soon as the person speaks again, the agent stops.
Language. Algerian Arabic, Darja, is a variety of Arabic spoken in Algeria, a language of daily communication, rarely written, with loanwords from French. Many callers switch from one language to the other within a sentence: that is code-switching. What that changes. Recognition accuracy varies with accent, dialect, speed and background noise; it is measured by the word error rate, which counts substitutions, deletions and insertions. A language that is mostly spoken, rarely written and mixed with French is a harder case for these systems; how much harder, only your calls can say. What we do. We do not promise a figure: we measure on two hundred of your calls, per language, before opening. If Darija dominates, the agent starts with the simplest requests, or offers the keypad. The method is described in measuring that an AI works.
«Darija is measured, not promised: two hundred of your calls, before opening.»
Eight weeks, starting by listening
Eight weeks in the example, before opening in the daytime.
Listening. Two weeks on two hundred real calls, with people's agreement according to your obligations: what people ask, in which language, with which words. That is where the three requests to automate and the test set come from. Telephony. Forwarding to the agent and transfer to a person are set on your switchboard. The tools. The diary and the files are connected for reading, then for writing appointments.
Opening. First the calls after closing, where nobody answered: the risk is lowest. Then daytime overflow. Ongoing listening. Every week a sample of calls is listened to again, and the words not understood are added. Where the model runs. At a provider or on your servers, depending on the confidentiality of your calls, see deploying a model on your own servers. What the agent does not do. It does not replace the reception, does not negotiate and does not decide: it takes what was being lost. The durations are those of a constructed example.
What is inside a voice agent
The figure stacks five layers; the third says what the agent is allowed to do.
- What you receive: forwarding of untaken calls to the agent; three requests handled end to end; the diary and files connected; handover to a person with a summary; the confirmation message; the table of calls resolved, handed over and abandoned; sampled listening.
- What the company avoids: the unanswered call at peak hour, the phone ringing into the void after closing, the same answers repeated all day, the appointment noted on a slip of paper, the caller repeating everything at each transfer.
- What is measured: calls taken by the agent; share resolved without handover; share handed over; abandoned calls; response time per turn; word error rate per language; requests not understood.
- What Symloop does: listening to your calls, connecting to your switchboard, the agent and its instructions, tools connected to your diary and your files, trials per language, weekly follow-up, in Arabic and in French.
Telephony and voice. The call arrives through your switchboard; voice covers recognition, end-of-speech detection, barge-in and synthesis. The agent. That is the accented layer: a model, written instructions, a closed list of requests, and guardrails. It gives no price that is not in its base, promises nothing, and hands over when in doubt.
Tools and follow-up. Diary, orders, files, answer base; summaries, resolution rate, words not understood. Ready-made or custom. Ready-made voice agent services exist, and suit a trial in French. Custom work is justified when your switchboard, your diary and your files have to be connected, your callers' languages handled, the calls kept on your premises, and what the agent understood measured week after week.
- What is an AI voice agent?
- A program that answers the phone: it transcribes what the person says, understands the request, consults your tools and answers aloud. It handles repetitive calls and hands the others to a person.
- Does an AI voice agent understand Darija?
- That is measured, not promised. A language that is mostly spoken and mixed with French is a harder case for speech recognition. We test on two hundred of your calls, per language, before opening.
- Which calls can it take?
- Those whose answer is fixed or checkable: hours, address, booking or moving an appointment, tracking an order or a file. Negotiated prices, complaints and emergencies go to a person.
- What happens when it does not understand?
- It asks the question once more, then hands over to a person with a summary of what was said. Outside opening hours, it takes the number and the request for a call-back.
- Does the switchboard have to be changed?
- Not necessarily. The agent connects to an IP switchboard through the usual signalling protocol; an older switchboard may need a gateway.
- How long does it take to answer?
- In the example, a second and a half between the end of the question and the start of the answer. The aim is to get under one second, stage by stage.
- Can the agent book an appointment?
- Yes, if it is connected to the diary: it offers free slots, holds one, and sends a confirmation by message.
- Do callers know they are speaking to a machine?
- Yes: the agent says so in its first sentence, and offers to speak to a person. It is a rule we set in its instructions.
- Where are the calls processed?
- At the model provider or on your own servers, depending on the confidentiality of your calls. The choice is made at scoping.
- Where do you start?
- With the call log: how many missed calls, at what hours, for which requests. Then with calls after closing, where nobody answered.