This page is for anyone who asks how it is built: how we know who each call is for, what separates one business from another, how long it takes to answer and what happens on the day something goes down.
1.2 seconds
From when you stop talking to when the agent starts speaking2 s natural · 4 s, too long
Each response has its time written next to itMandatory
Without a phone number to forward to, an agent cannot be publishedWithout cutting off calls
A machine is withdrawn once it finishes what it has in progressThe first thing the system looks at is not who is calling, but which number they have called. That number belongs to a specific business and a specific agent, and that settles everything else: which voice answers, which calendar is checked and which data can be looked up.
If you have several numbers, each one can lead to a different agent: one for appointments, another for faults. You decide which number goes where.
It is used to look up the record within that business and only within that business, to greet the caller by name and pick up where the last conversation left off.
They come in through the switchboard and are placed wherever there is room at that moment. If a machine is taken out of service, it stops receiving new calls before it shuts down.
In the panel, each number appears with the agent that answers it and the backup number.
The platform serves many businesses at once and none of them sees anything of the one next door. Every piece of data carries a label saying which business it belongs to, and that label travels with it wherever it is used.
Records, appointments, requests, knowledge and recordings are always requested with the business up front. A query without a business returns nothing.
The keys used to query your system or charge a card are yours, they are encrypted and they are not used to handle anyone else’s call.
Owner, editor, read only, billing and the platform team. There is a record of who changed what, and of every time someone listens to a recording.
We work to a target of 1.2 seconds between the caller finishing speaking and the agent starting to answer.
The panel records the time of each turn, so the target is checked call by call. Up to 2 seconds reads as a normal pause; above 4 it is too long and the turn is flagged, so that you can look at what slowed it down.
Each turn of the call, with what was said, how long it took and what the agent looked up.
While it talks, the agent checks your calendar and the systems you have connected. Those look-ups have a time limit, because on the other end there is a person with the phone to their ear.
Finding a slot, opening a request or asking about an invoice are different things and cannot take the same time. Each one is given a margin, and when it runs out the agent says so aloud and offers to call back later or to transfer the call to a person.
A look-up that ran over time is written into the turn, with how long it took.
If the data lives in your system (an invoice, a debt, the status of a repair, a contract), it is requested at that moment and passed on. No copy is kept here to go stale and end up saying something that is no longer true.
And for each piece of data that comes in, it is decided in advance what the agent does with it: says it as it is, only confirms it if the person says it first, or never brings it up on the call.
An agent that improvises a free time or makes up a price creates a problem for you at the front desk. That is why what it can say is tied to what actually exists.
Times come from your calendar at that very moment, with the length of the service and the right professional. If there is no slot left, it says so and suggests another day.
It answers with what you have given it and with what your systems return. When there is no answer, it says so and opens a request with its reference.
It checks who is accepting calls at that moment and transfers. If nobody is, it takes down the reason and leaves the request open with its deadline.
In “Ask your data” the model understands the question, but the sum is done by the server on your data. Nobody makes up the numbers.
Before an agent takes its first call, you have to say which phone number calls are forwarded to if the service is unavailable. Without that number, it cannot be published.
It can be the mobile of whoever is on duty, the front desk landline or your usual switchboard: whatever you prefer. The worst case is that the phone rings where it used to ring before.
Turning what the caller says into text, and the agent’s text into speech, runs on machines with graphics cards. When one needs updating or a part replaced, it is taken out of the call rotation: it stops receiving new calls, finishes the ones it has and only then shuts down.
A cloned voice is made from ten or twelve seconds of recording and the express consent of the person lending it.
Speech recognition
What the caller says is turned into text with NVIDIA Parakeet, on NVIDIA graphics cards in our own machines. Neither the audio nor the transcript goes out to a third-party service to be recognised.
What we do not do yet: the agent speaks those six languages and no others, not Catalan, Basque or Galician. Synchronisation with calendars from other providers is in preparation, as is the delivery side. We do not describe it as if it were done.
We show the panel, the transcript with the timings and how the agent connects to what you already have. And if there is something we do not do, we tell you on that same call.