Gestión Smart
EN
Request a demo

How the AI voice agent works

How the Gestión Smart AI voice agent works

For technical and data protection leads: how each call is assigned to its business, how data is isolated, how long the agent takes to respond and what happens if the service fails. Speech recognition runs on our own GPUs.

2 s: natural pause 4 s: excessive delay

Without interrupting calls

Speech recognition: on our own GPUs

Last updated:

The number dialled picks business and agent

The system first identifies the number that was called, not who is calling. That number belongs to a specific business and a specific agent, and it determines everything else: which voice answers, which calendar is checked and which data can be accessed.

  • One number, one agent If you have several numbers, each one can be routed to a different agent: one for appointments and another for faults. You decide which agent answers each number.
  • The caller’s number, to recognise the customer It is used to look up the customer’s record, only within that business, to address them by name and know their previous requests.
  • Calls distributed across available servers Calls come in through the switchboard and are assigned to the server with available capacity at that moment. If a server is taken out of service, it stops receiving new calls before it shuts down.

The panel shows the calls in progress, those waiting and what callers hear, including when all lines are busy.

Switchboard screen in the panel: the business’s number, one call in progress on the plan’s single line, another waiting and the welcome message callers hear.

Data isolated per business

The platform serves many businesses at once and none of them can access another’s data. Every piece of data is linked to its business, and that link goes with it wherever it is used.

Every query, limited to its business

Records, appointments, requests, knowledge and recordings are always queried with the business identified. A query without a business returns no data.

Credentials belong to each business

The access keys to your system and your payment gateway belong to you, are stored encrypted and are never used to handle another business’s calls.

Five roles and a change log

Owner, editor, read only, billing and the platform team. Every change is logged with its author and time, as is every playback of a recording.

Response time: what callers notice most on the phone

We work to a target of 1.2 seconds from when the customer finishes speaking to when the agent starts to respond.

The panel records the time of each turn, so the target is checked call by call. Up to 2 seconds is perceived as a natural pause; above 4, the turn is flagged so that the cause can be analysed.

Every call, transcribed and measured

Each turn is logged with what was said, how long the agent took to respond and what it looked up. Transcription is done on GPUs in our own infrastructure: the audio is not sent to third parties.

Real-time look-ups with a time limit

During the conversation, the agent checks your calendar and the systems you have connected. Each look-up has a maximum time, because on the other end there is a customer waiting for an answer.

Maximum time per tool

Finding a slot, opening a request or looking up an invoice are different operations and cannot take the same time. Each one has a margin; if it runs out, the agent says so and offers to call back or to transfer the call to a person.

Look-ups that exceed their margin are recorded in the turn, with their duration.

Live data: looked up, not copied

If the data resides in your system (an invoice, a debt, the status of a repair or a contract), it is looked up at that moment. No copy that could go out of date is stored.

For each piece of data, you decide what the agent does: it states it, states it only if asked, confirms only the last four digits, uses it solely to identify the account holder, or never takes it out of your system. You can also require callers to identify themselves before they are given any data.

The agent only offers what exists in your systems

An agent that improvises a free time or a price creates problems for your business. That is why everything it can offer is limited to the actual information in your calendar and your systems.

It only offers available times

Times come from your calendar at that very moment, with the length of the service and the corresponding professional. If nothing is available, it says so and suggests another date.

Without data, it does not improvise

It answers with the information you have provided and the information your systems return. If it does not have the answer, it says so and opens a request with its reference.

Transfer to a person when needed

It checks who is available at that moment and transfers the call to them. If nobody is available, it takes down the name, the confirmed phone number and the reason as a callback request.

The figures in the panel are calculated by the server

In the panel’s natural-language queries, the model interprets the question, but the calculation is done by the server on your data. The model does not generate any figures.

Mandatory backup number

Before an agent takes its first call, you must specify which phone number calls are forwarded to if the service is unavailable. Without that number, the agent cannot be published.

It can be the mobile of the person on duty, the reception landline or your usual switchboard. In the worst case, the call rings where it used to ring before.

Our own voice infrastructure

Transcribing what the customer says and synthesising the agent’s voice run on our own servers with GPUs. To update or repair a server, it is taken out of the call rotation: it stops receiving new calls, completes the ones it has in progress and only then shuts down.

Voice cloning requires between 10 and 12 seconds of recording and the express consent of the person.

Speech recognition

Speech recognition runs on GPUs in our own infrastructure. The audio is not sent to third-party services.

Current limitations: the agent speaks six languages (Spanish, English, German, French, Italian and Portuguese); Catalan, Basque and Galician are not available yet and have no date. Coming soon: synchronisation with calendars from other providers. In development, no date: logistics and delivery.

Answers for your technical team

How long does the Gestión Smart voice agent take to respond?

The target is 1.2 seconds from when the customer finishes speaking to when the agent starts to respond. The panel records the time of each turn, so it can be checked call by call: up to 2 seconds is perceived as a natural pause and, above 4, the turn is flagged so that the cause can be analysed.

What happens if the system fails during a call?

The call is forwarded to the backup number. That number is mandatory: without a forwarding number, Gestión Smart does not allow an agent to be published.

How do you prevent a call from being answered with another company’s data?

The system first identifies the number that was called, not who is calling. That number belongs to a specific business and a specific agent, and the assignment is checked against the database. In addition, every query for records, appointments, requests or recordings includes the business: a query without a business returns no data.

Can the system be updated without interrupting calls in progress?

Yes. A server that is taken out of service stops receiving new calls before it shuts down and completes the ones it has in progress.

Can the agent give out-of-date information?

Not if the data resides in the business’s system. Invoices, debts, the status of a repair or a contract are looked up at the time of the call, and no copy that could go out of date is kept. For each piece of data, the business decides what the agent can do: state it, state it only if asked, confirm only the last four digits, use it solely to identify the account holder, or never take it out of its system.

What languages does the agent speak?

Spanish, English, German, French, Italian and Portuguese. Each business chooses which language the call starts in, and the caller can ask for another one mid-conversation: it is still the same call, with what they have already said and with their appointments. The AI notice plays in the business’s language, and the panel is available in those same six languages. Catalan, Basque and Galician are not available yet and have no date: they must first pass the same tests with real calls.

Where is call audio processed?

On our own servers with GPUs. Speech recognition is performed on that infrastructure, so the audio is not sent to third-party services. When a server needs updating, it stops receiving new calls, completes the ones it has in progress and only then shuts down.

We answer your technical team’s questions in the demo

We show you the panel, the transcript with the response times and how the agent connects to your systems. If something is not available, we tell you clearly.

Request a demo

See privacy and security

  • Up and running in 1 day
  • Additional integrations within 3 weeks at most
  • Free integration consulting