Enterprise AI CallcenterOn-premises
How an AI call centre works: the whole call stays on your servers
The call arrives through your SIP provider or PBX. From there on, everything — speech recognition, deciding what to say, voice synthesis, recordings — happens inside your infrastructure.
Architecture
The components
Telephony gateway
Runs on your server and connects to your SIP trunk or PBX. Inbound and outbound calls, busy detection, transfer to a colleague.
Speech recognition
Our own speech recognition, running on the GPU server, tuned for phone audio and for each language we deliver.
Voice
Our own neural voices, optimised for phone conversations and correct stress. A custom voice for your brand on request.
Conversation control
Who speaks when, caller barge-in, pauses, repeat protection, AI disclosure and recording notice.
Language model
Phrases the answers and decides when to call a tool. Your existing provider via API, or a fully local model with no internet.
Control room
Agents and versions, campaigns, do-not-call list, calling hours, recordings, transcripts, data extraction, client portals, roles, audit log.
The path of one turn — with measured times
- The caller speaks; audio arrives over SIP at the gateway on your server.
- Speech recognition turns speech into text while the caller is still finishing.
- The language model decides the answer and calls a tool if needed (e.g. a balance lookup in the CRM).
- The voice synthesises the answer in pieces, so the first sound starts before the whole answer is ready.
- Call, transcript and extracted fields are stored in the hub; a webhook notifies your systems.
Times are lab and test-call measurements (October 2026). The end-to-end turn time depends mostly on the language model you choose.
What happens on your servers during a call
A call reaches your telephony gateway through your existing SIP trunk or PBX. From there it goes to our own speech recognition, which turns the caller's words into text inside your building. The conversation layer decides what to do next, with the language model of your choice: your existing provider through an API, in which case only the text of the turns is sent, or a fully local model with no internet connection.
If the agent needs data, it calls your systems through HTTP tools, using only the actions you allowed. The answer then goes to our own voice models, which speak it back into the call. Recordings, transcripts and logs stay on your servers, with the retention you set per client.
Because every step runs inside your infrastructure, your security team can check the data flow, the network rules and the audit log in the same way as any other internal system. Nothing about the call depends on a cloud speech service, so you keep control of the data as well as the cost.
During a pilot we measure these steps on your own hardware and with your chosen language model, because that model is the largest part of the delay. You get the numbers for your set-up, and your team sees how the agent behaves on real calls before anything is rolled out. The same control room then shows live calls, campaigns, recordings and the audit log for every agent.
Hardware and capacity
We supply enterprise GPU servers sized to your volume. This is what we measured on one of our lab servers:
| Configuration | Simultaneous calls | Voice first audio (median / p95) |
|---|---|---|
| Enterprise GPU server — standard setting | 50,000+ | 136 / 178 ms |
| Enterprise GPU server — maximum-accuracy setting | 16 | p95 156 ms |
| CPU-only server (small deployments) | 6 | 97 / 179 ms |
Measured with real speech recognition on recorded Bulgarian speech. Production servers are sized for your volume; higher volume means more servers.
Frequently asked questions
Does everything really run on our servers?
Yes: our own speech recognition and voice models, conversation control, the telephony gateway, campaigns and recordings. The language model is your choice: your existing provider via API (then only the text of the turns is sent) or a fully local model with no internet.
What hardware is needed?
We supply enterprise GPU servers sized to your volume. For reference, one of our lab servers handled 50,000+ simultaneous calls. See step 2 of the configurator.
How fast does the agent answer?
In our test calls with a cloud language model the median time from the end of the caller's sentence to the first sound of the answer was 1.4 s; the voice itself starts 136 ms after the answer text is ready. The language model is the largest part, so we measure it with your choice in the pilot.
Which telephony does it work with?
Any standard SIP trunk or PBX: your existing numbers and carrier stay in place.
See it working on your own scenario
Book a 20-minute call, listen to the agent, or let us build a free sample agent for one of your call types. No commitment.
- GDPR by design
- Runs on your own servers
- 24/7 operation and support
- A team of 5 in Sofia, EU
Already running in production at organisations in Bulgaria — names under NDA.