“We chose a local vendor, so our calls stay local.”
The sentence comes up in almost every voice agent procurement. It goes into the minutes, it reassures the data protection officer, and it is rarely checked again. It is often true of the company on the invoice. That just isn't the company that hears most of the call.
Why it is easy to believe
There are three good reasons the sentence survives.
The contract. You have one agreement, one invoice and one name to deal with. Everything that happens behind that name sits in a sub-processor annex that few people read after signing.
The marketing. Sovereignty has become a selling point in its own right. When OVH completed its acquisition of the French transcription service Gladia this summer, “deeper European data residency guarantees” was one of the headline arguments. It is a real argument – for the one link it applies to.
The telephony. The number is local, the carrier is local, and the call arrives on a local network. The first link in the chain is often exactly as local as it sounds. It is easy to let the first link stand for the whole chain.
What actually happens to a call
A customer calls and gives her name, date of birth and what she is calling about. Before the agent has answered her, the call has usually passed through these links:
- The phone line. The carrier receives the call and passes the audio on. This part is often local all the way.
- The platform. The company you have a contract with, which runs the conversation. This is where the configuration lives, and usually the logs.
- Speech recognition. The audio becomes text – with the name, the date of birth and the reason for calling in it. This is often a separate, specialised service.
- The language model. The text is sent to the model that decides what the agent will say. The most widely used models come from a handful of large US companies. Many of them offer processing in Europe, but that is a choice someone has to have made – it does not happen by itself.
- Speech synthesis. The reply is turned back into audio. Sometimes by yet another vendor.
- Storage afterwards. Recordings, transcripts and analytics can sit somewhere other than where the call was processed in real time.
If the agent runs on a speech-to-speech model, links three, four and five collapse into one. That makes the chain shorter. It does not automatically make it more local.
Each of the six links can be a separate company, in a separate country, with its own sub-processors. “Our calls stay local” is only true if it is true for all six. It is the same pattern that runs through the rest of a voice agent: every link in the call has its own owner, and capacity is the lowest of four numbers owned by four parties. Data flow follows the same logic.
