Cognera answers your phone and your website chat from a machine in your building. It listens, understands, answers from your own documents, and hands off to a person when it should — and the recording, the transcript, and the answer never leave the premises.
That is the whole point. Most businesses that would benefit most from an AI receptionist are the ones told they cannot use one.
Speech in, speech out, entirely on your own machine. Every stage of the round trip runs locally — there is no step where audio is uploaded to be processed.
A caller speaks. The audio is transcribed on your machine. The question is matched against your own indexed documents, answered by a local model, and spoken back in a natural voice. Nothing is sent out for processing at any point in that chain.
It knows what you have told it and admits what it has not. Asked something outside its material, it says so and offers to take a message rather than inventing a price, a policy, or an appointment.
Hours, services, pricing, policies, directions, availability — whatever your documents actually say.
Captures the caller’s details and the reason for the call when the answer is not something it should attempt.
Escalates to a person on the conditions you set, rather than trapping the caller in a loop.
The part people mean when they ask whether it “knows our business.” It does, and here is the actual mechanism — no magic, and nothing uploaded.
The things that already describe your business — service descriptions, policies, price lists, FAQs, manuals. Markdown, plain text, PDF, or Word.
Each document is split into passages and converted into a numeric form by an embedding model running on your machine. The result is one index file on your own disk.
An incoming question is converted the same way and compared against every passage. The closest ones are retrieved — and only those clearing a confidence threshold.
Those passages become the source material for the reply. If nothing clears the threshold, the honest answer is that it does not know.
Two settings decide how talkative it is: how many passages to retrieve, and how confident a match must be. Set them too tight and a well-stocked knowledge base will still answer with almost nothing — which is usually the real cause when retrieval appears to “stop working.”
The most common complaint about automated answering is not that it sounds robotic. It is that it forgets — making the caller repeat a name, a number, or the reason they rang.
What the caller actually rang about is held for the whole call, so the answer at minute four still relates to the question at minute one.
A caller who says “no, Tuesday” is not arguing with a system that keeps offering Monday. The correction sticks.
Name, number, email and the reason for the call are captured and kept, so a message arrives complete rather than as a fragment.
Anything outside the approved documents stays outside them. It says so and offers to take a message.
Currently in development, and listed here because it is fair to know what a system does not do yet. Memory today lasts for the duration of a call.
Three reasons, in the order they usually start to matter.
Some calls cannot go to a third-party API — not as a preference, but because of the industry you are in. Local execution is the only answer that survives that question.
Metered AI turns usage into a variable cost that grows exactly when business is good. Owned hardware is a fixed cost that stops growing.
Hosted models get deprecated, repriced, and have their behaviour changed underneath you. A model file on your disk does none of those things.
Secondary to the agent, and priced as additions rather than as separate products. Both reuse the hardware and the model runtime already installed.
Generates a complete marketing site from a short description of the business, with a chat concierge grounded in that business’s own information — the same retrieval mechanism as the agent.
Image and video generation and upscaling on your own GPU. No queue, no credit balance, no per-image cost once the hardware is in place.
Sold as a setup on your hardware, not as a per-seat subscription. Hardware is separate; a capable GPU you already own is used as-is.
The voice agent, installed
Added to an agent install
Individual capabilities
The first conversation is about whether local execution is actually right for your situation. Sometimes it is not, and that is a useful thing to establish early.
