The recording becomes text locally, every call is classified cheaply, and deep analysis runs only for the sales ones. That is how a hundred per cent of calls get reviewed without the bill running away.
The call is recorded by the phone system, turned into text on your own server, then passed through two AI steps: cheap classification of every call and deep analysis only of the sales ones. The deep pass extracts the customer's requirements, the products mentioned, what was agreed and the next step. Nobody listens to recordings by hand, and the bill stays small.
In a sales team the phone is the main channel and at the same time the only one that leaves no trace. Email is searchable, a quote is filed, a conversation disappears the moment you hang up.
The consequences are familiar: a promise nobody remembers; an enquiry that never became a quote because the salesperson was on the road; a question three customers asked in the same week that nobody noticed. Recordings on the phone system solve nothing — nobody has time to listen to twenty calls a day.
Audio on its own is unusable. The first step is transcription. There is a decision here with consequences: do the recordings leave your building.
A customer call contains names, phone numbers, addresses and sometimes prices that are not public. Local transcription settles the question at the root — the recording never leaves your infrastructure. Modern open models manage on a CPU, with no specialised hardware.
There is a second, purely practical reason: your own vocabulary. Transcription errs exactly where it matters — brands, product lines, part numbers. Feed it your own list of terms and those errors disappear.
Run deep analysis on every call and you pay a premium for material that is "when is the courier coming" eighty per cent of the time.
The split that works is two-stage. Every call goes through cheap classification — a category and a sentence or two of summary. Only the calls recognised as sales go on to deep analysis with a strong model. The result: a hundred per cent of calls reviewed, at cents a day.
Not a retelling. Structure: what the customer asked for, what quantities and budget were mentioned, which products were named, what was agreed, what the next step is, what objections came up.
The products mentioned are looked up in the live catalogue, so the result is a list of real items with prices rather than names in a block of text.
Recordings are collected from the phone system automatically. Transcription is local (faster-whisper, on CPU) with an editable company vocabulary — recordings never leave the server. Classification sorts calls into categories — sales, internal, service, supplier — and flags quote candidates. Deep analysis runs for those only.
A sales call can become a draft quote in one click: the customer is filled in, the proposed line items come from the analysis with catalogue prices, and the notes go into the terms. Nothing is sent on its own — the draft waits for a person.
The call is matched to a customer by phone number. Which salesperson handled it is determined by presence, by the extension mapping, or by how they introduced themselves at the start of the recording.
Dictations are the same mechanism for conversations outside the phone system. A salesperson who met a customer talks the meeting into a recorder on their phone, and the recording follows the same path — with a prompt for a monologue rather than a dialogue.
Everything runs through the shared AI cost accounting, with budgets and limits.
It is not a voice bot. The system does not answer the phone and does not talk to customers. It listens afterwards and organises what was said. The selling stays with the salesperson; what goes away is the note-taking and the forgetting.
Do the recordings leave the company? They do not have to, and they should not. In Kronixon PI transcription is local — it runs on your own server, on CPU. Only text goes out, and only for analysis.
Is every call analysed in depth? No. All calls are transcribed and classified cheaply; deep analysis runs only for the sales ones and for calls flagged by hand. That way a hundred per cent are reviewed without the cost growing.
What does this cost per day? Transcription is free because it is local — you pay only CPU time. Classification is under a cent per call. At twenty calls a day, deep analysis runs for a handful of them.
Is it a voice bot? No. The system does not speak to customers on the phone. It analyses conversations afterwards.
How does it know which salesperson handled the call? By presence on the extension, by the mapping of extensions to people, and where that is ambiguous, by how the person introduced themselves at the start of the transcript.
What about calls made outside the phone system? They are recorded as a dictation — from a web recorder or an uploaded file — and follow the same path, with a prompt for a retelling rather than a dialogue.
Is a quote created automatically? No. The analysis flags the call as a candidate and prepares a draft with proposed items. Sending it is a human decision.
Tell us what eats your team’s day. We will show you — on our own live system — how much of it can simply disappear.