Omniat Ai · Inside the assistant

One call. One paragraph in reply. Nine pairs of hands in between.

An Ai voice assistant sounds like a single thing on the phone. It isn't. Every answer is produced by nine small, separate jobs, each with one responsibility and one handover. Scan the chain below; read only the actor you care about.

The nine actors Meet Omniat

Why we publish this

The architecture is the answer.Where answers come from, what stops invention, who decides on a hand-over, and what happens when the line goes quiet.

One call, in four beats

around ninety seconds, two turns

01 · RingPick-up, first secondNumber matched to a configuration, a reference attached, greeting spoken 02 · AskOne sentence inTranscribed, corrected, topic named from a fixed catalogue 03 · AnswerOne paragraph outFacts pulled from the client's library, drafted, polished, spoken — interruptible 04 · EndOne of four exitsGoodbye, transfer with context, or a silent release — for a dropped line or a caller gone quiet

That is the whole experience from the caller's side: a welcome, an answer, a clean ending — around ninety seconds for a two-turn call. Everything below is what makes those ninety seconds feel effortless.

The chain

a horseshoe — in and out through the same door

01 Receiver02 Transcriber03 Topics04 Knowledge05 Writer06 Polisher07 Voicenext turn ↺

08 Listener — awake in parallel while the assistant speaks 09 Call ender — wakes on any of four end signals

Audio enters at the receiver and leaves at the receiver — a horseshoe. Each turn runs 01 to 07 and loops back for the next question. The listener stands guard whenever the assistant is speaking, and the ender waits, patiently, for any of its four signals. Nine jobs, one conversation.

The nine actors

one responsibility each

Each card reads the same way: what goes in, what comes out, and the three things worth knowing. The first seven run in order on every turn. The last two are the quiet ones — and they are what make the assistant feel like a conversation rather than a recording.

01 Call receiver

ring → audio, both ways

The only actor touching the phone network — matches the number to a configuration in the first second
One reference per call; concurrent calls share nothing
Notices a dropped line and tells the ender; writes the transcript to the call record at close

02 Transcriber

audio clip → one clean line

Sound to text, no interpretation — dictation, not understanding
A per-client correction dictionary fixes known mishearings: brand names, currencies, abbreviations
Swaps applied by judgement, not find-and-replace — the cheapest quality fix in the chain

03 Topic identifier

transcript → topic labels

Labels come from a catalogue the client's team writes — never invented mid-call
The top quality risk: a wrong topic makes a fluent, irrelevant answer — worse than a refusal
Two flags outside the list — "a person, please" and "we're done" — confirmed first when ambiguous

04 Knowledge selector

topic labels → a briefing

The rule: two pinned documents plus one section per matched topic — nothing else loads
Never the open internet, never another client's library; account data arrives by its own scoped route
The client owns every section — named owner, edit date. Changing answers is editing, not engineering

05 Reply writer

transcript + briefing → one paragraph

Five rules: answer first · bundle only · four sentences max · close with a question · no financial or legal advice
Every claim must trace to a loaded sentence — no plausible additions about prices or limits
Empty briefing → says so, offers a transfer. A graceful refusal is a correct outcome

06 Reply polisher

draft → speakable text

Currency and numbers rewritten the way they are said aloud
A per-client pronunciation guide: the dictionary fixes the ears, the guide fixes the mouth
Spots the end-of-call marker, strips it, wakes the ender; a needed rewrite is a defect sent back

07 Voice generator

speakable text → streamed audio

One voice per brand, chosen at setup and locked — no drift in mood across the day
Pronunciation hints override the model's own guess, always
Streams as it generates: the first words reach the caller before the last words exist

08 Listener

caller audio → interrupt or silence

Awake while the assistant speaks; a real interruption cuts it mid-word and yields the floor
Tuned for politeness: breaths, doors and background TV ignored without missing real speech
A hard cap on how long the assistant may speak before pausing to check the caller is still there

09 Call ender

any end signal → a clean ending

Four signals, four exits: goodbye (end marker) · transfer with context (escalation) · silent release (no response, or line dropped)
Transfers announce themselves, bridge the call, and send the transcript so the caller never starts again
Silence is an ended call, not a fault — lines don't sit open, the record stays honest

Why the separation matters day to day: when something goes wrong on a call, the fault sits with exactly one actor. A mishearing belongs to the transcriber, a confident irrelevant answer to the topic identifier, a mangled brand name to the polisher. A team that can name the actor can fix the problem in an afternoon — which is the difference between an assistant that improves weekly and one that stays charming and wrong.

Where the assistant stops

drawn by access, not by subject

Every question a caller asks lands in one of three columns — and the middle one moves. A first deployment answers from published knowledge alone; each integration moves the line outward, with no change to how the assistant speaks.

Answered from the library

How it works, what it costs, coverage, limits, hours — anything the client has written down for the public.

Answered from a connected system

Balances, statuses, charges — retrieved live where an integration exists, and spoken under the same five rules as everything else. The chain keeps its shape; one more source feeds the briefing.

Handed to the team

Anything it cannot see, anything needing judgement about money or contracts — and any caller who asks for a person, without argument.

What never moves: the assistant never claims to have looked something up when it has not, and never builds a personal answer from general facts. If the value is not in front of it, the call goes to someone who can see it — with the conversation already on their screen.

Why it is built this way

Diagnosable

Every fault has an address, so quality work is surgical rather than superstitious — tune one actor, leave eight alone.

Edited, not engineered

Answers live in documents the client owns. Changing what the assistant says is an editing job, where the subject knowledge already sits.

Refusal by design

The assistant is allowed to have no answer. The chain makes the graceful hand-over the easier path than an invented fact.

Consistent

Same voice, same shape of reply, same closing question — first call and ten-thousandth.

Get in touch

Hear the chain, not a demo reel.

Call Omniat on a live number, interrupt it mid-sentence, ask it something it cannot know. The architecture on this page is what you will hear.