Resources
How Katexs Works
A plain-language tour of what happens between a phone ringing and a booking landing in your calendar. No prior knowledge of speech systems required.
The short version
Katexs sits between your phone system and your business software. It listens to a caller, decides what to do using the instructions you wrote, does it by calling your tools, and speaks back — fast enough that the whole loop feels like a conversation rather than a transaction.
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Customer │────▶│ Phone or │────▶│ Katexs │
│ Calls or │ │ Chat │ │ Agent │
│ Messages │ │ Widget │ │ (AI Brain) │
└─────────────┘ └─────────────┘ └──────┬──────┘
│
┌──────────────────────────┼──────────────────────────┐
▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌─────────┐
│ Calendar│ │ CRM │ │ Transfer│
│ Book │ │ Push │ │ Human │
└─────────┘ └─────────┘ └─────────┘The loop, step by step
- 1
The call arrives
A caller dials your number, or a visitor opens the chat widget. Katexs answers within a ring and identifies which agent owns that number or key.
- 2
Speech becomes text
Audio streams in and is transcribed continuously, not in blocks. Partial words are available before the caller finishes a sentence, which is what makes fast responses possible.
- 3
The agent decides
Your instructions, the conversation so far, relevant knowledge base passages, and any variables are assembled into a turn. The model produces either something to say or a tool to call.
- 4
Tools do the work
Checking a calendar, creating a contact, looking up an order. Katexs calls your integration, waits, and feeds the result back into the conversation.
- 5
Text becomes speech
The reply is synthesised and streamed out while it is still being generated, so the caller hears the first words almost immediately.
- 6
The call ends and is recorded
Transcript, recording, extracted variables, outcome tag, and cost land in Call Logs, and your webhook receives the same payload.
The pieces you configure
| Piece | What it is | Where it lives |
|---|---|---|
| Agent | Instructions, voice, model, and tools — a single conversational worker. | Creator |
| Team | Several agents with routing and handoff rules between them. | Creator → Teams |
| Flow | A drawn conversation graph for deterministic paths. | Creator → Flows |
| Tool | A named action the agent can take, with a typed input schema. | Creator → Tools |
| Knowledge base | Documents the agent can quote from, chunked and indexed. | Assets |
| Number | A phone line bound to an agent, team, or flow. | Settings → Phone Numbers |
How the agent knows things
Three different mechanisms, often confused. Instructions are always present and shape behaviour. Knowledge is retrieved only when relevant and quoted for facts. Variables are injected per conversation and carry the specifics of this caller.
- Instructions — Short, imperative, about behaviour: who to transfer to, what to refuse, how to close. Not the place for your price list.
- Knowledge base — Long-lived facts: policies, product details, hours by location. Updated without touching the prompt.
- Variables — This caller, this order, this appointment. Fresh on every conversation.
Why latency is the whole game
People tolerate a pause of roughly half a second in conversation. Past that they start talking again, which produces the overlapping, stilted exchanges that make automated calls feel wrong. Katexs streams every stage and starts speaking before the full reply exists. The most common cause of a slow-feeling agent is not the model — it is a tool that takes three seconds to answer, or a prompt that has grown to two thousand words.
What Katexs does not do
- It is not your CRM — Katexs writes into your systems of record; it does not try to replace them.
- It is not a dialler by itself — Campaigns handle list-based outbound, but compliance and consent are yours to hold.
- It does not train on your calls — Your conversations are not used to improve any model. See Security & Privacy.