Every inbound call answered — 2am included

What an inbound call actually sounds like

The call arrives on your number — an existing SIP trunk, a Twilio number, or the voice widget on your site — and the agent answers immediately. It speaks natural Hebrew, and if the caller talks over it, it stops and listens (barge-in), exactly as a human agent would. When the caller goes quiet it gently checks whether they are still there instead of leaving dead air, and after a long silence it closes the call politely. Noise cancellation is tuned for telephony and strips background human voices too — not just fans and air conditioning — which is precisely what ruins transcripts in a noisy office or a call from the car.

• Answered on the first ring, calls in parallel, no hold queue
• Two engines to choose from: speech-to-speech for fast turn-taking, or STT→LLM→TTS for tighter script control
• Barge-in: the caller can interrupt the agent at any moment
• A gentle nudge when the caller goes silent, and a clean hang-up after a long one
• Noise cancellation that removes background human speech, not only mechanical noise

What the agent actually knows

The agent answers from your business's knowledge base — documents, Q&A, procedures, price lists — and every entry needs human approval before it can be used. An unapproved entry is simply invisible to the agent, which is the simplest possible guard against a wrong answer slipping in. Alongside the knowledge base you can upload a CSV or Excel file — a product catalogue, a branch list, a stock table — and the agent can filter and enumerate it. The difference matters: "which brands do you carry?" is a question semantic search cannot fully answer, because it returns similar passages rather than a complete list — a data table can.

• Retrieval-backed knowledge base (RAG) — only human-approved entries reach a call
• Data tables from CSV/Excel for exact filtering and enumeration
• External tools: the agent can call your systems mid-conversation
• Guardrails that block out-of-scope answers
• When guardrails block repeatedly, the call auto-escalates instead of looping the caller

A returning caller does not start over

When someone calls again, the agent recognises them by phone number and enters the conversation with the context of last time: what you discussed, what was agreed, what is still open. It does not read the history aloud — it refers to it the way an agent who remembers would: "last time we opened a ticket about the boiler, is that what you're calling about?". That memory crosses channels: someone who called yesterday and messages on WhatsApp today continues from the same point, because identification is by the digits of the number, not by the channel. Conversations that have been anonymised drop out of this memory entirely.

Handing off to a human, two ways

Some calls need a person, and the switch has to be seamless. A cold transfer hands the caller to the destination line and the agent steps out. A bridged transfer does something different: the system dials your rep — including a specific PBX extension — brings them into the live call, and only then does the agent leave, so the caller never gets a call-back ring or lands in an IVR menu. Handoff triggers on rules you define, on an explicit request from the caller, or automatically when guardrails have blocked answers repeatedly.

• Cold transfer: the call is routed to the destination line
• Bridged transfer: dial the rep, extension included, and join them to the same call
• Works for widget callers too, with no phone line on the caller's side
• Recording and summarisation keep running after the agent leaves

What is left after the call ends

Every call produces a full record: recording, transcript, summary, sentiment, an escalation reason where there was one, and a resolution status. That is the difference between "we got a lot of calls" and knowing exactly what callers asked and where they got stuck. Identifying details are detected and masked in the transcript by an engine tuned for Hebrew, alongside a full anonymisation flow for right-to-be-forgotten requests — including a reference number you can quote back to the person and an audit trail of who ran what and when. Access to recordings and transcripts is controlled by role.

When AI inbound answering is not the right fit

The agent is strong on calls with structure: recurring questions, collecting details, booking, checking status, routing and screening. It is weaker where the value of the call is the human contact itself — a negotiation, sensitive news, or an angry customer who has already tried three times; there the goal is for the agent to detect it early and transfer, not to persist. Two practical caveats: if you have no written content to answer from, the agent will not invent it — a real rollout starts with writing the knowledge base, not with pointing a number. And if your process requires reaching a system that has no interface at all, that has to be solved first rather than worked around.