Before you open the panel

A good agent starts on paper. Choose one type of call, for example appointment confirmations or handling quote requests, and write for it:

  • the purpose of the call, in one sentence;
  • what information the agent needs to find out, in the order a colleague would ask for it;
  • what it is allowed to say and what it is not (prices, deadlines, promises);
  • when it transfers to a person and to which number;
  • what must remain after the call: a result, a few fields, a note.

If you cannot write that on one page, the flow is too large for the first agent. Break it up.

1. Create the agent

New agents start empty, without a template. Give it a name that says what it does („Appointment confirmations”, not „Agent 1”) and set the conversation language.

2. Write the first reply and the instructions

The opening line is what the person hears immediately after answering or after the recording announcement. Brief: who is calling and why.

Instructions are the text by which the agent guides the conversation. A few rules that make the difference:

  • write as if for a new colleague, with example phrases, not lists of keywords;
  • ask for short answers, one idea per turn; an agent that talks a lot is hard to interrupt;
  • write correctly, with diacritics, because the voice reads exactly what is written;
  • say explicitly what it does when it does not know: it does not invent, it notes the question and promises a call back.

3. Choose the voice

The voice is chosen from the ElevenLabs library directly from the agent record. Listen to a few variants on your real replies, not on the test text. If you use a cloned voice, you need the consent of the person whose voice it is.

4. Add the knowledge

The information that changes (schedule, approved prices, delivery conditions, addresses) is not kept in instructions, but in the agent's knowledge base. Upload short and clear documents; when something changes, update the document, not the scenario.

5. Give it tools

The tools are what the agent can do during the call:

  • ending the call, when the conversation has ended;
  • the transfer to a number, for the situations you wrote at the beginning;
  • webhook tools, if the agent needs to read or write something in your system, for example the status of an order.

For each webhook tool, write a clear description of when it is used. The agent decides based on the description.

6. Say what you want to remain after the call

In the analysis settings, define the evaluation criteria (did the client confirm? did they ask for a human?) and the fields to extract (desired date, name, reason for refusal). These appear on every call in the history and can be sent to the CRM.

7. Test it like a difficult client

Start a test call to your own number from the agent’s profile. Do not stop at “hello, I want an appointment”:

  • have long conversations, of 6–12 turns;
  • interrupt it in the middle of a sentence;
  • respond vaguely (“maybe”, “I don’t know, we’ll see”);
  • ask for something it is not allowed to promise;
  • change the language or the topic halfway through.

After each test, read the transcription and adjust the instructions. Usually the first problems are replies that are too long and questions asked twice.

8. Put it on a line

Connect the agent to a number or a forwarding rule. On lines that pass through Kallina’s telephony server, the recording notice is played automatically before recording, so do not add it in the first reply.

Start with part of the calls, listen to a few conversations each day in the first week, and only then scale up.

Common mistakes with the first agent

  • One agent for everything. Reception, sales and complaints in the same scenario produce an agent that is unsure in all three.
  • Information in instructions. The price changed next month remains forgotten in the scenario.
  • Short tests. An agent that gets through three polite lines can fail at the first objection.
  • No human hand-off. Each agent needs a transfer or a call back for what it does not know.