Voice agent: from an empty canvas to the first outbound campaign
One end-to-end example — an agent for the Amanat medical center that confirms appointments and reschedules them. It covers every capability of the platform: graph nodes, variables, CRM calls, knowledge bases, transfer to a human, voice, telephony, an outbound campaign and post-call review.- Format: self-paced, ~4 hours including practice
- Access required: the Admin or Developer role in the cabinet
- Outcome: your own agent that has passed testing and been published
How to read the courseThe modules follow the order in which an agent is actually built: first the dialog map, then the call settings, then the launch. Each module is a goal, steps in the interface, a screenshot of that very screen and the pitfalls people run into most often.Click a screenshot to open it at full size. Tall settings panels are shown in full — scroll the image.
Two market profilesThe course works for both calling market profiles — Kazakhstan (kz) and USA — TCPA (us_tcpa). The profile is set in the organization settings and defines the legal calling rules.Building the agent (sections I–V) barely depends on the profile. Market differences show up when calling starts (section VI):
- Calling window. For
kz, it is the organization’s calling window, which the campaign working hours can only narrow. Forus_tcpa, it is the legal window in the contact’s local time: federal rules, state rules and the campaign working hours, and the strictest wins. - Contact file. For
us_tcpa, thecontact_ziporcontact_statecolumn is required: the platform uses it to determine the contact’s local time. - Frequency limits. For
us_tcpa, they are on by default and are complemented by the state’s legal limits; forkz, they are off by default. - Do not call registry. Automatic additions to the registry from a conversation work only for
us_tcpa.
Course outline
Section I · How an agent works
Before clicking anything, you need to understand the model: an agent is not one big prompt but a dialog map that the engine follows turn by turn.Section II · Dialog logic
Five modules after which the agent can speak, listen, compute, query external systems and branch. This is the core of an agent author’s work.- Conversation: a phrase or a prompt
- Transitions: five kinds and their priority
- Variables: where they come from and where they go
- External systems: Function and tools
- Branching without a model: Logic Split
Section III · Knowledge, reuse and people
A dialog rarely goes in a straight line: the customer asks an off-topic question, asks for a human or simply says goodbye. Four modules on how to plan for that.- Global node and knowledge base
- Components: one block for all agents
- Transfer to a human
- Ending: End and the “interruptible goodbye”
Section IV · Agent settings: voice and call
So far we have built what the agent says. Now — how it sounds and how it behaves on a phone line.- Agent prompt, models and history window
- Voice: recognition, synthesis, interruptions
- Call control and voicemail
- Call recording and CRM notifications
Section V · Conversation quality
An agent that “seems to talk” and an agent for which you know what it achieved and where it broke the rules are different products. Five modules on measurability.- Agent objectives
- Post-call analysis: what we learn after the call
- Observers and incidents
- Validation, chat test, voice test, Copilot
- Publishing and versions
Section VI · Launch: how the agent starts calling
A published agent does not call anyone on its own. It needs a channel (telephony) and a trigger (a campaign, an inbound call or a call from your system via the API). This section is where the differences between market profiles show up: the calling window, the contact file and the Do not call registry.- Telephony: trunk and routes
- Outbound campaign: from a file to the first dial
- A call from your system: the public API
- Calling compliance: Do not call
Section VII · After the call
The most useful part of the work begins once calls are already going out: reviewing a specific conversation and monitoring the health of the platform.- Call history: transcript, analytics, diagnostics
- Operations: system health and roles
Summary
Checklists, pitfalls, exam — keep this part at hand while working: it covers what to check before every publication and every launch.Where to start
Start the course
Module 01: how the engine decides what the agent says next.
Checklists
What to check before publishing an agent and launching a campaign.
The course is based on HubTalk AI · Conversation Flow. The screenshots show the interface in the Russian locale; the data and integrations in them are test ones. Questions about the course and access — to your organization’s administrator.