How AI agents work

Who this is for: whoever manages the account.

What it is for

AI agents are assistants that answer messages using artificial intelligence. They work 24 hours a day, resolve questions, hand off to a person when needed and close the conversation when the customer needs nothing else. You can choose the AI provider and model for each agent: see AI providers.

💡

Tip:An agent is not a fixed-answer bot: it understands the context of the chat, uses the business information you load and can run actions (transfer, notify, tag, query your API).

Two ways to create one

When you click New agent you choose how to build it:

Tell me what you need: an assistant interviews you and builds the agents for you. See Agent architect.
Manual: start from a template or from scratch. See Templates and gallery by industry.

What you configure

Step	What you configure
General	Name, type, provider and model, model parameters, prompt, closing rules, inactivity, relations with other agents, automatic notifications and tags
Tools	Actions the agent can run. See Tools
Channel	Which numbers or channels this agent answers on

A channel cannot be assigned to two agents at once: if you pick one already in use, TalkAI tells you it belongs to another agent.

Agent types

Conversational: handles a channel and talks to the customer.
Transactional: does not talk to customers. Other agents consult it for a specific task (read a document, verify a receipt, query your system) and the conversation stays with whoever asked.

ℹ️

Note:For a transactional agent, the first line of its prompt is what other agents see to decide whether to consult it. See Prompt.

Relations between agents

Each relation has its own switch and its list of allowed targets:

Transfer to another AI agent: hands the conversation over to another agent (for example, from Reception to Sales).
Consult a transactional agent: asks it a question without handing over the conversation.
Route to a department: moves the conversation to the human team.

⚠️

Warning:In the lists, empty means ALL, not none. To restrict, check only the allowed targets. To stop an agent from using a relation, turn its switch off.

All of this is shown on a map. See Relations and map.

The flow of a conversation

A message arrives on the assigned channel.
The agent analyzes the message with the full context and replies.
If needed, it runs a tool or consults another agent.
If it cannot resolve the issue, it escalates to a person (the conversation moves to "Queued" and then "Human").
When the customer needs nothing else, the agent closes the conversation according to its closing rules.

Activating an agent

A new agent is not activated directly: first you try it in the test chat and confirm it works the way you want. See Test chat and activation.

Common mistakes

Leaving a target list empty thinking it means none: empty means all.
Assigning a channel already used by another agent.
A transactional agent with a vague first prompt line: nobody consults it.