What a conversation flow is

The pieces an AI agent's conversation is made of, and how a written flow differs from the flow charts of the old generation.

The pieces an AI agent's conversation is made of, and how a written flow differs from the flow charts of the old generation.

The pieces

• Replies are the things the agent says.
• Dialogues group replies into an exchange about one subject.
• Use cases describe situations the agent should recognise.
• Generative replies are answers written from your articles rather than fixed text.
• Actions let it do something in another system mid-conversation.

What has changed

In the old generation you drew every path: this button, then that message, then these three options. Everything not drawn did not happen.

Now you describe the situation and what should result, and the agent works out the exchange. Fewer things to maintain, and less control over the exact wording.

Fixed text where it must be exact

Anything with a legal or commercial consequence should be a fixed reply rather than generated: cancellation terms, prices, guarantees, anything you would want to quote back.

Generated text is right in substance and variable in wording, which is fine for explanations and not for commitments.

Generated answers for everything else

The bulk of questions have an article behind them. Letting the agent answer from articles means every improvement to the knowledge base improves the conversation, with nothing to rebuild.

Design the ending, not just the middle

Most flows are written up to the answer and stop. What happens after matters: does it check whether that helped, does it offer a person, does it close cleanly?

See also

• Conversation design: how to build a good exchange
• Creating a dialogue
• Setting up generative replies