Conversation AI: single prompt or flow builder?

The difference in one paragraph

A prompt-based bot works from one main set of instructions. Everything it knows about tone, rules and the next step lives in that one place. HighLevel's newer flow builder splits the work across connected AI Agents instead, and each one can have its own instructions and its own tools. One agent can greet and sort, another can book, another can handle a support question.

How we choose

HighLevel doesn't force either option. This is the rule we use in our own builds.

• A single prompt when the bot answers questions from the knowledge base and steers toward a booking.
• A flow when the conversation has to branch on who the person is or what they want, when you're putting a setter script into the bot, or when objection replies have to follow a set path.

If the bot answers and books, one prompt is enough. Build a flow when the conversation splits.

Two flow builders, two sets of parts

This is where most confusion comes from. HighLevel's older flow builder and its newer multi-agent canvas are separate experiences, documented separately, with different parts. Mixing their node lists produces a plan nobody can build.

The newer canvas, documented in October 2026, is for inbound conversations. It starts at a fixed Start Chat entry point, uses AI Agent nodes for the work, a Router to send the conversation down the right path, and a Global AI Agent for requests that can come up on any path.

The older builder uses AI Message (the AI writes the words), Custom Message (your exact text is sent), AI Splitter (it sorts on what's already been collected and sends nothing) and an [END] marker. Some older notes, ours among them, call the splitter Conversation Splitter. Use the names on your screen.

Traps in the older builder

If your account still shows the older builder, three things catch people out.

• [END] ends an objective, not the conversation. If the chat keeps going after a flow reaches [END], that's how it's designed. Plan what happens next.
• Custom Triggers can't start the initial flow. If your plan opened the flow from a Custom Trigger, plan a different start.
• AI Message rewrites, Custom Message doesn't. When the wording has to be exact, such as a price policy or a legal line, use Custom Message.

Building a flow on the newer canvas

These steps follow HighLevel's own article for the multi-agent canvas.

1. Click Create Bot, choose Flow Based Builder, then Build.
2. From Start Chat, use Add New Node to add AI Agent nodes. Write each one's instructions and give it only the tools it needs.
3. Add a Router where the conversation splits, and decide what the Global AI Agent should catch on every path.
4. Connect the routes, then walk each one and check its waiting, transfer and ending paths.
5. Run Test with a test contact.
6. Set up Deploy, then switch the bot on through Bot Settings.

Test without touching real customers

HighLevel warns that test runs can use the bot's tools for real, which means a test can change contact records or book appointments. Use a test contact you own, and a calendar slot nobody needs, every time.

In our flow builds we also add a safety check after every step that waits for a reply, so the person's message is checked before the next AI reply goes out. That check is something we design into the flow. It is not a built-in HighLevel node, and it won't be there unless someone puts it there.