AI Flows
Build a team of AI steps that answers your customers on its own, with your models, prompts and knowledge.
What an AI flow is#
An AI flow is a small team of steps that answers on its own. Each step does one job, such as searching your knowledge, deciding what the customer needs, or writing the reply, and passes its result to the next step.
The flow runs from the top step down. The output of the last step that runs is the reply the customer gets.
Knowledge
Search your docs
LLM
Answer with context
Filter
Needs a human?
Text
Fixed reply
Where to find AI flows#
AI flows live inside your chatbots, so AI Flows is not in the sidebar. Open them from:
- The AI Flow step in Chat Builder. Under Which AI flow answers here, pick Create a new AI flow and click Create and open, or Use an existing AI flow and choose one. The flow opens right inside the chatbot. This is the usual way.
- The chatbots list. Flows that no chatbot uses yet are listed under AI flows not used by a chatbot, with Open and Add to a new chatbot (a new draft chatbot around the flow, with Undo).
- Search (⌘K, or Ctrl+K on Windows). Type the flow's name; results are grouped under AI flows.
On the AI Flows page, New AI flow asks What should your AI flow do? and offers these starts:
- Ready AI flows: click Use this on one that fits, then change it.
- From my website or files: Agentim reads your site or documents and builds a flow that answers from them.
- Describe it: say what it should do, and the Assistant writes the prompts and steps.
- Blank AI flow: one AI agent with the default model. It gets a random name such as Swift Falcon; rename it any time.
- Already have an OpenAI Assistant? Connect it: see OpenAI assistants and the Setup wizard.
The Team builder canvas#
Every AI flow opens on the Team builder canvas. The header shows the flow's name, its release status and how many agents it has, plus AI flow settings (the gear), Use in a chatbot, Preview and Publish.
- 1
Add a step
Click Add a step under the last step and choose what it does: AI agent, Knowledge, Python, Text, Filter, Jump or Exit.
- 2
Configure it
Double-click a step to open its settings. Change its kind with the type menu in the top right, give it a Name (English letters, numbers and _ only) and an optional description.
- 3
More actions
Right-click a step for Run once, Add branch (Filter steps only) and Remove.
- 4
Test and go live
Click Preview to chat with the flow, then Publish and put it in a chatbot with Use in a chatbot. See Run, preview and publish.
Step names matter: later steps read a step's result by its name, like {{Classify.intent}}. Rename carefully, then update the prompts that use the old name.
Only a Filter step can lead to more than one next step. A Jump step can't have any step after it.
Step types#
| Step | What it does | Guide |
|---|---|---|
| LLM | Calls an AI model with your instructions. Can answer with named fields. | LLM step |
| Knowledge | Searches your website pages and files and returns the most relevant passages. | Knowledge step |
| Filter | Checks conditions on earlier results and picks the next step. | Filter step |
| Text | A fixed text, with variables filled in. | Text step |
| Jump | Continues the flow from another step. | Jump step |
| Python | Runs your own Python code. | Python step |
Inside any step, type @ to insert a variable such as {{question}} (the customer's message), {{chat_history}} or another step's result. See Variables.
Models#
Pick a model for each LLM step. The picker lists current models from OpenAI, Anthropic Claude, Google Gemini and Cohere, plus open models on Together AI and your own Azure OpenAI deployment. Each model is tagged Recommended, Fast or Powerful.
AI replies use AI credits. On Pro they can run on your own AI key instead. On some accounts AI is included in the plan; AI usage & credits shows which applies to you.
Use a fast, low-cost model for simple jobs like classifying a message, and a stronger one for the final answer. Details in LLM step.
Structured answers#
Turn on Answer with a schema in an LLM step to get named fields (Text, Number, True/False or One of…) instead of free text. Later steps read each field as {{AgentName.field}}, which makes Filter conditions simple and reliable.
How to build a schema, presets and common mistakes: LLM step: answer with a schema.
Put an AI flow to work#
An AI flow goes live only inside a chatbot. Customers reach it through the chatbot's website widget, WhatsApp, other channels and the API, never through a link of the flow's own.
- From the flow: click Use in a chatbot, then Create chatbot (a new draft chatbot: Start, this flow, a step for each exit) or Add next to one of your chatbots. See Use it in a chatbot.
- From a chatbot: add an AI Flow step in Chat Builder and choose the flow. See AI Flow step.
Then publish the chatbot. Chat links and website codes of AI flows shared on their own before October 2026 keep working.