Chat Assistant
A Chat Assistant gives a flow a chat window. Someone types, the robot answers, and the conversation drives real automation — not a chatbot bolted on the side, but the flow itself talking to a person.
There are two modes, and the same nodes and the same robot power both. You choose per flow.
| Who drives | Use it for | |
|---|---|---|
| Conversational | The user | Open-ended questions, assistants, agentic sessions |
| Guided | The flow | Wizards, checklists, controlled data capture |
Conversational
Like ChatGPT, backed by your agents.
A free-form chat the user drives, usually backed by an AI Agent node. The agent reasons over the question, calls its RPA tools, and streams a clean answer back in real time.
- A free-form chat box, like talking to ChatGPT — backed by an AI Agent node
- The agent reasons, calls your RPA flows and tools, then streams the answer back live
- Best for open-ended Q&A, assistants, and agentic sessions that decide what to do
The shape of such a flow is short:
Chat In → LLM Agent (with your tools) → Streaming Text → Chat Out
That is the whole thing. The Chat Assistant is how an agent gets a face — the agent supplies the reasoning and the tools; this supplies the window someone can actually talk to.
Guided
The flow drives, step by step.
The robot presents structured widgets and waits for each answer before moving on. It is the deterministic way to capture controlled input and walk a person through a process.
- The robot presents structured widgets — buttons, dropdowns, date pickers, text, file upload
- Each step blocks and waits for the answer, then the flow moves on. Fully deterministic
- Best for wizards, checklists, controlled data capture, and "operate this system for me" flows
It is the better choice whenever the answer set is known. A dropdown of five valid options cannot be answered wrongly; a free-text question can, and then your flow has to cope with it. Reach for Conversational when you genuinely cannot enumerate what the user might want.
What it can do
| Rich input widgets | ButtonGroup, Dropdown, Checkbox, Radio, Datepicker and Textbox — each blocks and waits for the user |
| File exchange | Users upload documents to the robot, and the robot can offer files back — even as a public link |
| Streaming and progress | Stream an LLM answer token by token, or show a live progress indicator the flow updates in place |
| Pairs with AI Agents | Chat In → LLM Agent → Streaming Text → Chat Out |
| End-to-end encrypted | Browser and robot negotiate keys — RSA, then AES-256-GCM. The proxy routes ciphertext and never reads it |
| One canvas, both modes | Guided widgets or a free-form agent, from the same nodes |
The blocking model
The idea that makes Guided mode work: a widget node blocks. The flow stops at that node until the person answers, then carries on with their answer on the message object.
So a wizard is just a flow with several widget nodes in a row, and everything you already know applies — branch on an answer, loop until the input is valid, catch when something downstream fails.
Encryption
Worth knowing if anyone asks where the conversation goes. The browser and the robot negotiate keys directly — RSA for the exchange, then AES-256-GCM for the session. The proxy in between routes ciphertext and cannot read it. See Security.
Chat Assistant or a Form?
Both collect input from a person, and they suit different shapes of task.
| Reach for | |
|---|---|
| Form | Everything is known up front. One page, submitted once, and the flow runs afterwards |
| Chat Assistant | The conversation is the process — later questions depend on earlier answers, or the flow reports back as it goes |
A form is a request. A chat is a session.
See also
- AI Agents — what sits behind Conversational mode
- Form Builder — the one-shot alternative
- Security — the encryption model