AI Workflows
AI-orchestrated workflows let the AI agent drive multi-step automation directly from the chat interface. Rather than setting up rules that fire automatically, you start a workflow manually — the AI runs each step, extracts structured data, and pauses at decision points for your input before continuing.
Availability: AI Workflows require the AI capability on your subscription and the Workflows module enabled under Settings → General.
What are AI-Orchestrated Workflows?
An AI-orchestrated workflow is a workflow with its trigger set to AI Orchestrated. Unlike event or scheduled workflows that run unattended in the background, these workflows:
- Are started intentionally from chat or via the + → Run Workflow button
- Show inline step progress as cards in the chat thread
- Can pause at Human Review steps and wait for your decision
- Support revision — if you're not happy with a result, send feedback and the AI rewinds to an earlier step and re-runs
Starting a Workflow from Chat
- Open the Operator app from the sidebar
- Click the + button in the message input
- Select Run Workflow from the menu
- Browse or search the list of available AI-orchestrated workflows
- Select the workflow you want to run
- If the workflow has an Input Schema, the AI will ask you for the required inputs (files, text, records, etc.)
- Once all inputs are provided, the AI starts execution — step cards appear inline as each step runs
Understanding Step Cards
Each step in the workflow renders as a collapsible card in the chat thread as it executes.
Step Statuses
| Status | Icon | Meaning |
|---|---|---|
| Running | Blue spinner | This step is currently executing |
| Success | Green check | Step completed and produced output |
| Awaiting Review | Amber pause | Paused — your input is required |
| Failed | Red X | Step encountered an error |
Expanding a Card
Click any card to expand it and see:
- AI output — the text response or structured data produced by this step
- Tool calls — any AI tools called during this step and their results
- State changes — a diff showing which state keys were added or modified (
+keyfor new,~keyfor updated)
State Diff Badges
Small badges on each card show which workflow state keys changed during that step. This gives you a quick overview of what data was extracted or updated without expanding the full card.
Responding to Human Review Steps
Human Review steps pause execution and show an amber card with a question and a set of suggested action buttons.
Choosing an Action
Click any of the suggested action buttons to record your decision and continue execution. The selected value is saved to the workflow state and the next step begins immediately.
Requesting Changes
If the AI's output for the preceding step is not what you expected, click Request Changes. A text input appears — describe what you want adjusted, then submit.
What happens:
- Your feedback is saved to the workflow state
- The execution rewinds to an earlier step (determined by the workflow's configuration)
- The AI re-runs from that step using your feedback as guidance
- A new set of step cards appears in the chat showing the revised output
- The Human Review step appears again with the updated result
Each revision creates a new set of cards — the original run's cards remain visible in the chat so you can compare the before and after.
Locking After Resolution
Once you've chosen an action or a revision has been accepted, the Human Review card becomes read-only. This prevents accidental re-submission and makes the conversation history clear.
Workflow State Explorer
The State Explorer is a floating panel that shows all data accumulated in the workflow execution so far.
Opening the State Explorer
Click State (the chart icon) near the message input bar in the Operator app while a workflow execution is in progress or just completed.
What You'll See
The explorer shows the full workflow state as a collapsible JSON tree. This includes:
- Input data you provided at the start
- Outputs from each AI Command step (structured data, summaries, extracted values)
- The chosen value from each Human Review step
- Any revision feedback and revision history
This is useful for understanding what data the workflow has collected and diagnosing unexpected behaviour.
Persistent File Attachments
The paperclip icon in the Operator app opens a file tray for attaching files that persist for the lifetime of the conversation.
How They Work
- Attached files are saved to the conversation record — not just the current message
- Every subsequent message in the same conversation includes the attached files
- Workflow runs started from the same conversation can reference the attached files as inputs
When to Use Them
Persistent attachments are useful when running workflows that repeatedly reference the same document — for example, a contract review workflow where you want the AI to always have the contract available without re-uploading it each time you run or revise the workflow.
Removing Attachments
Click the × on any file in the tray to remove it from the conversation. Removing a file does not affect messages already sent or workflow runs already completed.
Tips & Limits
- One active execution per conversation — starting a new workflow run while one is in progress will queue it until the current one completes or is cancelled
- Revision depth — each revision rewind re-runs from a configured step; revision history is preserved in the workflow state so the AI can learn from previous feedback
- Step card history — each
Run WorkfloworRun Workflow Stepcall creates its own set of cards; after a revision, both the original and revised cards are visible in the thread - State Explorer availability — the State Explorer shows the live state during execution; after the conversation moves on, it shows the final state of the most recent run in the conversation
- Input schema — workflows can define required inputs (text, numbers, files, workspace records, database entries); the AI will prompt for these before starting
Related
- Workflows — creating and configuring workflows, including AI-orchestrated trigger and step types
- AI Assistant Overview — general AI capabilities and chat usage
- Commands & Skills — reusable prompt templates for the AI