Workflow Assistants
Workflow assistants are an advanced form of AI assistants. Instead of a fully open conversation, the employee follows a defined sequence of steps guided by the AI. This makes it possible to standardise and automate complex, multi-step processes — consistently and repeatably.
What is a Workflow Assistant?
A workflow assistant is based on a visual flowchart of nodes connected by arrows. Each node performs a specific task:
| Node type | Description |
|---|---|
| Start / End | Entry and exit point of the workflow |
| Evaluate | Analyses the context and stores information in the workflow memory |
| Condition | Evaluates an expression and routes the flow in the right direction |
| LLM | Calls a language model for a specific task (optionally: waits for user input) |
| Approval | Asks a yes/no question and routes the flow based on the answer |
| Pass-through | Connects routes without an AI task or visible output |
| Subagent | Runs the same task in parallel for multiple items |
The workflow automatically saves its state after each step. When the employee provides input, the workflow resumes exactly where it left off — even after closing the browser.
How do I use a Workflow Assistant?
Start a new conversation and choose a workflow assistant from the list. The assistant automatically guides you through all steps. In some steps the assistant waits for your response; in others it works independently without input.
On the left side of the conversation you can see the workflow steps and which step is currently active.
You can close the chat at any time and continue later. The workflow will resume where you left off.
Example: Deep Research Assistant
The Deep Research Assistant is an example of a workflow assistant available for employees. It guides through a thorough research process:
- Define the topic — the AI asks about the research topic and desired depth
- Formulate search queries — the AI creates targeted search questions
- Conduct research — the AI carries out the research and gathers relevant information
- Generate report — the AI writes a structured research report
- Wrap up — the employee reviews the report and can ask follow-up questions
Creating a Workflow Assistant
As an administrator, you can build and publish workflow assistants for employees in the administration environment. Choose Assistants → New assistant → Workflow assistant. You can start with a blank canvas or use an existing preset as a starting point.
Workflow assistants can be shared with groups and colleagues, just like regular assistants. When sharing with colleagues, AI-Corporate can automatically send an email notification so recipients know the workflow assistant is available.
On the right side of the chat input you will find the Workflow button. Click it to open a panel showing all steps and their current status (pending, active, done, awaiting input, skipped, error).
You can close the chat at any time and continue later. The workflow will resume where you left off.
Creating a Workflow Assistant
As an administrator, you can build and publish workflow assistants for colleagues in the admin panel. Go to Assistants → New assistant → Workflow assistant. Start from a blank canvas or use a preset.
Global Settings
When creating or editing a workflow, the settings panel configures the following for the whole workflow:
- Model: The AI model used by all steps.
- Tools: Which tools are available (e.g. Recording, Internet Search). LLM steps can disable globally enabled tools per step.
- MCP servers: External connections.
- Documents and images: Files that LLM steps can reference.
Creating a Triage Node
An evaluate node analyses user input and stores structured values in the workflow memory (e.g. department, urgency). Configure:
- A system prompt describing what should be analysed.
- An output schema: fields with name and type (
string,number,boolean,enum).
Stored values are available as state.fieldname in condition nodes.
Configuring a Condition Node
A condition node reads the workflow memory and routes the flow using CEL expressions:
| Desired behaviour | Expression |
|---|---|
Field is true | state.approved == true |
| Field has value | state.department == "sales" |
| Numeric threshold | state.priority > 3 |
Each branch has its own connection to the next node.
Using a Pass-through Node
A pass-through node does not run an AI task. It simply forwards the workflow to the next step. Use it to bring routes together neatly, for example when one path skips a step and then needs to rejoin the main route.
A pass-through node must have at least one incoming and one outgoing connection. It is not shown as a separate content step in the workflow progress panel.
Recording in Workflows
Enable the Recording tool in the global settings and add an instruction in a step’s system prompt to save results. Users see recorded documents via the Recording button (📄) in the toolbar.
Restoring a Conversation
Not happy with the direction the conversation has taken? The restore button on an earlier message lets you roll back the workflow to that point. All messages after that point are deleted and the workflow resumes from that exact moment.
Restoring is permanent — the deleted messages cannot be recovered.
Files and subagents in workflows
A workflow can ask the user to provide files, for example a source document, assessment model, contract, policy document or completed form. The user can upload a new file or select an existing file from file management. Files added through a workflow form become available to the assistant for that chat, but they are not automatically selected for ordinary chat questions.
New files are processed before the workflow can reliably continue. If a required file is still being processed or has failed, the user must wait or choose/upload the file again.
The Subagent node runs the same task for multiple items in parallel. Use it when a workflow needs to process a list, such as one document, participant, employee, client, case or form per run. The subagent receives the current item in an isolated context. Make prompts refer explicitly to the current item and avoid instructions that process multiple items at once.
