Workflows are where repeatable work stops living in someone's head. On WisebotAI a workflow is a drag-and-drop canvas — but you don't have to start from a blank one. Describe what you need and AI drafts the graph, then you edit any step in plain language.
Start by describing it
When a new conversation comes in, search our knowledge, classify it as billing / technical / sales, answer the easy ones, and route the rest to the right team with the transcript.
AI turns that into a connected graph of nodes. You review it, tweak wording, and publish. Editing later is the same: "add a step that posts escalations to the #support Slack channel" and the graph updates.
The node types you'll actually use
- Trigger — webhook, scheduled/timer, manual run, a public form link (
/f/{workflow}), or an agent. - KB search — pull grounded, cited context from your knowledge base.
- LLM / agent step — reason over the input with the model you choose (GPT or Claude), with tools attached.
- Manager / condition branch — classification-based or if/else routing.
- HTTP — call any REST API (GET/POST/… with auth).
- MCP tool — call a connected system (Slack, Notion, your CRM).
- Code step (E2B) — run Python, Node, or Bash in a sandbox when logic outgrows no-code.
- Output / hand off — return text, JSON, or a generated PDF/PPTX, or hand to a person.
A triage flow, end to end
- Trigger: new conversation.
- KB search → LLM: draft a grounded answer and a category.
- Manager branch:
billing | technical | sales. - Billing: call the CRM connector to pull the account; if balance is sensitive, approval gate before any change.
- Technical: if confidence is low, hand off to the engineering queue with context.
- Sales: post to #sales via the Slack connector and create a lead.
- Every branch leaves a run record you can open and review.
Why "visible" matters
Automations that hide in a black box are impossible to trust or debug. Here every published run leaves a trail — inputs, the path taken, each step's output — so ops and leadership get the audit story, not just developers.
Tips
- Keep branches shallow. Two or three clear paths beat a sprawling tree.
- Ground before you act. Put a KB step before any tool call so actions are based on real context.
- Gate the irreversible. Refunds, deletions, external messages — approval-gate them.
Next: tool calling with MCP for the connectors above, and writing agent system prompts for the LLM steps.