AI in Workflows

KnoxCall brings AI into workflows two ways: an AI step you drop into a graph like any other node, and an AI copilot that drafts workflows from a plain-language prompt and explains failed runs.

The AI Step

The AI node runs an Anthropic model with structured output — you declare the shape you want and get back typed data, no prompt-parsing glue. Four operations:

Operation	What it does	Key config
Classify	Picks one label from a list	`labels: [...]`
Extract	Pulls named fields out of text into an object	`fields: [{ name, description, type? }]`
Summarize	Condenses text to a target length	`maxLength` hint
Generate	Free-form generation	`maxTokens` cap (default 1024)

Example — classify an inbound message:

{
  "type": "ai",
  "data": {
    "type": "ai",
    "config": {
      "operation": "classify",
      "text": "{{trigger.data.message}}",
      "labels": ["billing", "technical", "sales"]
    }
  }
}

The text field (and other inputs) support {{ ... }} variables. Output is a structured object you reference downstream, e.g. {{ai_1.label}} for a classification result, or the extracted fields for an extraction.

Model & Credentials

AI steps resolve Anthropic access tenant-key-first: if you’ve configured an Anthropic key under Integrations, it’s used; otherwise the platform key is used. The model is centrally configured (default claude-sonnet-5) — you don’t hard-code a model id in the node.

Metering & Limits

AI nodes count as weighted operations (×3) against your monthly operations quota.
AI-token usage is metered against a monthly AI-token budget per plan tier.
When either limit is reached, an AI step is blocked (recorded, not run) rather than silently overspending.

The AI Copilot

Open the AI panel in the builder toolbar to draft a workflow from a description.

Draft a workflow from a prompt

Describe what you want — for example:

“When a GitHub issue is opened, classify it as bug or feature, and post a Slack message to the right channel.”

The copilot returns a coherent, auto-laid-out graph (trigger → nodes → edges) that lands unsaved on your canvas. It goes through the same publish gate as a hand-built workflow: structural validation is enforced, and any per-node configuration gaps come back as warnings so you know exactly what to fill in (connections, secrets, endpoints) before publishing.

Explain a run

On any finished execution, click Explain in the run log. The copilot reads the run’s redacted step history and returns a plain-language explanation of what happened plus a suggested fix — for example, diagnosing an unreachable URL and pointing at the exact node to correct.

Metering

Copilot calls are rate-limited and draw from the same AI-token budget as AI steps (each call reports its own token cost).

Tips

Start Extract/Classify with a few clear labels or fields; add more only if results are ambiguous.
Use the copilot to scaffold, then refine node config by hand — it’s a starting point, not a black box.
Keep an Anthropic key in Integrations if you want AI usage billed to your own account.

​AI in Workflows

​The AI Step

​Model & Credentials

​Metering & Limits

​The AI Copilot

​Draft a workflow from a prompt

​Explain a run

​Metering

​Tips

​Next Steps

Creating Workflows

Templates