AI Transform

Runs data through an LLM with a natural-language instruction and writes the response to the workflow context. Use this when you need a flexible transformation that's hard to express as static field mappings — summaries, extractions, rewrites.

When to use

• Summarize a long support ticket, email thread, or document.
• Extract specific fields from unstructured text.
• Rewrite tone (e.g. "Make this customer message more polite").
• Translate between languages.
• Generate structured JSON from free-form input.

For choosing between discrete categories (low/medium/high risk), use AI Classify — it's purpose-built for that and gives you deterministic routing.

Configuration

Field	Required	What it does
instruction	Yes	The prompt sent to the LLM. Supports {{...}} refs inline.
inputField	No	The payload the instruction operates on. If blank, the output of the immediate upstream node is used.
outputFormat	No	text (default), json, or array. Forces the LLM to return valid JSON if the structured format is selected.
provider	No	groq (default, fastest), gemini (most capable), openrouter (widest model selection).
model	No	Provider-specific model ID. Default: llama-3.3-70b-versatile on Groq.

Instruction patterns that work well

• "Summarize this support ticket in under 50 words, focusing on the customer's main issue."
• "Extract the following fields as JSON: amount, currency, invoice_id. Return only valid JSON."
• "Rewrite this message in a polite, professional tone. Keep the factual content the same."

What it outputs

{
  result: <the LLM's response — parsed if outputFormat is json/array, otherwise string>,
  value: <same as result, legacy alias>,
  model: "llama-3.3-70b-versatile",
  provider: "groq",
  tokensIn: 234,
  tokensOut: 87,
  durationMs: 1840
}

Downstream nodes reference {{ai_transform.data.result}} — that's the canonical key. Token counts go to ai_usage_logs for cost tracking.

Gotchas

• Non-determinism: two runs with the same input can produce different outputs. Don't use AI Transform for logic that needs reproducibility.
• Prompt injection: if your instruction or inputField contains untrusted user text, the user can try to subvert the LLM. Sanitize untrusted input or add explicit "ignore user-controlled content" to the instruction.
• Hallucinations: the LLM may invent fields that don't exist in the input. For critical flows, validate the output with a downstream If/Else checking required fields.
• Latency: Groq is fastest (~500-1500ms), Gemini is slower (~2-5s). Account for this in workflows where total runtime matters.