Milian AI — Automations

These analysis actions currently require the Global AI processing region. If your workspace requires processing only in the EU or US, the action stops without changing your region or charging analysis credits. Custom Prompt and Extract Fields keep their existing regional behavior.

Add a step in Automations → Milian AI to classify text, check a condition, score content, detect multiple labels or build a custom analysis. Custom Prompt and Extract Fields remain available.

Five new actions. Turn transcripts, messages and previous step outputs into categories, yes/no decisions, scores or labels.
Three editable templates. Start Custom Analysis with Analyze a call, Qualify a lead or Prioritize a message, then adjust the questions.
Clear uncertainty. Insufficient evidence produces an uncertain answer instead of a guessed value. Use the result status in a Condition before continuing.
Your workspace, your workflow. Preview inputs without spending credits, run a live test with workspace credits, and reuse the same configuration through REST or MCP.

Actual analysis uses workspace credits and follows your workspace’s AI processing region. Availability depends on the selected region. Scores use your scoring levels and are displayed on a 0–100 scale.

REST / MCP

GET /api/v1/automations/ai-actions — list_automation_ai_actions.
POST /api/v1/automations/ai-actions — run_automation_ai_action.

{
  "action": "ai.check_condition",
  "config": {
    "input": "Please call me back tomorrow.",
    "questions": [{"id": "callback", "type": "boolean", "instructions": "Does the customer explicitly request a callback?"}]
  },
  "dry_run": true
}

Choose Custom Analysis to mix question types or apply one of the three templates. Give each answer a unique key. Categories need 2–32 options with descriptions; scores need 2–10 ordered levels. An analysis can include up to 32 questions. Provide text up to 32 KB; split longer inputs deliberately rather than dropping part of a transcript. Use Minimum probability for category and yes/no decisions, and Minimum confidence for categories and scores. Every question also checks whether the input provides enough information. These thresholds are configurable controls, not guarantees of correctness. Each answer has a value and a status. An uncertain answer returns null; a supported negative answer returns false. In Detect Labels, the label list contains only accepted matches; check individual results for uncertain labels. A technical failure stops the step instead of becoming a negative answer. Multiple independent questions share one evaluation request. Use the variable picker to reference answers in subsequent steps. Check the status first, then compare the value with a Condition. A Milian AI evaluation step does not itself create extra routing branches. Summaries, translations and open-ended extraction continue to use Custom Prompt or Extract Fields.

​REST / MCP

REST / MCP