Make.com Categorize Text: AI Toolkit vs HTTP → Jev for Typed Labels and Confidence Thresholds

According to Make’s AI Toolkit app page, Categorize Text takes “Text to categorize” and a list of categories (each with Name and Description). You choose whether to return an explanation and whether to “Assign only one category” (Yes = single label; No = multi-label). The module sits after a trigger and optional retrieval (CRM contact, sheet row), then feeds Filters and a Router that write to Slack, help desk, or CRM. Make states it does not store your data for training; errors such as token limits surface as RuntimeError; interaction logs follow standard scenario retention (documented as 30 days). Treat the module as a structured classifier inside the canvas—not as a portable API contract.

Categorize Text shines when classification is a routing problem inside Make: one bundle in, N Router routes out. Name+Description categories steer better than bare names; optional explanations help humans in the mid-band; Make’s built-in AI provider works on all plans so you can prototype without a third-party key (paid plans can attach OpenAI/Anthropic and similar). Teams usually hit limits when they need (1) the same label contract in apps outside Make, (2) Score/Noul-style rubrics beyond flat categories, (3) abstention calibrated across orchestrators, or (4) a decision API shared with Zapier, n8n, and backend jobs. Make’s guide illustrates a 0.7 confidence threshold before routing—treat that number as an example to calibrate on your tickets, not a universal default.

Neither path replaces evaluation on your labels. The table contrasts scenario shape and decision contract—not marketing scores.

Stay on Categorize Text when categories are few and stable, false routes are cheap to undo, everything already runs in Make, and you mainly need visual Router wiring plus optional explanations for reviewers. Write precise Name+Description pairs, prefer “Assign only one category” for exclusive queues, keep a fallback Router route for no-match bundles, and log category plus confidence (when present) with the model/provider id. Use Make’s built-in provider to prototype; switch to a custom provider only when you need a specific model or billing posture.

A common pattern is: Trigger (Webhook, Gmail, Sheets) → optional retrieval → HTTP module POST to JevModel’s /api/v1/systemone with Choice (2–12 labels) and optional Score/Noul → Filter on label and score-like fields → Router → side effects only on high-confidence routes → human-review queue otherwise. Keep secrets in Make connections—never hard-code keys in module fields or blueprints committed to git. Respect product budgets: validated state ~8000 characters after validation, 1–8 questions per request, free verified accounts get 5 successful runs per UTC day. JevModel hosts typesafe/jev-1.13 via OpenRouter; independent of TypeSafe AI; not multimodal on this site; no downloadable weights.

// Illustrative only - configure the HTTP module in Make; keep secrets in connections.
POST /api/v1/systemone
{
  "state": {"channel": "email", "body": "Invoice #992 is wrong - please credit us."},
  "questions": [
    {
      "id": "route",
      "type": "choice",
      "prompt": "Support category",
      "labels": ["billing", "technical", "general"]
    },
    {
      "id": "auto_ok",
      "type": "noul",
      "prompt": "Safe to auto-tag without a human?"
    }
  ]
}

Make’s LLM integration walkthrough recommends validating category membership and confidence after Categorize Text, then routing: high band to auto-action drafts, mid band to Slack/human review, low band to escalate or log. The 0.7 figure in the guide is an illustration—measure precision/recall on a holdout of your own mail or tickets before you trust any number for write actions. The same band design applies when the classifier is Jev: Filter on Choice label + Noul/Score, then Router. Wire mid and low routes before connecting destructive modules. See confidence-and-abstention and support-triage.

1) Freeze a label dictionary with edge cases; avoid a vague catch-all as the only fallback. 2) Decide single vs multi assign. 3) Pick Categorize Text vs shared Jev API based on reuse and abstention. 4) Store keys only in connections. 5) Add validation Filters and a no-match fallback route. 6) Measure holdout precision/recall after every model or category-description change. 7) Gate destructive actions behind thresholds and a human route. 8) Do not claim ranking or accuracy gains you have not measured.