Power Automate Text Classification: AI Builder Category Models vs HTTP → Jev

In Power Automate, classification is usually a routing step: take inbound text (SharePoint item, Outlook email body, Teams message, Forms answer), ask a model to pick a category, then branch with Condition / Switch before you write to Dataverse, Service Bus, or a ticket system. Two first-party AI Builder paths dominate the docs: (1) a custom text classification model you train on labeled examples; (2) the prebuilt category classification model when you need a quick standard classifier. A third pattern is the HTTP connector that POSTs to an external decision API such as JevModel. None of these replaces a labeled holdout from your own traffic.

Microsoft Learn documents the action “Classify text into categories with one of your custom models.” You select the published AI model, map Text and Language, and the action returns Classification plus a Confidence score between 0 and 1. Strengths: labels and training data stay inside your Microsoft tenant; designers can wire Conditions without leaving the maker portal. Limits: you own labeling, model quality, and republishing when categories drift; confidence is not a portable Choice/Score/Noul contract for non-Power apps; credit consumption follows advanced text analysis rules (per 1k characters) and the flow is premium. Treat the score as a vendor signal until you calibrate on your own labeled sample.

The prebuilt action “Classify text into categories with the standard model” takes Text and Language and returns the same Classification + Confidence score shape—without you publishing a custom model first. Prefer it for prototypes and coarse routing when Microsoft’s standard categories are close enough. Prefer a custom model when your taxonomy is domain-specific (refund vs shipping vs product) and false routes are expensive. Either way, AI Builder advanced text analysis still consumes credits per 1k characters on eligible plans, and AI Builder flows remain premium. For automation stacks outside Microsoft 365, compare Zapier, Make, n8n, and Pipedream patterns.

The table contrasts workflow shape and decision contract—not marketing accuracy scores. Calibrate any confidence on your own labeled sample.

Stay on custom or prebuilt AI Builder when labels are few and stable, false routes are cheap to undo, the whole automation already lives in Power Platform, and you mainly need maker-portal field mapping plus Microsoft credit accounting. Keep an Other/ambiguous branch, log Classification and Confidence with every route, and recalibrate after taxonomy or model republish. Prefer prebuilt for speed; prefer custom when security wants training data to stay in-tenant.

A common pattern is: Trigger → Compose/Parse to normalize text into state → HTTP POST to JevModel’s /api/v1/systemone with Choice questions (2–12 labels) and optional Score/Noul → Condition on the chosen label and score-like fields → side effects only on high-confidence paths → human-review queue otherwise. Store the API key in Azure Key Vault or a secure connection reference—never hard-code secrets in action inputs or commit them to git. Respect product budgets: validated state is capped (about 8000 characters after validation), 1–8 questions per request, and free verified accounts get 5 successful runs per UTC day. JevModel hosts typesafe/jev-1.13 via OpenRouter; it is independent of TypeSafe AI (which develops Jev), is not multimodal on this site, and does not ship downloadable weights. For a similar HTTP pattern on event-driven workflows, see pipedream-text-classification.

// Illustrative only — configure the HTTP action in Power Automate; keep secrets in Key Vault.
POST https://jevmodel.app/api/v1/systemone
Authorization: Bearer <API_KEY_FROM_CONSOLE>
Content-Type: application/json

{
  "state": "Customer asks for a refund on order #1842 after late delivery.",
  "questions": [
    {
      "id": "route",
      "type": "choice",
      "prompt": "Support route",
      "labels": ["billing", "shipping", "product", "other"]
    }
  ]
}

Whether you use AI Builder’s Confidence score (0–1) or Jev Score/Noul, calibrate on a labeled sample before auto-applying tickets. A practical policy: auto-apply safe labels above a high band, send mid-band items to a human queue with the model’s top labels as suggestions, and force review below a floor. Wire mid and low Conditions before any write action. TypeSafe’s Choice confidence formula is (p − 1/n)/(1 − 1/n); that value is not calibrated accuracy. See confidence-and-abstention and support-triage.

1) Freeze a label dictionary with edge-case examples. 2) Decide Other vs discard for unmatched items. 3) Pick AI Builder custom, prebuilt, or shared Jev API based on reuse, tenant posture, and abstention needs. 4) Store keys only in Key Vault / secure references. 5) Confirm AI Builder credit and premium-flow licensing with your admin (docs evolve; do not hard-code dollar costs). 6) Measure precision/recall on a holdout after every model or taxonomy change. 7) Gate destructive actions behind thresholds and a human branch. 8) Do not claim ranking or accuracy gains you have not measured.