---
title: 'AI Classify'
description: 'Route execution to one of several branches based on LLM classification.'
---
# AI Classify

Sends data to an LLM along with a list of labels, and routes execution to the
matching label's outgoing branch. If no label matches confidently, the
`fallback` branch fires.

Unlike AI Transform (which returns free-form text), AI Classify forces the
LLM to pick one of N predefined outcomes, so downstream routing is
deterministic.

## When to use

- Risk scoring: classify a customer as `low_risk` / `medium_risk` / `high_risk`.
- Intent detection: is this support ticket a `billing`, `technical`, or `complaint`?
- Sentiment routing: `positive`, `neutral`, `negative`.
- Any "categorize this and branch on the category" use case.

## Configuration

| Field | Required | What it does |
|---|---|---|
| `instruction` | No | Extra context for the LLM. Keep short — the labels themselves should be self-describing. |
| `inputField` | No | The payload to classify. If blank, the upstream node's output is used. |
| `labels` | Yes | JSON array of `{key, description}`. Each becomes an output handle. |
| `provider` | No | `groq` (default), `gemini`, `openrouter`. |
| `model` | No | Provider-specific model ID. |

### Labels format

```json
[
  { "key": "low_risk",    "description": "No prior issues, clean history" },
  { "key": "medium_risk", "description": "Some flags but nothing severe" },
  { "key": "high_risk",   "description": "Chargebacks, complaints, or fraud signals" }
]
```

Each label's `key` becomes an outgoing handle on the node. The `description`
is what the LLM reads when deciding — so make it specific.

## Handles

- One handle per label (e.g. `low_risk`, `medium_risk`, `high_risk`).
- **Fallback** — fires when the LLM response doesn't match any declared
  key. Always wire this to something (even just an Emit Event for
  logging) so you don't silently drop traffic.

## What it outputs

```
{
  label: "medium_risk",
  rawResponse: "medium_risk",
  confidence: null,
  model: "llama-3.3-70b-versatile",
  provider: "groq",
  tokensIn: 128,
  tokensOut: 4,
  durationMs: 640
}
```

## Gotchas

- **Label keys must be JSON-safe identifiers**. Lowercase with underscores
  is safest. No spaces, no dashes (they confuse some LLM response
  parsers).
- **Overlapping descriptions**: if two labels sound similar, the LLM
  becomes non-deterministic. Make each description clearly distinct.
- **Cost**: classification is cheap (low token count), but runs per
  execution, so high-volume workflows add up. Watch `ai_usage_logs` to
  track spend.
- **Don't use for branching on exact field values**: if you already have
  a field like `customer.tier === 'pro'`, use **Switch** or **If/Else** —
  they're faster, free, and deterministic. AI Classify is for the cases
  where the category is implicit in free-form data.