> For the complete documentation index, see [llms.txt](https://oodlz-ai-studio.gitbook.io/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://oodlz-ai-studio.gitbook.io/docs/inbound-agent/the-brain/six-intent-classification.md).

# Six-Intent Classification

Every message classified into 1 of 6 intents before the agent runs — sales pressure only on active product intent, empathy on complaints, silent drop on harmful input.

> **Read the room before the agent speaks — sell only when the user is buying.**

Generic AI assistants treat every message as a potential sale. Compliments get redirected into product pitches. Genuine questions get answered with "have you seen the latest deals". Complaints get followed by "did you know about the weekly offers". The result is an assistant that sounds desperate, tone-deaf, or unhinged depending on what the user actually said. The Oodlz inbound agent classifies every incoming message into 1 of 6 intents before the main agent runs — and the matched posture controls how the agent behaves.

## How It Works

A cheap dedicated classifier runs on every message before the main agent is even invoked. It returns a structured result: intent, posture, confidence, and sentiment. The main agent reads that result and adjusts its behaviour accordingly. 6 intents and 6 postures.

| Intent               | Posture            | What the agent does                                                                                                     |
| -------------------- | ------------------ | ----------------------------------------------------------------------------------------------------------------------- |
| `on_brand`           | `hard` or `medium` | Active product intent. Full commercial response with brand cards, exact offers, direct deeplinks.                       |
| `casual_engagement`  | `acknowledge`      | Pure reaction, compliment, or agreement. Reply in 1–2 warm sentences. No tool call. No product mention.                 |
| `question_inquiry`   | `inform`           | Genuine question about how the platform works. Answer from FAQ knowledge with minimal commercial mention.               |
| `off_topic`          | `soft`             | Unrelated to the brand. Acknowledge the user's topic first, gentle pivot only at the end if at all.                     |
| `negative_complaint` | `empathy`          | Frustration, anger, public criticism. Empathy first, zero sell, ready to escalate to a human.                           |
| `harmful`            | `blocked`          | Profanity, abuse, threats, jailbreak attempts. Silently dropped with a polite fallback. The main agent is never called. |

The `acknowledge` posture has Rule 0 in the agent's instructions: do not call any tools, just reply in 1–2 warm sentences and stop. The `harmful` posture exits the pipeline before the main agent runs at all — saving every cent of LLM cost on bad-faith input.

The critical distinction is between `casual_engagement` and `on_brand`. A pure compliment with zero product intent is `casual_engagement`. A message expressing any interest in saving, deals, or products — even passively — is `on_brand` with medium or soft posture. Only `on_brand` hard fires when the user is clearly and actively requesting something specific.

## What You Get

* No sales pivots on compliments — the agent matches the user's energy and stops.
* No commercial nudges on complaints — empathy first, with [human escalation](/docs/inbound-agent/the-knowledge/human-escalation.md) ready.
* Zero LLM cost on harmful input — the cheap classifier drops it before the main agent fires.
* Predictable agent behaviour — intent and posture are routing decisions enforced by code, not prompt judgement.
* Same classification on every channel — website widget, social comment, DM, voice all run through the same 6-intent system.