> 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/why-this-is-not-a-chatbot.md).

# Why This Is Not A Chatbot

Live-data inbound agent that answers across every channel in one consistent brand voice.

> **Real product data, real intent, real brand voice — across every channel, in seconds.**

Most AI assistants answer from training data. The Oodlz inbound agent connects to a live product catalogue, classifies intent before responding, retrieves real offers with real deeplinks, and replies in a consistent brand voice across every channel simultaneously — voice or text, social comment or DM, chat widget or API call.

## What You Get

* Live retrieval from a brand-specific knowledge base — never frozen training data.
* [Six-intent classification](/docs/inbound-agent/the-brain/six-intent-classification.md) before any agent call — sales pressure only when the user is buying.
* Brand-ranked results — duplicates removed, exact matches promoted, highest cashback first.
* Cross-channel conversation memory — the same thread continues from Instagram DM to website widget.
* [Voice and text](/docs/inbound-agent/every-channel/voice.md) from one pipeline — markdown for the screen, clean prose for the ears.
* [Human escalation](/docs/inbound-agent/the-knowledge/human-escalation.md) with full context attached — session reference, original query, issue description.

## How It Works

A single webhook receives every inbound message — website widget, Instagram DM and comment, Facebook DM and comment, LinkedIn comment, X reply, BlueSky reply, voice client. Filler phrases are stripped, the cleaned query is classified into one of 6 intents, and the agent's posture is set before any expensive call fires. Harmful messages are dropped silently. Casual compliments get a 1–2 sentence reply with no tool call. Active product intent gets the full commercial pipeline.

When the agent does engage, it routes to the right specialist tool — brand catalogue search, general FAQ, how-to walkthrough, or human escalation — and returns a structured response. A re-ranking pass deduplicates results, promotes exact matches, and orders the rest by featured status and cashback rate. Voice requests fork in parallel: the response is stripped of markdown and URLs for natural speech, while the full markdown version is cached against a query ID for the client to retrieve separately.

Conversation memory lives in Redis under a `{platform}_{senderId}` key. Every channel uses the same scheme, so a user who asks on Instagram and follows up on the website widget hits the same thread.

## Quality Baked In

* Six-intent classifier runs before the main agent on every message — bad-faith input never reaches an LLM tool call.
* The `acknowledge` posture skips tool calls entirely — no "have you seen the latest deals" pivots after compliments.
* Brand results post-process: duplicates removed by name, exact matches promoted, featured brands prioritised, remainder sorted by cashback rate descending.
* Voice path strips markdown, removes raw URLs, and reformats brand listings into natural spoken sentences.
* Human escalation collects and validates email and issue description before the tool is called — no half-formed handoffs.
* One session-key scheme across every channel — shared memory enforced at the key level, not requested via prompt.

## Inputs

* Inbound message text or voice binary, plus channel and sender identifiers.
* Brand profile fields — name, tone, prohibited language, product type — pulled from the brand profile at runtime.
* Live brand catalogue — embedded into a per-brand vector index.

## Outputs

* Markdown response for text channels — formatted brand cards with name, cashback, description, categories, deeplink.
* Cleaned spoken text for voice channels, plus audio binary from text-to-speech.
* Structured human escalation event when the user requests a real person.
* Per-message log entry — channel, intent, posture, tool used, reply status.

## Cadence And Volume

Runs once per inbound message. The cheap classifier plus posture routing keeps cost at a fraction of routing every message through a generic frontier model directly. Volume scales with the brand's inbound traffic.

## What This Replaces

* The bolted-on chatbot answering from training data, not the live catalogue.
* The chat widget that loses memory the moment a user switches channels.
* The auto-reply tool that pivots every message — compliment, question, complaint — into a sales pitch.
* The voice add-on that reads markdown asterisks and hashtags out loud.

One agent, intent-aware, brand-voiced, channel-shared — replacing every one of those.

***

*This was built. No agent ships this out of the box.*