Introduction

This manual walks you through every part of the Breezee AI dashboard — from the moment you sign in for the first time to the moment a working chat agent goes live on your website. It is written for the admin who configures the product, not for end users who chat with the agent.

Each chapter exercises a real workflow against the live application. The screenshots show a fictitious B2B consultancy tenant being configured end-to-end. The chapter prose focuses on what each control does in general; the screenshots show one specific tenant's configuration as a worked example.

What Breezee AI is

Breezee AI is a multi-tenant platform for building, configuring, and deploying AI chat agents. You configure an agent in the dashboard, point it at your knowledge base, give it a personality and a job to do, then deploy it as an embeddable widget on your website. Visitors who interact with the widget become prospects in your dashboard; the agent qualifies them, captures information, books meetings, and (optionally) syncs the result to your CRM.

The platform is designed for B2B and growth-stage businesses where the chat agent is doing real sales and support work — not a toy chatbot answering trivia.

Key concepts

These are the terms the manual uses throughout. You don't need to memorise them now; this section is a glossary you can come back to. Most of them are also covered in Appendix A.

Organisation

The top-level tenant. Everything you do in the dashboard belongs to an organisation. Most customers have one organisation per business; larger customers occasionally have several (separating a parent brand from acquired sub-brands, for example).

Team

A workspace inside an organisation. Teams hold agents, content, prospects, and deployments. Most organisations start with one team; multi-brand or multi-region customers split into several. The product internally calls these "projects" — you'll occasionally see that word in URLs.

Agent

The AI chatbot itself. An agent has a personality (persona, profession, tone), a job to do (its goal), a set of skills, and one or more deployments. An agent does not exist in isolation — it belongs to a team and draws on that team's knowledge and configuration.

Skill

A modular capability you attach to an agent. Skills are the agent's repertoire — what kinds of conversations it can have. Out of the box the catalogue includes:

• greeting — opens the conversation
• informational — answers questions using your knowledge base, with a fast-path for FAQ-style content
• advisory — runs a consultative discovery flow before making a recommendation
• product_finder — helps a visitor find the right product (or service) from a catalogue
• book_meeting — drives a visitor through to a booking
• contact_request — captures a request for someone to follow up
• qualify_prospect — captures structured information to score and qualify the visitor
• a few others for edge cases (unrelated, unknown, emergency_contact_request)

You configure which skills an agent has and tune each one to your business.

Action

A tool the agent can call during a conversation — display a card, inject context, or trigger a downstream action. You don't configure actions directly any more: they are auto-instantiated when you configure a skill. This manual doesn't dedicate a chapter to actions because there is no longer a separate UI for them; relevant detail appears inside the Skills chapter.

Memory type

A piece of information the agent remembers about a visitor — things like company size, industry, growth stage, current challenge. Memory types are configured per team and consumed by skills. You decide which memory types are enabled, which are required, and how forcefully the agent should ask for each one.

Prospect profile

A named pattern that groups prospects by their memory state — for example, "Growth-stage SaaS lead" might be defined as company_size between 20 and 200 and growth_stage == scaling. Prospect profiles are consumed by routing rules.

Routing rule

A deterministic rule that nudges the agent toward or away from a skill when certain conditions are met. There are two kinds:

• suggest — push the agent toward a target skill when criteria match (e.g. "if this looks like a Growth-stage SaaS lead and they haven't booked a meeting yet, suggest book_meeting")
• discourage — push the agent away from a skill (e.g. "discourage product_finder when intent is unrelated")

Routing rules give you control over agent behaviour without writing any prompt text or model code.

Content / knowledge base

The documents and web pages you ingest so the agent can answer questions accurately. Content is mapped from a website (URL scraping) or uploaded as files (PDF, DOCX, Markdown). After ingestion it is chunked, classified, attribute-extracted, and embedded so the agent can retrieve it during a conversation.

Deployment

A configured instance of an agent embedded somewhere. A deployment has its own appearance (colours, theme, position), its own allowed domains, and its own embed snippet. One agent can have multiple deployments — for example, different themes on different brand websites.

Prospect

A visitor who has interacted with the chat widget. Each prospect carries the conversation history, the memories the agent captured, and any contact information they shared. The Prospects screen in the dashboard is where you triage and follow up on leads.

How a chat turn works (admin-level view)

You don't need to understand the internals of the chat engine to configure the product well, but a one-page mental model helps when something behaves unexpectedly.

When a visitor sends a message, the platform runs five things in sequence:

1. Intent planning — a classifier decides what the visitor is trying to do (asking a question, looking for a product, asking to book, etc.).
2. Tool planning — a separate decision about whether to search your knowledge base for this turn, and which search policy to use.
3. Slot extraction — a model pulls structured information out of the visitor's message (company size, industry, etc.) and updates the conversation's working memory.
4. Routing evaluation — your routing rules are evaluated against the updated memory state. If a rule fires, it adds a directive to the agent's instructions for this turn.
5. Response — the agent assembles its prompt (greeting + skills + retrieved content + routing directive + the conversation so far) and streams a response back to the visitor.

A sixth step runs in the background after the response is sent: the platform summarises the turn, updates the conversation summary, and persists the captured memories.

The reason this matters for you, the admin: most of what you configure in the dashboard — skills, memory types, routing rules, content — shows up in one or more of these steps. When the agent behaves in a way you didn't expect, working back through the steps usually identifies what to change.

How to read this manual

The chapters are designed to be read in order. Earlier chapters set up the state that later chapters depend on — for example, you can't configure routing rules until you have prospect profiles, and you can't have prospect profiles until you have memory types. The order is:

1. Account & onboarding — sign in, create your organisation
2. Organisation & team management — set up the team workspace
3. Building your first agent — create the agent and shape its personality
4. Ingesting knowledge — feed the agent your content
5. Memory types & prospect profiles — define what the agent remembers and how it segments visitors
6. Configuring skills — give the agent its repertoire
7. Routing rules — steer the agent's behaviour
8. Playground testing — test the agent before deploying
9. Deployments & widget embed — put the agent live on your website
10. Prospects — triage leads as they come in
11. Analytics — measure what the agent is doing
12. Integrations — connect HubSpot and other downstream systems

Two appendices at the end:

• Appendix A — Glossary
• Appendix B — Troubleshooting

You can dip into any chapter as a reference, but if you're configuring a fresh tenant from scratch, go in order — the screenshots and example data follow a continuous narrative.

A note on the screenshots

Every screenshot in this manual is captured in light mode against the live application. If your dashboard is in dark mode and you'd like to match the screenshots, use the theme toggle in the top-right of any dashboard page.

Ready? Start with Chapter 1 — Account & onboarding.