> ## Documentation Index
> Fetch the complete documentation index at: https://docs.usefini.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Introduction
> Build AI support agents with deterministic workflows, full observability, and the controls that regulated industries demand.
Fini is a platform for building AI agents that handle customer support, reading conversations, answering from your knowledge base, running multi-step workflows, and escalating to a human only when they should.
What lets teams in regulated industries (fintech, banking, insurance, healthcare) put AI agents in front of real customers is the combination of fine-grained control over what the agent does and full visibility into what it did. The sections below break down how Fini delivers both.
## The architecture
```mermaid theme={null}
---
title: How Fini sits in your stack
---
flowchart LR
subgraph KNOW["KNOWLEDGE SOURCES"]
TICKETS["Past tickets"]
HC["Help center"]
DOCS["Internal docs
Notion · Confluence · SharePoint"]
API_IN["Custom APIs
(your data)"]
end
FINI(("Fini
AI agent layer"))
subgraph DEPLOY["DEPLOYMENT SURFACES"]
HELPDESK["External helpdesks
Zendesk · Intercom · Front · HubSpot
Salesforce · Gorgias · LiveChat · Freshdesk"]
WIDGET["Chat widget
web · authenticated app · mobile"]
INTERNAL["Internal chat
Slack · Discord"]
HELPCENTER["Help center"]
BROWSER["Chrome extension"]
VOICE["Voice"]
API_OUT["API
(anywhere else)"]
end
HUMAN["Your human team
(escalation backstop)"]
TICKETS --> FINI
HC --> FINI
DOCS --> FINI
API_IN --> FINI
FINI --> HELPDESK
FINI --> WIDGET
FINI --> INTERNAL
FINI --> HELPCENTER
FINI --> BROWSER
FINI --> VOICE
FINI --> API_OUT
HELPDESK -.->|"escalate"| HUMAN
WIDGET -.->|"escalate"| HUMAN
INTERNAL -.->|"escalate"| HUMAN
VOICE -.->|"escalate"| HUMAN
classDef source fill:#F7F7F7,color:#131415,stroke:#E8E8E8
classDef agent fill:#131415,color:#FFFFFF,stroke:#131415,stroke-width:3px
classDef surface fill:#FFFFFF,color:#131415,stroke:#131415
classDef human fill:#C3EE5E,color:#131415,stroke:#131415,stroke-width:2px
class TICKETS,HC,DOCS,API_IN source
class FINI agent
class HELPDESK,WIDGET,INTERNAL,HELPCENTER,BROWSER,VOICE,API_OUT surface
class HUMAN human
```
Read the picture left to right. Knowledge flows in from any combination of sources. Fini holds that knowledge, applies your prompts and rules, and answers on whichever deployment surfaces you've connected. When the agent can't resolve a conversation, by policy, by confidence, or because the customer asked for a person, it escalates to your team with the trace and context already attached.
## What makes Fini different
If you've evaluated other AI support tools, five design choices are worth knowing about up front.
### Deterministic Rulebook execution
Most AI support products use the LLM to decide *what to do next*. Fini uses the LLM only inside behavior tree nodes, Check (intent matching), Reply (answer composition), Form (data collection), and uses a deterministic tree walker to decide which node runs when. The result: identical inputs produce identical execution paths. You can test a flow once and trust it to behave the same way in production.
See [Rulebook](/en/automations/rulebook) for the full execution model.
### Full observability per conversation
Every conversation Fini handles produces an AI Steps trace, an ordered log of every node that executed, what it received, what it returned, and what the agent did with it. There is no opaque "the AI decided to do X." If your agent gave a wrong answer, the trace tells you exactly which step produced it.
See [Inbox](/en/testing/inbox) and the [AI Steps trace](/en/automations/rulebook#observability-the-ai-steps-trace) for how this works in practice.
### Self-improving from every conversation
Fini watches the conversations its agents handle and proposes improvements automatically. It identifies **gaps in your knowledge base** (questions the agent couldn't answer), **conflicts between sources** (where two articles contradict each other), and **prompt-level issues** (intents where conversations consistently end in negative sentiment). Proposed updates show up in [Review](/en/knowledge/review), you approve or reject; nothing ships to production without a human in the loop.
### Reply guardrails
Configure per-agent [Guardrails](/en/configuration/guardrails) to check generated replies for banned terms, confidential attribute values, disallowed links, or custom criteria. Review hits and individual verdicts in Inbox; a hit can lead to a rewrite rather than an escalation, and check errors can leave the original reply unchanged. Use [Reply Behavior](/en/automations/reply-behavior) for separate routing controls such as internal notes or escalation on defined patterns.
### Controls built for regulated industries
Fini's customer base skews toward fintech, banking, insurance, and healthcare, industries where the cost of a wrong answer is high and the audit requirements are non-negotiable. That shapes the product:
* **Tags and Reply Behavior** let you say "never reply, only leave an internal note" for sensitive intents like account closure or chargebacks.
* **Test Suite** runs regression checks against your full Rulebook before every deploy.
* **Compliance posture**, SOC 2 Type II, GDPR, ISO 27001, BAA for healthcare deployments.
## How an agent works
Every agent in Fini has four aspects:
```mermaid theme={null}
graph LR
A[Agent] --> K[Knowledge
what it knows]
A --> B[Behavior
how it acts]
A --> D[Deployment
where it lives]
A --> O[Observability
how you watch it]
```
You configure each aspect independently, and they compose into one running agent. The five sections of the docs below map to these four aspects, plus the testing and analytics tools that make observability concrete.
What the agent knows. Connect your help center, upload files, link Notion or Google Drive, or write articles by hand. [Magic Articles](/en/knowledge/magic-articles) drafts new content from real conversations the agent couldn't answer.
How the agent thinks. [Prompts](/en/configuration/prompts) shape tone and high-level guidelines. [Tags](/en/configuration/tags) classify every conversation. [Attributes](/en/api-reference/attributes) pull in customer context. [Actions](/en/api-reference/actions) let it call your APIs.
What the agent does. [Rulebook](/en/automations/rulebook) builds step-by-step workflows like cancellations or refunds as behavior trees. [Reply Behavior](/en/automations/reply-behavior) decides whether to reply, leave an internal note, or stay silent.
Where the agent shows up. Embed the widget on your site, connect a helpdesk like Zendesk or Intercom, or publish a hosted help center.
How you keep it sharp. Audit real conversations in [Inbox](/en/testing/inbox), run regression checks in [Test Suite](/en/testing/test-suite), and track performance in [Analytics](/en/analytics).
You can run as many agents as you need. Each agent has its own knowledge, configuration, and deployments, so one workspace can power separate flows for support, sales, internal IT, or per-product teams.
## What a workflow looks like
Here's what a cancellation flow looks like as a behavior tree in Fini:
```
Steps
├── Check: intent is "cancel subscription"
├── Read: customer attributes (plan, status, billing date)
├── Check: plan is cancellable
├── Form: confirm reason for cancellation
├── Tool: call cancel-subscription API
└── Reply: confirm cancellation and next billing date
```
Each node is a small, well-defined primitive. The tree walker runs them in order, short-circuiting on the first failure. If the plan-is-cancellable Check fails, the customer is told why and the conversation is escalated. If the API Tool fails, the agent retries or escalates. Every step shows up in the AI Steps trace.
The full [Cancellation flow walkthrough](/en/walkthroughs/cancellation-flow) builds this end-to-end, wiring Knowledge, Attributes, Actions, Rulebook, and Reply Behavior together.
## Where to start
Get your first agent live in under 30 minutes. The shortest path from zero to a working deployment.
The home page in the dashboard. Create new agents, see active deployments, and pick which agent to configure.
A concrete end-to-end example that ties Knowledge, Attributes, Actions, Rulebook, and Reply Behavior together.
Widget, helpdesk, or help center. Pick the surface that matches how your customers reach you today.
## How the docs are organized
| Section | What's in it |
| - | - |
| [Knowledge](/en/knowledge/overview) | Sources, articles, Magic Articles, review queue. |
| [AI Agents](/en/configuration/agent-home) | Creating agents, agent cards, the per-agent selector across the app. |
| [Configuration](/en/configuration/prompts) | Prompts, Tags, Attributes, Actions. |
| [Automations](/en/automations/rulebook) | Rulebook and Reply Behavior. |
| [Walkthroughs](/en/walkthroughs/cancellation-flow) | End-to-end recipes that wire several features together. |
| [Deploy](/en/deploy/overview) | Widget, help center, and helpdesk integrations (Zendesk, Intercom, HubSpot, Salesforce, Front, Gorgias, LiveChat, Slack). |
| [Testing](/en/testing/inbox) | Inbox, Test Suite. |
| [Analytics](/en/analytics) | Volume, deflection, CSAT, sentiment, escalations. |
If you're brand new, start with the [Quickstart](/en/quickstart). If you already have an agent up and want to deepen it, jump straight to the section you care about.
## Related topics
- [Quickstart](/en/quickstart.md)
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.