Sharing the first iteration of our Discourse AI Enterprise Getting Started Guide. All feedback and questions will help make it better. Thank you!
Overview
Discourse AI is a modular suite of AI-powered features built directly into Discourse. It enhances community engagement, streamlines moderation, and helps users find and create content more effectively. Each feature has its own independent toggle and can be enabled or disabled individually. Your team controls what is active, when, and for whom. Some features, such as Related Topics and Sentiment, are enabled by default on hosted sites and can be disabled at any time.
Feature Summary
The table below provides a quick reference for every Discourse AI feature, including what data it processes and who can see AI-generated output.
| Feature | What it Does | Data Processed | Who Sees it | Default |
|---|---|---|---|---|
| AI Bot | Conversational chatbot via DMs or mentions; supports custom AI agents | User messages + community content (via search) | Users in allowed groups | Off |
| AI Search | Semantic search | Topic content (vector embeddings) | All users | Off |
| AI Triage | Automated post classification via Discourse Automation: auto-tag, auto-flag, auto-categorize. Requires rule configuration. | Post content when automation triggers | Staff-defined automation rules | Off |
| Helper | Assists composing posts: proofreading, title suggestions, explaining text | Text selected/written by user in composer | The individual user | Off |
| Related Topics | Provide a list at the bottom of a topic | Topic content (vector embeddings) | All users | On |
| Sentiment | Classifies post sentiment and emotion for moderation dashboards | Post content (automated classification) | Staff only (admin dashboard) | On |
| Spam Detection | Scans new posts from low-trust users for spam; one-click enable | Post content from new/low-trust users | Staff only (behind the scenes) | Off |
| Summarize | Generates summaries of long topics or chat channels | Topic/channel post content | Users with access to the topic/channel | Off |
| Translation | Translates posts and topic titles-requires target language configuration | Post text, topic titles | All users viewing translated content | Off |
Data & Privacy
Enterprise customers typically need clarity on data handling before evaluating features. Here are the key commitments.
No Training on Your Data
Discourse does not use customer data to train AI models, nor does it share customer data with third parties for that purpose. Customer conversations and content belong to the communities that create them.
Privacy by Design
-
All AI data is stored alongside your Discourse content: you own it, always
-
You can configure any supported LLM provider you choose where AI processing happens.
-
The hosted LLM uses open-weight models: you are not locked into any single vendor or model architecture
-
Prompt customizations are site-specific: changes stay on your instance
-
Group-based access control: most features can be restricted to specific user groups
-
Usage monitoring: admin dashboard tracks token usage and costs by model and feature
Modularity & Control
Each Discourse AI feature operates independently. You can enable a single feature without affecting any other. This means:
-
Turn on spam detection without enabling any user-facing AI features
-
Enable translation without exposing the AI bot
-
Roll out search improvements before introducing content generation
-
Pilot any feature with a trusted group before enabling site-wide
Every feature toggle is a simple admin setting. The following table shows the master switch and all individual feature controls:
| Setting | Feature | Default |
|---|---|---|
| discourse_ai_enabled | Master plugin toggle | Off |
| ai_translation_enabled | Content translation | Off |
| ai_bot_enabled | AI chatbot | Off |
| ai_helper_enabled | Composer/post AI assistance | Off |
| ai_summarization_enabled | Topic/chat summarization | Off |
| ai_sentiment_enabled | Sentiment analysis | On |
| ai_embeddings_enabled | Embeddings (search, related topics) | On |
| ai_spam_detection_enabled | Spam detection | Off |
Note*: AI Triage does not have a single toggle. It is configured via Discourse Automation rules.
Recommended Rollout Strategy
The exact rollout depends on your community’s priorities, constraints, and data policies. At a high level, we recommend three steps:
-
Decide what LLM and embeddings service to use and get it provisioned
-
Enable high-value, low-risk features first
-
Enable other features next, one at a time, prioritized by your assessment of value, effort, and risk. If desired, start with a trusted group before wider rollout.
Step 1: Choose LLM and Embeddings Services
You are not locked into any single AI provider. Discourse AI supports a wide range of LLM backends. Your options include:
| Option | Description | Setup | Best For |
|---|---|---|---|
| Discourse Hosted LLM | Pre-configured open-weight model, included with your hosting plan. Usage subject to your plan’s credit limits. | No API keys needed; usage limits by plan tier | Simplest path; ideal for getting started |
| AWS | Use AWS-hosted LLMs and embeddings within your own AWS account | Requires AWS account and API configuration | Data stays within your AWS environment |
| Bring Your Own (BYO) | Connect any supported LLM provider: OpenAI, Anthropic, Google, Azure, and more | Requires API keys from your chosen provider | Maximum flexibility; use preferred models |
In practice, the three common configurations are:
-
Metal, with Discourse hosted services for LLMs and embeddings: the simplest path, no API keys needed
-
AWS, with AWS hosted services for LLMs and embeddings: data stays within your AWS environment
-
AWS or metal, with BYO services for LLMs and embeddings: maximum flexibility with your preferred provider
Supported third-party providers include OpenAI, Anthropic, Google Gemini, Microsoft Azure, Amazon Bedrock, Cloudflare Workers AI, HuggingFace, vLLM, SambaNova, OpenRouter, Together AI, Groq, and more
Step 2: Enable High-Value, Low-Risk Features First
These features operate behind the scenes or enhance existing functionality without changing user workflows. They deliver immediate value with minimal change management:
-
Related Topics: suggests relevant discussions to users at the end of every topic
-
Semantic Search: improves how users find content, going beyond basic keyword matching
-
Spam Detection: scans new posts from low-trust users for spam, entirely behind the scenes
Step 3: Enable Other Features One at a Time
Prioritize based on your assessment of value, effort, and risk. If desired, start with a trusted group before wider rollout. Features to consider next:
-
AI Bot: conversational chatbot users interact with via DMs or mentions; supports custom AI agents
-
Translation: automatic AI-powered translations for your global community
-
AI Triage: automated post classification, tagging, and flagging (requires Automation rule configuration)
-
Summarize: generates summaries of long topics or chat channels
-
Sentiment: classifies post sentiment and emotion for moderation dashboards
-
Helper: assists composing posts with proofreading, title suggestions, and more
Tip: Start Simple You don’t need to plan for all features up front. Many Enterprise customers start with just the Discourse Hosted LLM and spam detection enabled, delivering immediate value with minimal configuration. From there, expand as your team is ready.
Getting Started
Ready to begin? Here’s what to do next:
-
Review the LLM provider options above and decide which path fits your organization’s data policies
-
Enable Discourse AI via Admin → Plugins → Discourse AI (the master toggle)
-
Configure your LLM connection on the LLMs settings page (if using BYO provider)
-
Start with high-value, low-risk features: Related Topics, Semantic Search, and Spam Detection
-
Monitor usage and impact via the AI Usage dashboard in Admin
Your Discourse account team is available to help plan and support your rollout at every step.