Most support forums built on Discourse eventually hit the same wall: the answer to a question already exists somewhere on the forum, three years and two hundred topics deep, and nobody asking today is going to find it. So they post again. A moderator answers again, or locks the new topic and points at the old one, which only works if they remember the old one exists. Multiply that by every category on an active forum and moderation stops being about rule enforcement and turns into search.
The Discourse integration is built around that specific problem rather than around moderation in general.
What it actually does
You pick the categories to watch — or leave the field blank and it watches all of them. Every new topic and every reply in those categories gets pulled through the same pipeline every other channel uses: triage, retrieval against your knowledge base, a confidence score, and then either a posted reply or a held draft. If the AI is confident, it replies in the topic as a bot account you control. If it isn't, the answer sits on the dashboard for someone to approve, edit, or dismiss — nothing goes out under your forum's name without a person having seen it, unless you've turned that off deliberately.
That last part matters because Automatic Deflections is off by default here, same as on every channel. You review real drafts before the AI is allowed to post unsupervised.
Setup, in practice
There's no OAuth screen and no shared platform app for Discourse — you create the connection from your own admin panel, which means you're not waiting on approval from anyone but yourself:
- Create a bot account in Discourse (something like
answerloops-bot) with Trust Level 1 or higher. - Generate an API key under Admin → API → New API Key, scoped to that bot account with
web_hooks,posts, andcategoriesaccess (a Global key also works). - Paste your forum URL, the key, the bot username, and optionally a comma-separated list of category IDs into Integrations → Discourse.
- Click Register webhook and answerLoops creates the webhook on your forum for you. If your instance restricts webhook management through the API, you get the payload URL and secret to add it by hand instead.
A category's ID is right there in the URL — /c/support/6 is category 6 — so there's no digging through settings to find it.
Why this is different from just having a bot reply
The point isn't a bot that answers questions. Plenty of forums have tried that and ended up with a bot confidently making things up in a thread where a wrong answer costs someone real time. The retrieval step is grounded in your actual knowledge base, and the confidence threshold — 0.8 by default, adjustable per organization — is the gate between "reply directly" and "flag someone." Set an escalation user and a low-confidence topic mentions them by username instead of leaving the asker with silence.
The bot also recognizes its own posts and skips them, so it never enters a loop replying to itself, and it ignores anything under ten characters, so a one-word bump doesn't trigger a fresh answer.
What this changes for the people running the forum
The topics that once ate a moderator's morning — "how do I configure X," asked for the fortieth time in a slightly different phrasing — get handled before a person has to type the same answer they've typed all year. What's left in the queue is the stuff that actually needs a person: edge cases, bugs, questions the knowledge base genuinely doesn't cover yet. Those gaps show up as low-confidence tickets, which is itself useful — it's a running list of what your documentation is missing, generated by the questions people are actually asking instead of a guess at what they might ask.