How to Build Your First AI Assistant in Dust (Step by Step)
Dust (dust.tt) is a workspace tool for building AI assistants that read your company’s data and answer inside Slack. You define an assistant once — its instructions, its data sources, its tools — and everyone in the workspace can use it without touching a prompt.
This guide to Dust walks through a first real assistant, not a demo. We build one that answers product questions from a Notion space and posts to a Slack channel. Expect 30 to 45 minutes including connector setup.
Does Dust replace ChatGPT for a team?
Partly. It replaces the pattern where five people each paste documents into a chat window. The assistant holds the connectors, so answers are consistent and auditable.
It does not replace the model. Dust routes to models behind the scenes — OpenAI, Anthropic, and others depending on your plan — and you can usually pick per assistant which one it uses.
Step 1: Set up the workspace and roles
Sign up and create a workspace. Invite the two or three people who will actually use the assistant before it goes live; testing alone hides permission problems.
Roles matter more than people expect. Admins can see and edit every assistant and every connector. Regular members can use assistants shared with them but cannot open the underlying data connections. Decide who is an admin on day one — retroactively removing admin access after someone connects a private drive is a headache.
Step 2: Connect the data sources
Open the workspace settings and go to connections. Each connector authorizes through OAuth with the third-party service.
- Notion — pick specific databases or pages, not the whole workspace. Whole-workspace authorization means your assistant can retrieve HR pages.
- Slack — the Slack integration is how answers get delivered; enable it early, because channel-level permissions are configured here too.
- Google Drive — optional, but common. Scope it to a folder.
Authorize with an account that has exactly the access you want the assistant to have. The assistant inherits the permissions of the connecting account, which is the most common cause of both over-exposure and mysterious empty answers.
Step 3: Write the assistant’s instructions
Create a new assistant and give it a role statement, not a personality statement. Four lines do most of the work:
- What it is. “You answer product questions for the Acme customer support team.”
- What it reads. “Use only the Notion ‘Product Docs’ database.”
- What it does when data is missing. “If the answer is not in the docs, say so and suggest who to ask.”
- Output shape. “Answer in under 120 words, then list the source page.”
That third rule is the one people skip, and skipping it is why assistants invent answers at the edges of your documentation.
Step 4: Add tools and test
Dust assistants can call tools — a web search, a data query, a Slack channel lookup. Add one at a time and test after each. Two tools wired at once makes it impossible to tell which one produced a bad answer.
Run 10 to 15 real questions collected from your team before writing prompts. Include three questions the docs do not answer. If the assistant confidently answers those, your instructions need a stronger fallback rule.
Step 5: Publish to Slack
Publish the assistant and choose which Slack channels it should be available in. From then on a mention in that channel triggers it.
Restrict publishing to two channels at first. A widely published assistant with a bad fallback rule generates wrong answers at scale, and people trust a bot inside Slack more than they trust a chat window.
What does Dust cost, and where does it fall short?
Pricing is per seat on a workspace plan, with usage-based limits on top. The free tier is workable for evaluation with a small team but caps message volume and connector count. Verify current numbers on the pricing page — plans have been restructured at least once.
The real weakness is connectors that only expose what the underlying app’s API exposes. A wiki built in a tool with no real API will not be readable no matter how you configure Dust. Also, assistants are document-oriented: for spreadsheet math or multi-step data analysis, you want a different tool.
Quick Summary
- Dust is a workspace platform for AI assistants connected to company data, primarily delivered through Slack.
- Scope connectors narrowly — specific Notion databases, specific Drive folders — rather than authorizing a whole account.
- Assistant instructions need four things: role, allowed sources, a missing-data fallback, and an output format.
- Add tools one at a time and test with 10-15 real questions, including three your docs cannot answer.
- Pricing is per seat with usage limits; the free tier fits evaluation, not a support team of 20.
