AI data analyst in Slack. Democratize data.
Hey everyone — This is Maceo and Jason, founders of Datost.
TL;DR: Datost is an AI data analyst your team @mentions in Slack. It keeps a semantic layer of your business definitions so it understands what questions mean before writing SQL. On the hardest public text-to-SQL benchmark, it scores 75.2%. The best frontier model alone gets 33%.
Why I built this
We saw the problem while I was at Traba, a staffing marketplace. Traba had a lot of data and not enough analysts. Someone would need a number to make a decision, and the path to that number ran through SQL they couldn’t write or an analyst who was busy. The question would sit for hours. By the time the answer showed up, the decision had already been made on a guess.
There are AI analytics tools out there now. They all have the same problem: they don’t understand your business. When your PM asks “which accounts are at risk of churning,” there’s no column called that. “At risk” has a formula that lives in someone’s head, or a doc, or nowhere. These tools guess an interpretation and commit. Then someone in a Slack thread acts on the answer before anyone checks whether it was right.
The semantic layer
Datost maintains a layer of your team’s definitions, what “churn risk” means at your company, how you calculate NRR, which tables are trustworthy, what that internal acronym stands for.
When it hits a term it hasn’t seen, it looks it up the same way a new analyst would ask a coworker. When the semantic layer doesn’t have it either, it asks the user before writing any SQL.
It learns from mistakes too. When it gets something wrong, it leaves notes on the schema so it doesn’t make the same mistake twice. Teams that have been using it for a few months barely correct it anymore.
What it actually does
Your team @mentions Datost in any Slack channel. It searches the semantic layer, explores your schema in a sandbox, writes and tests SQL, then posts the answer in-thread with the work shown. A second model reviews every answer read-only. If the definition says one thing and the SQL does another, the reviewer catches it.
It also runs on its own. You can set it up to post a pipeline coverage update to #sales every Monday, or flag when a metric moves outside a range you care about. Or if youre old fashioned it can make a beautiful dashboard for you :)
It connects to Postgres, MySQL, BigQuery, Snowflake, Databricks. Also Datadog, Sentry, PostHog, Coda, Notion, GitHub, and anything else via MCP. It reads your Slack history to see how things were answered before. It reads your codebase to understand how a metric is actually calculated. Charts, dashboards, PDFs, exports, whatever format you need. Most answers come back in under a minute.
The benchmark
We ran Datost against BIRD-Interact, the hardest public text-to-SQL benchmark. 600 deliberately ambiguous questions across 22 messy and realistic PostgreSQL databases, published at ICLR 2026. Every question uses terms that aren’t in the schema.
Claude Opus 4.6, the model Datost runs on, gets 33% on its own. Datost scored 75.2%. Same model. The gap is the semantic layer. On analytical questions, 91%.
Full breakdown here: https://datost.com/blog/bird-interact
Ask
If your team has a Slack, Teams, or Google Chat workspace and data people keep asking about, we can get you running in a day.
- Book a call
- Email me or Jason: maceo@datost.com or jason@datost.com
- Or reply here
