GitHub's Qubot lets staff query the data warehouse in plain language, with hundreds of users
GitHub built Qubot, an internal Copilot-powered analytics agent available in Slack, VS Code and the Copilot CLI. GitHub found that structured, curated context made it more accurate and three times faster at reaching the right answer.
- Organization
- GitHub
- Industry
- Developer Tools
- When
- Jun 2026
- Evidence
- Company reported
- Agents and tools
- GitHub CopilotSlackKustoTrino
Reported outcome
Hundreds of users running thousands of queries. In GitHub's experiments, curated context made Qubot three times faster at returning the right answer.
- hundreds of users, thousands of queriesAdoption
- three times faster to the right answerEffect of curated context (experiments)
Caveats
Stated by the company that uses the agent (blog, earnings call, interview). Not independently audited.
- GitHub gives no exact user count, accuracy rate or time-saved figure.
- 'Three times faster' compares context setups in GitHub's own experiments. It does not compare Qubot with a human analyst.
- GitHub sells Copilot, the product this agent is built on.
- Qubot is for exploratory questions. It does not replace dashboards, and complicated questions still go to the data team.
What the agent does
Qubot lets any GitHub employee ask about any data model in the data warehouse in plain language. It runs in Slack, VS Code and the Copilot CLI.
It defaults to Kusto for recent event data and switches to Trino for complex joins and historical analysis, so users do not need to pick an engine.
What happened
GitHub says Qubot has hundreds of users running thousands of queries, and that questions in its data and analytics Slack channels fell sharply. It gives no number for that drop.
The main lesson was about context. In GitHub's experiments, structured and curated context made Qubot more accurate and three times faster at returning the right answer.
What to copy
- Treat business definitions and data context as a maintained product.
- Let the teams that own the data write its context.
- Send each context change through a pull request and an offline eval.
- Say what the agent is not for. Qubot does not replace dashboards.
Starter prompt
Written by us, not by GitHub.
Data context (tables, definitions, business rules): {context}
Question: {question}
Use only tables and metrics listed in the context.
Show the query, the result, and the definitions you used.
If the context does not define a metric, say so and ask. Do not invent a definition.
Sources
- How we built an internal data analytics agent — GitHub Blog, Jun 19, 2026
Added Nov 9, 2026 · Sources last checked Nov 16, 2026.
Get new use cases weekly
One short email. Sources included.
The newsletter is paused. Past issues stay in the archive.
Related use cases
Airtable's Slack data agent is said to do work that would take analysts an extra 200 hours a week
The vendor says the agent handles what would take Airtable's data team an extra 200 hours a week. A typical example answer arrives in about three minutes.
Insight Health: 97% of critical-channel alerts closed without an engineer stepping in
Since Claude Tag went live, 97% of alerts in Insight Health's critical alert channel have closed without an engineer having to step in.
Tennr made an AI agent the maintainer of an internal tool; recruiters shipped 15+ tickets in a month
In roughly a month, non-engineers shipped 15+ tickets on the internal offer portal through Claude, and got production changes back within the hour.