Sally Xu of xAI's data science team published a post on the company's website on November 2 describing Data Bot, a shared bot the team built inside Grok Bot. It can query the data warehouse, run analyses, return results in the format the person asking prefers, and join Slack channels. The post estimates that last week it handled roughly 70 hours of work that would otherwise have landed on the data science team.
That figure is an internal estimate, and no methodology is given. The byline on the page reads SpaceXSI, as xAI merged with SpaceX this year. The post describes the team only as "small" and gives no headcount, so outsiders cannot tell how large a share of the workload 70 hours represents.
Four sets of rules it was taught
Data Bot runs on the team-bot feature of Grok Bot. Its abilities come from shared skills maintained by the data team, which fall into four categories:
- Metrics and analysis: unified definitions for core metrics such as active users and revenue, with source tables and query logic spelled out. When a metric moves, the bot first breaks it down by segment and region, then checks holiday calendars and product change logs, and only then draws a conclusion.
- Experiments: A/B test setup, power analysis, checks for exposure balance and sample bias, and rollout planning. A conclusion report is written only once the test has enough power.
- Charts and dashboards: line charts for trends and horizontal bar charts for category comparisons, following the team's theme and style guide. Charts from a conversation can be turned into Databricks or Hex dashboards.
- Data engineering: standards for writing and reviewing dbt models, conventions for pipeline alerts, and how to add instrumentation when a new feature ships.
The other piece is shared memory. The bot records things like which team owns a given pipeline, or that a particular metric drop was actually caused by a regional holiday. New memories need administrator approval before they take effect for all users. What one person teaches the bot, the next person does not have to teach again, and wrong memories have a checkpoint to get past.
Who is using it
The post describes how three colleagues outside the data function use it.
Product manager Kevin Niparko treats it like a coworker he can ask, gets answers within minutes, and does not need to rerun an analysis when he asks a follow-up. He likes short conclusions paired with a chart, and Data Bot outputs in that style. His description:
"having a data scientist in his pocket"
Jon Brenner, who works in lifecycle marketing, gives it a Braze campaign ID and asks what people who opened the email did afterward. For the X Subscription promotion, he used this method to count how many recipients actually switched plans, a question Braze's built-in metrics could not answer.
Katy Abbaszadeh, director of revenue growth, uses it to measure how campaigns move the sales pipeline and existing deals, to track weekly targets, to find where progress is stuck, and to turn the results into charts for presentations.
Once in Slack, it can also run tasks on a schedule, such as tracking a metric according to the data team's standards or sending a weekly report. Colleagues can ask follow-up questions directly on the report or have it look into unusual swings. The post says complex analyses that require more judgment still belong to the data team, and the team admits it is still working out what extra guidance the bot needs.
Internal tool or product: two routes
Compared side by side, there are a few reference points from the past year.
In June, Anthropic described its internal approach: connecting business data directly to Claude gave only 21% accuracy, but after metric definitions and rules were fed in first, Claude could answer about 95% of data analysis questions on its own. In September, OpenAI added a Data agent to ChatGPT Work, taking the product route: it connects to data sources such as Snowflake and BigQuery and builds dashboards directly in six tools including Tableau and Power BI.
The xAI post is closer to the Anthropic approach: an internal team writes tacit knowledge into skills, and the model is just the executor. Unlike Anthropic, xAI has not published an accuracy figure, either as a before-and-after comparison or as an account of how wrong answers get caught. The product managers and marketers asking the questions often cannot verify a SQL query themselves, and that gap matters more than the 70 hours when it comes to whether other companies can copy this.
The same goes for permissions. The post mentions that memory updates require approval, but not who can query which tables. For a company-wide bot wired into a data warehouse, how access to revenue and user data is scoped is not explained.
The post also does not say whether Data Bot itself will be sold externally. The Grok Bot page offers a download and a contact-sales option, and in October it began assigning users dedicated email addresses. Whether outside customers will get the team-bot feature to the same degree depends on xAI's later product documentation.
Sources: xAI official blog, CocoLoop, Grok Bot official account. The 70-hour figure is the team's internal estimate; the Anthropic and OpenAI comparisons draw on figures both companies disclosed earlier.