Can you really get data answers just by asking in Slack? Anthropic shows how

Hi, it's shii-chan! Today I want to share an internal case study from Anthropic — how they made Slack data analytics as easy as just asking a question.

Claude Blog claude.com

What was announced?

Anthropic's Claude Blog published a case study about how its data team uses Claude Tag to power self-service analytics in Slack. Building on their earlier rollout of Claude Code internally (which reportedly reached around 95% accuracy), they extended Claude Tag to answer ad-hoc data questions right inside Slack.

Why it matters

"I just want a quick number, but I have to wait for the data team to get back to me" is a familiar pain point at a lot of companies. What makes this case study interesting is that it's not just "drop an AI agent into Slack and you're done" — it walks through permission design and data infrastructure just as much as the AI part. Because it's a real production deployment, there's enough concrete detail here for other teams to actually learn from.

Dive Deep

The technical foundation rests on three pieces:

• A governed semantic layer
• Markdown-based "skill files"
• An evaluation suite

One point the post really emphasized: skill files should be treated as continuously-updated shipped content, not a one-time knowledge dump. Since data models can change several times a day, skills need to be updated in near real time to keep up.

Five implementation recommendations stood out:

• Update skills dynamically by reading markdown from disk in real time
• Go beyond static knowledge — bake in actual analytical skills like predictive analytics, cohort analysis, and funnel analysis
• Connect the bot to business context: incident trackers, Slack threads, and internal doc search
• Take permissions seriously (service account restrictions, column-level PII classification, audit logging)
• Continuously measure adoption rate and accuracy

The permissions section was especially detailed. The guiding principle is "whoever can mention the bot has the bot's data access," implemented through five mechanisms: access limited to governed data only, column-level PII classification, clear connection paths, channel membership doubling as access control, and every query being labeled.

For rollout order, they recommend: design permissions first, verify the delivery mechanism, log telemetry from day one, wire in the knowledge index, then gradually add analytical skills.

What changes

In practice, the data team's Slack channels reportedly saw a response rate above 75%. One example: someone asked about a dashboard issue and got it identified and fixed within 90 seconds. Day-to-day analytics work — weekly summary reports, multiple daily experiment monitoring checks, pipeline and dashboard monitoring, and triaging data questions — is now largely automated.

Wrap-up

• Anthropic's data team built self-service Slack analytics on top of Claude Tag
• The foundation is three-pronged: a semantic layer, markdown skill files, and an eval suite
• Skill files work best when treated as continuously-updated shipped content
• Permissions follow "whoever can mention the bot gets the bot's data access"
• The data team's channels saw a response rate above 75%

If you're on a data or platform team thinking about deploying an AI agent in Slack, this post is packed with practical, ready-to-use lessons!