Back office ops · Production

Sweetgreen transforms unstructured data into conversational analytics with dbt and Claude

The problem

Sweetgreen's data was fragmented across multiple ingestion paths and databases that produced inconsistent metric values for the same question, every new business question required building a new script or pipeline from scratch, and manual Google Sheets ingestion allowed upstream errors to propagate downstream into dashboards — leaving the data team as a permanent bottleneck to timely insights.

First attempt

Multiple business intelligence tools including Tableau, PowerBI, and ThoughtSpot had been tried but failed due to a cultural adoption barrier — business users did not want to learn new tools.

Workflow diagram · grounded in source
1
Build fact and dimension models
integration
“Fact tables captured ‌core business events and served as the source of truth for measures (e.g., dollars sold, units sold). Dimension tables added standardized context and hierarchies (e.g., store → city → region).”
2
Standardize KPIs in Semantic Layer
integration
“dbt's Semantic Layer built on top of these tables to standardize KPIs and aggregation logic so metrics stayed consistent across dashboards and conversational analytics”
3
Quality checks catch bad data
validation
“The data team also added custom, model-specific quality checks on top of dbt's built-in tests to catch issues early and prevent bad data from reaching the business”
4
User asks question in plain English
trigger
“business teams can leverage conversational AI to self-serve data insights by asking questions in plain English and getting consistent, reliable answers”
5
Claude queries governed semantic metrics
ai_action
“Claude was reading directly from the dbt Semantic Layer (which enforced consistent metric definitions), users received the same accurate, reliable answers grounded in version-controlled, tested metrics, not hallucinations or unverified n…”
6
Conversational analysis delivered
output
“Claude reviews the available semantic models and provides multiple analyses (e.g., by channel, venue, time, or customer behavior) that stakeholders may not have considered”
Reported outcome

Self-service data analysis dropped from a two-week wait to a 30-minute job; business teams now query data in plain English through Claude, and the data team shifted from gatekeeper to enabler with faster, more consistent insights.

Reported metrics
Self-service analysis time30-minute job
Previous data team response waittwo weeks
Insights speed and metric consistencyfaster insights and more consistent metrics
Business confidence in datagained a lot of confidence in the data
Show all 5 reported metrics
self-service analysis time30-minute job
previous data team response waittwo weeks
insights speed and metric consistencyfaster insights and more consistent metrics
business confidence in datagained a lot of confidence in the data
data team time spent answering questionsfreed us up from spending so much time just answering questions
Reported stack
dbtdbt Semantic LayerClaudeMCPClaude Desktop
Source
https://www.getdbt.com/case-studies/sweetgreen
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Self-service data analysis dropped from a two-week wait to a 30-minute job; business teams now query data in plain English through Claude, and the data team shifted from gatekeeper to enabler with faster, more consist…

What tools did this team use?

dbt, dbt Semantic Layer, Claude, MCP, Claude Desktop.

What results were reported?

Self-service analysis time: 30-minute job; Previous data team response wait: two weeks; Insights speed and metric consistency: faster insights and more consistent metrics; Business confidence in data: gained a lot of confidence in the data (source-reported, not independently verified).

What failed first in this deployment?

Multiple business intelligence tools including Tableau, PowerBI, and ThoughtSpot had been tried but failed due to a cultural adoption barrier — business users did not want to learn new tools.

How is this back office ops AI workflow structured?

Build fact and dimension models → Standardize KPIs in Semantic Layer → Quality checks catch bad data → User asks question in plain English → Claude queries governed semantic metrics → Conversational analysis delivered.

Related back office ops cases