Back office ops · Production

BlaBlaCar builds a Data Copilot to shift data analysis left with LLMs and GitHub Copilot

The problem

Software Engineers at BlaBlaCar were blocked from self-service data analysis by unfamiliar tooling — SQL dialects, DBT, and Airflow — while Data Analysts were overwhelmed with ad-hoc 'quick questions', producing unmaintainable spaghetti SQL scattered across notebooks and sheets, and losing institutional knowledge in ephemeral Slack messages.

First attempt

Generic AI assistants and existing Text-to-SQL tools were found unsuitable: generic AI tools lack business-specific context, and text-to-SQL products are designed for non-technical business users rather than engineers.

Workflow diagram · grounded in source
1
Engineer asks question in IDE
Trigger
An engineer asks a data question directly inside VS Code.
▾ source quote
“When an engineer asks, 'How do I calculate monthly active users?'”
2
Semantic codebase search
Ai action
GitHub Copilot triggers #codebase to perform a semantic search across the workspace, identifying the intent behind the query.
▾ source quote
“It triggers #codebase to perform a semantic search across the workspace, identifying the intent behind the query.”
3
Literal string search
Ai action
GitHub Copilot utilizes #textSearch to scan for literal string matches within the generated documentation.
▾ source quote
“It utilizes #textSearch to scan for literal string matches within the generated documentation.”
4
Context injection from curated files
Integration
Relevant SQL and schema files are surfaced and their content injected directly into the chat context.
▾ source quote
“It surfaces files like tables/kpis_mau/usage.sql or tables/carpool_members/ddl.sql and injects their content directly into the chat context.”
5
Data Health Card generated
Validation
A Data Health Card is generated alongside every analysis, running heuristic checks as a soft warning system.
▾ source quote
“generating a Data Health Card alongside every analysis. It runs heuristic checks providing a Soft Warning system.”
6
Analysis generated as code artifact
Output
Every analysis is generated as a Python script with auto-generated unit test assertions.
▾ source quote
“every analysis is generated as a Python script with auto-generated Unit Tests (assertions)”
7
DA reviews pull request
Human review
The Data Analyst reviews the pull request, seeing the code and passing tests rather than a static chart.
▾ source quote
“The Data Analyst reviews the Pull Request. The DA doesn't just see a chart; they see the code and the passing tests.”
8
Repository indexes PRs as memory
Feedback loop
The Copilot indexes every merged pull request so the repository effectively becomes its long-term memory for future queries.
▾ source quote
“The Copilot indexes every merged Pull Request, meaning the repository effectively becomes its long-term memory.”
Reported outcome

Questions that used to sit in a Jira backlog for 3 weeks are now answered in 10 minutes inside VS Code; engineers validate feature ideas with data autonomously, and Data Analysts are freed from the support queue to focus on deep-dive modeling.

Reported metrics
Question resolution time3 weeks to 10 minutes
Engineer data autonomyEngineers are Autonomous
Analyst scalabilityAnalysts are Scalable
Reported stack
BlaBlaCar Data CopilotBigQueryGitHub CopilotVS CodeLLMsDBTAirflow
◆ Does this fit your context?

Compare to your context

Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.

Compare to your context →
~30 seconds · free
Source
https://medium.com/blablacar/why-we-built-blablacar-data-copilot-shifting-data-analysis-left-b4cc246faf52
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Questions that used to sit in a Jira backlog for 3 weeks are now answered in 10 minutes inside VS Code; engineers validate feature ideas with data autonomously, and Data Analysts are freed from the support queue to fo…

What tools did this team use?

BlaBlaCar Data Copilot, BigQuery, GitHub Copilot, VS Code, LLMs, DBT, Airflow.

What results were reported?

Question resolution time: 3 weeks to 10 minutes; Engineer data autonomy: Engineers are Autonomous; Analyst scalability: Analysts are Scalable (source-reported, not independently verified).

What failed first in this deployment?

Generic AI assistants and existing Text-to-SQL tools were found unsuitable: generic AI tools lack business-specific context, and text-to-SQL products are designed for non-technical business users rather than engineers.

How is this back office ops AI workflow structured?

Engineer asks question in IDE → Semantic codebase search → Literal string search → Context injection from curated files → Data Health Card generated → Analysis generated as code artifact → DA reviews pull request → Repository indexes PRs as memory.

Related back office ops cases

WHAT TO DO WITH THIS

Now compare it to your context

This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.

Browse the Back office ops cases →