How a Small Data Platform Team Serves 2,000 People — with Self-Service and AI

Qonto is a European business bank with nearly 2,000 employees and a data platform team you could fit around one table. That ratio only works if the team stops being the bottleneck — if it ships products instead of pipelines, and lets the rest of the company serve itself.

This talk walks through how we got there, and what it cost us.

Self-service starts with an interface. Ours is Qontract, a data contract layer built on ODCS: any team declares the data it wants ingested — source, schema, freshness, ownership — and gets a production pipeline without a data engineer writing code. Contracts became the boundary between producers and the platform, which means schema changes stop being 2 a.m. incidents and start being pull requests.

Then the orchestration reality: dozens of dbt repositories owned by different domain teams, all scheduled through a single Airflow instance via declarative cross-repo sensors. We'll show how the multi-repo setup gives teams autonomy without forking the platform, where it breaks (dependency graphs nobody owns end-to-end, scheduler contention, warehouse routing), and the guardrails we added to keep it honest.

Next, trust as a product: an internal data pipeline status page, visible company-wide, so freshness and failures reach consumers before they reach us.

Then AI, which for us means three different things.

We build with it. Every product above was shipped faster because the team uses coding agents daily — a small team's throughput is now a function of how well it wields them.

We unblock it. AI without data is nothing, and the entire company's AI ambitions ran straight through us. That made a small platform team the bottleneck for something far bigger than analytics. Contracts, ownership metadata, and MCP access to the warehouse are what got us out of the critical path.

And we run on it. Our support and on-call load is increasingly handled by an AI duty agent that triages incoming data requests, investigates pipeline failures, and answers what used to be a human's whole morning.

Takeaways: what to standardize, what to hand over, and what a small team should refuse to own.