Why Canva Built an AI to Query Snowflake, Stripe, and Code

Key Takeaways

• Canva built an internal system connecting Snowflake, Stripe payment records, and code repositories to let executives query operations in natural language.
• Cliff Obrecht first attempted this project two years ago under the code name Maestro, but it stalled because company data structures were incompatible.
• Direct data queries uncovered operational waste that status reports hid, including hundreds of abandoned code experiments still running live.
• The AI system delivers insights that are about 90% accurate, removing the personal ambitions and biases that shape traditional middle-management reporting.
• Collison notes that querying ground truth directly shrinks large companies by removing the layers between leadership and actual code.

The Maestro Failure and Clean Data

Two years ago, Canva co-founder Cliff Obrecht tried to build a universal business dashboard called Maestro. It failed almost immediately. The idea was simple: build a central console that gave executives instant visibility across every team. The reality was messy. Canva's data lived in disconnected formats, scattered across different databases, spreadsheets, and third-party tools.

“I got super excited about this two years ago and I like design this system I used to call Maestro that gave me insights into everything,”▶ 37:20 Obrecht said. “Then we ran into the problem all our data structures weren't compatible and it was a bit of a mess.”▶ 37:25

The fix was not prettier business intelligence charts. Canva unified its foundational data layers, linking code repositories directly to Snowflake data warehouses and Stripe billing records. Once the structured data aligned, Canva layered language models over the top. Instead of waiting for an analyst to write custom SQL or build a dashboard, executives can ask plain-English questions across code, finance, and product metrics.

Sidestepping Human Spin and Stale Experiments

Every growing startup develops an information filter. As headcount climbs, the reports reaching the executive team get polished. Bad news gets softened, pet projects get protected, and metrics get cherry-picked.

“Historically every bit of data you got was also sort of clouded with human interpretation, human opinions, biases, sort of like their own ambitions,”▶ 35:28 Obrecht explained. “Now I can just natural language query... We've got this internal system that spans all that data and I can pull out really, really deep insights and they're generally 90% correct.”▶ 35:02

That 90% accuracy is enough to expose hidden drag. When Canva began querying its code repositories alongside active product analytics, leadership discovered hundreds of stale code experiments running undetected. These abandoned tests were cluttering the product, burning compute, and slowing down deployments. Nobody had reported them because nobody was incentivized to audit old failures.

Stripe president John Collison pointed out that this direct access changes executive work. It “effectively makes organization feel smaller cuz you can be closer to the actual ground truth of the work or the code.”▶ 35:57

Language models struggle with complex arithmetic, but they excel at semantic search across structured tables. When leadership can ask, "Which features have had zero commits and declining usage for six months?" without going through three layers of managers, company bloat has nowhere to hide.

What to Do With This

Pick your single largest operational blind spot this week: abandoned feature flags, inactive recurring subscriptions, or dormant customer accounts. Write a script or connect an LLM agent to your database and pull every project or user account that has seen zero activity in 90 days. Run the query yourself instead of asking team leads for a status update.