Medallion architecture, in plain language

If you've sat in a meeting about data platforms recently, someone probably said 'medallion architecture' and everyone nodded. Here's what it actually means, minus the vendor gloss: it's a promise about how raw your data is at each shelf in the warehouse.

Bronze: keep the receipts

The bronze layer is your data exactly as it arrived: every export, every API response, every log, untouched. It feels wasteful to store data you'll never show anyone, but bronze is your insurance policy. When something looks wrong downstream, bronze is how you prove whether the source sent bad data or your pipeline broke it. Teams that skip bronze end up debugging blind.

Silver: make it trustworthy

Silver is where the cleaning happens: duplicates removed, types fixed, names standardized, records validated against rules you've written down. The key property of silver is that it's defensible: every transformation from bronze is code, versioned and reviewable, not a mystery script on someone's laptop. Silver is the layer your data team actually works with.

Gold: answer the question

Gold tables are shaped for consumption: revenue by region by month, churn by customer segment, the inputs your ML model trains on. They're small, fast, and named in business language. When a dashboard is slow or wrong, the fix is almost always 'build the right gold table', not 'make the BI tool query harder'.

Why this matters to a non-engineer

The medallion pattern isn't really about technology. Databricks didn't invent layered cleaning, they named it well. It's about accountability. When the layers exist, 'why does this number look wrong?' has a checkable answer: you trace gold back to silver back to bronze and find where reality diverged. When they don't, the same question turns into a week of archaeology. If your reports can't be traced like that today, then that, not a new AI tool, is probably the highest-value fix in your data estate.