Medallion architecture bronze silver gold is a three-layer way to keep raw data, a cleaned unified model, and purpose-built marts apart. Bronze lands source data as-is, silver conforms it into reusable entities, and gold serves a specific job — not the other way around.
Strip the marketing off the metals
The names are a memory aid, not a product feature. You can implement the same idea in a warehouse, a lakehouse, or a pile of well-named schemas. What matters is the rule for each layer, and the rule about what each layer is forbidden to do.
Microsoft Fabric's medallion best-practices series on 3 September 2026 put those forbids in writing: bronze must not clean, silver must not serve dashboards, gold must not compensate for upstream quality. If you remember nothing else, remember those three sentences.
Bronze: land it, do not "fix" it
Bronze is the landing zone. Keep source names, source types, and source mistakes. Add only what you need to replay: ingest time, file name, batch id, producer version.
Cleaning in bronze destroys evidence. When a dashboard later disagrees with the operational system, you need the raw payload to see who was right. If bronze already "fixed" nulls and recast dates, you have no court of appeal.
Bronze may partition and compact for cost. That is storage hygiene, not business logic. If you find case statements on status codes here, they belong further downstream.
Silver: the unified data model
Silver is where source systems become company language. Customers, invoices, shipments, and employees get stable keys. Types are made honest. Duplicates are resolved on purpose, with a rule you can explain.
This is the layer Citi Global Wealth treated as the unified data model in the cloud move Computer Weekly described on 8 September 2026. They spent eighteen months leaving five regional datacentres for one platform, and reporting went from weeks to days. The compression did not come from a new chart tool. It came from one conformed model.
Silver should be reusable. If only one dashboard can read a silver table, you have built gold and mislabelled it. If every dashboard re-implements "active customer" on top of silver, you have not finished silver.
Field notes on unifying data, governance and getting analytics teams to one source of truth.
Every Monday, five minutes.
Gold: one job, one grain, one audience
Gold is a mart. It answers a job: finance close, growth weekly, warehouse operations. It may aggregate, pivot, and apply the filters that audience has agreed.
Gold must not invent quality that silver lacked. If keys are still colliding, the fix is silver. If a source field is wrong, the fix is the producer or bronze quarantine. A gold view that papers over those issues will be right until the week it is disastrously wrong.
Keep gold thin. Wide gold tables that try to serve every team become a second warehouse with no owner.
A plain path to implement it
Create three schemas or three datasets with boring names. Put producers at bronze. Put modelled entities at silver. Put each executive dashboard on a gold object that reads silver, not bronze.
Write the layer rules in the repo, not in a slide. The Fabric series is a useful checklist even if you are not on Fabric: no cleaning in bronze, no dashboards on silver, no quality theatre in gold.
When a request arrives as "just add this column to the executive mart," ask which layer it belongs in. A source field belongs in bronze first. A conformed attribute belongs in silver. A metric for one audience belongs in gold.
What medallion is not
It is not a licence to copy every table three times. Some reference data can land and be used with almost no silver work. Some event streams need a richer bronze than a nightly dump.
It is not an excuse for slow reporting. Citi's public story was about shortening reporting, not adding ceremony. If your gold build takes so long that people extract to spreadsheets, the architecture has failed its job.
It is not vendor-specific. Lakehouse vendors will sell you medallion as a feature. The feature is discipline.
How LensAIght fits without the slideware
LensAIght is a unified data intelligence platform: ingest, modelled datasets, and the chat or dashboards that read them. That only helps if the modelled layer is actually silver — shared, named, and not a private extract per notebook.
Use bronze for replay, silver for agreement, gold for a job. If a conversational query and a dashboard disagree, they are not sharing silver.
Frequently asked questions
What is medallion architecture bronze silver gold in one sentence?
Bronze keeps source data as-is, silver conforms it into a reusable unified model, and gold builds thin marts for a specific audience or job.
Can dashboards read silver directly?
Not if you want silver to stay a model rather than a pile of one-off queries. Microsoft Fabric's 3 September 2026 guidance said silver must not serve dashboards. Give the dashboard a gold object.
Why not clean data in bronze?
Because you will need the original payload when sources disagree with reports. Cleaning belongs in silver, with a rule you can replay.
How many gold tables should we have?
As many jobs as you are willing to own, and no more. Each gold object needs an owner and a grain. If you cannot name them, do not create the table.
Is medallion only for lakehouses?
No. The same three-layer rule works in warehouse schemas. The storage engine is not the architecture.
Where do data contracts sit?
On the producer interface into bronze, and on the silver entities consumers depend on. Gold should not be where you first discover that a field changed.
Book a demo to see a managed data platform running on your own sources
LensAIght
