Medallion architecture

Three steps from messy to trusted.

Medallion architecture is a simple way to organise your data. It moves through three stages: raw, cleaned, and ready to use. Each stage has one job, so problems are easy to find and the final numbers can be trusted.

Finance
CRM
Core systems
Spreadsheets
Power BI
Dashboards
AI agents

Silver · Cleaned

Duplicates removed, formats fixed, records matched across systems. This is where the numbers start to agree.

  • Remove duplicates
  • Fix dates, currencies and names
  • Match records across systems

Checks at this stage

  • No duplicate keys
  • Dates and amounts parse
  • Every row traces to a source row

What the data looks like

LineRegionCurrent
inv-101North£120
inv-102South£80
cr-103North−£20
inv-104North£95

Four current transaction lines. The £95 revision replaces £90.

The work in each stage

This is what I do at each step.

Bronze

Raw

Get the data in, and keep an untouched copy.

  • List every source system and who owns it
  • Set up the connections and the loading schedule
  • Store each load exactly as it arrived
  • Add alerts so a failed load is noticed the same day

Silver

Cleaned

Fix the data and make the systems agree.

  • Fix dates, formats and missing values
  • Remove duplicates
  • Match the same customer, property or case across systems
  • Agree one shared structure (schema) for each subject
  • Run quality checks and set aside rows that fail

Gold

Ready to use

Turn clean data into the figures people actually use.

  • Define each business figure once, like revenue or cases closed
  • Shape the data so reports are fast
  • Build Power BI reports or custom dashboards
  • Set who can see what
  • Write a plain-English data dictionary

Around all three stages

Pick the platform

Snowflake, Databricks or Microsoft Fabric, chosen to fit your team and your tools.

Security and rules

Access, privacy and audit trails set up from the start, not added later.

Handover

Your team learns how it works, with documentation they can follow.

AI on top, when ready

Agents that answer questions from the ready-to-use data, with the source rows shown.

With and without it

Where do the numbers come from?Different spreadsheets and exportsOne agreed place
Do two reports agree?Not alwaysYes, they use the same definitions
What if a source changes?Reports break and nobody knows whyRaw copy is kept, so it can be traced and fixed
How long to answer a new question?Days of manual workA new report on data that is already clean
Can you add AI later?Poor results on messy dataYes, on data you already trust

Not sure where you are today?

A short call is enough to see which stage is hurting most.

Or send a message