Data platforms are often sold as if every company were a global technology firm: streaming pipelines, dozens of tools, a large engineering team. A mid-size company with an ERP, a CRM, a few spreadsheets and an e-commerce site needs something much simpler, and the simpler version is usually the one that gets finished.
The goal is not the technology. It is one version of the truth for revenue, margin, customers, stock and cash, available every morning, trusted by finance and used by operations.
Five layers, each with one job
- Ingestion: copy data from the source systems on a schedule, without changing it. Managed connectors cover most ERPs, CRMs and marketing tools; a few custom scripts cover the rest.
- Storage: one cloud data warehouse where all the copies land. It separates analysis from the operational systems, so reports no longer slow down the ERP.
- Transformation: SQL models, under version control and tested, that turn raw tables into clean business entities: customers, orders, products, invoices.
- Semantic layer: the official definitions of every metric, written once. Revenue, gross margin and active customer mean the same thing in every report.
- Consumption: dashboards, decision products, spreadsheets connected to the warehouse and, increasingly, AI assistants that answer questions on top of the same definitions.
What a mid-size company can skip
- Real time everywhere: most decisions are daily or weekly. Reserve streaming for the few processes that truly need it, such as fraud or inventory in e-commerce.
- A data lake before a warehouse: unless you have large volumes of files, images or logs, a warehouse is enough.
- Building connectors in-house: managed connectors are cheaper than the engineers needed to maintain your own.
- Ten dashboard tools: choose one, and retire the rest.
People and ownership matter more than tools
A small team can run this stack: one or two analytics engineers who own the models and the semantic layer, and analysts embedded in finance, sales and operations who turn data into decisions. Each core entity needs a business owner who decides what it means, and every quality problem should be fixed at the source, not patched in a report.
A realistic roadmap
- Months 1 to 3: ingest the ERP and the CRM, model sales, margin and customers, publish the first trusted management report and retire the spreadsheet it replaces.
- Months 4 to 6: add inventory, purchasing and cash; define the semantic layer; give each department its own decision product.
- Months 7 to 12: forecasting, customer profitability and the first machine learning or AI use cases, now built on data the business already trusts.
A modern data stack is not a technology project with a go-live date. It is the foundation that makes every later investment in analytics and AI cheaper, faster and more believable.
Bartor S.
Founder
