Spreadsheets are excellent tools for analysis. They are poor foundations for shared numbers. A data platform moves the foundation into a place where definitions are agreed once, data is refreshed automatically, and everyone works from the same figures.
What a single source of truth actually is
A single source of truth is not one giant database. It is an arrangement in which:
- data from the operational systems is collected automatically, on a known schedule;
- each business metric has one agreed definition, implemented once;
- reports and analyses read from that implementation, not from private copies;
- anyone can see where a number comes from and when it was last updated.
Excel does not disappear. People keep using it for exploration — connected to the platform, instead of to a manual export.
The building blocks
Ingestion from source systems
Data is extracted from the ERP, the CRM, e-commerce, operational applications and external sources. Where possible, through APIs or change data capture rather than files dropped on a shared drive. The quality of this layer depends on the quality of your system integration: systems with clear ownership of data are far easier to extract from.
Storage: warehouse or lakehouse
The data lands in an analytical store designed for queries over large volumes. Common options include cloud data warehouses such as Snowflake, BigQuery or Microsoft Fabric, lakehouse platforms such as Databricks, or open-source databases such as PostgreSQL or ClickHouse for smaller volumes or when data must stay on your own infrastructure. The choice depends on volumes, existing skills, budget and where the data is allowed to reside.
Transformation: where definitions live
Raw data from source systems is not ready for reporting. Transformations clean it, join it and compute the business metrics. Writing these transformations as versioned code — with a tool such as dbt, or plain SQL in a repository — rather than as formulas hidden in spreadsheets brings the benefits of software engineering to data: review, history, tests and documentation.
A useful structure is in layers:
- Raw: data as it arrived, untouched.
- Cleaned: types corrected, duplicates removed, codes standardised.
- Business: entities and metrics as the business understands them — customers, orders, margin, stock rotation.
Semantic layer and reporting
On top of the business layer, a semantic model defines metrics and dimensions once, so that "revenue by region" means the same thing in every dashboard. BI tools such as Power BI, Tableau, Looker or open-source alternatives such as Apache Superset or Metabase connect to it.
Data quality is tested, not hoped for
Trust in numbers is lost quickly and regained slowly. Data quality checks run automatically with every refresh:
- completeness: are yesterday's orders all there?
- uniqueness: is every invoice counted once?
- validity: are the dates, amounts and codes plausible?
- consistency: do totals match the source system and the accounting?
When a check fails, the pipeline stops or flags the affected report, and the data owner is notified — before the management meeting, not during it.
Governance without bureaucracy
A data platform needs a small set of rules, applied consistently:
- An owner for each data domain — sales, finance, operations — who decides definitions and answers questions.
- A short glossary of business terms and metric definitions, maintained alongside the code.
- Access by role, with personal and sensitive data protected and access logged, in line with data protection obligations.
- A change process for definitions: when "active customer" changes meaning, everyone should know from when.
The foundation for AI
Clean, documented data is also what AI projects need. Forecasting, anomaly detection and AI assistants that answer questions over company knowledge all depend on data that is complete, consistent and well described. Companies that skip the data platform usually discover this in the middle of their first AI project.
Where to start
Do not start by moving everything. Choose one decision that matters — monthly margin by product line, stock levels, sales pipeline — and build it end to end: ingestion, transformation, tests, one report that replaces the spreadsheet. Agree the definitions with the people who use the numbers. Then extend, domain by domain. Our data platform and analytics service works in exactly these increments, so that each step replaces a spreadsheet people actually used.