The first 90 days of a data platform that people actually trust

Most organisations that describe themselves as “data-driven” are, in practice, running on a handful of spreadsheets maintained by three people who are not on speaking terms about definitions. The reporting is not wrong so much as it is unresolvable: nobody can trace a number back to its source, so every meeting re-litigates the same figure.

Start with the decisions, not the sources

Data platform projects fail when they begin with a source inventory. A source inventory tells you what exists; it does not tell you what must be true for someone to make a decision on Monday morning. Begin instead with the five to ten decisions the business makes repeatedly — dispatch routing, credit exposure, plant throughput — and work backwards to the minimum data required to make each one defensible.

That single reframing typically cuts the first release scope by more than half, because it removes the temptation to model every column of every legacy table before delivering anything useful.

Treat the semantic layer as a product

The semantic layer is where most of the value and nearly all of the politics live. If “active customer” means four different things across four dashboards, the warehouse is not the problem. Give the semantic layer an owner, a change process and tests, exactly as you would an API. Publish metric definitions in writing and make changing one a reviewed event.

Put tests at every hop

• Freshness — assert that each source arrived within its expected window.
• Volume — alert on row counts that move outside a learned band.
• Referential integrity — fail the build rather than publish orphans.
• Business rules — encode the invariants finance already checks by hand.

Every one of those tests is cheaper to write than the meeting required to explain why the number changed overnight.

Make lineage a deliverable

When an auditor or a customer asks where a figure came from, the answer should be a query and a diagram, not an afternoon of archaeology. Lineage is not a nice-to-have bolted on at the end; it falls out naturally when ingestion, transformation and semantic layers are built with the same discipline as application code.

Do those four things — decisions first, owned semantics, tested hops, documented lineage — and the platform stops being a cost centre argument and starts being infrastructure nobody debates.