Before the dashboard: what a mid-sized company should build first in its data platform
Everyone wants dashboards and AI. Few want to hear that both depend on a boring layer underneath. Here is the order that keeps a data project from collapsing under its own weight.
Three reports, three different revenues
It is the monthly management meeting. Sales presents revenue for the month. Finance presents a noticeably lower number. Operations has a third. The first twenty minutes are spent not on decisions but on arguing about whose spreadsheet is right.
This is the problem a data platform exists to solve, and it is almost never a problem of tools. Each team pulled data from the same systems at different times, applied slightly different rules and saved the result in their own files. There was never one agreed place where the numbers were prepared.
A data platform is not a place to store data. It is a place where the company agrees, once, what its numbers mean.
Layer one: land the raw data and never touch it
Start by copying data from each source system into a central store exactly as it arrives: orders from the ERP, tickets from the helpdesk, events from the website. No clean-up, no renaming, no clever fixes. Keep it append-only, with a record of when each batch arrived.
It sounds wasteful, but it is the most valuable decision in the project. When someone discovers six months later that a rule was wrong, you can replay history from the raw layer and rebuild everything correctly. Without it, the mistake is baked in forever.
Layer two: clean it once, in one place
The second layer applies the unglamorous fixes: consistent types, duplicates removed, one customer ID across systems, one time zone, one currency convention, empty strings turned into proper missing values. Every rule is written as code and documented, so a new colleague can read what was done and why.
The key word is once. If three teams each clean the same data in their own way, you are back to three different revenues. The platform exists so that cleaning happens in a single, reviewed pipeline.
Layer three: define the business once
The top layer is organised around business questions rather than source tables: customers, orders, products, and the metrics built on them. This is where you settle the arguments. What is an 'active customer'? Is revenue measured at order, shipment or payment? Do returns reduce this month or the original month?
Write each definition down and put it in code that is reviewed like any other change. When someone proposes a different definition, it becomes a visible decision with a date, not a silent difference between two spreadsheets.
Quality is cheaper at the door
Trust in data disappears quickly and returns slowly. The cheapest protection is to check data as it enters, not after a director has quoted a wrong number.
- Freshness: did last night's load actually arrive? Alert if not.
- Volume: today has a tenth of yesterday's rows. Something broke upstream.
- Shape: a column that was a number is suddenly text, or a required field is empty.
- Agreements with source teams: when the team that owns the ERP renames a column, the platform should hear about it before the dashboard breaks.
The same foundation feeds AI
There is a second payoff. When the company later wants an AI assistant that answers questions about sales or stock, it needs exactly this: trustworthy, well-defined data in one place. An assistant that queries the curated layer gives the same answer as the dashboard. An assistant that reads raw exports gives a fourth revenue figure and a fresh argument.
For that reason, teams that invest in the boring layers first usually find that their AI projects go faster, not slower.
Start with one question
Do not try to model the whole company. Pick one question that crosses teams and hurts when answered inconsistently, such as monthly revenue by product line. Build the thin path from raw data to that single number, with tests, and let people use it. When the argument about that number ends, you have proof the platform works and a pattern to repeat.
Keep reading
Connecting ERP, CRM and your online shop: the quiet mistakes that cost the most
Integrations rarely fail on day one. They fail in month four, when an order appears twice, a price is stale and nobody can say which system is right.
Read articleYour RAG demo is brilliant. Your data is why production won't be.
The demo answered every question. Then it went live and quoted last year's price with total confidence. The model was fine. The documents were the problem.
Read article