Getting your data ready for AI when it lives in spreadsheets

The most common data situation in an SME is not "no data". It is too much data, in too many places, with three slightly different versions of the same customer name.

Fix in this order

• Pick one system as the source of truth for each entity — customers, jobs, invoices. One each, no exceptions.
• Standardise identifiers before you standardise anything else. A consistent customer ID solves more problems than a tidy address field.
• Make sure you can export. If a supplier controls your data and will not give you a CSV or an API, that is a strategic risk, not an IT annoyance.
• Only then worry about completeness. Perfect data is not a prerequisite; retrievable data is.

What good enough looks like

For most first use cases, good enough means: the last twelve months of records, one row per event, consistent dates, and a field that reliably links records to a customer. That is it.

Teams spend months cleaning data for a use case they never validated. Validate the use case with messy data first.

The privacy step people skip

Before any dataset leaves your systems, decide what is personal data and whether it needs to be there at all. Stripping names and contact details from a dataset used for forecasting removes most of the risk at almost no cost to accuracy.