When I helped build a finance function from a blank page across more than 20 markets, some of the most important decisions were also the least glamorous: the chart of accounts, the cost centre structure, the reporting hierarchies, and who owned customer, vendor, SKU, pricing and costing data.
None of that looks like an AI project, yet it decides whether one will work.
Why AI raises the stakes
A person who sees two versions of a customer name will usually notice. A model may simply treat them as two customers, then give you a confident answer about both. Weak master data was always an inconvenience. With AI it can produce a clear, well-written and wrong conclusion, and that is harder to spot.
Four questions to ask first
- Who owns each data set? Customer, vendor, SKU, pricing and costing each need a named owner and a clear process for changes.
- Is there one hierarchy? Market, channel, brand and product groups should roll up the same way in every report. If each region uses its own version, cross-market analysis is unreliable.
- Do definitions match? Gross sales, allowances and variable margin must mean the same thing in every market. Otherwise a comparison mixes a real change with a change of method.
- Can you tie back? Every dashboard total should reconcile to the ledger. If it does not, fix that before adding any layer on top.
Start small
A company does not need perfect data everywhere before it starts. Pick one valuable question, such as margin by brand and market, and clean the data behind that answer. You get a useful result quickly, and you show the business why the groundwork matters.
This is the less exciting side of AI in finance, but it is often what separates a good demonstration from a tool people trust every month. Teams that invest in structure now will get far more from AI later.
Newer: AI in finance needs controls tooOlder: Where AI earns its place in FP&A