Scope to one process
Cleaning all company data first is how AI projects die. Instead, pick your first use case and prepare only what it needs. A support chatbot needs your FAQ and product docs; a proposal assistant needs your five best proposals; an invoice workflow needs a folder of examples. Days of preparation, not months.
Consolidate before you clean
AI answers are only as good as the source you point it at. The pattern that works: one authoritative location per knowledge type. One current price list, one product documentation folder, one procedures handbook. Delete or archive the outdated copies; conflicting versions confuse AI exactly the way they confuse new employees.
Write down what lives in heads
Every small company runs on undocumented knowledge. The fastest capture method: record a senior person explaining the process on a call, let AI transcribe and structure it, then have them correct the draft. An hour of talking becomes a usable document the same day.
Set access rules as you go
Decide per source who and what may read it. Salary data, customer personal data and legal files stay out of general-purpose AI access. A simple green, orange and red classification beats a governance project you never finish.
FAQ
How much historical data does AI need?
For most SMB use cases: current documents plus a handful of good examples. Recency beats volume.
Should we build a data warehouse first?
Not for getting started. Warehouses matter for analytics at scale; document-level AI use cases need organized folders, not infrastructure.
What format should documents be in?
Anything text-based works; searchable PDFs and standard office formats are fine. The structure of the content matters more than the file type.