01Garbage in, confidently out
An assistant grounded in three contradictory price lists will pick one and defend it. The model is not wrong; the organisation is.
Before any AI work, inventory the sources, mark the authoritative one and retire the rest. This is unglamorous and it is where the outcome is decided.
02Structure beats volume
A small, well-labelled corpus outperforms a large messy one. Consistent product attributes, clean customer records and dated policies improve every downstream feature at once.
Fix duplicates and missing identifiers early — they are the reason cross-system automation stalls later.
03Make quality continuous
Add validation at the point of entry, not as a quarterly cleanup. Required fields, controlled vocabularies and automated checks stop the drift returning.
Publish a small data-quality dashboard. Visible numbers are the only thing that keeps this work funded.