Clean the data before you add AI.
AI repeats whatever your item, customer and transaction data already say. Clean those three before you fund a model.
Every serious AI conversation inside a company arrives at the same obstacle. The model is ready. The data is not. Teams then try to prompt their way around a dirty item master. It does not work.
What "clean enough" means
You do not need a perfect enterprise data platform to start. You need the records a decision depends on to mean one thing. One customer, not four spellings. One item, with a unit of measure the warehouse and finance both accept. Transactions posted in the period they happened, not in a batch when someone had time.
If two reports can disagree on last month's sales without anyone being surprised, you are not clean enough for a forecast.
Ownership is the cleanup
A data project owned only by IT will standardise a file and leave the behaviour unchanged. The next order will create the next duplicate. Put a business owner on customers, items and the documents that move them. Give them the right to reject a record that does not follow the rule.
Then make the rule visible in the system: required fields, duplicate checks, a short list of codes. A policy in a slide will not survive a busy afternoon. A field that will not save will.
Thirty days that are worth it
Pick one master, usually items or customers. Merge the obvious duplicates. Close the dead records. Agree the codes you will actually use. Measure the before and after: how many duplicates, how long a lookup takes, how often someone exports to fix a report.
That measurement is also your business case for AI. You will know which decision is finally safe to assist, because you can see the record it will read.