The data is there. The shape is wrong.
Years of orders, invoices and customer history sit inside an old ERP. It's real, valuable data — and it's almost impossible to ask a question of. The tables were designed for a screen from 2009, codes mean different things in different modules, and the meaning lives in the heads of people who've used it for a decade. None of that is a reason to keep it locked up.
What "AI-ready" actually means
An AI can only answer well if the data underneath it is clean and structured. That's less exciting than the model and more important. Before any AI touches your history, the work is to make the data mean one consistent thing.
- One definition per thing: a customer, a product, an order status that mean the same across every record.
- Explicit relationships: which order belongs to which customer, stated in the data rather than implied by a report.
- Preserved context: the notes and exceptions that explain what the raw numbers don't.
Clean data beats a bigger model
It's tempting to think a more capable model will paper over messy data. It won't. A strong model on confused data gives confident, wrong answers — the worst kind. A modest model on clean, well-structured data gives answers you can trust and check. The advantage is almost always in the data.
History that answers questions
Once the history is in a shape an AI can query, it stops being dead weight in an old system. You can ask what actually happened last quarter, why a customer left, where margin quietly leaked — and get an answer grounded in your real numbers, not a guess. That's the point of the migration: your past starts doing work for you.