AI does not fix broken ERP processes
A model trained on a process nobody follows learns the exceptions — not the process.
The most common AI proposal we see in ERP environments is invoice automation. It is also the proposal with the highest success rate — and the reason is telling: invoices are structured, the matching rules already exist, and a person makes the automated decision the same way every time.
The second most common proposal is a pricing or forecasting model. These succeed far less often, and that reason is telling too: the underlying process is not executed consistently, so the historical data encodes fourteen behaviours instead of one.
Data readiness is a process question, not a storage question. Teams judge data readiness by whether the data is accessible. The better question is whether the data describes a process that was actually followed. A discount field that regional sales teams use differently is accessible, complete and worthless as a training signal.
What to check before committing. Is the process standardised across the entities whose data you want to use? Is there a measured touchless-posting rate today, so the improvement can be named? Is there a named owner in the business who can define what a good outcome is before the model is built? Can the AI service write back into the process — or only produce an observation someone has to act on manually?
The last question ends more use cases than any technical constraint. A recommendation that lands in a report instead of in the transaction where the decision is made changes the decision in only a minority of cases.
The sequence that works. Standardise the process where the value is. Measure the current state so the benefit can be proven. Assign ownership. Then automate — and only then apply a model. The reverse order produces pilots, and pilots produce slides.