AI use cases that belong inside ERP.
Start with finance, sales, inventory and purchasing tasks a person already checks. Rank them by value, risk and how clean the data is.
The useful AI inside ERP is rarely a new module with its own name. It is a faster version of a check someone already does, sitting on transactions the company already trusts.
Where it earns its place
In finance: explain a variance, draft the commentary on a late customer, match an invoice to a receipt when the references almost agree. In sales: summarise an account before a call, flag a quote that breaks the discount rule. In inventory: point at items whose movement no longer matches the forecast. In purchasing: suggest a reorder from real lead times, not from a default the system shipped with.
Notice the pattern. A person still decides. The system prepares the case.
Where to wait
Do not begin with anything that posts a journal, promises a customer, or changes a price without a review. Those are high-cost errors, and they teach the organisation to switch the feature off. Also wait if the master data is still argued about. A model cannot reconcile two item codes that your branches have not reconciled.
Rank before you build
Score each idea on three things only: the value if it works, the damage if it is wrong, and whether the underlying data is already good enough to test this month. Pick the one with real value, limited damage, and data you can see today.
That first use case is not the strategy. It is the proof that the company can tell a good suggestion from a fluent one. Everything larger depends on that skill.