AI in ERP gets pitched two very different ways, and it’s worth separating them early, because only one of them is something we’d actually recommend building on right now.
The first pitch is that AI reads the data your ERP already has and does something specific and useful with it. The second is that AI becomes the interface: you ask it questions instead of using forms and reports, and it handles the transactions for you. The first is real and working today. The second isn’t something a business running its actual books should be betting on yet.
Where it works: narrow, data-grounded tasks
Document extraction. Supplier invoices, delivery notes and customs paperwork arrive as PDFs and scanned images with no structure to them. A model trained to pull line items, quantities and reference numbers out of these documents, feeding directly into a Purchase Receipt or Material Request, removes a genuinely tedious and error-prone manual step. This works because the output gets checked against a known schema before it’s committed anywhere, not trusted blindly.
Anomaly detection. With years of transaction history sitting in Stock Ledger Entries and GL Entries, a model can flag transactions that deviate from an established pattern: a purchase price that’s out of range for that supplier and item, a quantity that doesn’t match historical order patterns. This is useful precisely because it’s a flag for a human to check, not an automatic action taken on its own.
Forecasting. Demand forecasting and reorder point suggestions, built from historical Sales Order and Stock Ledger data, genuinely improve on the “gut feel and a spreadsheet” baseline most businesses are actually running on today.
Where it doesn’t work yet: replacing the ledger’s own guarantees
The reason a Stock Entry writes its Stock Ledger Entry and its GL Entry in the same transaction is that stock and the books physically cannot disagree. That guarantee comes from the platform’s transactional logic, not from anyone’s judgment, and it doesn’t transfer to a model inferring what should have happened. An AI interface can be a very good way to query a system whose numbers you already trust. It isn’t yet a substitute for the deterministic logic that makes those numbers trustworthy in the first place.
The actual question to ask
When AI gets proposed as part of an ERP implementation, the useful question isn’t “does it use AI.” It’s “what specific, bounded task is it doing, and what happens when it’s wrong.” Document extraction with a human review step has a clear answer to that. A conversational agent making transactional decisions on its own usually doesn’t, at least not yet.