Is AI a reason to switch ERP systems?

Almost never on its own. The features that deliver value today — document capture, forecasting, anomaly detection — are largely available as add-ons to systems you already run, and the ones that are genuinely platform-native are early enough that waiting a release costs little. If you are replacing your ERP for other reasons, AI capability is a reasonable tiebreaker between finalists. As the primary driver it is a weak case.

Does AI in ERP create compliance problems?

It creates control questions. An AI that codes an invoice has made an accounting judgment, and your auditor will want to know who reviewed it and what happens when it is wrong. The practical answer is threshold-based review: automatic posting below a dollar limit with sampling, human approval above it. Vendors who cannot explain how their automation logs its decisions have not thought about the audit conversation.

How much training data does forecasting need?

Enough history to cover the patterns you want predicted — which for a seasonal business means several years, not several months. More important than volume is stability: a company that changed its product mix or channel structure last year has history that partly describes a business it no longer runs. That is a data problem no model solves.

What about AI features for smaller companies?

The document-capture case works at almost any size and is the one to start with, because the value scales with invoice volume rather than company size. Forecasting rarely pays for itself below a few hundred SKUs. Assistants are mostly a convenience feature at small scale, where the person with the question can usually just ask the person who knows.

Artificial Intelligence in ERP — Use Cases and Benefits

Every ERP vendor now has an AI story. Some of it is real and shipping, some is a roadmap slide, and telling them apart during a demo is genuinely hard — which is the point of this page. What follows is sorted by how much evidence exists that the feature works in production, not by how prominently it appears in vendor marketing.

What works today

Document capture and coding. The most mature application by a wide margin. Incoming invoices, purchase orders and remittances are read, matched against POs and coded to accounts, with confidence scoring that routes uncertain cases to a human. Accuracy on standard formats is high enough that the economics work at moderate volume, and the failure mode is benign: the system asks instead of guessing. If you evaluate one AI feature, evaluate this one.

Anomaly detection. Flagging postings that deviate from established patterns — duplicate invoices, unusual vendor payments, quantities off by an order of magnitude. This is pattern matching rather than anything exotic, and it works well because the underlying question ("is this transaction unlike the others?") is one statistics has answered for decades.

Demand forecasting. Genuinely better than the moving averages built into older systems, provided you have several years of clean history and a business whose shape has not changed. Improvements of a few percentage points in forecast accuracy translate into real inventory reduction at scale, which is why distribution and process manufacturing systems were first to ship it.

What disappoints

Conversational assistants. Every major vendor has one — SAP's Joule, Microsoft's Copilot in Dynamics 365, Oracle's assistants in NetSuite and Fusion. They answer questions about data you could already query and draft text you could already write. In deployments so far, the demonstrable gains sit with occasional users who never learned the interface, not with the power users whose time is expensive. Useful, not transformative, and rarely worth a tier upgrade on its own.

Autonomous process execution. The promise that the system will run procurement or scheduling without human decisions. In practice the exceptions — the very cases where judgment matters — are exactly what the automation handles worst, and the review effort for edge cases can exceed the effort saved on routine ones. Treat any "touchless" claim as a question about your exception rate.

The prerequisite nobody sells you

AI features are only as good as the master data underneath them. Duplicate customer records, item masters with inconsistent units of measure, a chart of accounts that grew by accretion — these degrade AI output more thoroughly than any model choice. A company whose data is in that state gets more value from three months of data cleanup than from any AI module, and the cleanup is a prerequisite for the module anyway.

The uncomfortable version: if your current reporting is unreliable because of data quality, AI will produce confident output from the same unreliable data. Confidence is not accuracy.

Questions worth asking a vendor

Which of these features are generally available today, and which are in preview or on the roadmap? Is it included in my tier or priced separately, and is the pricing per user, per transaction, or consumption-based? Where is the data processed, and does it leave my tenant? Is customer data used to train shared models? What does the audit trail record when the system makes a decision — and can I set confidence thresholds myself?

That last question separates products built for accounting environments from products built for demos. An AI that posts without a reviewable decision log will not survive contact with your auditor, whatever it does for throughput.

A reasonable position

Start with document capture, because the case is clear and the risk is contained. Add anomaly detection if you have the transaction volume for it to matter. Treat forecasting as a project with a data prerequisite rather than a feature you switch on. Ignore the assistants until someone in your organization asks for one by name. And do not let AI capability drive an ERP selection — the systems that fit your processes will have caught up on this before your implementation finishes.