Where generative AI actually belongs in your ERP

Every ERP vendor now has an "AI" badge on the box. Most of what's behind it is either a chatbot bolted onto the help docs or a roadmap promise. That doesn't mean generative AI is hype inside an ERP — it means the value is narrower and more specific than the marketing suggests. After two decades implementing Oracle ERP and a couple of years shipping production LLM systems, here's the honest map.

Where it earns its keep

Large language models are good at a handful of things that happen to be expensive in an ERP context:

• Unstructured-to-structured. Turning emails, PDFs, and supplier documents into clean records — invoice capture, order entry, contract metadata. This is the highest-ROI use, full stop.
• Search and retrieval. "Find me the PO that matches this exception" across systems that were never designed to talk to each other.
• Drafting and summarization. First-pass variance explanations, customer responses, internal write-ups — with a human approving before anything ships.
• Guided workflows. Helping an occasional user complete a task they do twice a year without filing a ticket.

Where it quietly burns budget

The failures are predictable:

• Anything requiring exactness without a check. Financial postings, tax determination, and reconciliations need deterministic logic. An LLM that's right 97% of the time is a 3%-error machine in a domain where errors compound.
• "Ask your data" dashboards that replace a report nobody trusted with an answer nobody can audit.
• Net-new infrastructure for a problem a workflow rule already solved.

The right question isn't "where can we add AI?" It's "where are people retyping, re-keying, or re-reading the same information — and can a model do the first draft while a person keeps the final say?"

The pattern that works

Keep the system of record deterministic. Put the model at the edges — at intake, at retrieval, at the first draft — and keep a human in the loop wherever money or compliance is on the line. Measure it against a real baseline (hours saved, error rate, cycle time), not a demo. Start with one workflow, prove it, then expand.

That's far less exciting than "AI-powered ERP." It's also the version that survives contact with an auditor.