Every ERP vendor now ships a copilot, and every earnings call mentions AI. Underneath the noise, a small set of use cases are genuinely paying back inside ERP landscapes today. A larger set are promising but early. And a few popular ideas deserve a polite refusal.

Paying back today

Support-desk intelligence. The most mature use case by a distance. Ticket classification, known-error matching and drafted resolutions over your incident history — the corpus already exists, the outcomes are measurable, and payback is typically inside two quarters. If you run an AMS desk and haven't done this, it's the first cheque to write.

Document-to-transaction processing. Supplier invoices, purchase orders, delivery notes, onboarding packs — extracted, validated against master data, and posted with confidence-based routing. This existed before LLMs; language models pushed straight-through rates from "good" to "most of the volume" because they handle the messy, non-standard documents that used to hit the exception queue.

Explaining the system to its users. A retrieval-grounded assistant over your process documentation, training materials and configuration answers the eternal question — "how do I do X in this system?" — without a ticket. Adoption is instant because the pain is universal. Ground it in your documents, insist on citations, and keep it read-only.

Promising but early

Conversational reporting. "Show me margin by region last quarter" as a chat prompt works impressively in a demo and dangerously in production if your semantic layer is weak — the model will confidently query the wrong revenue definition. The prerequisite isn't better AI; it's the governed metrics layer you should have anyway.

Custom-code analysis for migrations. Using LLMs to explain and triage thousands of legacy ABAP objects ahead of an S/4HANA move genuinely accelerates the assessment phase. Treat the output as a first pass for humans to verify, not a verdict.

Skip, for now

  • Autonomous transaction posting. An agent that posts journal entries or changes master data without human approval is an audit finding waiting for a date. Approval-gated automation, yes. Autonomy in the ledger, no.
  • Chat-based configuration. Letting a model modify system configuration conversationally combines maximum blast radius with minimum reviewability.
The pattern across everything that works: AI reads, drafts and routes. Humans approve anything that changes the books.

How to sequence it

Start where the corpus already exists and the metric is obvious: the support desk, then documents, then user assistance. Each pays for the next, and each teaches your organisation how to evaluate the harder cases when the vendors come selling them.