AI in ERP: Where It Helps, Where It Doesn't, and How to Start

AI is now part of almost every ERP vendor's roadmap. Some of it is genuinely useful; some is a chatbot bolted onto a menu. This article separates the two from an implementer's perspective.

What AI is good at in ERP

Large language models excel at reading unstructured text, matching it to structured data and drafting responses. ERP operations are full of that work:

• Document processing — supplier invoices, purchase orders, delivery notes, certificates. AI extracts header and line data and matches it to purchase orders and products.
• Order intake — emailed customer orders turned into draft sales orders.
• Email triage — classifying shared-inbox messages and routing them to the right queue.
• Support assistance — drafting replies grounded in the customer's orders, shipments and warranties.
• Sales assistance — summarising account history before a call, suggesting follow-ups.
• Natural-language reporting — answering plain-English questions from curated data views.

What AI is not good at

• Being the system of record. The ERP's transactional integrity — double-entry accounting, stock moves — must remain deterministic.
• Reliable arithmetic and pricing. Prices and totals must come from ERP rules, not model output.
• Accountable decisions. Approving credit, posting to the ledger, releasing a batch — people decide.
• Working on bad data. Matching documents against a catalogue full of duplicates produces poor results.

Controls that make AI safe

• Human review — AI creates drafts; people confirm. Auto-acceptance only for low-risk, high-confidence cases you define.
• Permission awareness — AI sees only what the requesting user can see.
• Logging — prompts, outputs and actions recorded against the related record.
• Evaluation — accuracy measured on a held-out set of real documents before and after changes.
• Provider choice — data processing terms and regional requirements checked.

How to choose a first use case

Pick something with:

1. Volume — happens many times a day.
2. Measurable accuracy — you can count correct extractions or accepted drafts.
3. Clear review step — a person already checks the result.
4. Clean underlying data — or a plan to clean it.

Supplier invoice processing usually ticks every box, which is why it's the most common starting point.

A sensible rollout

1. Sample real documents and test candidate models.
2. Build the workflow with review queues and logging.
3. Pilot with one team; measure accuracy and edit rates.
4. Expand to more document types or teams.

The prerequisite nobody wants to hear

Most AI projects in ERP succeed or fail on master data quality. If vendor names, product codes and units of measure are inconsistent, fix that first. It improves everything else too.

See AI ERP automation and AI document processing for how we approach these projects.