Large language models (LLMs) do more than generate text. Set up correctly, they significantly reduce the repetitive, time-consuming and error-prone work in business processes. But the "let AI do everything" approach usually ends in disappointment. In this article I explain where AI really helps, where it does not, and how to make a safe start.
Where AI really helps
LLMs are very good at turning unstructured text into structured data. In enterprise processes that translates into:
- Email classification: sorting incoming customer emails into orders, complaints, price requests and so on, and routing them to the right team
- Data extraction from documents: reading fields from supplier invoices, order forms and bills of lading and converting them into a structured format
- Context-based reply drafts: producing reply drafts grounded in internal documentation for an employee to approve
- Report summaries: summarising variances in ERP reports in plain language and listing likely causes
Where it does not help (or is risky)
For work that requires exact calculation (prices, tax, stock quantities) an LLM is the wrong tool; the ERP already does these calculations deterministically. Likewise, decisions with a high cost of error and no human oversight should not be left to AI. Instead of saying "I don't know", a model can produce an answer that looks plausible but is wrong.
Four parts of a solid architecture
- Structured output: Ask the model for output with predefined fields (such as a JSON schema), not free text, so it can be validated automatically.
- Context (RAG): Feed the model relevant passages from internal documents so it relies on your rules, not its general knowledge.
- Human approval: Especially at the start, output is approved by a user before it is written to the ERP.
- Logging and measurement: Every request, response and user correction is logged. Accuracy and time saved are measured from these logs.
Integrating with the ERP
On the SAP side, the safest path is for AI to start with read-only access. Approved data is written to SAP through standard interfaces (BAPIs, OData services); tables are never updated directly. This keeps the audit trail intact and ensures SAP's own business rules still apply.
Security and data protection
Documents that contain personal data or trade secrets can be masked before they are sent to a model. Enterprise APIs that do not retain data, or on-premise models for suitable scenarios, are options. API keys and system instructions must never be sent to the browser; they belong on the server.
Four steps to get started
- Pick one frequent task with a measurable outcome (for example supplier invoice entry).
- Measure today's baseline: time per document and error rate.
- Run a narrow pilot and use the output with human approval.
- Once the success criteria are met, gradually increase scope and the level of automation.
Successful AI integration happens where technology meets business knowledge. Positioning AI correctly as a tool and keeping human oversight at the centre of the process is the key to long-term success.
Continue with measuring document-processing accuracy and enterprise AI architecture, or see AI solution architecture for your own scenario.
