Knowledge
Integrating AI into legacy systems and SAP: how to do it without a rebuild
Existing systems and SAP are no barrier to AI: the AI runs as its own building block alongside the system and connects through interfaces or a database link — the core system stays unchanged. One pattern has proven itself: start read-only, generate suggestions, then write back under control with human approval.
Why your existing system is no barrier
The common worry is: “Our system is too old for AI.” In practice, that is rarely the problem. The AI is not built into the existing system — it runs as its own service alongside it: it reads data through interfaces, processes it externally and returns results under control. The core system stays unchanged — updates, maintenance contracts and the ability to install releases remain untouched.
Common integration patterns
- Reading through an interface: data comes from the existing system via an API, a database link or a regular export. For many use cases, read-only access is already enough.
- Processing outside the system: the AI runs as its own service — EU-hosted or on your own infrastructure — and extracts, checks, summarizes or suggests.
- Controlled write-back: results go back into the system as a draft or pre-entry, not as a direct posting.
- Human approval: an employee reviews and approves. Only once quality holds up over time is more of the work gradually automated.
- RPA as a last resort: if no interface exists at all, a software robot can operate the user interface. That works, but it is maintenance-heavy — more of a stopgap than a target state.
What is concretely possible with SAP and ERP
SAP and most ERP systems ship with documented interfaces (in SAP’s case, OData services, BAPIs and IDocs, among others). These let you connect typical tasks:
- Read incoming invoices and purchase orders and pre-enter them for posting.
- Check master data and suggest corrections for discrepancies.
- Create orders or documents from emails and files — for approval, not final.
- Generate text using data from the system, such as line-item descriptions or draft replies.
Risks you should plan for
- Data quality: AI makes poor master data visible, not better. A data check belongs at the start of every project.
- Access rights: the AI needs its own technical user with minimal permissions — not a personal account with broad access.
- Protecting your investment: no modifications to the core system. The connection runs through official interfaces so your system stays update-capable.
A step-by-step approach
- 1. Scope the use case: one process with a clear beginning and end, such as incoming invoices of a single document type.
- 2. Start read-only: the AI first gets read access only and generates suggestions.
- 3. Pre-enter instead of posting: results land in the system as a draft.
- 4. Approve and learn: an employee reviews and corrects — every correction shows where the system is still off.
- 5. Expand gradually: only once quality holds up do you add more document types and more automation.
FAQ
Frequently asked questions
Do we have to replace our ERP or SAP to use AI?
No. The AI runs as its own building block alongside the existing system and connects through interfaces. The core system stays unchanged and update-capable.
What if our legacy system has no interface?
There are usually still ways: a read-only database link, regular exports, or as a last resort RPA that operates the user interface. RPA is more maintenance-heavy and is better suited as a stopgap.
Does the AI write back directly into SAP?
Only under control. The proven approach is for the AI to pre-enter documents or transactions as a draft and have a human approve them. Postings without approval only make sense once quality has proven itself over time.
How long does such an integration take?
A scoped use case — such as reading and pre-entering incoming invoices — goes live in 6–8 weeks as a fixed-price pilot from €39,000.