Häufig gestellte Fragen
Which AI applications in ERP are in productive use in 2026?
Application areas established and in productive use in 2026 include automatic document recognition and account assignment (often combined with OCR), demand and sales forecasting for materials planning, anomaly detection in postings and inventories, and natural-language assistants that deliver analyses without a reporting screen. Where machine and sensor data are connected, predictive maintenance is added, i.e. anticipating maintenance needs in advance. These use cases have in many cases moved from pilot operation into productive day-to-day use in recent years, but remain tied to the quality of the underlying data. Which functions are actually available and useful depends heavily on the ERP system, the industry and the existing system landscape.
What is the difference between generative language models (LLMs) and classical machine learning in ERP?
Large language models such as GPT or Claude are generative and are suited to unstructured text — for example summarising documents, answering questions in natural language or producing drafts. Classical machine-learning methods such as random forest or gradient-boosting models, by contrast, work better with tabular ERP data and deliver forecasts, for instance of sales volumes or payment defaults. In practice the two approaches complement each other, since ERP systems process both structured transaction data and unstructured documents. Both, however, build on clean master data and a central data repository, which is why data quality determines the actual benefit.
Is AI in ERP worthwhile for mid-market companies?
For many mid-market companies, the entry point lies in well-defined use cases with high data quality — above all document processing and demand forecasting, where initial effects often become visible within the first quarter. Reported figures from the industry range, depending on maturity level, from significantly lower costs per invoice to forecast accuracies of around 90 percent; such values are, however, reference points and not guaranteed outcomes. What matters is not full automation but relief from routine work and faster, better-informed decisions. The specific implementation varies by industry, company size and the customisation depth of the respective ERP setup and should be validated against measurable KPIs.
Which ERP vendors have the most mature AI integration in 2026?
In 2026, SAP with its assistant Joule and Microsoft with Copilot are regarded as the most technically advanced; both are deeply embedded in their respective cloud suites and can work together via an agent-to-agent (A2A) connection. SAP has expanded Joule into an agentic platform with numerous specialised agents and a large number of embedded capabilities across S/4HANA, SuccessFactors and other modules, while vendors such as Oracle NetSuite, Infor and Sage continue to develop their own AI components. A blanket statement about the best vendor is of limited value, however, since suitability depends on the specific use case, the data situation and the regulatory framework. It is sensible to look for traceable recommendations rather than autonomous decisions, an end-to-end audit trail and open interfaces such as a REST API or OData.
What data protection and compliance obligations apply when using AI in ERP?
Since ERP systems regularly process personal data, the GDPR requirements apply to AI use: a sound legal basis, a data processing agreement with the respective AI provider and, where the risk is high, a data protection impact assessment are required. Added to this is the EU AI Act: prohibited practices have already been banned since February 2025, while the obligations for high-risk systems originally scheduled for 2 August 2026 are expected to be postponed following the agreement on the “Digital Omnibus” reached at the end of 2025 (for standalone Annex III systems to December 2027) — the final adoption and publication in the Official Journal is decisive. The upper fine limit of 35 million euros or 7 percent of global annual turnover applies only to prohibited practices; violations of high-risk and provider obligations are capped at 15 million euros or 3 percent. In practical terms, it is also important to clarify contractually whether inputs feed into model training, as reputable enterprise offerings generally rule out such training with customer data.
What prerequisites must be met for AI in ERP?
The most important prerequisite is high data quality, because forecasts and suggestions are only as good as the underlying data; erroneous postings or incomplete master data lead to misleading results. Well-maintained master data and a central, consistent data repository in the sense of a single source of truth are therefore recommended, complemented by open interfaces for data transfer. Organisationally, recommendations should remain traceable and correctable and be documented via an audit trail so that auditability and accountability are preserved. A step-by-step approach has proven effective, starting with clearly delimited use cases in an isolated test environment and expanding only once results are robust.