AI in procurement fails due to data quality and missing standards

Almost 70 percent of medium-sized companies are planning or testing the use of algorithms in procurement, yet only 6.0 percent rate their implementation as highly mature. The reason for this gap is not the technology, but the foundation: Without reliable data, clear process standards, and qualified employees, automation and machine learning have no impact. Anyone wanting to digitize procurement must first clean up master data, eliminate isolated solutions, and standardize document processing.

Why AI in procurement often gets stuck in practice

It fails due to a lack of resources, poor data quality, and unclear processes. According to the Einkaufsbarometer Mittelstand 2026 by Onventis and the BME, 69.5 percent of the surveyed specialists and executives use, test, or plan the use of artificial intelligence. However, the reality is sobering: Only 6.0 percent classify their AI use as highly mature. Even with classic automation, only 18.6 percent see a high level of maturity. The hurdles are structural. 64.8 percent of respondents cite a lack of personnel or financial resources as the biggest obstacle. 56.0 percent fail due to data quality, and 46.2 percent complain about a lack of process standards. Although 60.6 percent of companies have earmarked budgets for technology and digitization measures, these funds are only clearly defined in 20.5 percent of cases. This shows: The will to modernize is there, but the basic prerequisites for feeding algorithms with meaningful information are missing.

Where the greatest potential for automation really lies

Invoice processing offers the greatest leverage for efficiency gains and is the logical starting point. 95.1 percent of procurement managers see the greatest automation potential exactly here. A modern document and accounting system turns an unstructured document archive into a largely automated, DATEV-compatible accounting system. The process from receipt to booking works without media breaks. The system reads documents via OCR and AI vision – even handwriting – and creates a finished booking proposal including accounts, splits, and tax keys. It is important that the system is not based on rigid rules, but learns from every human correction. Every confirmed or corrected booking trains the software, which measurably increases the proposal quality per creditor. A safety net of rules, history, machine learning, and large language models ensures that an orchestrator decides in three stages: automatically, for review, or manually. Nothing is waved through blindly. The result is a GoBD-compliant database right up to payment, which allows questions to be asked of the archive and answers to be received in seconds with the appropriate source link to the document.

How reliable data changes procurement

Only when inventories, prices, and delivery times flow together into a normalized truth can software make well-founded decisions for reorders. In many companies, data is scattered: Marketplaces, shops, ERP systems, returns, and inventories are not linked. A central analysis center pulls this information together and makes it joinable. This ends the situation where a number is in six systems and none of them is correct. On this basis, humans and machines work together. The system models the actual business, including multi-year seasonality, supplier lead times, and backward-calculated order deadlines. This results in concrete recommendations with justifications for prices, reorders, or pre-orders. The approval-first principle applies: Nothing fires without human approval. If parameters change, audit trails and drift detectors report immediately if something tips over. This data flows directly into business monitoring, which merges raw data into a continuous P&L and liquidity calculation. Cost modeling down to the individual invoice line – such as shipping, commission, and storage costs – is versioned and exactly allocated. Every key figure remains traceable and auditable back to the source.

Which legal requirements apply from August 2026

From August 2, 2026, the transparency obligations of the EU AI Act will apply, while the strict high-risk rules have been postponed to the end of 2027 by a change in the law. Anyone using chatbots for supplier inquiries or voice agents in procurement must disclose from this deadline according to Article 50 that a person is communicating with a machine. The obligations for high-risk systems (Annex III) originally planned for August 2026 were postponed to December 2, 2027, by the Digital Omnibus Regulation. Classic applications in procurement, such as invoice processing or spend analytics, generally do not fall under the high-risk classification anyway. However, if systems for automated credit checks of suppliers are used, these postponed deadlines apply. In Germany, the BSI is not responsible for monitoring, as is often falsely claimed, but according to the KI-MIG, the Bundesnetzagentur is.

How to successfully enter automated procurement

The path leads via a clean foundation of structured documents and a central database, not via isolated pilot projects. The various tools in the company need a shared memory. Each module must record processes, behavior patterns, and feedback and learn from them. An overarching memory brings this information together in a common knowledge base. In this way, correlations across tool boundaries can be recognized – for example, when a marketing campaign leads to a short inventory reach and reordering must take place in good time. The system reports strategy conflicts when knowledge and action contradict each other, and independently identifies knowledge gaps. Management configures processes and rules top-down according to their ideas, and these roll directly into the departments without translation losses. Only in this way does a theoretical plan become a measurable relief in daily business.