AI in SMEs: Why Companies with 50+ Employees Are Falling Behind

AI in SMEs: The origins of the 47.2 percent lag

Traditional medium-sized enterprises use artificial intelligence less frequently than large corporations and small businesses because entrenched processes clash with a lack of IT resources. According to a survey by the ifo Institute from June 5, 2026, 54.5 percent of German companies across all sectors use AI, up from 40.9 percent the previous year. However, the distribution by company size shows a clear dip: While large enterprises stand at 67.2 percent and small companies at 51.2 percent, the mid-market only reaches 47.2 percent. Those employing between 50 and 249 people often have processes too complex for simple standard software, but lack their own development department to train custom models. This gap causes established companies to fall behind technologically, while start-ups and corporations increase their efficiency.

What the ifo figures reveal about actual usage

The 54.5 percent quota merely measures the spread of the technology, not its depth or economic benefit. A company whose marketing department occasionally generates texts is included in the statistics just like an industrial company that has automated its entire production control. The data clearly shows that in-house developments remain the exception: Only 18.7 percent of AI-using companies build their own systems. Nearly three-quarters rely on paid external solutions. The technology has become part of everyday life, especially in industry (58.7 percent) and the service sector (56.2 percent), while the main construction trade shows the strongest dynamic with a jump from 7.1 to 39.8 percent since 2023. Other surveys, such as the one by Bitkom from February 2026, see usage among companies with 20 or more employees at 41 percent. These deviations arise from different samples but confirm the general trend: Adoption is rising rapidly, but the depth of integration often remains superficial.

Why the modular principle closes the resource gap

Companies without their own IT staff must rely on a ready-made foundation that only needs to be configured by management. This is exactly where many isolated AI projects fail because they try to train a model from scratch on their own data. A functioning modular principle reverses this path: The system already provides the architecture, the interfaces, and the governance rules. The management level defines the processes and approval stages, and these guidelines roll directly into the departments without translation losses. An overarching memory connects the individual tools. It records processes, behavioral patterns, and feedback, and learns from them. For example, if a campaign performs extremely well but inventory reach becomes tight, the system recognizes this correlation across tool boundaries and reports the conflict. This solves the core problem of lack of time and scarce personnel resources that causes SMEs to fall behind in the statistics.

How an integrated system operates in daily business

A productive system connects isolated data silos via a common knowledge layer to form a consistent truth. In business monitoring, this means that raw data from ERP, marketplaces, and shops flow together into a real P&L and liquidity calculation. Costs for shipping, commissions, or storage are allocated down to the individual invoice line. Every metric remains traceable to the source and auditable. In the area of accounting, such a system reads receipts via OCR and AI vision, creates finished booking proposals with accounts and tax keys, and learns from every manual correction. An orchestrator decides in three stages whether a receipt is booked automatically, submitted for review, or processed manually. Nothing is waved through blindly, and the system remains GoBD-compliant up to the DATEV export.

How AI-supported customer databases and analytics drive decisions

A normalized database ends the situation where the same metric is reported differently in six systems. An analysis center pulls together data from marketplaces, shops, returns, and inventories. On this basis, concrete, justified recommendations for prices, reorders, or clearances are created. The principle of approval applies here too: The AI suggests, the human decides. In the customer data platform (CDP), all traces from ERP, newsletters, and support are combined on one data record. The system enriches these profiles independently, officially validates VAT IDs, and extracts contacts from crawled websites. Every enrichment is backed by a source. This enables precise action proposals, for example when a regular customer threatens to churn, and makes the knowledge usable for human employees and AI agents alike.

Which legal deadlines apply after the Digital Omnibus

The obligations for high-risk AI systems under Annex III of the EU AI Act will not apply until December 2, 2027, while the general transparency obligation for generative AI is already active. Many older publications still name August 2026 as the deadline for high-risk systems. This information is outdated. On July 24, 2026, the Digital Omnibus Regulation (EU) 2026/1744 was published, which postponed this deadline by over a year. However, Article 50 was not postponed: Anyone using AI to generate content or communicate with customers via chatbots must disclose this. A legally compliant system has already firmly anchored these transparency obligations, risk classes, and human oversight in the form of tiered approvals (L0 to L4) in its architecture. The processing of sensitive data takes place via European endpoints to maintain data sovereignty.