Why AI Projects Fail: The Myth of the 60 Percent Quota

Anyone looking for the causes of unsuccessful digitalization initiatives inevitably comes across a seemingly set-in-stone figure: 60 percent. A supposed analysis from the summer of 2026 claims to prove that exactly this proportion of AI projects in medium-sized businesses fail due to inconsistent formats, gaps in master data, and missing history. The conclusion of many consultants is simply that data quality must be improved. But anyone making decisions based on such distorted quotes is investing their budget in the wrong place. The problem runs deeper than a few typos in the customer database. It is about the architectural inability of many systems to seamlessly document a metric.

Why AI projects fail and what lies behind the 60 percent figure

Why AI projects fail cannot be answered with a blanket quota for the SME sector, as the frequently cited 60 percent figure is based on a forecast taken out of context. The origin of this number is not a retrospective analysis from July 2026, but a Gartner press release from February 26, 2025. In it, the analysts predicted that by 2026, 60 percent of AI projects globally will be abandoned if they are not supported by AI-ready data. It was a prediction for global organizations, not a measured abandonment rate in the German SME sector.

The fact that blogs embellish this number later with specific reasons like "inconsistent formats" or "missing history" and explicitly relate it to SMEs is a free interpretation. The reality of the failures is much more complex and documented much more drastically in other surveys. For example, a meta-analysis by the RAND Corporation from 2025 of 65 documented enterprise AI initiatives shows that 80.3 percent of the projects deliver no business value. Of these, 33.8 percent are abandoned before production. A survey by S&P Global Market Intelligence among more than 1,000 companies in 2025 found that 42 percent of companies abandon most of their AI initiatives. The failure is real, but the cause is not what is stated in superficial summaries.

What poor data quality means in business practice

Poor data quality in business practice means that the same metric is in six different systems and none of them is reliably correct. If the ERP system, the online shop, the POS system, and the marketplace billing output different truths about a day's revenue, the foundation for any automation is missing. An artificial intelligence cannot make intelligent decisions if it operates on contradictory raw data.

The problem is exacerbated when companies try to bridge these discrepancies through manual exports and Excel islands. In many companies, the monthly financial statement or liquidity planning is created by manually copying tables together. This might work for a human controller because they know from experience which column to ignore. An AI agent, on the other hand, requires a normalized truth. If marketplaces, shops, ERP, ads, returns, inventories, and prices are not cleanly pulled together and joinable, the system produces hallucinations or simply wrong recommendations for action.

How a lack of traceability blocks automation

Successful automation through artificial intelligence requires that every calculated metric is seamlessly auditable down to its original document. If a system reports that energy costs per operating day have risen by 18 percent and this depresses the result by 2.1 points, a user—or a downstream AI agent—must be able to verify this claim. If the result is a black box, trust is missing, and the project is stopped.

This is exactly where most initiatives fail that quickly unleash a chatbot on unstructured corporate data. The AI provides answers but cannot prove the calculation path. Without a clear chain from the source to the output, the system is useless for business-critical decisions. The governance that ensures actions remain traceable is missing.

How auditable business monitoring lays the foundation

A functioning business monitoring system merges raw data from all sources into a continuous calculation where every value remains traceable to the source. This means the end of Excel islands. Instead of manually aggregating data, it is automatically imported and allocated. The result is a continuous P&L cascade, a clean monthly closing, and a rolling 13- to 52-week liquidity forecast from a single source.

The decisive difference to conventional dashboards is the depth of auditability. Every metric must follow a strict logic: Source field → Mapping → Tariff → Formula. If cost modeling reaches down to the individual invoice line—including shipping, commission, platform, and storage costs, which are versioned and allocated per channel, category, and country—then the system is auditable instead of a black box. Such a system scales even with 1.2 million invoices and 2 million invoice lines with a history spanning several years. Via protocols like MCP (Model Context Protocol), AI agents can then use this clean database to generate reliable answers.

How to check your own systems now

Managing directors best check the AI readiness of their data by tracing the path of a central metric from the dashboard backward to the raw data export. The test is simple: Take the contribution margin of a specific product or service from the last month. Ask IT or controlling to show the exact calculation path down to the original document.

As soon as the sentence "We pull this as a CSV here and adjust it in Excel" falls in this chain, the foundation for AI automation is not given. Every media break and every manual intervention destroys the chain that an algorithm needs to work reliably. Before budgets flow into complex AI models, the database must be normalized so that marketplaces, shops, and ERP systems form a single, actionable truth. Only when concrete recommendations are delivered with justification and nothing fires without approval (approval-first) does a system emerge that humans and machines can trust.