Data Quality as the Foundation for AI: Why Good Data Determines Success or Failure

Artificial intelligence is the dominant technology topic in the German mid-market. Yet while companies invest in AI tools, models, and platforms, many overlook the decisive success factor: the quality of their data. Current research published by the Frankfurter Allgemeine Zeitung shows that 85 percent of failed AI projects do not fail because of the technology, but because of inadequate data quality. At the same time, a study by the German Chambers of Industry and Commerce (DIHK) from March 2026 reveals that 71 percent of small and medium-sized enterprises consider access to high-quality data the biggest hurdle when introducing AI. This article provides you with a practice-oriented guide to systematically improve your data quality and thereby lay the foundation for successful AI projects.

Garbage In, Garbage Out—Why Data Quality Is the Number One AI Success Factor

The principle is as old as computer science itself: when a system is fed with erroneous, incomplete, or inconsistent data, it cannot deliver reliable results. With traditional software, poor data quality leads to incorrect reports and flawed processes. With AI systems, the problem is amplified because machine learning detects patterns in data—and it learns faulty patterns just as reliably as correct ones.

The Economic Consequences of Poor Data

The numbers are clear: according to the current Gartner Data Quality Report from March 2026, poor data quality costs companies an average of 12.9 million US dollars per year. For mid-sized companies in Germany, this does not always mean losses in that magnitude, but the relative impact is comparable. Incorrect inventory levels, duplicate customer records, outdated supplier information, and inconsistent product data create a cascade of consequential costs: manual rework, incorrect deliveries, missed business opportunities, and customer churn.

Gartner also forecasts that organizations actively investing in data quality will experience roughly 70 percent fewer problems with AI implementations by the end of 2026 than companies that skip this step. This is not a theoretical prediction but a calculation based on real project data.

Why Mid-Sized Companies Are Particularly Affected

Large corporations employ specialized data engineering teams, Chief Data Officers, and data governance departments. The German mid-market, however, frequently operates without these structures. Data is maintained in various departments, systems, and formats—often without central accountability. The DIHK study impressively documents this gap: only 18 percent of the SMEs surveyed have a documented data strategy. In reverse, this means that more than four out of five companies manage their data without a strategic plan.

At the same time, the pressure to deploy AI is increasing. The current Bitkom survey from March 2026 shows that 54 percent of companies plan higher investments in data management for 2026. The three most important drivers: AI requirements (82 percent), regulatory requirements and ESG reporting obligations (63 percent), and rising customer expectations (51 percent). The willingness to invest is there—but it must be channeled properly.

The Five Most Common Data Problems in Mid-Sized Companies

The DIHK study identifies three main problems: data silos, missing standards, and manual data entry. In practice, we regularly encounter five recurring patterns at SMEs that prevent AI success.

1. Data Silos Between Departments

The most classic and at the same time most persistent problem: sales works with the CRM, accounting with the ERP, marketing with its own spreadsheets, production with an MES system. Each department maintains its own data world. Customer numbers do not match, product names vary, and master data exists in multiple versions. When an AI system accesses this fragmented data, contradictory results emerge that create more confusion than value.

A typical example: sales records a customer as “Mueller GmbH,” accounting as “Fa. Mueller,” the warehouse as “Mueller Industriebedarf.” For the AI, these are three different customers—with correspondingly incorrect analyses of ordering behavior, revenue forecasts, and customer lifetime value.

2. Manual Data Entry and Media Breaks

In many mid-sized businesses, data is still manually transferred into systems. Delivery notes are typed in, orders from emails are manually entered into the ERP, measurement values are read from machine displays and recorded. Every manual step is a potential source of error. Transposed digits, typos, forgotten entries, and time delays add up to a significant quality problem.

3. Missing Data Standards and Naming Conventions

Without binding rules for data entry, a thicket of formats and designations develops. Date formats switch between DD.MM.YYYY and MM/DD/YY, units of measurement are sometimes kilograms, sometimes kg, sometimes KG. Article numbers follow no uniform schema. Free-text fields are filled in differently. These inconsistencies make automated evaluations and AI training extremely error-prone.

4. Outdated and Redundant Data

Databases grow over the years. Former employees, long-dissolved supplier relationships, discontinued products, and historical price lists remain in the system. Without regular cleansing, data volume increases while data quality decreases. AI models trained on these legacy records learn patterns from a past that is irrelevant or even misleading for current business decisions.

5. No Clear Data Responsibilities

Who owns the customer data? Who is responsible for maintaining article master data? Who decides on retention periods? In companies without defined data responsibilities—and according to DIHK, that is the vast majority of SMEs—nobody systematically takes care of data quality. The consequence: everyone assumes someone else is handling it. In reality, nobody is.

The 5-Step Framework for Better Data Quality

Computerwoche presented a field-tested framework in March 2026 that is specifically tailored to the needs of mid-sized companies. It consists of five sequential steps, which we present here with concrete measures, recommended tools, and a realistic timeframe.

• Step · Measure · Tools and Methods · Timeframe
• 1. Inventory · Complete inventory of all data sources, formats, and responsibilities in the company · Data mapping workshops, data flow diagrams, interviews with departments · 2 to 4 weeks
• 2. Cleanse · Identification and correction of duplicates, gaps, errors, and outdated entries in all core systems · Data profiling tools, duplicate matching, rule-based validation · 4 to 8 weeks
• 3. Standardize · Introduction of binding naming conventions, data formats, field specifications, and input masks · Master Data Management, data catalogs, style guides for data entry · 3 to 6 weeks
• 4. Automate · Replacement of manual capture processes with automated data pipelines and interfaces · ETL tools, API integrations, n8n workflows, automated validation rules · 4 to 10 weeks
• 5. Monitor · Ongoing monitoring of data quality through dashboards, alerts, and regular audits · KPI dashboards, automatic anomaly detection, quarterly data quality reviews · Ongoing

Why the Sequence Matters

Many companies make the mistake of jumping directly to step 4—automation. They buy expensive integration tools, build data pipelines, and automate processes without first having inventoried and cleansed the existing data. The result: they automate the error. Faulty data now flows faster and more reliably through the system—but it remains faulty.

The FAZ puts it succinctly: data cleansing must take place BEFORE AI implementation, not in parallel or afterward. Those who maintain this sequence save significant costs for rework and error correction in the long run.

Typical Pitfalls During Implementation

Beyond the wrong sequence, there are other common mistakes in data quality projects:

• Perfectionism over pragmatism: Not every data field needs to be perfect immediately. Start with business-critical data—customer master data, product data, financial data—and expand step by step.
• One-time project instead of continuous process: Data quality is not a project with a beginning and end. Without step 5 (monitoring), quality deteriorates back to its original level within a few months.
• Technology without organization: Even the best tool is useless if nobody is responsible. Define data owners for every critical data area.
• Lack of executive support: Data quality must be a leadership priority. Without a mandate from the top, there is no budget, no priority, and no authority to enforce changes across departments.

Case Study—Logistics Company Reduces Incorrect Deliveries by 34 Percent

Computerwoche reported in March 2026 on a mid-sized logistics company that impressively demonstrated the impact of systematic data cleansing. The company was struggling with a high rate of incorrect deliveries that both caused costs and burdened customer satisfaction.

Initial Situation

The logistics company operated several parallel systems: a warehouse management system, a transport management tool, and a separate order management system. Address data was partly entered manually, article numbers were not uniformly coded, and inventory data regularly deviated from actual warehouse stocks. Initial attempts to introduce AI-driven route optimization and demand forecasting failed due to inadequate data quality. The algorithms delivered unusable results because the underlying data was full of inconsistencies.

Measures

Instead of continuing to invest in AI tools, the company decided to first restore the data foundation. Over a period of approximately three months, the following steps were implemented:

• Data Audit: Complete inventory of all data sources and data flows between systems.
• Duplicate Cleansing: Matching and merging of over 15,000 duplicate address records.
• Standardization: Introduction of uniform article numbers and address formats across all systems.
• Automation: Implementation of automated interfaces between warehouse management, transport management, and order management.
• Quality Control: Establishment of weekly data quality reporting with automatic alerts for anomalies.

Results

The results were measurable and convincing:

• Incorrect deliveries: Reduction by 34 percent within the first six months after completion of the cleansing.
• Inventory accuracy: Increase from previously 78 percent to 96 percent.
• Rework: Manual correction effort decreased by more than 40 percent.
• AI readiness: After data cleansing, AI-driven route optimization was successfully introduced and immediately delivered reliable forecasts.

This example shows: the investment in data quality not only pays off directly through fewer errors but creates the very prerequisite for successful AI deployment. Without the prior data cleansing, the logistics company would have continued to try in vain to get AI projects running.

Frequently Asked Questions

What does a data quality project cost for a mid-sized company?

Costs vary considerably depending on company size, number of systems, and the state of existing data. As a rough guide: a structured data quality project for an SME with 50 to 250 employees and three to five core systems typically ranges between 15,000 and 60,000 euros. The most common mistake is viewing this investment as a pure cost factor. Gartner puts the annual cost of poor data quality at an average of 12.9 million US dollars. Even if you scale this down to a fraction for an SME, the costs of inaction far exceed the investment in quality.

How long does it take until we have AI-ready data?

That depends on the starting condition. With the 5-step framework described, you should plan a period of three to six months for steps 1 through 4. Step 5—monitoring—is an ongoing process. Important: you do not have to wait until everything is perfect. Once business-critical data areas are cleansed and standardized, you can start initial AI pilot projects and improve data quality in other areas in parallel.

Do we need specialized IT staff for this?

Not necessarily. For the inventory and the definition of standards, you primarily need domain expertise from the specialist departments—the people who work with the data daily. For the technical implementation (duplicate cleansing, interface development, automation), external support can be useful, especially if there is no internal data engineering expertise. What is decisive is the role of the data owner: one person per data area who is bindingly responsible for quality.

How do we concretely measure data quality?

Data quality can be measured along five dimensions: completeness (proportion of filled mandatory fields), correctness (agreement with reality, e.g., through random sample checks), consistency (agreement of the same data across different systems), timeliness (age of records), and uniqueness (proportion of duplicate-free records). Define a target value for each area and measure against it regularly. A pragmatic starting point: bring completeness of customer master data to at least 95 percent.

Do we have to cleanse all data at once?

No, and you should not try to. Prioritize by business relevance. Start with the data that is most important for your planned AI use cases. If you are planning an AI-driven demand forecast, start with order, inventory, and supplier data. If you want to build an AI assistant for customer service, begin with customer master data and communication history. Expand the scope step by step.

References

The following sources form the basis of this article and were published or evaluated in the period from March 1 to 7, 2026:

• Frankfurter Allgemeine Zeitung (approx. March 4, 2026): “Data quality is vital for AI projects”—https://www.faz.net/aktuell/wirtschaft/digitec/
• German Chambers of Industry and Commerce (DIHK) (approx. March 2, 2026): Study on AI barriers for SMEs—Access to high-quality data as the biggest challenge—https://www.dihk.de/de/themenfelder-und-branchen/digitalisierung
• Gartner Data Quality Report (approx. March 1, 2026): Costs of poor data quality and forecasts for AI implementations—https://www.gartner.com/en/newsroom
• Computerwoche (approx. March 5, 2026): Best practices for data quality in mid-sized companies—5-step framework and logistics case study—https://www.computerwoche.de/
• Bitkom (approx. March 3, 2026): Survey on investments in data management 2026—https://www.bitkom.org/Presse