Despite the promises of Artificial Intelligence, a striking 70–80% of AI initiatives fail—more than double the rate of traditional IT projects. Contrary to common assumptions, these failures rarely stem from algorithmic limitations. Instead, they are overwhelmingly caused by poor data quality.
Data Quality: The Bedrock of AI
Machine learning systems rely entirely on the data they are trained on. If the data is flawed, incomplete, biased, or inconsistently labeled, even the most advanced AI models will underperform or produce harmful outcomes. Key data-related issues include:
- Inaccurate or incomplete data: Leads to ineffective learning regardless of algorithm sophistication.
- Biased datasets: Result in AI systems that perpetuate or amplify biases (e.g. societal inequalities).
- Data scarcity or excess: Either leads to oversimplification or overfitting on noise.
- Inconsistent labeling and siloed data: Prevent meaningful pattern recognition and limit the AI’s scope.
Case Studies of AI Failure
- IBM Watson for Oncology: Despite a $62 million investment by MD Anderson Cancer Center, Watson failed to provide clinically useful cancer treatment recommendations due to being trained on hypothetical, rather than real, patient data. Its opaque “black-box” decision-making also reduced physician trust, leading to project termination [1][2][3][4].
- Amazon’s AI Recruiting Tool: Trained on historical hiring data skewed towards male candidates, the tool developed a systemic gender bias. It downgraded resumes mentioning "women's" activities and favored male language patterns. The project was abandoned after attempts to de-bias it failed [3].
- Air Canada’s Chatbot: A customer received incorrect information about refund policies from an AI chatbot, resulting in legal liability. A tribunal ruled Air Canada responsible for the misinformation, forcing them to honor the refund and exposing the risks of deploying AI with flawed data [3].
- Apple Intelligence: In 2025, Apple’s news summarization AI fabricated false information and misattributed it to credible sources like the BBC. Public backlash forced Apple to suspend the feature and reassess how it labels AI-generated content [3].
Broader Costs of Poor Data
- Financial: Failed AI projects waste millions in investments and resources that could have been used elsewhere.
- Time: These initiatives often consume years of development, delaying more viable solutions.
- Reputational Damage: High-profile errors damage public trust in both the company and AI as a technology.
- Compliance: Data quality is crucial for regulatory compliance, ensuring accurate reporting, transparency, and risk management. In sectors like banking or insurance, and under rules like GDPR, CSRD, or e-invoicing, poor data can lead to misreporting, privacy breaches, and penalties.
Solutions for AI Success
Organizations must treat data as a strategic asset, implementing robust Data Quality Frameworks before building AI solutions. This includes:
- Assigning data quality ownership: The right person for the job is most likely already working for the organization. This person must have in-depth knowledge of the company's data and business. These are two essential pillars that cannot be obtained by recruiting from outside the company.
- Involving business users in data quality processes because they are the ones who understand the anomalies most likely to impact the proper functioning of the company.
- Accepting that Data Silos are inevitable, as they reflect the disparities in the use of data within the different departments of a company. Don't waste time or money trying to break down these silos. However, there are software solutions that make it possible to reconcile data from heterogeneous sources in real time, upstream of AI algorithms.
- Monitoring Data Quality on a daily basis (Data Content Observability): Your automatic monitoring system must send relevant alerts to the right people as soon as a situation likely to impact the operation of the company occurs.
- Using modern tools for Data standardization and Remediation operations as opposed to python scripts that end up accumulating and are completely opaque to business users.
Conclusion
No AI system can outperform the quality of its input data. Organizations that wish to realize AI’s potential must prioritize clean, relevant, and unbiased data from the outset. "Garbage in, garbage out" remains a harsh but inescapable truth.
External links:
