Expert Analysis

Informants consistently emphasized that high-quality data is indispensable across the AI lifecycle, representing a shared concern among both C-level executives and compliance officers. Maintaining high-quality data for AI systems involves high operational costs and is extremely expensive.

Poor data quality was described in concrete terms—missing, incomplete, or partial data, inaccurate, irrelevant, duplicate, or inconsistent records, and incorrectly labeled or annotated datasets. This can lead to a "garbage in, garbage out" situation, introducing a lot of noise into the models and severely limiting their accuracy. The consequences extend to delayed time-to-market, increased operational costs, eroded trust in AI systems, and regulatory non-compliance.

Concerns around data privacy, leakage, and cybersecurity threats were prominent across interviews. Employees may inadvertently expose sensitive information through AI tools. Some vendors claim to be AI providers, but in reality, they're collecting everyone's data to train their models and sell them to big tech companies.

A company purchasing LLMs must be absolutely certain that its data is securely handled by the supplier. Managers often struggle to control what employees input into AI tools, so it's crucial for every organization to have a clear policy on this matter. Deployed AI models may be exposed to adversarial attacks, where malicious users attempt to manipulate predictions or access sensitive information.

This underscores the need to ensure that every department complies with national privacy regulations, particularly GDPR. There's so much data that it's hard to ensure none of it leaks or gets lost. Strong security measures are absolutely essential.

Unintended algorithmic bias was consistently identified as a critical challenge in AI system development. Even well-trained models can produce biased results, which may have unintended social or ethical consequences that are often difficult to detect in real time. Model drift—over time, AI models may become less effective due to changes in underlying data or external factors, leading to inaccurate predictions.

Real-world examples illustrate how contextual factors can distort model outputs. For instance, a senior data science lead at a U.S.-based company specializing in smart water sensors shared that during the Super Bowl in the U.S., people's water usage patterns change dramatically. The AI-based sensors misinterpret this as a leak, introducing bias into the model.

Bias can also emerge from the early stages of data preparation. Even slight pixel-level inconsistencies in training data can cause major disruptions, leading to bias and rendering the results irrelevant. Humans are biased, and therefore so are the developers of these models. There's a deep lack of understanding about what AI actually does and how it works. In practice, it's a black box.

Our organization conducts rigorous quality assurance processes based on multiple logic layers throughout the training phase, because we don't fully trust the model outputs. We flag errors as they arise and perform unique model validation for each dataset—we don't just feed data into the model blindly.

The topic of AI regulation generated considerable discussion, particularly among participants from highly regulated sectors such as finance, healthcare, legal services, and the public sector. Many customers frequently ask about the compliance of AI-based products with regulatory and legal standards. Despite growing awareness, many acknowledged that regulatory adaptation is still in its early stages.

Most current efforts appear to focus narrowly on privacy compliance. Some companies are just beginning to explore the broader implications of the EU AI Act. For now, most are focusing on transparency and documentation of training processes. A data protection consultant in Italy confirmed the limited practical readiness: "Although we're seeing a shift toward AI Act compliance, we currently lack the practical tools to address it. For now, it's mostly about documentation." There's no clear way to enforce these regulatory frameworks, so in practice, anyone can do whatever they want. The burden of responsibility is largely pushed back onto the major suppliers, like OpenAI. However, some organizations are beginning to adopt more structured and proactive approaches, including continuous monitoring and flexible data governance frameworks.

This study investigated how strategic professionals in AI and data science conceptualize the role of data in shaping AI-enabled solutions. Drawing on in-depth interviews with 74 senior experts, the findings offer a grounded, practice-oriented perspective that disrupts dominant model-centric paradigms in AI research. While much academic focus remains on algorithmic innovation and model development, this study repositions data as the principal site of complexity, uncertainty, and strategic decision-making in real-world AI development.

The model delineates a data-centric process that spans from initial collection and preparation to deployment, monitoring, and explainability; each phase involves distinct professional roles and tightly coupled interdependencies. Rather than presenting a linear pipeline, the model emphasizes the recursive and evolving nature of data work, portraying data not as a static input but as an active infrastructure that is continuously shaped by, and shaping, technical and organizational decisions. Whether through feature engineering, error correction, or interpretability practices, data emerges as both the foundation and connective tissue of AI systems.