The future of AI is built on data bricks

Artificial intelligence has become one of the central pillars in enterprise digital transformation strategies. AI is widely expected to automate processes, enhance decision-making, and unlock new business models. Yet, when observing AI deployments across organizations, a persistent gap between expectations and real-world value remains evident. Many AI initiatives are technologically well-funded but struggle to scale into operations or fail to generate tangible business impact.

In many cases, the issue does not lie in the algorithms or model sophistication, but in the underlying data foundation.

Through my hands-on involvement in deploying AI and data-driven initiatives in the financial and banking sector at FPT, I have had the opportunity to closely observe the foundational relationship between data and AI’s ability to generate real business value. The perspectives shared in this article are not intended to offer a universal formula for all enterprises, but rather to reflect insights and lessons accumulated from working with organizations at different levels of data maturity.

AI is only as intelligent as the data it learns from

At its core, AI is not “intelligent” in the human sense. It is a system that learns from data, identifies patterns, and generates predictions based on historical inputs. This means the quality, completeness, and representativeness of training data directly determine the reliability of AI outputs.

In practice, it is not uncommon to see models achieve high accuracy in controlled testing environments but perform poorly once deployed. The root cause often lies in training data that fails to reflect real operational contexts—or data that looks technically clean but lacks business depth. In such cases, AI still produces outputs, but those outputs are not trustworthy enough to support real decision-making.

When data fails to reflect operational reality

A common challenge arises when data is collected under ideal conditions, while real-world environments are far more complex and volatile. In computer vision or behavior recognition use cases, for example, training data is often standardized in terms of lighting, angles, or context. Once deployed in uncontrolled environments, model performance can deteriorate rapidly.

This highlights a critical point: AI itself is not “wrong”—the data simply isn’t real enough. If training data does not represent the world in which AI operates, expecting sustainable value creation becomes unrealistic.

Data bias and the risks of AI-driven decisions

Data bias is one of the most subtle and dangerous challenges in AI implementation, as it does not always surface in technical performance metrics. When historical data reflects biased decisions or behaviors, AI models trained on that data tend to replicate—and amplify—those biases.

In human resources, this is particularly visible in recruitment use cases. If historical hiring data shows preferences toward certain backgrounds, genders, age groups, or profiles, AI screening systems may interpret these traits as positive signals. As a result, qualified candidates who do not resemble past patterns may be eliminated early in the process.

A similar issue arises in workforce optimization and cost reduction. When performance data is measured purely through quantitative indicators—such as hours worked, tasks completed, or cost per employee—AI may recommend cutting roles deemed “inefficient.” Yet such data often fails to capture intangible contributions like experience, leadership, or tacit knowledge. Decisions made solely on these metrics risk eroding core capabilities that are invisible to the data.

These examples demonstrate that the real risk lies not in AI itself, but in how data is selected, interpreted, and applied within specific business contexts.

Beyond immediate resource and trust losses, biased AI decisions can lead to significant financial damage—such as missed business opportunities, operational inefficiencies due to inaccurate demand forecasts, or unnecessary inventory costs. They can also weaken long-term competitiveness by eliminating experienced personnel based solely on incomplete metrics. From a legal and reputational standpoint, biased AI may expose organizations to discrimination claims, erode trust among employees, candidates, and customers, and cause lasting harm to corporate culture and brand equity. At a broader level, unchecked data bias risks reinforcing social inequalities by embedding historical biases into automated decision loops.

The reality of enterprise data in vietnam

Through hands-on work with multiple organizations, one recurring observation stands out: data remains the biggest barrier to AI adoption. Enterprise data is often fragmented across disparate systems, created for isolated operational purposes, and lacking shared context. Data access and integration still rely heavily on manual effort, making analysis slow and difficult to scale.

In this context, “leapfrogging” directly into AI without a solid data foundation typically results in prolonged pilot projects that are hard to replicate and fail to generate meaningful business impact.

When AI gets the right answer—but is still not used

Another common scenario is that highly accurate AI models never make it into production. In many cases, models function as black boxes, offering predictions without clear explanations. This lack of explainability becomes a critical barrier in sensitive domains such as finance, healthcare, or legal services, where users and regulators must understand the rationale behind decisions to ensure trust and compliance.

Additionally, AI outputs may conflict with organizational culture, regulatory requirements, or human judgment. While AI may propose statistically optimal solutions, it often lacks consideration for qualitative factors that humans value.

This underscores the need to position AI within a broader ecosystem—where data, people, and processes coexist and interact.

Generative AI and a new perspective on data

The rise of generative AI has brought data challenges into sharper focus. Gen AI requires not only large volumes of data, but data enriched with context, knowledge structures, and meaningful connections. When data lacks proper governance or contextual grounding, Gen AI can easily produce outputs that sound plausible yet are fundamentally unreliable.

This forces enterprises to rethink how they store, organize, and leverage data—not just for traditional analytics, but to enable meaningful interaction between humans and AI.

Conclusion

AI is neither the starting point nor the final destination of digital transformation. It is a tool, and its value depends heavily on the data foundation beneath it. Well-constructed data “bricks”—governed, contextualized, and aligned with business realities—enable AI to support decision-making and deliver sustainable enterprise value.

In today’s AI landscape, the data challenge no longer revolves around storing or querying billions of records. The more important question is how to transform those records into a knowledge network—one where humans and AI can interact, reason, and co-create insights. When data evolves from information into knowledge, AI becomes an organic part of enterprise operations, rather than an add-on technology layered onto existing systems.