Before More Models, AI in Commerce Needs an Infrastructure Revolution

The implementation of artificial intelligence in commerce is moving at a remarkable pace. AI investment in retail is accelerating rapidly, with the market projected to grow from $11.6 billion in 2024 to over $40 billion by 2030.

Similarly, new models are released in rapid succession, vendors continue to expand their capabilities, and retailers are under increasing pressure from boards and investors to demonstrate that they are actively integrating AI into their operations.

On the surface, this creates the impression of an industry undergoing a fast and decisive transformation. Yet, beneath that activity, most of these initiatives are being layered onto systems that were never designed to support them.

The limitation is neither the availability nor the quality of models. It is the condition of the infrastructure underneath them.

When implementing AI, the first budget should not go into AI itself. It should go into what we call “data trust”, which is data infrastructure. These are connections between systems, data deduplication, and real-time updates.

Failing to do so is why AI fails in commerce, and why so many efforts begin with momentum but struggle to translate into sustained impact. To successfully implement AI, we need to start with infrastructure.

In contrast, in commerce, the conversation itself tends to start one layer too high, focusing on choosing models or identifying high-potential use cases. What receives far less attention is whether the underlying data architecture can support any of these decisions in a meaningful way.

If the underlying data is fragmented, AI will only generate outputs that can actually be harmful or misleading.

In larger organizations, this tendency is reinforced by the need to signal progress. “Doing something with AI” becomes an objective in itself, and introducing a model is often the most immediate way to achieve that, regardless of whether it produces measurable value.

At the board level, expectations around AI are accelerating in the form of better personalization and measurable gains in efficiency and margin. At the system level, readiness is not keeping pace. The longer this gap persists, the more expensive it becomes to correct, as layers of technology are added on top of foundations that remain unchanged.

What “not ready for AI” looks like in practice

A main issue is the lack of coherence across systems. For instance, a single customer may exist simultaneously in an ERP platform, an e-commerce environment, and an offline or loyalty system, each instance carrying different histories and assumptions.

Product data, while technically complete, is limited in terms of providing context that would allow a model to interpret or reason about it effectively. Poor data quality alone is estimated to cost organizations nearly $13 million per year, underscoring how foundational these constraints have become.

Human teams learn to navigate these inconsistencies over time, building informal bridges between systems. A model, in contrast, operates strictly within the structure it is given, and the outputs it generates are constrained accordingly.

When AI is introduced into this environment, the result is not a lack of output but a gradual degradation in its quality.

A model can process data and produce outputs, but it will not connect systems on its own.

In practice, this may show up as a personalization engine that begins to offer highly specific product recommendations that categorize a customer in ways that do not reflect their actual preferences or intent.

In AI-driven systems, customer experiences often become more precise but less relevant.

This shift from general relevance to specific irrelevance can be more damaging than it appears, as it alters how customers perceive the brand itself. In the worst cases, these systems do not simply underperform. They actively degrade the customer experience.

Why most AI pilots stall, and what it costs

This dynamic is one of the primary reasons why a large proportion of AI pilots in commerce fail to progress beyond experimentation. Industry data suggests that only around half of AI initiatives make it from pilot to production, leaving a significant proportion stalled at the experimentation stage.

The issue is rarely the absence of investment or intent. The constraint emerges when the outputs generated by these systems do not justify scaling.

At that point, organizations face a choice between continuing to invest in optimization or abandoning the initiative altogether. In many cases, they choose the latter because the environment in which it is being deployed cannot support consistent results.

Across the industry, the same infrastructure gaps repeat themselves. Customer data remains fragmented rather than unified. Product information exists, but lacks the depth required for interpretation. Systems operate in parallel rather than as a single, coherent flow of information. When models are introduced into this environment, they scale these problems.

The cost of this approach extends beyond direct expenses. It manifests as a system-wide accumulation of recommendations that fail to convert, personalization that feels inconsistent, and automation that amplifies existing errors.

Over time, this creates a subtle but significant erosion of trust within the organization. Teams begin to question the reliability of the systems they are building, and additional layers of technology are introduced in an attempt to compensate. The result is a delay in reaching the clarity required to build systems that work.

The reason why many companies build AI pilots but never go beyond them is that they’re building these pilots on top of weak foundations, which is like cooking with bad ingredients.

You might have a Gordon Ramsay recipe and excellent equipment, but if the ingredients are poor, the result will still be negative.

The difference between startups and large corporations

Smaller organizations, particularly those in the mid-market range, often demonstrate a greater ability to implement AI effectively. This is because their systems are less fragmented and their decision-making processes are more direct.

In larger organizations, the structure itself introduces potential hindrances. Initiatives are frequently driven by the need to respond to external expectations, and AI becomes a category that must be addressed instead of a capability that is carefully integrated.

Saying “we want to implement AI” is like bringing electricity into a factory without machines. The lights turn on, but production doesn’t improve, because you invested in technology without knowing what it will actually do.

The shift from models to infrastructure

An alternative approach is beginning to take shape among organizations that have encountered these limitations directly.

Instead of allocating initial budgets to AI models, they are investing in what can be described as data trust, which is the creation of a foundation where data is consistent, connected, and continuously updated. The objective is to ensure that the data being used can support meaningful decisions.

Data preparation can take up to 80% of the time spent on machine learning projects, underscoring how much of the work sits in the underlying data layer rather than in the models themselves.

This process typically begins with a detailed mapping of existing systems and the identification of gaps in data flow and integrity. The first 30 to 60 days are often dedicated to understanding how information moves or fails to move across the organization. This is followed by a period of integration and standardization, during which data is cleaned, de-duplicated, and aligned across platforms.

Over the next three to six months, companies focus on building reliable data pipelines and connecting high-value systems in ways that support real use cases. Only once this foundation is in place does it become possible to introduce AI into workflows that can produce consistent, measurable outcomes.

Within a six- to nine-month horizon, organizations that follow this approach typically begin to see meaningful results in the form of AI-driven processes that influence metrics such as conversion, retention, and margin improvement. At that point, models stop being experiments and become part of operations. Their performance stabilizes and their outputs become actionable to the point that teams can begin to rely on them.

The acceleration of AI investment in commerce is unlikely to slow in the near term. What remains uncertain is whether infrastructure readiness will keep pace and be able to support it. The widening gap between what organizations expect AI to deliver and what their systems can support is already visible in the growing number of pilots that fail to reach production. Addressing this gap requires a shift in focus to building this critical infrastructure.

Moving the conversation from models to infrastructure does not carry the same immediacy or visibility. But it is where many of the current constraints reside. And it is here where the next phase of meaningful progress in AI-driven commerce is likely to emerge.