Your AI Project Is a Data Problem First

The Pattern We See Everywhere

Many companies reach a point where they decide to “do something with AI.” A pilot is defined, a tool is demonstrated, and a few months later the conclusion is disappointing: the answers are unreliable, and the team stops using it.

In most cases, the root cause is not the AI itself. The issue lies in fragmented operational data. Information is spread across ERP systems, spreadsheets, SharePoint, PIM platforms, and personal notes. The same customer, product, or order exists in multiple versions, with no clear source of truth.

An AI system built on top of this does not resolve the inconsistency. It scales it.

A Single Source of Truth

The common reaction is to aim for one system that does everything. In practice, this rarely works. Even organizations with strong ERP systems rely on multiple tools, and full consolidation is a long and costly process.

A more practical approach is to define a single source of truth per data type. Each dataset has one system where it is maintained and one clear owner. For example, customer data lives in the CRM, stock in the ERP, and product documentation in a PIM. Other systems reference this data rather than duplicating it.

This is primarily an organizational decision, not a technical one. The structure can be defined quickly, but enforcing it consistently is where the real value is created.

Add a Layer, Not More Complexity

Becoming AI-ready does not require moving all data into a central warehouse. For many SMEs, that approach introduces unnecessary complexity.

Instead, a more effective strategy is to build a governed access layer on top of existing systems. This layer translates raw data into business-ready formats, routes queries to the correct sources, and manages access control. The underlying systems remain unchanged.

This approach also creates flexibility. AI applications connect to the layer rather than individual systems, allowing data sources to evolve without breaking existing solutions. Data improvements and AI initiatives can progress in parallel.

What This Means in Practice

If AI initiatives are not delivering results, the solution is rarely a better model or a new tool. The key questions are more fundamental:

• Who owns each type of data, and which system is the source of truth?

• Where is data manually copied between systems?

• Which decisions rely on data that is not fully trusted?

By answering these questions and introducing a governed access layer, AI use cases such as assistants, reporting, and planning become reliable and scalable.

The Right Order

The sequence matters. Reliable data comes first, AI follows. This approach may seem less impressive at the start, but it delivers significantly stronger results over time.