AI doesn't solve your data problem. It amplifies it.

In practice, however, we see that this expectation does not always become reality. A pilot delivers promising results, but as soon as you try to scale, questions begin to emerge. Can you trust the answers? Where does the information come from? And why does the same system sometimes produce different outcomes?

The root cause usually isn't the AI model itself. What AI exposes are the data challenges that were already there.

The pattern behind AI projects that stall after a successful pilot

The barrier to getting started with AI is low. Connecting a chatbot to SharePoint, implementing Microsoft Copilot or testing an AI agent can often be done in a relatively short time. That is exactly why the first results often look promising.

But what happens when you move beyond the pilot phase? Many organisations encounter a different reality. Users begin to notice that answers are not always correct, or nobody can clearly explain how the model arrived at a particular conclusion.

At that point, AI starts revealing issues that have often existed for years.

The same customer appears multiple times across different systems, each with a slightly different spelling. Five versions of the same document circulate within SharePoint without a clear owner. Marketing and Finance use different definitions of an "active customer". As long as people search manually and interpret information themselves, these inconsistencies often remain hidden. Once AI starts using this information, they immediately become visible.

We see this not only with our own clients. Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned before the end of 2025 after the proof-of-concept phase. The main reasons include poor data quality, inadequate risk management, rising costs and a lack of demonstrable business value.

Research from Harvard Business School also shows that the quality of unstructured data—such as emails, contracts and internal documentation—is a major factor in the reliability of AI-generated outputs.

AI sounds convincing, even when it's wrong

The principle of garbage in, garbage out has existed for decades. AI simply makes its consequences far more visible.

When a model generates answers based on outdated, incomplete or inconsistent information, it often does so with exactly the same level of confidence as when the information is correct.

AI doesn't reduce existing data problems. It makes them more visible, faster and more expensive.

For personal use, that usually isn't a major issue. An overly ambitious holiday itinerary is little more than a minor frustration. Within an organisation, however, the consequences are far greater. Answers must not only sound plausible—they must also be accurate, verifiable and traceable.

Generative AI and machine learning rely on the same foundation

Although AI applications serve different purposes, they all depend on one thing: trustworthy data.

Generative AI using enterprise data

Tools such as ChatGPT, Claude and Microsoft Copilot are already trained. Within organisations, they are connected to sources such as SharePoint, CRM platforms and document repositories.

The model itself does not change, but the answers do.

For example, if multiple versions of a price list exist in SharePoint, the model may confidently reference the wrong one. Likewise, if access permissions are not configured correctly, users could gain access to information they would not normally be authorised to see.

Custom machine learning

In applications such as predictive maintenance, demand forecasting and pricing models, data plays an even bigger role. Here, data is not only the input—it is also the training material.

If errors exist in the training data, those errors become embedded in the model itself. As a result, they are often harder to detect and significantly more expensive to correct.

In both scenarios, the same principle applies: the greatest limitation is not the technology itself, but the quality of the underlying data.

The data foundation ultimately determines the pace of AI

From our experience in Data Science, Data Engineering and Data Governance, we have seen this pattern for many years. Long before AI became mainstream, unclear definitions, fragmented data sources and a lack of ownership were already slowing down digital transformation.

AI simply magnifies these existing challenges.

Organisations that successfully scale AI therefore do not begin by investing in the latest model. They begin by building a solid data foundation.

1. Clear ownership of data sources

Every critical data source should have a clearly assigned owner who is responsible for its quality.

When no one owns the data, errors often remain unresolved, reducing the reliability of AI-generated outputs.

2. Understanding data origin and meaning

Data only becomes valuable when it is clear where it comes from and what it means.

This requires consistent business definitions, shared terminology and visibility into how data is created, used and exchanged across departments.

Without that context, it becomes difficult to validate, explain or trust AI-generated insights.

3. Continuous attention to data quality

Data quality is not a one-off project.

New systems, business processes and users continuously introduce changes. Organisations that actively monitor and improve data quality prevent the same issues from recurring over time.

The business value lies in trust, not in the model

Many discussions about AI focus on models, agents and the latest features.

In practice, we see that the biggest success factor lies elsewhere.

Only when employees can trust that AI-generated answers are based on accurate, consistent and traceable information does wider adoption become possible. That is when productivity gains, better decision-making and new business applications become achievable.

The key lesson is simple: AI is not a solution to existing data problems. It exposes them.

What has remained hidden for years in systems, spreadsheets and documents suddenly becomes visible in every prediction, recommendation and answer generated by an AI model.

Practical example: how an AI assistant exposed existing data issues at a grid operator

A grid operator introduced an AI assistant to help maintenance engineers access technical documentation more quickly. The initial results were promising. Employees found information faster and the pilot appeared to be a success.

However, questions emerged during the wider rollout. The AI assistant sometimes produced different answers to similar questions. Further investigation revealed that multiple versions of the same maintenance procedure were stored across different systems. At the same time, a predictive maintenance model relied on asset data that was recorded differently across various applications.

The AI assistant was not the problem. It simply used the information that was available. What became visible were inconsistencies in documentation, business definitions and data sources that had existed for years. Only after assigning ownership, harmonising definitions and continuously monitoring data quality could the organisation scale the solution with confidence.

The most important lesson? An AI assistant does not solve existing data problems. It reveals them the moment you start relying on its answers.

Build AI that is both valuable and trustworthy

The organisations bringing AI projects into production today are not necessarily those using the most advanced models. They are the organisations with a reliable data foundation.

Many organisations begin their AI journey by focusing on technology. Our experience shows that lasting results are achieved when technology, data and governance evolve together.

Are you wondering how mature your data foundation is for AI? Or would you like to better understand the risks and opportunities within your organisation? We'd be happy to discuss the steps needed to make AI not only work, but work reliably and at scale.

Receive data insights, use cases and behind-the-scenes peeks once a month?

Sign up for our email list and stay 'up to data':