Data Foundation for AI: An Executive Checklist Before You Invest

Get a diagnostic of your data foundation and discover the next steps to take.

The promise of AI in question: from experimentation to the scale challenge

Recently, we have seen growing interest and maturity in AI adoption, which has shifted from experimentation to the core of business. Companies are no longer just testing AI in isolated pilots, but are looking to embed the technology in the core of their operations and deliver real impact.

However, even with successful cases, it is clear that many AI initiatives are still stuck in the experimentation phase. According to Gartner, worldwide investment in AI is expected to grow 44% in 2026, but a Deloitte study found that only a quarter of IT leaders have managed to move their pilot projects into production.

If AI is delivering value across different parts of the business, why can't we see the impact at scale and in a consistent way for organizations?

The answer is not simple, and it points to a challenge greater than technology alone. Beyond the choice of tools and models, the real obstacle is often the lack of a data foundation prepared to support AI.

Building a data foundation for AI in practice

What is the level of readiness of your data?

To take the first step, it is essential to understand where your organization actually stands. Thinking about it is not enough. Start with a structured diagnostic.

The questions below help you do this in a guided way. The idea is to capture, honestly, the maturity of your company with respect to its data and identify where conversations need to be prioritized internally.

Answer the following questions based on your own experience:

Access, consistency and context

How in practice do you access relevant data?

1. Every request becomes a negotiation, and approvals are so slow, unpredictable, and frequently tied to bureaucracy that they often delay everything.
2. There are rules, but the application is bureaucratic, unpredictable, and sometimes creates delays.
3. The flow works, but there are still many manual steps.
4. Access is predictable and governed. Permissions and restrictions are clear and do not hold up the project.

Beyond raw data, how much business context exists?

1. Almost none. It is hard to understand what the data means, its limitations, and how it is used.
2. Some sources exist, but they are incomplete or poorly documented.
3. There are context sources for the main datasets, but they are not always up to date.
4. The standard is clear: every main dataset has description, classification, rules, limitations, and examples of use.

Are only structured data considered, or also content such as documents, emails, and transcripts, which are usually governed assets?

1. The majority of the organization is structured. Unstructured content remains dispersed and lacks governance.
2. There are repositories for this content, but without the same catalog, classification, and control applied to structured data.
3. Part of the unstructured content is organized and governed, but only partially.
4. Unstructured content receives the same rigor: cataloging, classification, owner, and usage rules.

Trust, quality and traceability

When leadership looks at a piece of information, do they trust what they see or question it?

1. They question it. Numbers diverge, and it is not clear where they come from or whether the update is current.
2. It depends on the subject. Some numbers are reliable, others need validation every time.
3. In general, they trust it, but there are still gray areas that delay decisions.
4. Trust. Definitions and sources are stable, and trust is high.

How often does poor data quality hinder the progress of AI initiatives?

1. Frequently. Inconsistency, lack of data, duplication, and outdated data make progress unpredictable.
2. It happens regularly and requires manual validation that consumes too much time.
3. It happens occasionally. There are mechanisms that detect and contain problems.
4. Rarely. Quality is continuously monitored and treated as a priority.

When an inconsistency is found, is it possible to trace its origin?

1. It is difficult. Finding the origin can take a long time, and it often happens after transformations have already been applied.
2. Yes, but it requires a lot of effort and investigation.
3. Only for critical data, not for everything.
4. Yes, with ease. Origin, transformations, and dependencies are visible and auditable.

Governance, responsibility and reuse

Who is responsible for the data?

1. It is not clear. Responsibility is diluted, and problems keep coming back.
2. It is centralized in the IT area, with little involvement from business areas in definitions.
3. There are responsible people in some domains, but practice is inconsistent.
4. Responsibilities are defined for each domain, with defined roles and management routines.

When different areas look at the same metric, what happens?

1. The numbers diverge, and the discussion turns into "whose number is right?"
2. There are definitions, but they are not always followed consistently.
3. Critical metrics are consistent, but there are still exceptions and workarounds.
4. Definitions are stable and shared, with a single source of truth.

How much of what the company builds in data for one initiative can be reused in another?

1. Nothing. Each initiative needs to be rebuilt from scratch.
2. There are attempts, but it is difficult to discover what exists, so reuse becomes an exception.
3. Reuse happens in part, but there are still many duplicated efforts.
4. Reuse is intentional, with clear definitions to support this practice, including standards, components, documentation, and governance.

And on AI?

How does the company handle sensitive data in the context of AI?

1. It is not a case. The risk of inadequate use is unknown and creates constant obstacles.
2. There are policies, but their application is inconsistent and poorly audited.
3. Governance exists, but there are still gaps.
4. Governance is automated and systematic: classification, control, and auditing are part of the flow.

Was the data foundation built with human consumption in mind, or was it also prepared to be consumed by AI systems?

1. It is structured for consumption by humans. AI initiatives have to adapt everything, with mixed results.
2. It covers the needs of humans and AI initiatives that have already been reworked to be used by AI, with some adaptations.
3. In some domains, the foundation is considered a consumer by AI (formats, metadata, permissions, and rare cases), but it is not yet general.
4. The foundation is designed for consumption by both humans and AI: context, permissions, and traceability follow the same standards for both, and AI can be used in a fully reliable way.

What might stop the scaling of AI even when the pilot works?

1. The data does not exist in the same way across different sources and business units, is incomplete, or is unreliable.
2. Data quality is limited, and the context of each data source varies, so reuse depends on personal effort.
3. Governance does not keep up with expansion, becoming a bottleneck or even being ignored in some cases.
4. The data base supports expansion and adaptation in a predictable way.

Is your data driving or sabotaging your AI strategy?

AI is only as intelligent and efficient as the data it consumes.

Building a data foundation for AI is no small challenge for companies racing to adopt the technology. According to Gartner, 60% of AI projects will be abandoned due to a lack of AI-ready data. In addition, only 32% of organizations are able to address the issue.

To close the diagnostic, consider which level best describes your organization at each point. The answer usually comes from the predominant level, not from the best practice you have in one area.

Level A: Starting stage

The base is still being built. The data is difficult to access and standardize, and the trust of the business areas in what the data says is low. Quality and business context are issues that delay every initiative, and the lack of governance makes every AI project more expensive and risky. In this case, the highest priority is to build the foundation before scaling anything.

Level B: Emerging stage

There are rules, processes, and some structures, but applying them is inconsistent across areas and domains. Quality and the context of data vary, and reuse happens more by personal effort than by design. Some initiatives advance, but each one still needs substantial manual effort to be built.

Level C: Structured stage

The foundation is already predictable enough to support a good share of the company's cases. Access, quality, and responsibilities are clearly defined in most domains, and the main metrics are consistent. Some exceptions and gray areas still exist, but they are not an obstacle to the majority of initiatives.

Level D: Advanced stage

The foundation was designed to serve both human and AI consumption. Data, structured and unstructured, is treated with the same rigor, including context, permissions, traceability, and controls. Trust is high, reuse is intentional, and the foundation can be expanded or adapted with predictability.

Next steps to build a data foundation for AI

If your answers point to the first levels, the goal now is to turn these findings into concrete priorities.

A good starting point is to gather the leaders who answered these questions and compare the results. Where the answers diverge, there is usually an important conversation to have about the business and about the barriers that repeat themselves. From there, choose the fronts that will attack the biggest impact first.

Building a data foundation for AI is not about accumulating large volumes of data. It is about making the data you have trustworthy, understandable, and ready to be consumed by AI. Without that, the expansion of AI remains dependent on individual effort, and its results remain unpredictable. With it, the value of AI becomes measurable, repeatable, and sustainable over time.