Across every industry, AI is now a boardroom-level priority. Organizations correctly see it as a driver of efficiency, innovation, and competitive advantage. But despite the excitement, many AI initiatives stall in practice.
In fact, a recent AI Report (2025) found that only 87% of generative AI projects reach production, highlighting the difficulty in translating ambition into results. Meanwhile, an Informatica study found that less than two-fifths of AI projects had made it to production. 67% of firms said they’d released less than half of their AI projects.
Why does this happen?
AI projects typically fail because businesses don’t focus on the right problem. Many assume that implementing AI is a purely technical problem. In fact, AI adoption is a business change management problem mixed with a technical problem. Recognizing this connection is an important first step in planning for the success of an AI project.
Luckily, this problem parallels other types of adoption in the data world, particularly data science projects.
What AI projects and data science projects have in common
Although AI adoption may feel entirely new, the connection between business change management and data technology is deeply rooted in the history of data technology itself.
The same factors that led data science projects to fail also become stumbling blocks for AI initiatives.
In this article, we’ll take a look at what those factors are. Then, we’ll dive into a roadmap you can use to avoid them by better preparing your organization to be AI-ready.
Beyond the data swamp
AI solutions differ from traditional data workloads, such as those used in analytics. We’ve discussed this in detail before. In brief, the non-deterministic nature of AI systems introduces additional complexity in areas such as data pipeline management, model selection, testing, and governance.
At its root, however, the same problems that can delay data projects also threaten AI projects.
Data access
AI applications work best with access to large volumes of high-quality data. However, accessing this data is often not straightforward.
Data silos keep data isolated from the rest of the company, making it difficult to find and leverage it for business value. The need for data to fuel AI solutions often leads companies to realize that they lack fundamental tools for discovering and utilizing data.
Data readiness is more than data quality
We’re accustomed to considering data readiness as synonymous with data quality, but the truth is more complicated than that. Most organizations already possess the necessary data.
The real challenge is whether it’s ready to be used, and that problem is not entirely technological. It’s also organizational. Data may exist across multiple systems, but if it takes months to track down the right sources or subject matter experts to validate the business logic, projects stall before they show value.
In practice, messy or undocumented data doesn’t always stop an AI program, but it forces teams to spend valuable time wrangling instead of building. That causes delays. To combat this, AI-powered tools can ease this burden by classifying, transforming, and documenting datasets. But they cannot replace an organizational approach to governance. The fastest path to production-grade AI is making data both findable and usable across the enterprise, so teams can focus on modeling instead of searching.
A roadmap for adopting AI in your organization
While these sound like technical challenges, in reality, they aren’t. We have the technology to solve them. However, overcoming the connected organizational challenges can prove more difficult.
Data access and data quality become issues because companies don’t have a single, unified approach to managing data. This means more than shared tools. It refers to a shared understanding of the data lifecycle, encompassing how data is collected, curated, monitored, and maintained.
Again, data science is a good parallel. In that domain, the data science lifecycle outlines how data projects interact across the technical and business domains. This same theory applies equally to AI projects as it does to data science projects. The data science lifecycle provides a roadmap for assessing how your teams approach data at each stage, from ingestion to transformation to utilization.
Here’s what to focus on in each phase of the data science lifecycle to ensure that your teams have a shared understanding of data that will provide a firm foundation for all your AI initiatives.
Business understanding
Understanding the logic behind your business problems and their root impact on your business is critical to any successful data science initiative. If you get this step wrong, everything that follows is meaningless.
Unfortunately, companies often struggle at this critical stage. That’s because data engineering teams often end up making data decisions without input from the business stakeholders who best understand the data.
Getting this stage right involves delving into both the “how” and the “why” of each business problem before technology or data are even involved. To help with this, a good practice to adopt is to engage key business stakeholders early and seek feedback often. This will help align the logic that will eventually become encoded in your data science project correctly.
And while this is true of data science, it’s equally true of AI. The same need to understand business logic first flows through both types of projects. The only difference is one of implementation and technology.
Data understanding
Business understanding needs to map directly to data understanding. The more the technologists setting up the technological side of the project understand the business problems at stake, the better.
This often goes one of two different ways.
Sometimes, this can be easy. This typically occurs when the business problem has been previously researched, and well-documented datasets already exist.
At other times, the problems are very niche or new, and they require practitioners to embark on a mission of data discovery. These missions can be daunting if datasets are not well-documented, especially in large organizations, where it’s either hard or impossible to track down the creators or owners of the dataset.
Again, although the technology differs, many of the same problems impact AI projects. Both project types rely not only on data, but also on understanding that data.
Data preparation
Data preparation comes next. This typically involves joining data from various sources. To do this, data teams leverage the knowledge your teams have accumulated in the previous two stages to begin applying transformations to the data. This stage ensures that data is available, easy to query, and of high quality.
In the past, data preparation was actually a ‘Herculean’ task. Joining data from various sources was daunting for several reasons. Data was difficult to access and typically spread across multiple systems, with no easy or cost-effective way to consolidate it into a single location. A significant amount of time was spent in the preparation stage, determining how to bring it all together in an efficient manner.
Today, luckily, data preparation is easier. Technologies like Starburst make data access nearly universal. Forced data centralization is now a thing of the past.
Data preparation sets the stage for building AI applications. It provides the foundation upon which AI app developers build their solutions. The good news is that, if you get the business and data understanding phases right, this phase flows naturally from the previous two.
Exploratory data analysis
Next comes analyzing data and deriving key features from it. Previously, this step was another major pain point. Today, this is an area where you can utilize AI successfully to discover and extract insights from your data stores.
In many ways, this phase is where the ‘fun’ begins, as data practitioners are finally in a position to start gaining insights that will help solve the business problem.
Data governance is a significant consideration at this stage, as the right data should always be accessed by the right person at the right time. Once you have unified and connected data, data scientists can explore it much faster.
Overall, this is where every data scientist wishes they could start, but are typically bogged down in prior phases, finding the right data, understanding it, and then prepping it.
Data modeling
Data modeling is the part of this process that everyone tends to focus on. Truth is, however, everything before this phase is the more time-consuming part of the process.
At this point, a proper understanding of the business problem helps ensure the most suitable model type is chosen (classification, regression, clustering).
Modeling requires a lot of domain-specific business knowledge. That’s why it’s critical to have the business stakeholders closest to a given dataset involved in the earlier phases. Upstream process errors inject misunderstandings and issues that are almost impossible to unravel at this point.
Model evaluation
Finally, after all of this, you can finally move to the last phase. Model evaluation determines whether the models you’re using produce accurate answers. For most AI applications, this will mean selecting the right foundational model or Large Language Model (LLM), as well as ensuring you’re feeding in the proper domain-specific context using techniques such as retrieval-agumented generation (RAG).
Here again, business context is critical. It helps data scientists to understand the level of tuning required as well as the success criteria for the solution. For example, in some cases, a model with 90% accuracy in a classification problem is sufficient, but in other cases, 99% accuracy is a requirement. It all depends on the business problem and what the business requires from the solution. Specifically, it depends on how it will be used, who will use it, what decisions it will be used to make, and what the stakes are associated with those decisions.
This is one area where AI agents are having a significant impact, due to their ability to recognize patterns and look past outliers.
Model deployment
Model deployment makes your AI application available to end users. The work isn’t done here, though. Deployed models require continuous monitoring for performance, accuracy, quality, and tone.
When fine-tuning on a narrow dataset, especially one focused on a highly specific subject, models can begin to drift from their original alignment. For example, a tone that was once friendly may become overly terse. In other cases, subtle issues in the fine-tuning data, such as hidden trigger words, can lead to harmful outputs like biased language or insecure code.
In the data industry, this is often known as data drift. It means that the nature of the data used to build the model changes; therefore, the model needs to be changed. Monitoring for data drift is a part of many monitoring solutions. Additionally, for AI solutions, feedback from end-users on the outputs is a key dataset that drives improvements in the model and solution.
Data gathered from monitoring a deployed model can be used to drive improvements in its next iteration. This data should feed back into another round of the data science lifecycle, with this process repeating iteratively to produce models of increasingly better quality and accuracy.
All data projects involve many of the same pitfalls
Businesses stumble with AI because they tackle the technological problems before tackling the business ones. The two are interrelated.
Solving for the business problem requires an understanding of the underlying business logic. Without it, any technological solution is unlikely to succeed. That’s as true of data science projects as it is of AI projects. Although the technology differs in meaningful ways, the methodology needed to approach the problem is similar and should be considered as such.
Companies are more likely to succeed when they flip the script. Properly understood, AI adoption is not merely a technology project, but instead a business project with technology attached. Once you break down the business barriers to data discovery and collaboration, the rest falls into place.
