In the first part of our series, we explored why IT modernization is essential to AI readiness. This article takes the discussion a step further and looks at another obstacle that often prevents promising machine learning prototypes from making it into production: the data platform behind them.
A model may perform well in a controlled environment, yet struggle when faced with the demands of daily operations. That’s because successful AI deployment depends less on the model itself than on the platform supporting it. Here’s what matters.
The data platform can make or break an AI project. Getting AI into production takes more than choosing the right model. It requires a data foundation that scales with operational demands, supports reuse across applications, and provides full visibility into where data comes from.
True data sovereignty requires both flexibility and governance. In daily operations, where your data is hosted matters less than whether you can move it, control access, and track how it is used. These capabilities allow companies to scale without unnecessary vendor lock-in while retaining long-term control over their data.
Gartner predicts that more than 40% of agentic AI projects could be canceled by the end of 2027. To make AI agents sustainable across an organization, companies need a central AI Operating System built on a shared data foundation. This prevents each department from having to build every new AI application from scratch.
KUKA and a German machine manufacturer with 2,500 connected pumps in the field showcase how a well-designed industrial data platform enables AI applications to run reliably in production and scale across operations.
It’s a familiar situation in German industrial companies. According to Bitkom’s Artificial Intelligence 2025 study, 53% of companies cite a lack of technical expertise as a barrier to AI adoption, while 36% point to insufficient quality of results.
These findings reflect what we see in practice. The real challenge is not building the first proof of concept. It’s turning a promising pilot into a reliable production system.
Interestingly, poor data quality is not always the main culprit. Often, the underlying data infrastructure simply wasn’t designed to handle the demands of live operations.
A pilot typically runs on a manageable, carefully curated dataset. Production is a different story. Data streams in continuously from machines, plants, and sensors, often at different frequencies and with varying levels of quality.
Unclear responsibilities add another layer of complexity. Who owns the data? The business department, IT, or the data science team? Without a clear answer, coordination becomes difficult at every interface.
Then there’s the problem of isolated solutions. An AI application that works well at one plant may be difficult to roll out at another because the underlying data platform was never designed for reuse.
How can companies tell whether their existing data platform is ready to support AI in production, potentially across multiple sites? Four warning signs are particularly revealing:
If any of these issues sound familiar, it’s worth taking a closer look at your data platform before investing in additional AI use cases.
Sound familiar? If you’ve spotted any of these warning signs, our Executive Guide is a great place to start. Discover how Data, Operations, and IT teams can get their data platforms ready for AI, with real-world success stories and practical migration examples to guide you.
Data sovereignty is a recurring topic in discussions about data platforms. It’s often viewed through a political or regulatory lens, particularly when it comes to cloud provider selection or the EU Data Act. But when designing a data platform, sovereignty has a very practical meaning. First, there’s data portability. Can the company move its data to another cloud or platform provider without major obstacles? Equally important is the ability to control who can access which data, scale the platform without becoming dependent on a single vendor, and maintain a complete record of who has accessed or modified information.
In daily operations, these capabilities often matter far more than the physical location of cloud infrastructure. They determine how much control a company truly has over its data and AI solutions.
Few topics generate as much debate in digital transformation projects as platform selection. Data lakehouse, Microsoft Fabric, Databricks: the number of technologies and competing claims seems to grow faster than most organizations can meaningfully evaluate them. But choosing a technology shouldn’t be the starting point.
Before committing to a platform, companies need to answer two questions:
Governance should take priority over vendor selection. Who can access which data? How will quality be maintained? How can data and its transformations be traced? For a closer look at different lakehouse approaches, see our article on data lakehouse architectures.
One factor is particularly easy to underestimate when building AI solutions for industry: industrial data is fundamentally different from conventional business data. It includes time series and sensor readings, often with timestamps that aren’t synchronized across machines and systems. Images and text from maintenance records or quality inspections add further complexity. Making sense of this variety requires consistent semantics: a shared understanding of what data means and how it should be interpreted and transformed. Without that common foundation, even the most modern platform becomes a patchwork of disconnected solutions, leading to costly rework down the road.
For a data platform to perform reliably beyond the pilot stage, three governance capabilities are essential:
These aren’t measures you can implement once and forget. They require ongoing attention and improvement, much like the industrial equipment they ultimately support.
What does successful AI deployment look like when it’s built on a solid data platform? Two examples illustrate the possibilities.
KUKA uses its iiQoT Industrial IoT platform, built on Microsoft Azure, to detect anomalies across robot fleets. Today, around 100,000 machine learning models run in production on the platform. Managing that scale would be virtually impossible without a well-organized, scalable data foundation.
A leading German machine manufacturer needed a new platform for its 2,500 connected pumps in the field after Weidmüller discontinued its existing solution. Device Insight developed a platform designed from the outset to support future applications in agentic AI, machine learning, and service operations.
Both examples point to a useful pattern. Anomaly detection is often an excellent starting point for companies taking their first steps with AI. It makes use of sensor data that’s already available, can deliver measurable results quickly, and builds confidence for more advanced applications. For a closer look at how to measure the business value of these applications, read our article on Return on Data.
Looking ahead, there’s a compelling reason to make the right architectural decisions today. According to Gartner forecasts, around 40% of agentic AI projects could be canceled by the end of 2027 due to rising costs, unclear business value, and inadequate risk controls.
That isn’t an argument against AI agents. It’s a warning about what happens when companies introduce them without a common foundation. Without shared governance, organizations risk creating a growing collection of incompatible AI agents. Each department develops its own solution, repeating the same mistakes that have held back so many early AI pilots.
What companies need instead is a unified AI Operating System, built on a shared data platform that serves every business function. Rather than starting from scratch with each new application, teams can build on existing data connections, capabilities, and governance structures.
For companies planning their next AI initiative, three principles are worth keeping in mind:
How ready is your data platform for AI? Talk to our experts to find out how AI-ready your data platform really is.
A structured readiness assessment gives you a clear picture of where you stand and helps identify the most valuable next steps. Find out in just 5 minutes with our self-assessment.
About the author:
Ingo Forstner combines a background in mechanical engineering with extensive experience in portfolio and operations management, specializing in industrial AI projects. He is currently Senior Portfolio Manager at Device Insight, a Munich-based company within the KUKA Group. Previously, he held positions at Linde Group and MAN Diesel & Turbo.
A model may perform well on clean, controlled data during development, only to encounter infrastructure that cannot reliably handle continuous, changing data streams in production. The real bottleneck is often the data platform rather than the model itself.
Every new use case starts from scratch, the platform struggles with growing data volumes, data lineage is unclear, or missing values are interpreted differently across teams. Even one of these issues is a good reason to review your existing platform.
Neither is inherently better. The right choice depends on the problem you’re solving and who will use the platform. Are you primarily supporting data engineers who work with complex technical architectures, or business users who need straightforward access to information? Those requirements should guide your decision.
In day-to-day operations, data sovereignty is less about choosing a particular cloud provider and more about data portability, granular access controls, and comprehensive audit trails. These capabilities give companies greater independence and control over their data.
Anomaly detection is often a strong first choice. It builds on existing sensor data and can deliver measurable results relatively quickly, creating confidence for more complex applications.
Because the second and third applications can reuse the infrastructure established for the first, rather than starting from scratch. This reduces costs, shortens development time, and lowers the risks associated with scaling AI across an organization.
