The 73% problem: Why AI investments fail without data leadership

The technology industry has a running joke these days: LLM stands for "lose lots of money" rather than large language model. Our research suggests there's substance behind the sarcasm. Organizations are building out AI capabilities and burning through tokens without the underlying data foundation to make it worthwhile.

Our latest Cloud Insights Report quantifies the problem: 73% of organizations are investing significantly in generative AI-enabled data management. An equal percentage of business executives believe that a failure to adopt agentic AI in the next 12 months will harm business. Yet many lack the foundational capabilities and infrastructure to see real returns.

Organizations are applying technology before fixing the data and wondering why they're getting disappointing results.

 This is a leadership problem that requires organizations to think differently about data itself.

The distinction that changes everything

I make a clear distinction between data leadership and data management because they are fundamentally different. Historically, organizations have conflated the two.

Data management is operational. It's about how data is processed through various systems, the mechanics of storage and retrieval, and the day-to-day operations of keeping the data flowing.

Data leadership is strategic. It's about establishing what you're going to capture, why you're going to capture it, how you're going to use it, how it should be structured, and how you're going to secure it. Data leadership defines your master data management strategy and determines which elements need to be enriched, why and how.

This distinction becomes critical when you consider emerging technologies, especially AI adoption. How well your data is organized and structured in a manner that allows it to be dynamic, enriched, and comprehensive is amplified by AI.

The speed and scale problem

 And that's exactly the benefit as well as the cost: AI amplifies everything. Good data architecture delivers better results at unprecedented speed. Conversely, poor data architecture delivers bad results at unprecedented speed and scale.

Consider a simple example from our own operations. We implemented a tool to generate RFP responses. It can produce 75% of an RFP response to shorten our cycle time. But it requires a centralized repository of data that aligns with our solutions, understands the questions, and responds with standardized data. 

A human comes in at the back end to verify: Did we get this right? But here's the crucial part: that feedback needs to go back into the tool. If the question requires a different output, you need to update the data on a dynamic basis. Then the next time you run it, the answer is better from the start. 

Without this feedback loop, without the underlying data architecture, you're getting bad results at scale and speed, potentially taking actions that no one has vetted. When you consider agentic AI making autonomous decisions, the stakes get even higher.

The sovereignty maze

Data leadership today also means working through an increasingly complex web of data sovereignty requirements. These go beyond regional differences. They’re country-specific, and in some cases, even state-specific within the US.

The way we think about this challenge is no different from how we approach security: build sovereignty into the design of technology solutions, just like you build it into security. If you can build enough into your standard design to meet the highest criteria, then you don't have to rework multiple systems for different clients or regions.

Yes, it's more costly to maintain the highest level of compliance across the board, but it's not as costly as having seven different models embedded in one client relationship or one region. This design-first approach becomes essential as sovereignty requirements continue to become more stringent and locality-specific.

We're also seeing increased discussion around repatriation — controlling where data sits, whether in local privatized cloud, public cloud, or on-premises environments. This is really about solving data sovereignty issues while maintaining the computing capabilities organizations need. 

This connects to a broader principle our research identified: reducing complexity by rebalancing systems and consolidating cloud platforms. When you optimize your infrastructure this way, you free up budget for innovation while creating the flexible foundation AI requires. The most effective approach uses hybrid cloud management to dynamically place workloads where they perform best, whether that's for compliance, cost, or performance reasons.

Bridging the business-IT divide

Our research reveals a significant disconnect between business executives and IT executives, particularly around agentic AI readiness. Business executives are reading and hearing about AI’s potential for savings or significant business growth. They're change agents and early adopters looking for the next thing to drive the business forward.

IT executives face the implementation reality: incredibly complex, non-interoperable environments with decentralized systems, different third-party providers, and inconsistent data structures.

It's an aspirational vision versus pragmatic implementation reality. 

Data leadership bridges this gap by starting with the business outcome and working backward. You have to start with data leadership and data architecture before you even think about what technology you want to apply. The implementation of technology is actually easy, but fixing the underlying data so that the technology drives the right results and return is the heavy lift.

The Innovation Leaders advantage

 Our research identified a group we call "Innovation Leaders," organizations that are getting data leadership right. What sets them apart? They prioritize industry cloud platforms and data capabilities to free up budget for innovation. They build data architectures designed for reuse and scalability. These organizations solve specific data problems while creating foundations that support future challenges.

 Start with the end in mind

 My advice for actionable steps leaders can take to build data leadership capabilities in the next 12 months is simple: stop, breathe, and think.

• Keep the end in mind. Look at the business outcome you're trying to achieve before you do anything from a data perspective. Build out your ecosystem, understanding how you'll accomplish the goal and how the data structure needs to be orchestrated.
•  Start with fundamental questions: Do I have the data? Is it clean and standardized? How dynamic is its input and feedback loop?

 Don't try to solve everything all at once. Solve a specific data problem, but build your model to be reusable for the next challenge.

Looking ahead

Data leadership will only become more critical as organizations advance their AI adoption. The fundamentals – clean data, proper architecture, security, and governance – remain the same. But the application changes dramatically when you're dealing with unstructured data, autonomous decision-making, and real-time processing at scale.

Organizations that get the data foundation right will capitalize on AI's potential. Those that don't will continue to lose lots of money on their LLMs.

The choice is clear: invest in data leadership now, or keep putting the cart before the horse and wondering why AI isn't delivering the results you expected.

Explore how Unisys can help elevate your data strategy or provide insights to help modernize cloud and IT for what comes next by downloading the Unisys Cloud Insights Report 2025.