Why Your AI Investments Keep Failing (And How to Fix It)

Semarchy’s Craig Gravina offers commentary on why your AI investments keep failing and how to fix it. This article originally appeared in Insight Jam, an enterprise IT community that enables human conversation on AI.

Enterprise leaders are asking the wrong questions about AI. They want to know which models to deploy and how quickly they can scale. The question they should be asking is: can AI access the same data, in the same way, as our people and systems?

Technology companies are projected to spend close to $700 billion on AI this year. Yet Semarchy’s 2026 AI report reveals a striking confidence gap at the heart of these ambitions. Ninety-nine percent of US enterprises consider themselves AI-ready, and 88% believe they’re ahead of their competitors. But when you look closer, 60% of those same organizations cite data management and governance as their number one challenge. That contradiction tells you everything you need to know about where AI investments are going wrong.

The instinct is to try to avoid these pitfalls by cleaning up the data before deploying AI. But the problem isn’t actually that your data needs cleaning. It’s that you’re not delivering your data as a product that any consumer – human, system, or AI – can trust and use.

What “Data Friction” Really Reveals

There’s a pattern I’ve seen repeat across industries. Companies invest millions in cutting-edge AI capabilities, only to see projects stall or fail entirely.

When technology innovation leads, without data strategy as an equal partner, you get impressive demos that don’t translate into sustainable business value. These initiatives tend to be driven from the top – CTOs (33 percent) and CEOs (32 percent) lead most AI initiatives, reflecting how central AI has become to business strategy. But when business confidence outpaces data infrastructure, you end up with exactly the kind of over-confidence gap our research identifies. US organizations are pushing AI into production based on ambition that isn’t yet matched by the data foundations needed to support it.

But here’s what most organizations miss: the fix isn’t to have CDOs prepare the data better for AI. When organizations talk about “friction” between data and AI, they’re describing symptoms. The root cause is that the data was never packaged to be consumed. It was stored. Maybe catalogued. Perhaps cleaned periodically. But no one thought about delivery.

Now think about what happens in a typical AI initiative. Your data science team identifies a promising use case and requests datasets. Then the questions begin. Where did this data originate? Is this version current? What transformations has it been through? Is it even complete?

Sound familiar? These are the same questions you hear from any team wanting to work with your data. The fact that AI surfaces them more acutely doesn’t mean data for AI projects needs special preparation. It just reveals that your data isn’t being delivered as a usable product – and never was.

There’s a reason for that. Data management has traditionally been all about periodic batch reporting, prioritizing control over speed. AI demands both simultaneously – but so does modern business. The fix, then, is to always deliver data in a way that all consumers – humans, systems, and AI alike – can use. It’s a fundamental shift in how organizations think about data delivery.

From Data Management to Data Products

Data cannot remain a byproduct that gets managed reactively. It needs to become a Data Product.

A Data Product is more than a clean dataset. It includes:

• Semantic models that explain what the data means – not just its schema, but how to interpret and reason about it

• Built-in governance – role-based access control (RBAC), quality rules, and lineage are intrinsic, not applied after the fact

• Clear ownership – one accountable team, not fragmented responsibility across a dozen systems

• Multiple access patterns – apps for humans, APIs for systems, and AI-native endpoints for agents and models

This thinking connects directly to DataOps principles, where governance and agility work together, not against each other. Continuous delivery ensures trusted data flows reliably. Embedded quality checks catch issues before they propagate. Clear lineage provides visibility into data origins and transformations. And federated ownership with centralized standards ensures consistency without creating bottlenecks.

The organizations that get this right don’t need special AI data preparation projects. They simply make their existing Data Products available to AI consumers through endpoints with the same RBAC and policy enforcement governing human access.

Time to value then accelerates, as teams spend less time investigating data and more time discovering insights. Stewardship overheads decrease because quality is built in, rather than inspected in. And, critically, AI goes from special case to first-class consumer – accessing the same Data Products that humans and systems already trust.

Are Your Data Products AI-ready? Three Diagnostic Questions

Before your next AI investment, answer three questions honestly:

1. Are you delivering data as a product, (not just extracting it as a dataset?
   If teams ask for data and receive files or exports, you’re still in extraction mode.

2. Does your data include semantic context (not just schema)?
   AI needs to understand meaning, not just structure. It doesn’t matter how clean your data is; if it lacks semantic models that explain relationships, rules, and interpretations, AI will struggle to reason correctly.

3. Can AI consume data through the same interfaces as humans?
   If AI needs separate pipelines or special permissions to access your data, you’ve created a governance gap. AI should go through endpoints with the same accountability as any other consumer.

If you can’t answer yes to all three, your priority isn’t cleaning data for AI. It’s evolving from managing data to delivering Data Products.

The technology is ready. The models are capable. But only organizations that deliver Data Products will scale AI successfully and responsibly. Those who keep treating data as raw material will continue the expensive cycle of failed pilots and unrealized potential.