AI’s real bottleneck isn’t models — it’s the data fabric underneath

By the end of 2025, half of all companies had AI running in at least three business functions — finance, supply chains, HR, customer ops. Copilots, agents, predictive systems, you name it. The rollout has been fast, and honestly, faster than I expected. But now that AI is embedded in real workflows, the pattern is clear: the bottleneck isn’t model performance or compute. It’s the data.

Specifically, it’s the lack of business context attached to that data. AI can crunch numbers and spit out answers at lightning speed, but without understanding why something matters — which customers are strategic, what tradeoffs are acceptable during a shortage — it makes the wrong call. Irfan Khan, president and chief product officer of SAP Data & Analytics, put it bluntly: “Speed without judgment doesn’t help. It can actually hurt us.”

That’s the core problem. Traditional data strategies have been about aggregation — extract from operational systems, dump into a warehouse or lake, run dashboards. That worked fine when humans were in the loop to fill in the missing context. But AI doesn’t just display data; it acts on it. If the system doesn’t know which customer is a strategic account or what contractual obligations apply, it’ll optimize for the wrong outcome. Technically correct, operationally flawed.

The context premium

Consider two companies using AI to manage supply-chain disruptions. Both have access to the same raw signals — inventory levels, lead times, supplier scores. One adds context: business policies, metadata, which customers are strategic accounts, what tradeoffs are acceptable. The other doesn’t. Both systems analyze data quickly. Only one moves in the right direction. Khan calls this the “context premium,” and it’s a good way to frame the advantage of a well-designed data foundation.

Most companies know they’re not ready. Only one in five consider their data approach highly mature. Only 9% feel fully prepared to integrate and interoperate their data systems. Those numbers are sobering, but they also explain why the conversation is shifting from “how do we build better models” to “how do we build a data fabric that preserves business semantics.”

Don’t consolidate, integrate

The emerging answer is a data fabric — an abstraction layer that spans infrastructure, architecture, and logical organization. For agentic AI, the fabric becomes the primary interface. Agents interact with business knowledge rather than raw storage systems. Knowledge graphs play a central role, letting agents query enterprise data using natural language and business logic.

The value comes from three components working together: intelligent compute for speed, a knowledge pool for business context, and agents for autonomous action grounded in that understanding. It’s not about moving everything into one place. It’s about connecting information across applications, clouds, and operational systems while preserving the semantics that describe how the business works.

This is a real shift from the old data strategy playbook. For years, the mantra was “consolidate everything into a single source of truth.” That approach stripped away context in the name of simplicity. Now we’re realizing that context is the whole point. Without it, AI is just a fast engine for bad decisions.

I’ve seen this play out in enough enterprise projects to know that the data fabric approach isn’t a silver bullet. It requires investment in metadata management, governance, and the kind of cross-functional collaboration that most organizations find painful. But the alternative — letting AI run on context-free data — is worse. Speed without judgment doesn’t just fail to deliver ROI. It actively undermines it.