The Hidden Flaw Holding Back Enterprise AI: It All Comes Down to "Context"

Enterprises run a wide range of AI tools in parallel, from customer service AI bots and conversational assistants to process automation platforms, each seeming quite intelligent on its own. But once fully launched in production, their outputs can be erroneous, inconsistent, or completely unaligned with actual business operations. A collaborative report from MIT Technology Review Insights and Atlan quantifies this reality: all surveyed organizations are using AI to some extent, yet only 16% have delivered transformative value from their AI deployments.

The report's dataset paints an even more detailed picture. 40% of respondents said they only see moderate gains limited to specific tasks or teams, 21% reported meaningful returns, and 23% saw little to no ROI at all. Most companies are still stuck at the stage of partial efficiency improvement, far from reaching that "transformative value" threshold.

Real-world deployment patterns also confirm this gap. 38% of respondents are still running AI use cases in pilot phases, 35% use tools like copilots and chat assistants to support their employees' work. Only 12% of enterprises have deployed AI agents to production environments, and just 3% said AI handles most of the workload for critical business processes. The number of AI tools keeps growing, but full process transformation has not kept pace.

Prukalpa Sankar, co-founder of Atlan, puts the issue very directly: there is a massive gap between the raw capability of AI models, and how practically useful those models end up being within an organization. She makes an even sharper observation: for most real-world use cases, general artificial intelligence is already here — it just isn't very useful at the vast majority of companies. She uses a vivid intern analogy: skipping the work to build a dedicated context layer for AI is roughly equivalent to hiring the smartest intern on the planet, blindfolding them, and dropping them straight into your office with no orientation.

The report names this missing piece the context layer: a piece of architectural infrastructure that sits between business systems and AI platforms, that translates all distributed organizational intelligence across the enterprise into governed, machine-readable context that AI agents and models can use, and that allows business knowledge to compound in value over time. Put in simpler terms: AI has to understand business-specific terminology, metrics, data sources, permission boundaries, and the relationships between all these elements before it can take reliable, appropriate actions. Many enterprises buy a whole lineup of seemingly smart standalone tools, only to discover each works well on its own, but none understand their actual business when combined.

This also explains a very common discrepancy: AI that looks dazzling in demos frequently makes mistakes once it is integrated into an organization's workflows. Raw model capability is a generic commodity, but enterprise context is a private, proprietary asset. The context layer may sound like an extension of data governance, but in practice it operates more like organizational engineering. It won't produce the flashy, viral demo moments, but it is the deciding factor in whether AI can progress from supporting employees to owning and running end-to-end business processes.

This report was produced by MIT Technology Review Insights as custom content, separate from the editorial team's independent work; all research and writing was completed by humans, with AI tools only used in production workflows under human supervision. The full report is available for download here: MIT Technology Review Insights x Atlan Joint Report; Original article: Creating AI value with enterprise context.