Enterprise AI Depends on Data People Can Trust

In 2025, companies invested an estimated $30 billion to $40 billion in generative AI, yet MIT’s Project NANDA found that only 5 percent of the more than 300 enterprise deployments it examined were producing measurable financial returns.

I have spent more than 25 years working inside enterprise data programs. Well before generative AI entered the conversation, I saw how quickly technology lost credibility when people could not rely on the data behind it. Generative AI raises the stakes because it can turn questionable data into a confident answer.

We used to say garbage in, garbage out. Now it is garbage in, disaster out. When data looked suspect, a person could use their judgment and contact a data specialist. The specialist would conduct a root-cause analysis, identify the problem, and apply a fix. Today, agents can act on whatever data they encounter. Without knowledgeable human oversight, the result can become part of operations, be used in decisions, and scale across the organization.

The same problem can surface in a chatbot. Skipped governance can turn bad or poorly understood data into an answer that sounds authoritative, immediately and at scale.

Trust Is the Real Adoption Bottleneck

When an internal chatbot gives someone a confident but wrong answer, that person may not trust it again. In a large organization, AI will also find the data you would rather it did not find. I compare it to the dirty socks under the bed: that data may be out of sight, but AI is going to find that dark data and use it.

Context has to travel with the data and be the foundation that LLM’s draw upon. Before an AI model uses data, people need to know what it means, where it came from, who owns it and what business rules and policies surround it.

Leadership can miss this because the system rarely crashes. AI can be confidently wrong; it can suggest glue for your pizza (very old example). Consider a chatbot that gives a confident but wrong answer, and the workforce stops trusting it. Rebuilding that confidence takes much longer than establishing the right controls up front.

That erosion shows up in the data, too. Employees at more than 90 percent of the companies MIT studied keep using personal AI tools on the side, even after the official pilot their employer built failed to earn their confidence. When the approved tool has not earned their confidence, employees find another way to get the work done.

Much of this year’s AI research points to the same problem: organizations are moving ahead with AI before they have done the work to make their data trustworthy. Gartner has estimated that nearly two-thirds of organizations either lack the right data management practices for AI or aren’t sure whether they have them. Separate research from IDC found a similar gap on the trust side: 78 percent of organizations say they fully trust their AI, but only 40 percent have actually invested in the governance and explainability work that would justify that trust.

The mismatch between confidence and investment helps explain why so many AI initiatives stall before they reach production or fail to align with the intended value.

AI Governance Is Now the Competitive Advantage

For most of my career, data governance was one of the least glamorous parts of enterprise IT. Cataloging metadata, defining ownership, profiling data and managing workflows could feel like janitorial work. People often skipped it because the effort seemed greater than the value it produced.

AI has changed that. It is becoming a governance capability in areas such as designing data blueprints, creating industry-curated glossaries, and recommending and activating data policies. Even data lineage can now be inferred and streamlined faster than before, making the value of managed data pipelines and observability easier to realize.

Regulation is also making data transparency more urgent. The EU AI Act became broadly applicable on August 2, 2026, including transparency rules for certain systems and AI-generated content. The Act also establishes separate requirements for high-risk systems, including documentation and information that helps deployers interpret their outputs, although some high-risk provisions follow later timelines.

For companies, the practical lesson is straightforward: clear lineage and governance make it easier to document what data was used, where it came from, and how the system is meant to operate.

The U.S. regulatory picture remains unsettled. Without a comprehensive federal AI law, companies must navigate different state requirements, even as the White House seeks to limit states’ ability to create and enforce their own AI rules. Colorado’s AI Act, the most comprehensive state framework on the books, has already been delayed and rewritten once amid that fight, with its core obligations now pushed into 2027. For a lot of US companies, the temptation is to wait the uncertainty out.

I’d argue that’s the wrong instinct. Whichever regulatory model eventually succeeds in the US, the underlying requirement isn’t going away. Know where your data came from. Know whether anyone actually validated it before the AI used it. Developing that discipline now will help companies prepare for new regulation and turn promising AI pilots into tools the business can use with confidence.

What Separates the 5 Percent

Let me tell you about one client. Their general ledger team had to comb through 300 separate spreadsheets every single time a discrepancy showed up in the books. Three hundred. Someone finally rebuilt that entire mess as a single governed data asset, with the calculations, the reports, and the underlying data packaged together and validated up front. The next time an issue came up, the fix took a fraction of the effort it used to. Across that one process alone, it saved roughly 750 days of labor. That is millions of dollars of cost savings just from cleaning that up.

The improvement came from making the right data easy to find and reliable enough to use without second-guessing it. Collaboration and reuse made that possible. Trusted data products can be aligned with corporate standards, extended and upgraded over time, scored for trust and managed for value. This approach reduces the need to create hundreds or thousands of products tied to individual physical database tables.

Strong programs make that trust visible. In the data marketplaces I work with, assets can be classified as gold, silver or bronze based on factors such as quality, curation, usage and ownership. A business user can see how reliable an asset is before using it. That visibility helps people stop guessing and makes them more willing to reuse data that already exists.

The 95 percent figure is a warning about how casually many companies have treated the data behind generative AI. The organizations making progress are giving people a clear view of where their data came from, how it was validated, and whether it is fit for the job. After years of data management being treated as background work, AI is finally making its value visible.