How AI Raises the Standard for Data

For most of the history of enterprise data, data informed a decision. A person still made it.

Now AI queries the data, interprets what it finds, decides what should happen next, and acts across enterprise systems. That raises the standard for the data underneath it. A questionable number in a dashboard can be investigated. A questionable number used by an agent can become an action before anyone realizes there was a problem.

That growing tension between what AI is being asked to do and the data supporting it is showing up clearly in our latest research. We are currently gathering responses for our third annual Modern Data Survey, and with more than 500 data leaders and practitioners having shared their perspectives so far, we had seen enough to release an interim view rather than wait for the final report.

We just released those findings in 5 Emerging Trends in Enterprise Data & AI. Across the five trends, one pattern stands out: AI adoption is moving faster than the data foundations required to support it. Nearly 60% of organizations are already piloting or running AI agents in their data and analytics workflows, and another 15.6% expect to begin within six months. Yet just 8.4% say the data feeding their AI systems is trustworthy enough for production.

Organizations have shown that they can put agents to work. The harder question is whether they are ready to trust what those agents will do.

The Human Buffer Is Disappearing

Data quality problems are not new. Organizations have always dealt with inconsistent definitions, stale tables, missing lineage, unclear ownership, and business rules that live in people’s heads.

For years, people compensated for those weaknesses. An analyst questioned an unusual result. A business leader recognized that a number did not look right. Someone knew that one definition of revenue applied to finance while another applied to sales.

Human judgment became an unofficial part of any interaction with data. Agents do not automatically inherit it. If the system does not make a definition, exception, or policy explicit, the agent has to infer it or proceed without it. The data does not need to become worse for the risk to increase. It simply has a more direct path to action.

That is why trust is no longer a general aspiration for the data team. It is becoming a production requirement for AI.

The Data Is the First Constraint

When we asked what is preventing agents from reaching production, the answers were clear. Data quality and trust ranked among the top three barriers for 76% of respondents. Missing context and lineage followed at 64%, ahead of security concerns. The explanations the market often reaches for, including immature tools and a shortage of skills, ranked much lower.

This matters because much of the AI conversation still centers on models, tools, and talent. Those investments may improve what an agent can do, but they cannot tell the agent whether a customer record is current, which definition of margin is authoritative, or whether a particular action is allowed.

The model can only work with the understanding the organization has made available to it.

Context Has to Be Engineered

Business context is the widest gap in the findings so far. Most organizations understand its importance. Nearly 61% say a reliable context layer is necessary for AI agents. But only 16% deliberately design and engineer that layer as a product. One in four has no formal context layer at all.

By context, I mean the definitions, relationships, lineage, policies, and quality signals that explain what enterprise data means and how it can be used. Much of this knowledge already exists, but it is usually scattered across catalogs, documentation, semantic models, governance tools, dashboards, and conversations.

AI cannot reliably work from context that remains scattered around the organization. That understanding has to travel with the data.

The organizations furthest along with agents appear to recognize this. Those running agents in production are nearly four times as likely to have intentionally developed a context layer than organizations that are interested in agents but have not started using them. Organizations with an engineered context layer are also roughly five times as likely to say they can connect their data work to measurable business outcomes.

The survey shows correlation, not causation. But the pattern is difficult to ignore. The organizations making more progress with AI are also doing more deliberate work on the meaning and reliability of their data.

Governance Has to Reach the Action

Governance also changes when AI moves from producing an answer to taking an action. It is no longer enough to know where the data came from and who can access it. Organizations must also know what information an agent can use, what decisions it can make, what systems it can change, and who is accountable for the outcome.

Our findings show a significant gap between the governance organizations say they need and the mechanisms they have in place. While 65% believe AI-enabled decisions must be explainable, traceable, and defensible, only 10% maintain both an audit trail of AI inputs and outputs and a link from decisions back to the underlying data. Just about 18% have a clear, documented AI accountability framework.

That leaves a basic operational question unresolved: if an agent takes the wrong action, can the organization explain what happened and determine who was responsible?

As the range of actions available to AI expands, governance has to expand with it.

Data Foundation Is Now Part of the AI System

What stands out to me is not that enterprises have unresolved data problems. We have known that for years. What has changed is the distance between those problems and their consequences.

An agent can be launched in weeks. Trusted data, reusable business context, and enforceable governance take longer to build. That difference in pace is creating the gap we see in the research.

The answer is not to slow every AI initiative until the entire data estate is perfect. It is to stop treating the data foundation as a separate modernization effort that can be addressed later.

For every agent moving toward production, organizations should be asking:  
What data will it use?  
Does that data carry the context needed to interpret it?  
What decisions and actions are permitted?  
Can those actions be traced back to their source?  
And who is accountable for the result?

Those questions are no longer adjacent to AI strategy. They are the work required to make AI operational.

Until next time,

Saurabh