Companies are rushing to put AI agents to work, but there is a rather uncomfortable problem lurking underneath all that enthusiasm: many businesses apparently do not trust the data those agents are using.
Interim findings from the third annual Modern Data Survey show that 57.3 percent of organizations are already piloting AI agents or using them in production. Another 15.6 percent expect to begin within the next six months. Just 4 percent reportedly have no plans to use them.
That sounds like enterprise AI adoption is moving along quite nicely. Unfortunately, the numbers surrounding the data behind those agents tell a very different story.
Just 8.4 percent of respondents say the data feeding their AI systems is trustworthy enough for production. Even among organizations that already have AI agents running in production, only 21.7 percent describe their data as highly trustworthy. Overall, 59 percent report low or mixed confidence in the trustworthiness of the data behind their AI systems.
In other words, companies are increasingly comfortable letting AI agents do things, but they are far less comfortable with the information those agents are using to decide what to do. That seems like a pretty serious disconnect.
The problem becomes more concerning as AI agents move beyond answering questions and generating summaries. An AI assistant that produces a questionable answer is one thing. An agent capable of interacting with enterprise systems, workflows, and data can potentially act on that questionable information.
The survey reflects that concern. Data quality and trust now rank as the number-one barrier to putting AI agents into production, with 76 percent of respondents placing the issue among their top three barriers.
Business context presents another problem. Sixty-one percent of respondents consider a reliable context layer critical or very important for AI agents, but only 16 percent say their organization deliberately designs and engineers that context layer as a product. For many companies, definitions, policies, relationships, and other institutional knowledge remain scattered across tools, documentation, and employees’ heads.
That might have been merely annoying when a human was ultimately responsible for interpreting the data. It becomes considerably more important when software is expected to understand that information and potentially act without someone reviewing every decision.
Governance does not appear to be keeping pace either. Only 18 percent of respondents have a clear, documented AI accountability framework. Meanwhile, 62 percent rank security and governance concerns among their top three barriers to getting AI agents into production.
The findings are not entirely gloomy. Organizations already running agents in production appear to be further along with the underlying infrastructure too. According to the report, they are 3.6 times as likely to have deliberately engineered their context layer and three times as likely to report being somewhat or very confident in the data behind their AI. The researchers correctly point out that this is correlation, however, and does not establish causation.
There is another caveat worth emphasizing. This is an interim report, and the third Modern Data Survey is still underway. More than 540 qualified responses had been collected from data leaders and practitioners across more than 65 countries as of August 19. The researchers also say questions have changed between survey waves, so comparisons with previous reports should be viewed as directional rather than apples-to-apples measurements.
Still, the early numbers expose an interesting contradiction in the enterprise AI boom. Businesses clearly want AI agents, and many are already experimenting with them or putting them into production. But enthusiasm for autonomous AI seems to be moving much faster than confidence in the information underneath it.
AI agents can become more capable, autonomous, and sophisticated, but none of that magically fixes bad enterprise data. If an agent is going to make decisions and take actions on a company’s behalf, trusting the data feeding it seems like a pretty important place to start.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz