The State of Data Observability: How Organizations Are Preparing for Agentic AI

Precisely’s Cam Ogden offers commentary on the state of data observability and how organizations are preparing for agentic AI. This article originally appeared in Insight Jam, an enterprise IT community that enables human conversation on AI.

The agentic AI revolution is beginning to take shape. AI and ML models are transitioning from being merely generative, like answering user questions or creating content, to agentic, where the AI model can make complex, autonomous decisions without additional user input. A data integrity strategy, including a robust approach to data governance, quality, and observability, is the only way agentic AI can do this accurately, consistently, and at scale.

Observability is our window into how our AI/ML models perform. Data observability provides continuous insight into the quality and reliability of data pipelines, while AI observability focuses on monitoring model health, behavior, and performance over time. Together, these concepts help teams understand how data is stored, processed, and influence model outputs – information that is essential for delivering relevant, trustworthy AI outcomes.

Until now, most observability processes have focused on structured data. But agentic AI models require richer, more contextual information to make intelligent decisions. That context often lives in unstructured data, which comes from a variety of internal and external sources, such as emails, videos, audio files, PDFs, and much more.

As the volume and variety of both structured and unstructured continue to grow, many organizations are struggling to manage, interpret, and extract meaningful insights from the information flowing through their systems. Tracking data from these disparate sources is extremely difficult without modern observability tools and processes that centralize insights and unify visibility across systems.

Precisely partnered with BARC, a leading technology analyst firm in Europe, and surveyed a qualified panel of IT, management, and other tech professionals to learn more about the current state of AI and data observability and how organizations are tackling this critical challenge. Here’s what we discovered.

Organizations are Building a Solid Foundation, but There is Room to Grow

Many organizations are making progress in observing data, pipelines, and models to support AI and ML initiatives. Over two-thirds have formalized, implemented, or optimized observability programs for each of these disciplines, and a similar amount (68 percent) rely on quantitative and/or qualitative metrics to measure the impact of those programs. Modern analytics tools are starting to take hold, too, with one-third of respondents using tools like predictive machine learning and real-time analytics to gather necessary observability data. We’re also starting to see organizational buy-in, with nearly 50 percent of business process owners overseeing data quality initiatives.

While this is a fantastic start, there’s still significant room for improvement. Right now, the number one obstacle for observability is training and skills gaps, with over half of all respondents citing this as a primary concern.

Key takeaway: Close the skills gap and implement comprehensive data observability training programs for IT professionals and key stakeholders. This process will help to solidify observability as a crucial component for improving data governance and quality at a foundational level, informing future decisions regarding your agentic AI models.

A rising demand for unstructured data requires a renewed focus on observability

Only 59 percent of respondents trust the inputs and outputs of the AI/ML models they rely on. Training teams to write more effective prompts can help, but that alone isn’t enough. To improve model performance – particularly for agentic use cases – organizations need to integrate unstructured data sources that offer additional context.

62 percent of organizations are exploring semi-structured data, and 28 percent are already using it. Meanwhile, 60 percent are evaluating unstructured documents. This trend underscores the growing importance of observability across diverse data types.

40 percent say that observing and governing unstructured data is now vital to their workflows – suggesting a growing gap between those with robust observability and those at risk of blind spots.

Key takeaway: Unstructured data is becoming increasingly important for improving both generative and agentic AI capabilities. Investing in metadata management and quality metrics will help improve visibility and trust in how this data is used.

Organizations Rely on Legacy Solutions Rather than Dedicated Observability Tools

Many organizations rely on a combination of tools and technologies to provide insight into the disparate elements of their AI/ML infrastructure. Currently, 69 percent of respondents use their data warehouse or lakehouse tools, 67 percent use a business intelligence or analytics tool, and 45 percent rely on data integration tools.

Meanwhile, only 8 percent of respondents report using a dedicated observability tool to oversee operations. While these legacy systems offer limited visibility, they often fall short of delivering a full picture.

For example, a data warehouse may show you the health of your stored data, but not how that data flows through the pipeline or influences model performance. In contrast, dedicated data observability solutions provide full-lifecycle monitoring, anomaly detection, and drift alerts – capabilities that will become increasingly vital as models grow more complex and autonomous.

Key takeaway: Shift reliance from basic data-gathering and monitoring tools to dedicated AI observability solutions and integrate them into your AI governance strategy. You’ll get deeper and more comprehensive insight into what your data is doing, leading to more informed decisions about how to improve the health and performance of your AI/ML models.

By taking a proactive, holistic approach to observability, organizations can lay the groundwork for building high-integrity, secure, and reliable AI/ML models. That foundation will be essential as agentic AI moves from possibility to business-critical reality.