Why Intelligent AI Agents Fail and How to Close the Data Foundation Gap

Most enterprise artificial intelligence (AI) agent pilots fail, not because of the model, but because of the data substrate beneath them. Schemas drift, identity is fragmented across software as a service (SaaS) silos, unstructured content remains unindexed, and access policies don’t propagate to the runtime (Figure 1).

Figure 1: The data foundation gap illustrates why data availability alone does not ensure AI agent readiness. Reliable agent performance depends on harmonized data, governance, and runtime controls. Image courtesy of Sushree Mishra.

By the time a team wires an agent to production data, it has inherited every shortcut from the last decade of data warehousing. Drawing on patterns observed across enterprise architectures, the AI-Readiness Quadrant, a diagnostic framework spanning four foundational dimensions, includes data harmonization and identity, semantic grounding for unstructured assets, runtime access enforcement, and lineage and observability. Organizations that address and strengthen these four dimensions can reduce agent failure rates, accelerate deployment readiness, and build AI systems that scale with greater reliability and trust.

Reacting to Significant Structural Shifts

In the last decade, three structural shifts in enterprise big data environments have widened the gap between data availability and AI readiness in many organizations. First, enterprises consolidated onto cloud data warehouses and lakehouses optimized for business intelligence (BI) workloads (human consumers reading dashboards), rather than for autonomous consumers who act on the data.

Second, SaaS sprawl fragmented customer, product, and account identity across dozens of systems; entity resolution was rarely treated as foundational infrastructure. Third, the rise of large language models (LLMs) inverted the value calculus of unstructured data, such as call recordings, contracts, and tickets, but governance, indexing, and runtime access for that material lagged years behind. The result is that enterprises have unprecedented data availability, but the data is not in the shape, identity, or governance posture that intelligent agents require to act reliably.

Assessing Big Data Architecture

The AI-Readiness Quadrant is a diagnostic framework that helps organizations assess the strengths and weaknesses of their big data architecture (Figure 2). It maps an enterprise’s data foundation against four dimensions that determine whether AI agents can act on that data reliably and at scale. For example:

• Data harmonization and identity. In an organization, can AI agents reliably resolve the same entity, including customer, account, product, and asset, across all systems of record?
• Semantic grounding for unstructured assets. Are unstructured content stores indexed, versioned, and retrievable in a way agents can use as grounding material?
• Runtime access enforcement. Do access and governance policies propagate to the agent’s decision-time context, or only to query-time pipelines?
• Lineage and observability. Is it possible to audit what data an agent used, where it came from, and how fresh or trustworthy it was?

Figure 2: The AI-Readiness Quadrant provides a framework for evaluating the four foundational dimensions required for reliable and scalable enterprise AI agents. Image courtesy of Sushree Mishra.

The framework’s diagnostic value is that each dimension is rated independently. Most enterprises are strong in one or two dimensions and weak in the others; the weakest dimension determines the ceiling of agent reliability. Organizations cannot compensate for weak identity resolution with a better model.

Spotting Warning Signs

Fragmented data sources, inconsistent governance, and poor data quality contribute to AI and agent failures in production environments. Each maps to a specific failure pattern in production. Fragmented sources without identity resolution cause agents to act on partial or duplicated customer context, such as recommending a product the customer already canceled or surfacing the wrong account.

Inconsistent governance allows agents to retrieve material they should not access, leading to compliance and trust incidents. Poor data quality compounds invisibly. A stale field, a wrong-version contract, or a duplicated record produces a confidently wrong agent decision that is difficult to trace back to the root cause. The common thread is that agents amplify whatever shortcuts the data foundation took. And those shortcuts only become visible when the agent acts on them in front of a user, regulator, or customer.

There are early warning signs that a big data environment is not equipped to support AI agents, automation, and multi-agent workloads. These signs are recognizable once companies know what to look for, such as access policies residing in application user interfaces (UIs) rather than the data layer, or the lack of a semantic layer that mediates between raw data and consumers. It is also wise to look for data quality that is validated only nightly in batch instead of at runtime, unstructured stores that are searched but not semantically indexed, and lineage that stops at the warehouse boundary.

Finally, the clearest signal is when the first agent demonstration works, but the second fails inconsistently in ways that no one, including engineers, can explain. That inconsistency is almost always a foundational problem and not a model problem.

Avoiding Common and Costly Mistakes

Organizations routinely make the same recurring mistakes when modernizing their data foundations for AI. The first is treating AI readiness as “a tool problem” or buying a vector database and declaring victory, instead of treating it as an architecture problem. Architecture-first does not mean building everything at once. The most effective approach is to establish the data foundation, ship one functioning capability, and expand from there.

Organizations sometimes defer identity resolution until an agent in production fails publicly. In addition, enterprises might build agents before the foundation is ready, driven by demo cycles rather than architecture sequencing. Underinvesting in observability and flying blind once agents are deployed is also ill-advised.

The steps to mitigate these risks are the same in each case. Assess the Quadrant honestly, identify the binding constraint (usually identity or runtime enforcement), select one bounded workflow to anchor investment, and instrument deeply enough to see what is actually happening when something fails.

Data readiness, proper governance, and large-scale data integration can greatly improve AI outcomes and business performance. Several patterns are consistent across enterprises that have been successful in this regard. For example, a financial services firm that invests in entity-level identity unification can materially reduce agent-driven compliance incidents because the agent acts on a single, governed view of each customer.

A manufacturer that indexes two decades of maintenance logs into a semantically grounded retrieval layer turns a previously unusable archive into the most-cited grounding source for its field-service agent. The pattern in these and other successful cases is the same: foundation work that looks unrelated to AI is what makes the AI work.

Ultimately, the model is rarely the reason why enterprise AI agents fail in production. The data foundation is almost always the culprit. AI readiness can be diagnosed concretely, not merely felt. The AI-Readiness Quadrant provides architects with a structured tool to assess where their foundation is strong and where the binding constraint is.

Sequence matters. It’s critical that identity and harmonization come first, followed by semantic grounding of unstructured assets, then runtime access enforcement, and finally observability. Out-of-order investment wastes capital and delays results. The organizations that internalize these keys will avoid the most common and expensive mistake of the current AI moment: chasing models when the architecture is the problem.

About the Author

Sushree Mishra is a principal data and AI architect at Salesforce, focused on enterprise data foundations and agentic AI architecture. He has more than 17 years of experience designing data platforms and helping global enterprises operationalize AI at scale. His work sits at the intersection of data readiness, multi-agent systems and enterprise governance. Connect with him on LinkedIn.

Disclaimer: The authors are completely responsible for the content of this article. The opinions expressed are their own and do not represent IEEE’s position nor that of the Computer Society nor its Leadership.