Chloe Maraina is a distinguished expert in Business Intelligence, renowned for her ability to transform vast streams of big data into clear, actionable visual narratives. With a professional background rooted in data science, she brings a forward-thinking perspective to the complexities of data management and integration, particularly as organizations transition from experimental AI pilots to robust production environments. In this discussion, we explore the vital role of situational awareness and semantic modeling in ensuring AI accuracy, the current statistical landscape of project success versus failure, and the massive infrastructure overhauls required to support multi-agent networks. Our conversation covers the evolution of agents through context, the technical necessity of purpose-built data retrieval engines, and the emerging challenges of governing autonomous swarms in an increasingly agent-centric architecture.
Many organizations treat AI like a black box, but providing situational awareness is similar to what management responsibility?
It is essentially no different than providing context to a human employee who has just been hired to perform a critical job. Imagine bringing in a junior analyst and handing them a complex task without a single word of background information; they would be lost, and the AI agent is exactly the same in that regard. Context is what gives these agents the situational awareness required to deliver trustworthy and accurate outputs, and without it, the most sophisticated models are doomed to fail before they even start. When we provide that proper awareness through high-quality data and specific business logic, these agents can suddenly perform miraculous feats, from optimizing entire global supply chains to managing complex sales campaigns. It is the invisible thread that turns a “black box” into a functioning member of the team, capable of discovering and ingesting information independently to solve problems that were previously impossible for humans to oversee.
Given that so many AI initiatives remain stuck in the lab, what does the current data tell us about the success rate of moving these projects into production?
The reality is a bit of a double-edged sword right now, as we see significant progress dampened by a high volume of stalled efforts. According to the 2026 State of AI in the Enterprise report, only about one-quarter of the organizations surveyed have successfully moved 40% of their AI experiments into a live production environment. We are seeing an improvement in the general success rate, moving from a meager 17% a year ago to roughly one-third of the 1,200 projects examined in recent year-end research, but that still means more than half of these projects are failing to launch. These failures usually occur because the underlying data issues haven’t been resolved, leaving the agents without the context they need to be reliable. It is a laborious and often frustrating journey for enterprises that have invested heavily in experimental pilots, only to find that the gap between a chatbot and a real-world agentic system is wider than they initially anticipated.
How does the accuracy of an AI agent fluctuate when it moves from specific, contained tasks to analyzing broad swaths of enterprise data?
There is a startling drop-off in performance that can feel quite jarring for teams who have seen high success in controlled testing. When large language models are asked to reason through questions using limited, contained data sets, they are incredibly impressive, often reaching a 95% accuracy rate. However, a study from September 2025 demonstrated that once you ask those same models to derive outcomes from broad data spread across multiple disparate systems, that accuracy plummets to 50% or even lower. This is precisely why a basic retrieval-augmented generation, or RAG, approach is no longer viewed as sufficient for the complex demands of 2026. To bridge this gap, we are seeing that the addition of a semantic layer can push that accuracy back up toward 100%, effectively giving the agent a map to navigate the data landscape without getting lost in the noise.
With the recent 327% increase in the usage of multi-agent systems, what kind of infrastructure overhaul are enterprises currently facing to keep up?
The surge we have seen over the last four months is nothing short of explosive, but it has exposed deep cracks in traditional data architectures that were never designed for this level of autonomy. Enterprises are realizing they need to modernize or completely revamp their infrastructures to include vector embedding, reranking models, and data retrieval engines that are purpose-built for AI workloads rather than old-school BI tasks. We are shifting from an environment where data access was primarily the domain of humans to one where the majority of interactions are handled by agents, which requires a massive investment in governance and auditability. It is an expensive and time-consuming process to ensure that everything an agent needs is ready the very instant they call for it, especially as we move toward systems where agents are building other agents. The goal is to reach a point where the right context is accessible at the right time, but for many, the “environment” itself remains the biggest hurdle to achieving that 327% growth at scale.
As we look toward the future of “swarms” or multi-agent networks, how do we balance the need for collaboration with the risks of total autonomy?
This is perhaps the most daunting challenge on the horizon because managing a system that includes thousands of autonomous agents requires a level of oversight we are only just beginning to define. Currently, our research shows a divided landscape: 50% of agents in production are designed to assist humans, 40% work alongside humans with the goal of eventually becoming autonomous, and only 10% are fully autonomous today. The complexity arises when you have an accounting agent, a legal agent, and a procurement agent all trying to negotiate and hand off tasks to one another without a human in the middle to catch emergent anomalies. We have to figure out how to let them collaborate and “negotiate” while ensuring they don’t overstep boundaries, violate regulatory statutes, or expose sensitive data. It’s a delicate dance of giving them enough context to be effective while maintaining enough governance to prevent accidental consequences in these highly connected, high-speed swarms.
What is your forecast for the evolution of context-aware agents in the next several years?
I believe we are entering a phase of “gradual eating away” at the most difficult challenges, where the next several years will be defined by a slow but steady migration toward sophisticated multi-agent networks. We will likely see a significant shift where the 10% of fully autonomous agents we see today becomes a much larger slice of the pie as semantic layers and context-aware architectures become the standard rather than the exception. However, this won’t happen overnight; it will take years of devoted research and enterprise investment to solve the governance issues inherent in autonomous swarms. We will continue to see a rise in the AI success rate as we move past the “black box” mentality and start treating agentic integration as a core business logic requirement. Ultimately, the next few years will be about moving from agents that simply “answer” to agents that “act” with a deep, nuanced understanding of the enterprise’s unique operational DNA.
