Ben Yan, Senior Director Analyst at Gartner highlights that agentic AI promises enterprise value but faces real-world limitations. Success depends on customizing capabilities, managing expectations, and taking an agile, modular approach to development.
Agentic artificial intelligence (AI) is capturing significant attention as organizations seek new ways to drive efficiency, innovation and competitive advantage.
Unlike traditional automation or virtual assistants, AI agents are designed to independently interpret information, make decisions and act on behalf of users. This shift from human-driven to machine-driven decision-making is poised to redefine how businesses operate and create value across industries.
A recent Gartner survey conducted in January 2025, which included responses from 3,412 webinar participants, found that more than half of organizations (53%) are currently exploring agentic AI, while a quarter (25%) are testing these solutions through pilot programs. Only a small fraction—6%—have advanced to full-scale production. Additionally, 40% of respondents indicated plans to launch agentic AI initiatives within the next six months.
This growing momentum is accompanied by a wide range of interpretations and, at times, inflated expectations. As large language models (LLMs) continue to advance, many organizations are developing AI agents that can collect information, interact with various systems and complete assigned tasks. However, these agents frequently encounter challenges with enterprise contextualised decision-making.
The gap between hype and operational reality remains substantial. This can blind organisations to the real cost and complexity of deploying AI agents at scale, stalling projects from moving into production.
To realize the full potential of AI agents, organizations need to look beyond the hype. It is essential to focus on the foundational enterprise elements required to develop or implement AI agents that truly drive business value.
Customizing AI Agent Capabilities for Your Business
Current AI agents are a good option for organisations that need AI solutions capable of understanding user intent, retrieving and processing information from various data sources, and leveraging tools to complete tasks.
Organisations must take a flexible approach, by mixing and matching the right capabilities to suit their use case. This means configuring or building agents based on the data available, the tools and systems they need to interact with, and the LLM capabilities required. This level of customisation brings agents closer to the business context, increasing the value they can deliver.
Managing Expectations and Understanding the Limitations
Organisations must understand the limitations of AI agents to unlock their full business value. Doing so not only guides an implementation, but manages stakeholder expectations around scope, performance and impact.
One key limitation is the absence of critical components like world models, which is what allows AI agents to build an internal understanding of their environment and predict outcomes.
As human beings, we interpret how this world works via internal or abstract representations. For example, if a child sees apples falling from the trees several times, they will be able to predict how an apple falls next time. When we see something unusual or unexpected, such as a floating apple in the air, we will try to verify and may need to update our mental model or world state.
This active learning process is extremely important for AI agents to “understand” context, and update or improve as necessary. Current memory components of LLM-based AI agents are usually based on chat history and system logs. However, these cannot fully capture and store the dynamics of the agent itself, the environment or the world.
LLM-based AI agents also learn from data distribution, identifying correlations and probabilities rather than causations. This makes them not the most optical AI technique to use. For example, graph-based algorithms still outperform LLMs in areas like route planning, where accuracy and efficiency are critical.
It’s also important to recognise that an AI agent is not the same as an AI model. An agent is a composite AI-enabled system that combines multiple techniques to perceive, reason and act. Capabilities such as predictions or forecasting, planning and optimisation sit outside the strengths of LLMs and are better handled by other AI techniques.
These limitations are critical for organisations to understand as we’re still a long way from being able to entrust LLM-based agents with critical decision-making.
Focus on Core Components and Agility
Given the high uncertainty, technical complexity and rapid pace of AI agent development, taking an agile approach is essential. It will help organisations minimise latency from input to outcome, build trust and brand loyalty, and stay adaptable as the technology and market continue to evolve.
When building AI agent frameworks or solutions, consider “plug and play” components, technologies or models to avoid vendor lock-in. It isn’t recommended to build extensive frameworks and tools in-house. Instead, prioritise vendor solutions that are open and interoperable, or actively leveraging and contributing to open-source AI agent technology stacks.
