Analyst Perspectives

AI, like analytics, must lead to action. Too often, in both cases, too much of the exercise is left to the reader. We have tools to provide sophisticated analyses, including AI platforms that can be used to predict many types of behavior, but we fall short in helping the workforce know what to do with that information. Some examples are more obvious, such as fraud detection. If a transaction is predicted to be fraudulent, the transaction should be blocked. But even this example is not as cut and dry as you might think. I’m sure many of you have been frustrated standing at a hotel check-in desk or at a retail counter attempting to make a purchase when your credit card transaction has been denied. What is the appropriate action or set of actions that should be taken?

As AI continues to evolve, there is an increasing need for it to move beyond merely providing scores or recommendations to taking actionable steps. While scores and recommendations offer valuable insights, they often require decision-makers to translate the data into effective strategies. As enterprises invest in generative AI (GenAI), they recognize this need. ISG’s Market Lens research shows one-third of enterprises (32%) are using the technology to address business process workflow management, representing the third-biggest initiative to date based on investment. And, it is second on the list of GenAI use cases that will deliver the most benefit over the next two years.

Central to the concept of agentic AI is the development of Large Action Models (LAMs). Unlike traditional AI/ML models that focus primarily on prediction and classification, or large language models (LLMs) which focus on the next words to appear in a response, LAMs are designed to make decisions and execute a series of actions across a variety of environments. Just as LLMs leverage vast datasets of text, LAMs are developed based on a large body of action and outcome data. At their core, LAMs are optimized for function calling as the mechanism for taking action.

LAMs are not strictly necessary to create agentic AI. Agentic AI is defined as the ability to take autonomous actions, involving multiple processes or systems, based on understanding of the environment and the goals that should be achieved. We’ve begun to evaluate some of these capabilities in our Intelligent Automation Buyers Guide. [LINK] Here are a few ways agentic AI can function without LAMs. Agentic AI can operate on predefined rules and heuristics. By encoding specific guidelines for decision-making and action execution, these systems can function effectively in well-defined domains. Reinforcement learning frameworks can also be used to train AI agents to make decisions and take actions based on rewards and punishments. This approach allows agents to learn optimal behaviors from interaction with their environment without necessitating a large action model.

Agentic AI, while promising, faces its share of challenges, many of which overlap with other elements of AI. First of all, there are the data engineering issues associated with agentic AI which my colleague addressed in this perspective. There are also ethical concerns about the autonomous nature of agentic AI regarding accountability for actions taken by these systems. Determining who is responsible for mistakes or harmful outcomes—whether it's the developer, the user, or the AI itself—presents a complex dilemma. Bias, fairness, transparency and explainability issues must be tackled as well. Understanding how to balance human oversight with AI autonomy is a complex design challenge that can impact user acceptance. Seamlessly integrating agentic AI into existing workflows and systems can be complicated. Enterprises may face technical and operational obstacles when trying to adopt these systems alongside their traditional processes. Finally, the corpus of actions on which to train the agents may be limited and may not be relevant to the actions and decisions of a specific enterprise or process. Addressing these challenges is vital for the responsible development and implementation of agentic AI, ensuring it can deliver genuine value while minimizing risks and negative outcomes.

By advancing into action-oriented capabilities, agentic AI can facilitate automated responses and real-time adaptations, ultimately improving operational efficiencies and outcomes. This shift not only enhances user experience by reducing the burden of decision-making but also by leveraging the potential of AI's predictive analytics. Moving towards actionable AI can drive tangible results, streamline processes and foster a more proactive approach in various fields, from healthcare interventions to supply chain optimizations, thus fundamentally transforming the way we harness data in decision-making. As you evaluate various AI strategies and vendor offerings you should determine how and when to incorporate agentic AI into your information architecture.

Regards,

David Menninger