Agents are the 'third wave' of the AI revolution

Agentic AI may be moving artificial intelligence (AI) to a new level beyond generative AI, with the same characteristics and challenges — but also with some notable distinctions.

Marc Benioff, CEO of Salesforce, calls agentic AI the “third wave” in the rapid evolution of the field. “In just a few years, we’ve already witnessed three generations of AI,” he observed in a recent piece in the New York Times. “First came predictive models that analyze data. Next came generative AI, driven by deep-learning models like ChatGPT. Now, we are experiencing a third wave — one defined by intelligent agents that can autonomously handle complex tasks.”

Also: I’m an AI tools expert, and these are the only two I pay for

AI agents, or intelligent assistants, are intended to serve as digital co-workers, assistants, or customer service representatives, communicating via natural language processing. They “have the potential to augment human capabilities in ways previously unimaginable,” Benioff observed. 

“Imagine a world where businesses can deploy an AI workforce of agents to manage customer interactions, analyze data, optimize sales strategies and execute operational tasks in real time and with little human supervision.”

Across the industry, there is agreement that AI agents, with their narrow focus, bring new capabilities and ROI that wider AI cannot deliver effectively. “Agentic AI will be the next wave of unlocked value at scale,” Sesh Iyer, managing director and senior partner with BCG X, Boston Consulting Group’s tech build and design unit, told ZDNET. 

Also: We’re not ready to support autonomous AI agents, survey suggests

He added that this is “an opportunity to redesign processes fundamentally and unlock significant productivity gains.”

As with both analytical and gen AI, AI agents need to be built with and run along clear ethical and operational guidelines. This includes testing to minimize errors and a governance structure. As is the case with all AI instances, due diligence to ensure compliance and fairness is also a necessity for agents, Iyer said. 

As is also the case with broader AI, the right skills are needed to design, build and manage AI agents, he continued. Such talent is likely already available within many organizations, with the domain knowledge needed, he added. “Upskill your workforce to manage and use agentic AI effectively. Developing internal expertise will be key to capturing long-term value from these systems.”

There are notable differences between generative AI and agentic AI as well. “Agentic AI is specifically designed to make decisions autonomously, often without human intervention, which differs from how gen AI is typically used,” said David Brault, an expert at Mendix. There are a number of features and functions that separate agentic AI from gen AI, he noted, starting with context and focus. 

Also: 25% of enterprises using AI will deploy AI agents by 2025

While generative AI applications can be targeted across many capabilities and industries, agentic AI “is focused on specific environments and contextual situations,” he added. Accordingly, agentic AI’s current best use cases are “predictable and defined tasks with low risk of errors or low severity of impact when errors occur,” agreed Michael Connell, chief operating officer at Enthought. 

In addition, integrating agentic AI with existing systems differs that that of generative AI. “Leveraging the decision-making capabilities of agentic AI often requires modifications to existing systems and integrating with existing APIs to utilize established business logic to improve decision accuracy,” Brault said.

To prepare for the shift from gen AI to agentic AI, “start small and scale strategically,” he advises. “Identify a few high-impact use cases — such as customer service — and run pilot programs to test and refine agent capabilities. Alongside these use cases, understand the emerging platforms and software components that offer support for agentic AI.”

Also: What is Google’s Project Mariner? This AI agent can navigate the web for you

This includes looking beyond the technology and focusing on the user journey and associated workflows, Iyer urged. “Instead of grounding efforts solely in the technology, think holistically about the workflows agents will transform. Aim to reduce mundane tasks, improve productivity, and create better human-machine collaboration.”

“The challenge is applying agentic AI in the enterprise setting or in innovation-driven industries, like materials science R&D or pharma, where there is higher uncertainty and risk,” said Connell. “These more complex environments require a very nuanced understanding by the agent in order to make trustworthy, reliable decisions.”

Also: What is Google’s Project Mariner? This AI agent can navigate the web for you 

As with analytical and gen AI, data — particularly real-time data — is at the core of agentic AI success. It’s important “to have an understanding of how agentic AI will be used and the data that is powering the agent, as well as a system for testing,” said Connell. “To build AI agents, you need clean and, for some applications, labeled data that accurately represents the problem domain, along with sufficient volume to train and validate your models.”

Connell added that a growing reliance on agents “will necessitate new supervisory frameworks, especially in high-stakes fields where traditional oversight models will be inadequate.” This means human oversight is always needed — especially with a risk of unintended consequences if agents are misapplied.