Agentic AI: Decisive, operational AI arrives in business

A more operational, business-specific way of leveraging generative AI is beginning to take shape in the form of AI agents that quietly work behind the scenes, moving beyond gen AI’s creational capabilities toward autonomous decision-making in enterprise workflows.

Agentic AI, at its core, is designed to automate a specific function within an organization’s myriad business processes, without human intervention. AI agents can, for example, handle customer service issues, such as offering a refund or replacement, autonomously, and they can identify potential threats on an organization’s network and proactively take preventive measures.

Early examples of the technology include GitHub Copilot Workspace, an integrated code repository agent, and Google AI Teammate, an AI assistant that can manage projects by monitoring company processes, creating reports, and generating new tickets for programmers, notes Mikhail Dunaev, chief AI officer at ComplyControl, a banking technology provider.

“In the next couple of years, these two assistants could potentially replace an entire development department,” he says. “However, this wave is just beginning, and truly valuable assistants are not yet available to a wide range of users.”

A recent study from Capgemini found that 75% of organizations surveyed are looking to use AI agents in software development, making it a top early use case.

Cognitive AI agents can also serve as assistants in the healthcare setting by engaging with a patient daily to support mental healthcare treatment, and as student recruiters at universities, says Michelle Zhou, founder of Juji AI agents and an inventor of IBM Watson Personality Insights.

The AI recruiter could ask prospective students about their purpose of visit, address their top concerns, infer the students’ academic interests and strengths, and advise them on suitable programs that match their interests, she says.

Enterprises as varied as Aflac, Atlantic Health System, Legendary Entertainment, and NASA’s Jet Propulsion Laboratory are among those already pursuing agentic AI.

AI agents on the rise

According to a Capgemini’s survey of large enterprises, one in 10 organizations are deploying AI agents, with more than 50% planning to explore their use them in the next year. Forrester, in a recent blog post, named AI agents as one of the top 10 emerging technologies for 2024, with author Brian Hopkins, vice president of the Forrester emerging tech portfolio, calling them “perhaps the most exciting development” on this year’s list.

“AI agents are now leveraging advanced language models to perform complex tasks, make decisions, and interact autonomously on behalf of enterprises or individuals,” he writes. “This shift from purely generative AI to ‘agentic AI’ promises more sophisticated and less brittle automation capabilities.”

Agentic AI will also drive the evolution of other specialized AI tools, he adds, including TuringBot agents, which can generate software code.

Independent, decisive AI

The key to getting the most value out of AI agents is getting out of the way, says Jacob Kalvo, co-founder and CEO of Live Proxies, a provider of advanced proxy solutions.

“Where agentic AI truly unleashes its power is in the ability to act independently,” he says. “With agentic AI, organizations will be able to scale their operations and create innovation at incredible speeds.”

Live Proxies uses AI agents to detect and respond to cybersecurity threats. The company has built an AI agent to autonomously monitor network traffic and mitigate cybersecurity threats, without “constant human oversight,” he says.

“What this does is frees us to innovate and satisfies the clients by entrusting our security infrastructure to be robust and self-sufficient under the watch of an agentic AI,” Kalvo says.

Compared to gen AI, which focuses on generating new content such as text, images, or music, agentic AI focuses on decision-making, he says.

“Gen AI might then seem more creative, generating outputs similar to human-generated content, while agentic AI would be more operational, acting with direct implications on business processes or technological ecosystems,” Kalvo adds. “Essentially, agentic AI means autonomy and execution, and gen AI deals with creation and innovation.”

Agents driving ROI

Agentic AI can deliver value to organizations struggling to find the ROI in gen AI, adds Dunaev. In many cases, gen AI still requires significant human intervention, he says.

“In contrast, agentic AI has the potential to drive more tangible business value,” he adds. “It acts as an autonomous assistant, solving tasks independently with capabilities like autonomous decision-making and goal-oriented actions, which generative AI does not offer.”

Beyond the use of AI agents for specific and discrete tasks, it has the potential to do a series of tasks, in a step-by-step manner, says Mike Finely, CTO and co-founder of AnswerRocket, vendor of an AI assistant for enterprise analytics.

Agentic AI is the next evolution of AI, he says. “With this new superpower, language models break down complex problems and do their work in incremental steps,” he adds. “This means the model can handle harder problems, and we can interact and guide the solution, so that we understand how it got the answer instead of head-scratching in wonder at the apparent magic.”

Trust issues and reflection

While AI experts see great potential for agentic AI, many also acknowledge that users may have misgivings. Generative AI users have noted serious hallucinations, and users of AI agents may not trust them to act autonomously on their behalf.

To engender trust, developers of AI agents must make it easy for humans to check their work, says Live Proxies’ Kalvo.

“Agentic AI can be widely trusted as a function of transparency in its decision-making process and the system’s ability to explain its choices,” he says. “Trust is therefore established through intensive testing, clear communication about capabilities and limitations, with constant monitoring.”

Another approach is to use another AI to check the agents’ work, Finely says. A process called reflection uses one AI model to reflect the answer given by another.

“This is essentially as though we were having a human review of the output of a model, but instead, we are automating that task as well,” he says. “The net result may be slightly more time and expense, but if that is the price of trust, and it unlocks the enormous power of automation, then it will be the winning answer.”

Older AI-like technologies, including machine learning, have been used for years, and most organizations trust ML to do its job, he notes. Meanwhile, the fast adoption of gen AI shows that users are willing to trust it.

“Gen AI is an advanced application of machine learning, and companies are beginning to trust it when there are references to the sources that it used, when the specifics of any data that is being cited are provided, and when the model can explain the reasons for its conclusions,” Finely says. “Consumers are voting with their keystrokes.”