Imagine AI so sophisticated it could read a customer’s mind? Or identify and close a cybersecurity loophole weeks before hackers strike? How about a team of AI agents equipped to restructure a global supply chain and circumnavigate looming geopolitical disruption? Such disruptive possibilities explain why agentic AI is sending ripples of excitement through corporate boardrooms. 

Although still so early in its development that there lacks consensus on a single, shared definition, agentic AI refers loosely to a suite of AI systems capable of connected and autonomous decision-making with zero or limited human intervention. In scenarios where traditional AI typically requires explicit prompts or instructions for each step, agentic AI will independently execute tasks, learning and adapting to its environment to refine decisions over time. 

From assuming oversight for complex workflows, such as procurement or recruitment, to carrying out proactive cybersecurity checks or automating support, enterprises are abuzz at the potential use cases for agentic AI. 

According to one Capgemini survey, 50% of business executives are set to invest in and implement AI agents in their organizations in 2025, up from just 10% currently. Gartner has also forecast that 33% of enterprise software applications will incorporate agentic AI by 2028. For context, in 2024 that proportion was less than 1%. 

“It’s creating such a buzz – software enthusiasts seeing the possibilities unlocked by LLMs, venture capitalists wanting to find the next big thing, companies trying to find the ‘killer app,” says Matt McLarty, chief technology officer at Boomi. But, he adds, “right now organizations are struggling to get out of the starting blocks.” 

The challenge is that many organizations are so caught up in the excitement that they risk attempting to run before they can walk when it comes to deployment of agentic AI, believes McLarty. And in so doing they risk turning it from potential business breakthrough into a source of cost, complexity, and confusion.

Keeping agentic AI simple 

The heady capabilities of agentic AI have created understandable temptation for senior business leaders to rush in, acting on impulse rather than insight risks turning the technology into a solution in search of a problem, points out McLarty. 

It’s a scenario that’s unfolded with previous technologies. The decoupling of Blockchain from Bitcoin in 2014 paved the way for a Blockchain 2.0 boom in which organizations rushed to explore the applications for a digital, decentralized ledger beyond currency. But a decade on, the technology has fallen far short of forecasts at the time, dogged by technology limitations and obfuscated use cases. 

“I do see Blockchain as a cautionary tale,” says McLarty. “The hype and ultimate lack of adoption is definitely a path the agentic AI movement should avoid.” He explains, “The problem with Blockchain is that people struggle to find use cases where it applies as a solution, and even when they find the use cases, there is often a simpler and cheaper solution,” he adds. “I think agentic AI can do things no other solution can, in terms of contextual reasoning and dynamic execution. But as technologists, we get so excited about the technology, sometimes we lose sight of the business problem.”

Instead of diving in headfirst, McLarty advocates for an iterative attitude toward applications of agentic AI, targeting “low-hanging fruit” and incremental use cases. This includes focusing investment on the worker agents that are set to make up the components of more sophisticated, multi-agent agentic systems further down the road. 

However, with a narrower, more prescribed remit, these AI agents with agentic capabilities can add instant value. Enabled with natural language processing (NLP) they can be used to bridge the linguistic shortfalls in current chat agents for example or adaptively carry out rote tasks via dynamic automation. 

“Current rote automation processes generate a lot of value for organizations today, but they can lead to a lot of manual exception processing,” points out McLarty. “Agentic exception handling agents can eliminate a lot of that.” 

It’s also essential to avoid use cases for agentic AI that could be addressed with a cheaper and simpler technology. “Configuring a self-manager, ephemeral agent swarm may sound exciting and be exhilarating to build, but maybe you can just solve the problem with a simple reasoning agent that has access to some in-house contextual data and API-based tools,” says McLarty. “Let’s call it the KASS principle: Keep agents simple, stupid.”

Connecting the dots

The future value of agentic AI will lie in its interoperability and organizations that prioritize this pillar at the earliest phase of their adoption will find themselves ahead of the curve. 

As McLarty explains, the usefulness of agentic AI agents in scenarios like customer support chats lies in their combination of four elements: a defined business scope, large language models (LLM), the wider context derived from an organization’s existing data, and capabilities executed through its core applications. These latter two rely on in-built interoperability. For example, an AI agent tasked with onboarding new employees will require access to updated HR policies, asset catalogs and IT. “Organizations can get a massive head start on business value through AI agents by having interoperable data and applications to plug and play with agents,” he says. 

Agent-to-agent frameworks like the model context protocol (MCP) – an open and standardized plug-and-play that connects AI models to internal (or external) information sources – can be layered onto an existing API architecture to embed connectedness from the outset. And while it might feel like an additional hurdle now, in the longer-term those organizations that make this investment early will reap the benefits. 

“The icing on the cake for interoperability is that all the work you do to connect agents to data and applications now will help you prepare for the multi-agent future where interoperability between agents will be essential,” says McLarty. 

In this future, multi-agent systems will work collectively on more intricate, cross-functional tasks. Agentic systems will draw on AI agents across inventory, logistics and production to coordinate and optimize supply chain management for example or perform complex assembly tasks. 

Conscious that this is where the technology is headed, third-party developers are already beginning to offer multi-agent capability. In December, Amazon launched such a tool for its Bedrock service, providing users access to specialized agents coordinated by a supervisor agent capable of breaking down requests, delegating tasks and consolidating outputs. 

But though such an off-the-rack solution has the advantage of allowing enterprises to bypass both the risk and complexity in leveraging such capabilities, the digital heterogeneity of larger organizations in particular will likely mean – in the longer-term at least – they’ll need to rely on their own API architecture to realize the full potential in multi-agent systems.

McLarty’s advice is simple, “This is definitely a time to ground yourself in the business problem, and only go as far as you need to with the solution.”

This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff.

This content was researched, designed, and written entirely by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.