Agentic AI in 2026: More mixed than mainstream

Agentic AI is having its everything, everywhere, all at once moment. Or is it?

IT practitioners tracking what’s happening on social media, in vendor pronouncements, and throughout the thought leadership sphere where hyperbole runs hot are forgiven for thinking that the world’s businesses are soon to be running on AI agents, those bits of software that can automate a variety of tasks from code generation to content creation.

They see proclamations that the age of the agentic workforce is at hand (90% of IT support work executed by agents!) and anecdotes about enterprising programmers generating revenue from agent-based products. Meanwhile, others claim that the industry is being sold a bill of goods and that the reality on the ground does not match the hype.

As is often the case in tech, that reality is more nuanced. Even so, data clarifies. While 39% of organizations surveyed by McKinsey say they are experimenting with agents, only 23% have begun scaling AI agents within one business function.

AI agents explained

First, let’s take a step back: What exactly are AI agents and why are IT leaders excited about them?

Using contextual understanding and data available from large language models (LLMs) and other sources, AI agents sense and learn from their systems environment and reason their way through problems as they work to accomplish their objectives.

Agents also work with fellow agents, as well as other applications. They act like worker bees, bustling about to do work to please the queen. Achieving the goal efficiently is the end game.

Whereas a user might prompt an LLM toward a desired output with coaching via context, an agent figures out how to achieve its goal on its own, based on the problem-solving logic with which it was programmed and the data on which it was trained.

Salesforce, Workday, and Microsoft are among the many vendors that have already been embedding agents into applications that enterprises use for customer service and other functions.

Why agents are in hurry up and slow down mode

Adoption of multi-agent systems that work across platforms, however, has been slower going. High-profile failures of companies attempting to run their businesses on agents, as well as those whose agents experience catastrophic technical snafus, haven’t helped.

Multi-agent systems are technically challenging to build and operate, and vendors are hesitant to make such systems interoperable as they figure out how to monetize the data the agents generate and consume, says Nancy Gohring, senior research director, AI at IDC.

“It’s a tech question, as well as a competitive situation,” says Gohring, adding that vendors want to keep their customers within their ecosystems. For example, APIs for one vendor’s customer service platform don’t work with those of another vendor’s ecommerce software.

Cengage CIO Ken Grady agrees that vendors are working at cross-purposes as they seek to compete and protect their data moats. This is a big reason why very few companies have yet to reap real value from agents, he says.

Even developers end up on the agentic struggle bus

The twin specters of danger and risk even looms over software development.

While organizations of all sizes have successfully deployed agents to automate code, data exfiltration risks, vendor walled gardens, and rapidly accumulating fragility dampen agents’ prospects, says Voxel CTO Bryan O’Sullivan. This can lead to a “bunch of unreliable junk that doesn’t do anything but cost you a lot of money,” he says.

If even a fraction of agentic function is imprecise, it can derail the entire process, which likely explains why IT tends to be at the vanguard of agent adoption while other business units abstain. For this reason, “deep penetration of agents into other areas that are valuable” remains limited, O’Sullivan says.

One issue is rooted in memory — or rather, the lack thereof for agents. To fulfill their promise of autonomous operations, agents must have access to long-, medium-, and short-term memory, which are vital for learning from the tasks they execute. Without these capabilities they are essentially like LLM chat sessions; their shelf life is short.

Reason for optimism

Yet there is great belief that once the AI industry figures out how to solve these challenges — the collective business imperative will drive it there — agents will automate whole workflows, entire processes, and maybe even businesses.

IT leaders can reallocate staff to more strategic tasks or find new levers for innovation. It’s an idyllic dream perpetuated by technology’s great dreamers, including those who imagine businesses running around the clock, losing little as humans sleep.

For 2026, at least, up to 40% of all Global 2000 job roles will involve working with AI agents, which will redefine workstreams for many businesses, according to IDC.

Cengage’s Grady says he is confident that vendors will figure out the best path forward for themselves and their customers. Grady expects that organizations looking to create more efficient business processes will expand their agent implementations.

Business use cases that may have been operating at 2% may, for instance, grow to 20% as the technology and protocols mature, he says, adding that more companies will look to push successful pilots into production.

In the meantime, what should IT leaders do to prepare for the agentic enterprise?

As agents represent an emergent technology category, there isn’t a definitive playbook for adoption. At least, not one that is agent-specific.

Even so, IDC’s Gohring says that IT leaders must identify pilots to run and test. As they build confidence and comfort with their experiments, they should put mechanisms in place for control and visibility before they look to scale, she adds.

Companies should build the core abstraction layer and basic orchestration, test, learn, and look to incorporate governance and monitoring capabilities, especially as they scale to dozens or more agents.

Refining architecture based on real-world usage patterns is critical. And, as always, failing fast and learning is a critical part of the process.