Agentic AI poised for progress in 2026 — if CIOs get it right

The past year generated major hype about AI agents, with tons of experimentation and lots of failure, and some AI experts see only limited improvements in 2026.

Research firm Gartner, for example, has predicted that 40% of agentic AI projects will be cancelled by 2027, because of escalating costs, unclear business value, or inadequate risk controls.

Meanwhile, data about how many enterprises have actually deployed agents successfully is conflicting. A May survey by PwC found that 79% of companies represented had adopted agents in some capacity. But enterprise search vendors Lucidworks, which developed an agent to assess the AI capabilities of ecommerce sites, found that only 6% of the 1,100 sites it analyzed had deployed more than one agentic solution.

While some growth in agent deployments will happen this year, the technology may not yet hit the mainstream, says Don Schuerman, CTO of AI decisioning and workflow automation vendor Pega, in part due to hallucination problems with the LLMs that power agents.

“2026 will be the year that starts to separate the winning approaches from the failed approaches,” he says. “I don’t know if it’s the year where we actually see agents fully take over everything — that’s going to be potentially a little bit longer term of a transformation that maybe people estimate.”

Cloudy outlook

Part of the uphill battle many organizations face with agentic AI is that they have deployed LLMs expecting them to apply reasoning functionality to problems, creating an overinflated expectation of results, Schuerman says. “The LLMs aren’t reasoning machines, they’re just text prediction machines,” he adds.

Schuerman sees many organizations designing agents for predictable workflows where they don’t need to reason but can save employees time by taking over routine tasks. Truly succeeding with agentic AI, however, requires building reasoning into agent tasks at the design phase, he says.

“Really anchor what your agents are doing in those business processes, in those workflows, because most of what the enterprise is trying to do wants to run as a pretty deterministic workflow with a prescribed series of steps that you want to get performed in a consistent way every time, with high degrees of predictability, consistency, and audit,” he explains.

Schuerman argues that expectations about how agents should be used have been skewed by early rollouts. Instead, AI should be used to help redefine the business workflows that agents will take over, he says.

“What is a little bit of a myth is this idea that we’re just going to randomly deploy thousands of agents across our business and just let them go,” he says. “Instead, what we’re going to do is we’re going to use agents to define and design a lot of the workflows that we need in our business and do that with much more speed and completeness than we ever had before.”

Salesforce CIO Dan Shmitt agrees that hurdles remain for agentic AI’s outlook in 2026. For example, many organizations still lack clear roadmaps for how to start, scale, and define success, he says, adding that without high-quality data and a unified governance model, agents can produce unreliable results.

Still, Shmitt sees agents becoming more widely used as the year goes on, if not in the fully autonomous way some AI experts have predicted.

“We’re not likely to see fully autonomous systems deployed universally across organizations,” he says. “Instead, organizations will start to adopt agents as collaborative systems that work alongside people and other agents in day-to-day processes to augment employee productivity and decision-making.”

To get there, Pega’s Schuerman stresses CIOs’ need to stick to the basics of IT deployments when rolling out agents.

“We are seeing more and more of a realization that having new technology like agents doesn’t mean you get to forget about the important foundational work, like having the workflow right, having the data right, having the outcomes defined,” he says. “You’ve got to understand how they deliver outcomes. You’ve got to make sure agents are connected to the data.”

‘The sky’s the limit’

IBM CIO Matt Lyteson also expects more successful deployments of agentic AI in 2026, especially if IT leaders can focus on targeted outcomes when rolling out agents. CIOs also need to pay attention to data security and controls and better understand how agents interact with other IT systems, he says.

“Our focus is, how do we scale agents across more and more use cases to bring value to the organization, and how do I really understand the outcomes, the data that I’m going to need to give the agents, and then how to manage and control them?” he explains. “If organizations can do that, we’re going to see a lot more adoption and a lot more success.”

One impediment to deploying agents has been integrating them with existing systems and data, Lyteson says. Deploying agents without fully defining the targeted outcome is another, he adds.

“We’re used to, as an IT business, thinking about the process first and having that process accomplish something, instead of thinking about the outcome first and what I want an agent to achieve,” he says. “There are some things that start to get in our way and are tricky areas, and if you aren’t thinking through them in the right way, they can really result in stumbling.”

IBM has deployed hundreds of enterprise workflow AI agents and thousands of personal productivity agents, Lyteson says. For example, the company is using agents to triage IT support tickets and handle low-level support requests.

Lyteson recommends that CIOs stay open to the potential of agents, even if some pilot projects haven’t yielded good results.

“Every day, every week, we’re learning something new,” he says, adding that CIOs need to apply those learnings toward maximizing value for the organization. “We need to be continuously curious and to translate that into business outcomes. If myself and my peer CIOs are able to do more of that, then the sky’s the limit.”

Thinking about agent lifecycle

Like Lyteson, Asana CIO Saket Srivastava sees AI agent deployments growing in 2026, even as CIOs face several challenges.

Among those challenges is human resistance to using agents, but Srivastava also believes CIOs need to get a better handle on agent lifecycles, including tracking agents spun up by employees and deciding when to retire ineffective agents. Many CIOs will soon have to navigate a work environment with more agents than employees, and monitoring agent effectiveness will be essential.

In the meantime, reliability and trust concerns may still limit the number of agents deployed in the near term, he says.

“Trust comes from structure, trust comes from context, from the permissions, the visibility of how decisions are being made, how workflows are being progressed,” Srivastava adds. “And there certainly might have been a little bit of us all getting overexcited with all things AI and being in this pilot mode where we were trying out a bunch of things without having clarity. Do we have the right data; is it the right process?”

In some cases, organizations have added AI agents to workflows and processes that were flawed to begin with, he adds.

“In the simple days of automation, we used to keep saying that there’s no point in automating a bad process,” Srivastava says. “Are you looking at your processes more deeply? Are you looking at reimagining those processes and then applying AI to that? That’s what I see coming into the new year.”

While some agent pilot projects may have been premature, CIOs also need to keep experimenting and balance solid results with innovation, Srivastava says.

“Perhaps letting a thousand flowers bloom might not be the best approach, but create the right environment for experimentation to flourish, yet at the same time, you have higher confidence that AI is more ready to go solve for those problems,” he says. “Make sure that you’re solving the right problems, you’re solving it the right way, you’re measuring the outcomes, and then moving on to the next problem.”