In 2026, AI agents went from shiny object to top priority for enterprises, with 74% of leaders now expecting nearly half of their business processes to be redesigned around AI agents within four years, according to a Deloitte survey on agentic transformation published in August. Yet only 16% of respondents said they’re prepared.

As part of the 2026 AIQ 75 list, Fortune set out to find companies that are deep into rolling out AI agents and automating workflows at scale. This year’s survey included several new questions focused specifically on agentic AI use within companies.

Among the companies that ranked highest for agentic AI, Adobe, Freeport-McMoRan, Kyndryl, and Salesforce opened up about the ins and outs of their experiences thus far. They’re broadly tapping agentic AI for everything from invoicing to product research, software engineering, and both employee and customer help desks. These companies have all failed, pivoted, and made real progress, and they all came to the same conclusions: You have to work toward solving a specific business problem; bring tech and business stakeholders together; and focus relentlessly on ROI.

Fewer, more focused agents

Kim Basile, CIO of technology services company Kyndryl, doesn’t know how many of the company’s workflows now involve agentic AI.

“I’ll be honest with you. Year one, we were all about counting agents, and that wasn’t telling me the true story. So we realized that we needed to pivot away and really think about business process flows,” she said.

Salesforce—which sells the same agentic AI systems it’s using internally as its products—similarly pivoted from hundreds of agents to a few the company strongly believed could deliver specific benefits, according to SVP Andy White, who oversees adoption of Salesforce products for internal operations. He also doesn’t know how many agentic workflows are in place, but knows it’s happening in every department.

“We chased too much all at once and didn’t have enough focus,” he said.

International mining company Freeport-McMoRan didn’t have as much room for trial and error. Operating in an extremely capital-intensive industry where backend workflows translate to running multimillion-dollar machines, the firm was laser-focused on outcomes from day one. To narrow in on what would drive the biggest benefits, leadership looked at what workflows were simultaneously high-volume, manually intensive, and dependent on the types of data processes where agentic AI could excel.

The company is now implementing agents in around 10 areas, with sales invoice processing being the biggest standout use case, said SVP and CIO Bert Odinet. Freeport-McMoRan’s invoices involve variable pricing that’s dependent on the by-products found in each specific load of material, making them particularly complex and data-intensive. After six months using agentic AI to process approximately $500 million in invoices for just one line of product, the company has realized significant cost savings and is planning to scale to other product lines, according to Odinet.

What measuring ROI really means

Freeport-McMoRan has specific KPIs for each agentic process to ensure they’re delivering on the desired outcome and “not kidding [themselves] about the success of these agents,” said Odinet.

Toni Vanwinkle, VP of digital employee experience and cochair of AI at Adobe, another company building agents into its products, similarly stressed the importance of measuring if the intended outcome actually happened. What’s more, this has to go beyond productivity or efficiency.

“Did it help you sell faster? Did it help you land your client because you knew them more?” she said, adding that it’s about the “so what?” that comes after time saved.

For example, Adobe used agentic AI to reimagine its internal help desk, which serves more than 30,000 employees. The company didn’t only look at whether agentic AI could be used to close more tickets faster, but also the quality of responses (Were they accurate? Did employees say they were satisfied?) and whether the process resulted in giving back capacity to the support agents.

This has also been a major growing point for Kyndryl. Basile said the company first thought about agentic ROI in terms of cost cutting, initially focusing on worker productivity in year one and then department-level efficiency in year two.

“That isn’t the right answer,” she said. “I mean, yes, that is certainly an important component of an ROI, but really we needed to pivot our mindset and think about, ‘How … do I bring more value to the enterprise? How do I help finance close the books in three days versus five days, and [then with] those additional two days, allow them to get ready for investors?’”

Now in year three, Kyndryl is narrowing in on the impacts of the workflows themselves, and specifically, how they stitch together across the entire organization.

It takes two to reimagine

When Salesforce discovered some agents were meeting expectations and others weren’t, company leaders introspected to find out why. Beyond instances where the agent wasn’t attached to solving a specific problem, they found it was clearly the workflows with more business stakeholder support that were succeeding.

To solve for this, the company now pairs agentic builders from the IT team with business stakeholders to form small “rapid innovation squads,” running one for every business unit.

“You have to be partnered with an opinionated stakeholder that’s willing to disrupt their organization, because it will be disruptive,” said White.

For example, a rapid innovation squad for the legal team sought to make the company’s legal routing—assigning new cases to a lawyer with the correct experience to get it done—more efficient. According to White, the new agentic process increased the proportion of cases being resolved and routed to the right group the first time from 12% to 34%, saving the firm over $350,000 in a year.

Across these companies, all the leaders Fortune spoke with emphasized the need to bridge the gap between IT and business units more than ever before. It’s not as simple as just deploying the agent; it requires a lot of continuous feedback, oversight, and continual improvement, said White, and a lot of that falls on the business teams.

“That’s what we’ve asked of our leaders,” echoed Vanwinkle. “‘This is a place where we need you to lean in, like, y’all are the people that are going to help us point the arrow.’”

This is largely because, as all these leaders also agree, implementing agentic AI requires reimagining workflows from square one. Jumping into automating processes you have in place as they are can actually backfire.

“We put more energy on automating a process that wasn’t really validated in a way that it should have been,” said Basile, citing this as one of Kyndryl’s biggest mistakes in pursuing agentic AI transformation. “It was already clunky, but we just wanted to make it automated, and it didn’t fix anything. It actually made it more complicated.”

This is also why Odinet stresses the importance of getting fresh perspectives from people who are new to the processes, such as folks in other business areas. Functional stakeholders are used to doing something a certain way, and it takes deliberate intention to take a step back and ask, “Why?”

“You want smart thinkers,” he said. “But you want them to be open-minded to rethinking the whole thing.”