Businesses have systems of intelligence. Now they should move to systems of action

Agentic AI is getting into action. These autonomous systems promise to handle complex workflows end-to-end, and enterprises are eager to harness that potential. But here’s the catch: in the rush to deploy this technology, businesses risk breaking what already works. 

IBM’s Bruno Aziza, Vice President of Data, AI & Analytics Strategy at IBM, works with organizations to help them scale with emerging technologies to become leaders in their respective domains. According to Aziza, enterprises often overlook the reality that they already have AI and automation embedded in their workflows.

Rather than rebuild everything around agents, Aziza believes organizations should use them to build on what’s already working. “The path of success here is not to replace or eliminate those processes that have worked well,” he told IBM Think in an interview. “It’s about augmenting these existing processes.”

This augmentation approach becomes clearer when viewed against the broader evolution of enterprise technology. To understand how organizations can effectively integrate agentic AI into existing workflows, it’s essential to examine the progression from systems that analyze and recommend to systems that can autonomously act and deliver tangible business value.

Enterprise tech leaders have followed a familiar path: first building systems of record, then systems of engagement, followed by systems of intelligence. Now, with agentic AI, a new phase begins: systems of action.

“What we’re finding out is that while systems of intelligence are useful, they’re not sufficient,” Aziza said. “Thanks to gen AI, thanks to almost infinite compute and limited storage and great networking, we’re now in a phase where we can start thinking about how we accelerate the actions, which are really the components that are related to hitting the bottom line of the organization. We’ve always tried to do that, but now we’re in a phase of history where we can truly get to action.”

But building systems of action isn’t a simple, linear path. For many tech leaders, questions around autonomy, accountability and automation still loom large. And while the AI hype cycle has been driving headlines for several years, the real work begins once the spotlight fades.

“I think it’s something that mature businesses have to take with a grain of salt: everybody’s telling you it’s really going to be easy, right?” Aziza said. “You focus on great data, then you select your use cases, then you do some experimentation and then finally, you operationalize the use cases that are successful. But that’s not really what’s happening.”

A common misconception Aziza observes is the belief that building an agent and deploying it in enterprise will just work. “In the enterprise, there’s a lot more at stake.”

As more experimentation moves into production, business leaders are asking different questions about trade-offs, scale and performance.

“There’s maturity and enough examples in the industry so we can get to a level where we can accelerate the real production work being done there,” said Aziza.

One of the biggest shifts this year has been the push to build new AI agents. But according to Aziza, this rush may be misguided. Enterprises often already have AI and automation embedded in their workflows. Rather than rebuild everything around agents, Aziza said the better approach is to augment what’s already working.

He offers a clear example: loan approvals in banking. These are deterministic processes—adding probabilistic agents could introduce more uncertainty, not less.

“The ability to bring these two things together is really important,” he said. This, he explained, is agent minus: an agent that replaces working systems but delivers worse outcomes.

“That’s not something that’s being talked about enough, and organizations have to be aware of that,” he said. “It’s not going to be about elimination of automation and replacing it with a bunch of agentic work. It’s actually marrying the two.”

Bringing systems and agents together is what’s next. “We’re going to get to a place where more and more employees will be building their own agents, either because they’re using [a] pre-built agent and customizing it, or because you have high maturity in your employee base,” he said. “We’re going to see more agents. The question is, how do you orchestrate those agents? And so, there’s a set of capabilities that you need to develop in order to do this right.”

This orchestration doesn’t happen by accident. According to Aziza, organizations going into production with agents need to build five core capabilities to succeed: multi-agent collaboration, cross-ecosystem integration, alignment with existing tools and rules, supervision and control, and a dedicated operational layer known as agents ops. “These five capabilities are the key pillars of your agentic strategy moving forward,” he said.

Earlier this year, Model Communication Protocol (MCP) and other standards developed by Google (Agent2Agent Protocol), IBM (ACP) and other players emphasized the need for agent discovery and communication.

“Imagine that you wake up tomorrow in an environment of a swarm of analytics agents,” said Aziza. “Who’s doing what? Who’s getting the data? Who’s building the insight? Who’s supervising the process? Who’s verifying that the data is actually correct and the insight is actionable? There are a lot of these questions here that if you don’t have a standard, it’s just really hard to operationalize.”

Many compare the current state of agents to the internet pre-HTTP. But Aziza believes agentic AI brings something entirely different to the conversation. To him, the current evolution of AI could be compared to the invention of the car: yes, cars ultimately replaced horses, but they also drove innovation.

“There’s an incredible acceleration of innovation,” he said. “And there’s also this incredible opportunity of transforming ourselves to the next level. It’s just difficult to find a blueprint, but certainly we know that the unification of communication standards is going to be necessary if we want to multiply the impact of these individual agents.”