Artificial intelligence is no longer confined to experimental projects and isolated efficiency gains. Today, enterprise adoption of AI is rapidly expanding into critical business functions, with organizations expecting agents to increasingly take on the responsibility of coordinating work across applications, systems, and business processes.

Deploying AI into production has largely become a challenge of the past. Now, the way workflows are architected is also undergoing a shift. Enterprises are no longer relying on a single AI tool to complete a single task, but are entering a new phase where agents collaborate with each other, call tools and applications, exchange context, and hand off work to one another.

So, what does managing AI actually require when individual agents become part of an interconnected agent ecosystem? An upcoming research report from OutSystems and KPMG explores this: 71% of respondents expect enterprise applications to evolve toward a model where agents coordinate work across systems, teams, and business processes, or assist business users in completing more complex multi-step workflows.

From point adoption to orchestration governance

Isolated agents are typically evaluated based on their own outputs. However, while this approach can validate their technical capabilities, deploying them into production is an entirely different matter—especially because most real-world workflows involve multiple applications, datasets, and other agents. This introduces interdependencies: agents may exchange context, call APIs, or interact with enterprise applications, and rely on outputs generated earlier in the chain.

Gonçalo Borrêga, Senior Vice President of Products, AI, and Application Development at OutSystems, describes this as an orchestration challenge. "The higher the goal, the greater the coordination effort involved. This is where orchestration comes in," he says. "A mistake here can trigger a cascade of ten errors downstream."

This shift is already beginning to emerge. For example, a global logistics organization is moving from a single agent that analyzes application logs to a multi-agent model—where specialized agents collaborate in real time and share context.

Increased complexity, greater governance challenges

As these systems become increasingly interconnected, the focus of challenges has shifted from governing individual AI projects to maintaining control across the entire workflow. More than two-thirds of respondents rank security, permissions, and policy enforcement across agents as the top orchestration obstacle.

The implication is this: although most organizations do have governance mechanisms in place, this does not necessarily mean their processes can operate effectively across the entire workflow—especially in scenarios involving coordination among multiple agents and other tools. "You still need to ensure the security, observability, accuracy, and quality of all these agents, and you need to coordinate and orchestrate these efforts across multiple platforms," says Borrêga.

Some organizations have begun to adopt a more structured approach. For example, a large digital bank first defined an authorized set of AI tools and preconfigured workflows before allowing teams to deploy autonomous agents, thereby reducing the security risks of fragmented adoption while preserving room for innovation.

Stronger governance does not mean limiting choice

The agent-driven enterprise is inherently heterogeneous—different agents, applications, models, and data services will coexist across various platforms. Borrêga believes that when agents are added to an already fragmented application and data environment, simply adopting agentic AI does not equate to a smooth path to the agentic enterprise.

Organizations need this flexibility for a reason: the best tool for one use case today may not be the right fit for another scenario tomorrow. Therefore, rather than trying to standardize everything on a single vendor, enterprises should establish consistent engineering practices, orchestration mechanisms, and governance frameworks across their entire environment.

The goal is to achieve full visibility at the portfolio level and establish a consistent control layer—enforcing policies uniformly across the environment while preserving the freedom to choose the most suitable agents, tools, platforms, and models for each use case.

Business outcomes become the true test

As enterprise AI matures, organizations are increasingly measuring success by the impact on business outcomes, rather than narrowly focusing on activity-based metrics. The research shows that 64% of respondents evaluate AI's impact through error rate reduction, while 60% measure compliance or risk reduction.

Borrêga challenges the view of focusing solely on productivity gains and shifts attention to business outcomes: "Being able to generate more text or more code does not inherently lead to good results." The more relevant question is whether the redesigned agentic process has genuinely improved a business outcome—such as reducing costs, improving customer experience, lowering risk, or helping the organization respond more quickly to market pressures.

Combined with the research findings, it is clear: the next phase of enterprise AI will depend less on how many agents an organization runs and more on how effectively it connects those agents to trusted business systems. As applications themselves become increasingly agentic, organizations need to combine orchestration and governance, enterprise context, technology choice, and outcome-based measurement.

The full 2026 OutSystems and CIO Dive research report is coming soon, where all the data can be explored in depth.