CIOs trust their AI governance. Their tool count disagrees. When asked if their governance keeps up with AI adoption, 82% of executives say it does. In the same survey, 72% say they run more than 20 distinct AI tools and platforms.
Both numbers come from the same leaders, but they don’t paint a realistic picture. Governance and sprawl are on two ends of a spectrum, and they may even be on different ones.
OutSystems worked with KPMG and CIO Dive Studio to survey 858 executives in North America, Europe, Japan, India, and Australia about how they run agentic AI. The resulting report, Engineering Business Outcomes: The Shift to Agentic Systems, measures the distance between the control leaders believe they have and the control their environments allow.
Agentic AI has moved into the workflows that run the business, and 73% of organizations use it in risk and compliance operations. That was the most common use case, followed by 69% for customer service, 63% for IT operations, and 63% for finance and accounting.
These agents act across systems and hand work to one another. Gonçalo Borrêga, senior vice president of product management at OutSystems, described the stakes: "An error in one system can lead to 10 errors down the line." This will only grow, say 71% of respondents. They expect enterprise applications to become more integrated, with networks of agents coordinating tasks and handoffs across applications and workflows.
Leaders have reasons for their confidence. In fact, 83% govern agents through established mechanisms such as review boards and centralized policies.
Those mechanisms check an agent before it ships. However, agents keep changing after that. Michael Harper, managing director at KPMG, noted that tweaking a prompt or updating a foundation model "can completely alter an agent's behavior without ever triggering a formal SDLC gate."
Only 5% of enterprises have automated runtime controls and audit trails for their agentic systems. The figure drops to 3% in Europe and 2% in North America. Just 26% of leaders say they are "very well prepared" to resolve conflicts when agents collide in production.
Borrêga ties readiness to design. "Reversibility of an action is the key criterion when thinking about whether you should design escalation as a first-class path versus designing it as exception management," he said. An agent that can issue a refund or update a ledger needs that escalation path engineered in from the start.
Harper offered a startling calculation. "A 98% reliable agent invoked five times in a sequence is actually much less reliable than most executives assume," he said. Five steps at 98% each produce a correct result about 90% of the time. Add more agents and more steps, and the number keeps falling.
He also set a standard for oversight. Teams must be able to reconstruct why an agent made a decision and what context it used. "Without that level of observability, governance is just guesswork," he said.
Every AI tool brings its own permissions model and its own logs. Agents built in separate tools rarely see one another's actions. Borrêga views this situation as troubling for leaders. “We see in a lot of enterprises that they're building agents in silos, so managing conflicts that happen over real transactional data is a hidden problem that I would say many enterprises aren't even aware of," he said.
The survey shows the strain. Security and policy enforcement across agents is a leading orchestration hurdle for 67% of respondents. And 36% describe their approach to deciding where to use agentic AI as fragmented.
Sprawl ran highest in technology and insurance organizations. Organizations with fewer tools still carry the overhead of governing agents one tool at a time, and that overhead makes costs and interoperability harder to manage. A review board can approve each of those tools. It can't watch what happens when agents from five of them act on the same customer record.
The leaders closing this gap build governance into how their agentic systems run. Consistent policy enforcement across the AI ecosystem is a top priority for 87% of respondents, and 81% say the same about testing and validating agent behavior.
Context is part of control. Grounding agents in it with governance and control is a priority for 73% of respondents, and 74% rate integration with existing apps and systems as very or extremely important. Agents need live business data, with access bounded by the identity and permissions of the person or process they act for.
Harper described the target as a “central platform that makes the safest path the easiest path, giving business units the autonomy to deliver value securely.” The report turns that idea into a four-part action framework. It shows how leaders move from procedural oversight to runtime control while keeping their choice of models and tools.
Read Engineering Business Outcomes: The Shift to Agentic Systems to learn more about this framework.
Join OutSystems and KPMG on November 3, 2026, for The Control Paradox: Governance and Runtime Execution for AI ROI, a live discussion of what the findings mean for your agentic roadmap.
