Enterprise agents are delivering their clearest returns at developers’ desks, even as governance struggles to keep pace.
An OutSystems survey of 1,879 IT leaders found that 97% are exploring an agentic strategy and nearly half describe their capability as advanced or expert. Yet only 36% have centralised governance, while concern about uncontrolled “AI sprawl” reaches 94%.
The data suggests that software development is both the strongest proving ground and a warning about fragmented oversight.
Developer productivity leads realised value
Cost reduction is the most common expectation for AI, but only 22% say operational efficiency is their strongest result. IT development and productivity lead realised return at 40%.
Generative AI-assisted development is now common across nine of the ten surveyed countries, alongside traditional coding, outsourced work and SaaS customisation. Enterprises are layering AI onto proven development environments rather than replacing the entire stack.
IT operations and data analysis are the most explored uses, at 55% and 52%. These internal workflows provide measurable volumes, technical users and easier containment than customer-facing decisions.
Financial services and technology show the clearest movement into production. Slower sectors can learn from their focus on narrow, high-volume work whose failures remain visible and recoverable.
Integration matters more than a perfect data estate
Forty-eight percent identify legacy integration as the most important capability for expansion, and 38% say legacy systems stall projects between pilot and production.
The survey challenges the idea that every organisation must finish an enormous data cleanup before starting. Agents can work in complex environments if integration and governance improve alongside deployment.
That does not mean fragmented data is harmless. It means teams can address the specific systems needed for a workflow rather than postponing value until the entire enterprise is redesigned.
Trust is rising: 73% express high or moderate confidence in autonomous agents, and 67% trust code or workflows from third-party AI tools. Higher confidence makes technical control more important, not less, because adoption spreads faster.
Central orchestration remains the missing layer
Sixty-four percent lack centralised governance, and 41% rely on project-specific rules. Only 12% use a central platform to manage sprawl.
Human checkpoints are also difficult to implement. Two-thirds say pausing orchestration for review creates technical challenges. A genuine human-in-the-loop system needs more than an approval button; it must preserve context and resume safely after intervention.
Auditability should be designed into the workflow. Logs need to show generated code, tool activity, approvals and responsibility. In regulated or mission-critical settings, these breadcrumbs become evidence.
Central governance does not require one model or development platform. It requires a common inventory, risk classification and control layer across projects. Teams can continue choosing appropriate tools while security and leadership retain a coherent view.
The survey’s strongest lesson is pragmatic. Enterprise AI value begins inside IT because developers can absorb new tools and verify output. That success can seed wider deployment, but only if organisations solve orchestration before hundreds of local projects become unmanageable.
AI-assisted development is no longer an experiment. The next maturity step is ensuring every generated contribution, autonomous action and approval fits a system the enterprise can inspect and govern as one portfolio.
Portfolio oversight can also reveal duplication. Separate teams may buy similar copilots, connect the same repositories and invent different approval rules for equivalent risks. A central view allows successful patterns to be reused and weak projects to be retired. It should not become a committee that approves every prompt. Its purpose is to provide common infrastructure and evidence while keeping delivery close to developers. The balance matters: excessive friction recreates shadow AI, while no coordination turns early productivity into long-term operational debt.
