AI Sprawl Pushes CIOs to Reassert Control Over the Agentic Enterprise — Outsystems - August 26, 2026

The rapid spread of artificial intelligence across the enterprise is creating new technology management challenges as agentic experiments, embedded AI features and independently developed initiatives accumulate faster than organizations can track their costs, architectural dependencies and access to corporate data.

The pattern resembles earlier waves of technology sprawl, from departmental servers and shadow IT, which led to the unchecked adoption of software-as-a-service applications. The difference with AI and agentic systems today is the speed and scale at which they proliferate through organizations. AI tools can be acquired, developed and deployed by individual employees and business units with relatively little technical friction, while consumption-based pricing can turn modest experiments into significant operating expenses.

Gartner research underscores the scale of the issue. The firm predicts that the average global Fortune 500 enterprise will have more than 150,000 AI agents in use by 2028, up from fewer than 15 in 2025, creating what Gartner describes as significant agent sprawl and increased IT and management complexity. For CIOs, the combination is shifting the enterprise AI discussion from experimentation toward management, architecture and economic discipline.

Interview with Tiago Azevedo, chief information officer at OutSystems

“We went past the experimentation, the hype cycle in record time,” said Tiago Azevedo, chief information officer at OutSystems, during a recent BizTechReports executive vidcast interview. Organizations now have “a problem on their hands that they need to manage,” he said.

Azevedo said the proliferation is coming from several directions simultaneously. Major model providers are expanding their platforms, SaaS vendors are embedding AI into existing products, developers are building applications on top of foundation models, and employees increasingly have the ability to create their own agents.

The result is an AI environment that can grow without following conventional enterprise planning and budgeting processes.

“It’s like all of that management complexity multiplied by a factor of 10 or more,” Azevedo said, comparing agentic sprawl with earlier generations of shadow IT.

Experimentation Creates a New Cost Problem

The shift has complicated one of the central assumptions surrounding enterprise AI experimentation. Proofs of concept that never reach enterprise-wide deployment are frequently treated as failures. It would be a mistake, however, to make a binary ‘good or bad’ assessment of projects that fail to scale across the organization. In practice, Azevedo explained, some remain useful within individual departments or teams.

That creates a different management challenge for business leaders and enterprise technologists. While a pilot may never formally enter production across the enterprise, employees can become dependent on AI-enabled processes operating within discrete environments to perform their everyday work. When you multiply that pattern across departments, models and applications, organizations can accumulate a sizable AI portfolio by accident rather than design.

The economics amplify the problem. AI services increasingly rely on consumption-based pricing tied to tokens, compute or other measures of usage. That can make access easy to distribute while making total costs difficult to predict or control.

Azevedo said he has encountered individual instances in which an employee consumed $6,000 of AI resources in a single day. Smaller expenditures can become equally consequential when repeated across thousands of employees over the course of any given year.

The situation is compounded by the widening range of models available. Enterprises can choose among commercial foundation models, open-source alternatives, specialized models and models accessed through cloud platforms. Selecting the most capable model for every task can quickly become economically unsustainable.

“You have to really think through, in order to make it cost-effective, what are you going to choose to build each of the things you want to build,” Azevedo said.

The economics are bringing CFOs more directly into enterprise AI decisions as CIOs face pressure from boards and CEOs to demonstrate AI-driven transformation.

“It creates what I call a three-way squeeze. The CEO and board want to see results from AI. The CFO wants to understand the ROI and where the money is coming from. At the same time, the business wants access to more technology and more tools because it wants to change how the company operates.”

The tension also exposes a growing disconnect between business-led AI adoption and IT-led modernization. This occurs when business teams leverage AI to redesign workflows and customer experiences while technology organizations look for ways to apply AI to software development and infrastructure.

According to IDC, individual teams across different disciplines are deploying agents and automating workflows in an uncoordinated manner, resulting in ‘pockets of progress’ that remain disconnected from core enterprise operations. The fragmentation can produce incompatible AI tools, disconnected data and workflows, rising costs, inconsistent results and delayed returns on investment.

Those initiatives can advance independently even when they ultimately depend on the same data, security controls and budgets.

Enterprise Architecture Returns to the Center

Bringing AI sprawl under control will require CIOs to reapply some of the management disciplines developed during earlier technology cycles while adapting them to an environment in which simply restricting access is no longer practical. Azevedo argues that the goal is to give business teams room to experiment while establishing an enterprise architecture that provides visibility into what is being deployed, how much it costs, what data it can access and how those systems interact.

That puts enterprise architecture back near the center of the CIO agenda because every new agent introduces a series of dependencies that must be understood and managed before it can operate safely within the enterprise.

An agent designed to identify sales opportunities, for example, might need access to a customer relationship management system. That requires an organization to determine what data the agent can retrieve, what actions it can perform and whose authorization applies when the agent acts on behalf of an employee.

Those requirements multiply as enterprises deploy more agents. Organizations may eventually need dynamic catalogs of the tools available to agents, standardized interfaces for accessing enterprise systems, security controls governing those interactions and mechanisms for determining which models should perform particular tasks. In short, an architecture for AI and agentic orchestration must be established to provide the visibility and control necessary to manage the sprawl.

In this context, Model Context Protocol servers, gateways, application programming interfaces and reusable architectural components are becoming part of the emerging AI management layer.

Centralization could also help reduce duplication. Rather than allowing 100 employees to independently develop variations of the same AI capability, enterprises can identify common requirements and build reusable components.

“One emerging role of the CIO is to become the master architect that makes these things possible,” Azevedo said.

That architectural role also changes how organizations approach governance. As AI becomes embedded across applications, workflows and business functions, many of the controls traditionally addressed through policies and oversight must increasingly be designed into the technology environment itself.

Deterministic Systems Meet Nondeterministic AI

The architectural challenge becomes more consequential as agentic systems introduce nondeterministic behavior into environments traditionally built around predictable software.

Unlike the strict ‘if, then’ structure of conventional software, generative AI can produce different outputs from identical or similar inputs. This creates the potential for agents to misinterpret instructions or behave in unexpected ways as conditions change. It is a reality that requires organizations to determine when human oversight is necessary based on the potential consequences of an action.

Azevedo cautioned against assuming that enterprises can make agents themselves deterministic. Instead, organizations can create deterministic boundaries around parts of the environment in which those agents operate.

OutSystems applies that principle to AI-assisted software development. Its Mentor technology allows developers to use natural language to generate or review application components. The generative interaction is nondeterministic. The resulting application model, however, passes through a deterministic compilation process governed by established rules.

The distinction points to a broader enterprise design principle. Organizations may be able to capture the flexibility of generative AI while retaining predictable controls around data access, application behavior and execution.

Those architectural controls, however, do not make the agents themselves deterministic. Organizations still need to account for the possibility that an agent will produce an unexpected result or take an inappropriate action, making the consequences of those errors an important factor in determining how much autonomy an agent should receive.

That is why human oversight must remain a central part of the enterprise architecture, particularly when an agent can take actions that carry financial, operational or regulatory consequences.

Herding Cats

The emerging challenge for CIOs is therefore broader than controlling another wave of technology proliferation. AI is forcing enterprises to determine which capabilities should remain distributed, which should be centralized and where architectural standards can provide enough consistency to support both. When it comes to enterprise-wide AI, the ability to “herd cats” is shaping up to be a critical organizational competency.

The good news is that the industry has seen this phenomenon before. Earlier technology cycles eventually produced disciplines for managing servers, applications, cloud services and SaaS portfolios. Agentic computing is accelerating a similar process while adding variable consumption costs, autonomous actions and nondeterministic behavior.

Organizations that establish visibility into those systems early may gain an advantage that has little to do with possessing the most powerful model.

“I believe that the companies that will win are not necessarily the ones that have the best AI,” Azevedo said, “but are the ones that do it right.”

###

EDITOR’S NOTE: Click here to learn more about OutSystems