AI Sprawl Forces Enterprises to Rethink How They Build, Fund and Manage Agentic Systems — August 31, 2026
A Q&A with Outsystems CIO Tiago Azevedo
The adoption of generative and agentic AI is creating a familiar enterprise technology problem as organizations experiment with models, build agents and adopt AI-enabled applications faster than traditional technology management processes can evaluate their costs, dependencies, security requirements and potential overlap with other initiatives.
For CIOs, the challenge is increasingly shifting toward creating an environment where AI experimentation can continue without allowing fragmented technologies and unpredictable consumption costs to overwhelm enterprise architectures and budgets. In this BizTechReports executive Q&A, Tiago Azevedo, CIO of OutSystems, discusses why AI sprawl is emerging, how the economics of AI are changing the conversation with CFOs, why traditional approaches to technology governance are insufficient and how enterprise architecture can provide a foundation for scaling agentic systems.\
Here is what he had to say:
Q: Technology sprawl is hardly new. What makes the proliferation of AI and agentic systems different from earlier waves of shadow IT, SaaS and cloud adoption?
A: The market is hugely fragmented. Every SaaS vendor is including an AI component in its product. Then you have companies such as Anthropic, OpenAI, Google and Meta with large models and AI capabilities, and you have a huge number of products that extend those technologies or rely on them.
At the same time, everybody is experimenting because everybody is trying to learn. That creates pockets of technologies and use cases throughout the organization. Some experiments move into production. Others remain within individual departments. Meanwhile, new experiments continue to emerge elsewhere.
The other major difference is how accessible it has become for teams and individuals to build agents themselves. In the past, you might discover a server underneath someone's desk or find an application operating somewhere independently. Then SaaS arrived, and people could buy applications themselves. You might discover them when someone eventually came to IT asking for an integration. With agents, you have all of those dynamics occurring at a much greater scale.
Q: Does that mean enterprises should reconsider the assumption that an AI proof of concept has failed if it never reaches enterprise-wide production?
A: Absolutely. Sometimes something does not go into production across the entire enterprise, but it remains deployed within a particular department because that group finds it useful and uses it every day.
That creates an interesting management issue. When do you extend it to everyone else? Do you need to? And if you decide you no longer want it, how do you shut it down when people have incorporated it into their work?
That is one reason these environments become difficult to manage. You can end up with many useful departmental capabilities that never become enterprise applications, but organizations still need lifecycle visibility, governance, cost control and a way to scale, contain or retire these use cases. .
Q: At what point does widespread experimentation become an enterprise management problem?
A: We have seen that change very quickly. In a matter of a quarter, I went from rooms full of CIOs where perhaps 60% were experimenting to rooms where 80% were raising their hands and saying, essentially, “The sprawl is killing me. I need to take control of this.”
Particularly in large enterprises, there is usually enough money available to conduct individual experiments. The problem comes when you add all of those experiments together. You suddenly discover that you have created an entire technology stack that the CIO now has to account for, and it may not have been included the last time the organization prepared its budget.
Q: Why has the cost of AI become such an important part of the sprawl discussion?
A: Because it can become unaffordable very quickly if you do not make deliberate choices.
There are many models available, and they are appropriate for different things. You have to think carefully about which model you use for each task. That can range from optimizing an email or summarizing meetings to generating code or supporting an agent that performs business activities.
If the approach is simply to use the best or most powerful model for everything, the economics become very difficult. You need to manage the engine at each step and understand what you are using, why you are using it and whether the investment is producing sufficient value to justify the cost.
Q: You have described the early phases of AI adoption a bit permissive. What happened when enterprises allowed that experimentation to spread unchecked?
A: Yes. There was this ‘let 1,000 flowers bloom’ approach that we saw clearly in late 2025 and early 2026. What happened is that it just blew up the budgets of every company almost overnight. By the end of the first quarter, a different AI investment reality had taken root at the enterprise level.
One of the problems is that people did not necessarily understand the economics behind what they are consuming. They may use an AI service at home under a subscription and assume the economics work the same way inside an enterprise.
They did not, and do not. In an enterprise environment, consumption can become significant very quickly. I have seen situations where one person spent $6,000 in a day. Sometimes someone spends $300 and thinks that is not very much. But multiply $300 across 2,000 or 3,000 people and then across 12 months, and you begin to understand the problem.
Everybody thinks they are solving an important problem, but they may not realize that many other people across the organization are trying to address the same issue at the same time. Those resources might be better allocated to a common solution or a higher-priority initiative.
Q: How does that change the relationship among the CIO, CFO, business leadership and the board?
A: 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 CIO remains accountable for technology governance, costs and cost optimization. That becomes difficult when these initiatives are emerging from many different places and when consumption can increase much faster than traditional technology budgets.
Q: There also appears to be a disconnect between business-led transformation and IT-led modernization. How does that contribute to the problem?
A: You can have two very different conversations happening simultaneously. On one side, IT is looking at how AI can optimize the software development lifecycle. On the other, the business is thinking about how AI can change the way the company works.
Those conversations may be happening in different rooms or even different time zones. The business selects technologies to bring its ideas and use cases to life in an agentic format, while IT is applying similar capabilities to software development and technology operations.
Eventually, those worlds have to come together because they rely on enterprise data, systems, integrations, security controls and budgets.
Q: If simply blocking access is no longer realistic, what principles should CIOs use to bring that environment under control?
A: “Blocking” is almost a word I forbid my team to use. In the past, IT could simply prevent access to something. That approach does not work today because if you prevent people from using these technologies, you risk having the company fall behind competitors that are using them.
The answer is to make AI available within a managed environment. Some degree of centralization comes back into the picture. Instead of having 100 people independently trying to solve the same problem, perhaps you dedicate resources to solving that problem for those 100 people. You can also provide architectural components, best practices, modules and reusable pieces that people can incorporate into what they are building.
That allows experimentation to continue while reducing unnecessary duplication.
Q: Why does enterprise architecture become so important as organizations deploy more agents?
A: Consider something relatively straightforward. You build an agent that looks at sales opportunities and reminds sales representatives about opportunities they should pursue.
Where are those opportunities? Perhaps they are in large SaaS vendors’ database. Now you have to determine how the agent gets access to that data. You need tools that allow the agent to access the information it needs and other tools that allow it to perform authorized actions.
As you build more agents, you eventually need a catalog of those tools. You may have MCP servers providing access to systems. You need security controls. You need to determine whether the person behind the agent has authorization to access the information the agent is requesting.
Before long, you have multiple agents using multiple tools, security controls and models. All of those components have to be orchestrated.
One emerging role of the CIO is to become the master architect that makes these things possible.
Q: Does that also change how enterprises should think about AI governance?
A: Yes. You need to think about the architecture surrounding the agent, what it is permitted to access and what it is permitted to do.
Agents access company data. They may insert information into systems. They can perform actions and correlate data across systems. You have to make sure the appropriate rules are applied to those interactions.
That is why core enterprise architecture becomes very valuable again. You want people to have access to AI and you want them to innovate, but you also need an architecture that makes those activities manageable.
Q: Agentic systems are inherently nondeterministic. How can enterprises reconcile that with applications and processes that require predictable behavior?
A: You have to distinguish between the nondeterministic AI interaction and deterministic parts of the application environment.
For example, with OutSystems Mentor, natural language can be used to help generate application components. That natural-language interaction is nondeterministic. But the code generated from the OutSystems model is deterministic because it goes through a compiler and established rules.
It’s an important distinction. AI can help people build applications and improve as the models evolve, but there are parts of the environment where you still want deterministic rules and predictable controls to apply.
Q: Where does human oversight fit when the agents themselves cannot be made deterministic?
A: The agents are not going to be deterministic. We need to understand that sometimes agents will make mistakes.
If an agent can make a very dangerous mistake, you should include a human in the loop. The amount of human oversight should depend on the problem you are solving and the consequences if something goes wrong.
We also have to recognize that there are situations where agents may make fewer mistakes than humans. In those cases, an organization may benefit from having agents perform a specific activity while humans focus on higher-value work or on managing and overseeing what the agents are doing.
Q: What ultimately separates organizations that successfully scale agentic AI from those that simply accumulate more AI technology?
A: After what we have learned over the last three years, I believe the companies that will win are not necessarily the ones that have the best AI. They will be the ones that do it right.
Doing it right means having a structured approach to how you build agentic capabilities into the enterprise. You need to understand the agents, models, tools, data access, security controls and architectural components involved and how they work together.
We are moving beyond the period when experimentation itself was the objective. CIOs and organizations increasingly have something real that they have to manage. The organizations that recognize that transition and put the appropriate structure around AI will be in a much better position to turn those experiments into sustainable enterprise capabilities.
###