AI does not replace decisions; it reduces uncertainty

The real impact of AI shows up before the incident is resolved

Throughout the first series of articles, we tried to understand why increasingly sophisticated IT operations still struggle to respond quickly to incidents. We discussed how rising complexity increased the volume of available information, showed that monitoring more does not mean understanding better, explored the importance of observability as a new way of operating, and concluded that reducing MTTR depends, above all, on how fast we can build a shared understanding of the problem.

In the second series, we moved this discussion to Artificial Intelligence. First, we argued that specialized agents, on their own, do not solve the main challenges of modern operations, in “AI agents do not solve the main problem of IT operations”. Then we showed that any agent depends on shared operational context, in “Before agents, there is operational context”, that specialization does not replace a systemic view, in “An agent can know everything about a tool. Still, it may not understand the incident.”, and that this context is an asset unique to each organization, in “The next generation of AIOps will be built on operational context platforms”.

This sequence naturally leads us to a new question.

If we already understand that specialized agents depend on consistent operational context, what is, in the end, the real role of Artificial Intelligence within this architecture?

The most common answer is that AI will be used to automate tasks, replace repetitive activities, or even make decisions in place of operators. All of these applications are relevant and will certainly keep evolving. However, when we observe the dynamics of a large IT operation during a critical incident, we realize there is an even more important contribution.

Artificial Intelligence reduces uncertainty.

At first glance, this statement may seem abstract. In practice, though, it helps explain why some organizations manage to diagnose and resolve incidents significantly faster than others, even when using similar tools and having equally qualified teams.

Because, before restoring a service, every operation needs to answer a very simple question: what is really happening?

It is precisely in that interval, between the appearance of the first signals and the construction of a reliable explanation for the incident, that most of the time tends to be consumed. And it is in this phase that Artificial Intelligence can have its greatest impact.

Incident time is consumed by uncertainty

Whenever a major incident happens, there is a natural tendency to imagine that the MTTR clock is being consumed by executing the technical actions needed to restore the service. We think about restarting applications, rolling back deployments, expanding computing capacity, or replacing faulty components. These activities, of course, are part of the recovery process.

But anyone who follows critical operations knows they are rarely the longest stage. Long before the first corrective action is executed, there is a period in which the organization is still trying to understand the problem it needs to solve.

Teams start gathering evidence. Different tools record seemingly disconnected symptoms. Infrastructure specialists analyze servers and computing resources. Observability professionals investigate metrics and traces. The database team checks queries, locks, and processing queues. Networking looks for latency changes. The Service Desk tracks the growth in tickets. In parallel, managers try to estimate the impact on the business.

All of these activities have a common goal: to reduce uncertainty.

Each new piece of evidence eliminates some hypotheses and strengthens others. An alert initially considered the main cause turns out to be a consequence of another event. A change that seemed irrelevant gains importance when related to new information. Little by little, what looked like a disordered set of symptoms begins to form a coherent narrative about what really happened.

This process is much more than a simple technical investigation: it is a continuous process of uncertainty reduction. And understanding this may be the most precise way to grasp the real role that Artificial Intelligence will play in modern operations.

Every investigation is a process of eliminating hypotheses

There is a characteristic present in practically every operational investigation that rarely gets the attention it deserves.

When an incident begins, nobody has all the answers. Teams initially work with hypotheses.

Perhaps the problem was caused by a change made a few hours earlier. Perhaps it is related to a degradation in the infrastructure. It could be the consequence of an outage in an external service, an unexpected increase in load, a failure in a network component, or even a combination of small events that, on their own, would pose no risk to the operation.

In the first few minutes, all of these possibilities seem equally plausible. As new evidence emerges, some of them stop making sense.

Others come to explain the observed behavior better. The work of the operation consists precisely in continuously narrowing the set of possible hypotheses until only the one that best represents the cause of the incident remains.

This perception changes how we understand the diagnostic activity itself. Investigating an incident does not just mean looking for answers; above all, it means eliminating incorrect explanations.

Each log analyzed, each metric queried, each change reviewed, and each relationship identified between architecture components helps reduce the space of uncertainty the operation has to work in.

When we look at this process this way, we realize that the quality of the diagnosis depends not only on the experience of the specialists or on the amount of information available.

It depends on how fast we can turn scattered evidence into reliable knowledge. This is exactly where Artificial Intelligence starts to have an impact different from the one usually associated with automation.

The real role of AI is to accelerate understanding

Much of the discussion about Artificial Intelligence in operations focuses on task automation. People talk about agents able to open tickets, run playbooks, carry out recovery procedures, or trigger teams automatically. All of these applications have enormous value and will keep evolving.

However, before any of these actions happens, there is a much more important decision: the operation needs to understand the incident.

Without that understanding, any automation risks acting on symptoms instead of on the cause of the problem.

That is precisely why the main contribution of Artificial Intelligence is not necessarily in executing corrective actions, but in its ability to accelerate the construction of that understanding.

By correlating events from multiple sources, retrieving operational context, identifying recurring patterns, establishing relationships between recent changes and observed behaviors, highlighting relevant evidence, and eliminating less likely hypotheses, AI significantly reduces the cognitive effort required from teams.

Artificial Intelligence does not eliminate the need for investigation; it makes investigation more efficient. In other words, it does not replace the experience of specialists but amplifies it. It does not make decisions in isolation but creates better conditions for decisions to be made with greater confidence and in less time.

This difference may seem merely semantic, but it has profound consequences for the architecture of modern operations.

The shorter the time needed to reduce uncertainty, the sooner the organization identifies the likely cause of the incident. The earlier that understanding is built, the earlier corrective actions can begin.

That is why, in many situations, the greatest gain provided by Artificial Intelligence happens even before the first technical intervention.

AI does not directly reduce MTTR, but it reduces the time needed to understand the incident, and so MTTR drops as a consequence.

This relationship explains why organizations that use Artificial Intelligence more maturely do not necessarily automate more procedures than the others. They simply manage to turn information into understanding much faster.

It is precisely this capability that sets apart an operation that reacts to symptoms from one that quickly understands the dynamics of the incident and directs its efforts to what really needs to be solved.

Less uncertainty means less time to resolve incidents

Over the last few years, it has become common to associate the success of Artificial Intelligence with its ability to automate tasks. That perception is not wrong. In fact, we will see agents increasingly prepared to carry out operational activities, support processes, interact with systems, and speed up countless routines that still depend on human intervention today.

However, when we observe how large incidents are actually handled, we realize that the greatest value of Artificial Intelligence appears even before automation, at the moment the operation needs to understand the problem. That is exactly the idea we tried to develop throughout this article.

Every investigation begins surrounded by uncertainty. Several hypotheses compete with one another, different teams analyze complementary perspectives of the environment, and new evidence keeps emerging. Each relevant piece of information reduces part of that uncertainty until the organization can build an explanation consistent enough to guide decision making. It is precisely in this process that Artificial Intelligence shows its greatest potential.

By relating information from multiple sources, retrieving operational context, identifying patterns, and presenting relevant evidence, it significantly reduces the time needed to turn scattered data into shared understanding.

This may be one of the most important changes in how we see AI applied to operations.

For a long time, we imagined that its main goal would be to replace operators or to execute procedures automatically. Today we are beginning to realize that its greatest contribution is allowing people and agents to understand more quickly what is happening.

When that understanding comes earlier, the entire incident response sequence becomes more efficient. The team chooses more precisely which hypotheses to investigate, directs its efforts better, avoids unnecessary actions, and starts technical recovery much sooner than it would in a scenario of high uncertainty.

That is why reducing uncertainty and reducing MTTR are part of the same process, and not because Artificial Intelligence directly performs the service recovery, but because it reduces the time needed to discover which recovery should actually be executed.

This distinction seems subtle, but in practice it completely redefines the role of AI within modern operations.

The intelligence of an operation will not be measured by automation

For many years, operational maturity was associated with the volume of monitoring, the number of dashboards, the amount of automation implemented, or the speed at which technical procedures run. All of these elements remain important and will keep evolving.

But they no longer seem sufficient to explain why some organizations respond to incidents far more efficiently than others.

Throughout the latest articles, we showed that specialized agents depend on context, that this context needs to be shared, that it is an asset unique to each organization, and that context platforms take a central position in the architecture of intelligent operations.

This article adds one last piece to that reasoning: Artificial Intelligence produces value because it reduces uncertainty. And the lower the uncertainty, the shorter the time tends to be to understand the incident, make consistent decisions, and begin restoring the service.

Perhaps that is precisely why the intelligence of an operation should not be measured only by the degree of automation it has reached.

It should be measured, more and more, by how fast it can turn scattered information into shared understanding, because, in the end, every operational decision is a response to uncertainty, and every consistent reduction in MTTR begins long before the technical execution, that is, it begins at the moment the operation starts to understand, with confidence, what is really happening.

From vision to practice: ARGUS

One of the biggest challenges of IT operations is not the lack of information, but the difficulty of understanding what is happening when different systems, services, and tools present seemingly disconnected signals. This fragmentation increases uncertainty and requires specialists to rebuild the context before making decisions.

We developed ARGUS precisely to tackle this problem. Through its AIOps capabilities, the platform connects information from monitoring, observability, ITSM, infrastructure, and other sources, correlating events and relating components, services, changes, and dependencies in a shared operational view.

With this foundation, teams and automations can investigate incidents from more consistent information, reducing the effort needed to interpret isolated signals and identify the likely causes of problems.

The evolution of ARGUS to incorporate AgentOps and AI Gateway expands this proposal. In addition to making context available to AI agents, the platform now includes mechanisms to orchestrate, observe, and govern their actions, creating conditions for people, automations, and agents to take part in the same operation in a coordinated way.

The goal is not to replace specialized tools or existing agents, but to connect their contributions and reduce the uncertainty that arises when each participant works with a partial view of the environment.

Learn about ARGUS and see how it works in practice.

About the author

Paulo Florêncio de Menezes is an entrepreneur, partner, and co-founder of companies such as Target Solutions, BiPTT, and ARGUS.

He holds a degree in Data Processing from the Federal University of Amazonas (UFAM), an MBA in Business Management from Fundação Dom Cabral (FDC), and a Professional Master’s in Economics from Fundação Getulio Vargas (FGV).

In his Master’s research, he developed the IBEOP-IA (Brazilian Index of Potential Occupational Exposure to Artificial Intelligence), investigating the potential exposure of Brazilian occupations to AI capabilities by combining occupational data, artificial intelligence patents, and natural language processing techniques.