Written by: Nubank Editorial

Seeing an AI system do something impressive is one thing. Knowing where that capability can make a meaningful difference in your own work is another.

That gap matters when organizations move from experimenting with AI to applying it in everyday workflows. In this context, a demo can show what a technology is capable of, but identifying a useful application requires something else: enough understanding of how the technology works, combined with deep knowledge of the problem being solved.

This is one of the ideas behind the Enablement pillar of Nubank’s Ops AI Acceleration Program. Its mission is to give teams practical access to AI knowledge and tools while keeping that learning connected to the realities and challenges of their work.

A recent applied AI workshop with PJ Operations put that approach into practice. Rather than beginning with a ready-made AI solution, the session built a common foundation first and then asked participants to apply what they had learned to a real operational challenge.

The goal was to help bring technical understanding and operational expertise closer together.

Creating a shared language for AI

Understanding AI does not mean knowing how to build a model from scratch, but when people understand some of the concepts behind the technology, they can ask better questions about where it might be useful, what it would need to work, and how it could fit into an existing process. That is precisely why the workshop began well before the practical exercise.

The learning journey started with the broader acceleration of computing before moving into generative AI and agents. From there, participants explored how AI is becoming part of the software development lifecycle and how different building blocks can work together in modern AI systems: models, tools, context, skills, agents, and Model Context Protocols, or MCPs.

The purpose was to make the architecture behind AI systems less abstract. A model, for example, is only one part of what makes an AI application useful. Context helps it understand what is relevant to a particular task, while tools allow it to perform actions or retrieve information beyond what the model already knows, and skills can provide specialized instructions for recurring tasks. Agents, on the other hand, can combine reasoning and actions across multiple steps, while MCPs provide a standardized way for models to connect with external tools.

Knowing these distinctions changes the conversation, because instead of asking simply, “Can AI do this?”, teams can start asking more specific questions: What information would the system need? Which parts of the process require judgment? What tools or data would it need access to? Where could human expertise remain essential?

That shared vocabulary helps create a bridge between technological possibilities and the reality of operational work.

Check our job opportunies

Where Engineering meets Operations

AI adoption becomes more practical when two kinds of knowledge can meet. Engineering brings an understanding of what technology can do, how systems can be structured, and where technical constraints exist. Operations brings something equally important: detailed knowledge of how work actually happens.

Operational teams understand the exceptions, handoffs, recurring bottlenecks, business rules, and judgment calls that may not be obvious from the outside. They know which activities consume significant time, where information becomes difficult to interpret, and which parts of a workflow are repetitive without necessarily being simple. That context is critical for identifying useful AI opportunities.

The objective of enablement, then, is to create enough common ground for both perspectives to connect. When operational expertise is paired with a better understanding of AI capabilities, people can move beyond thinking about AI as a general-purpose technology and begin looking at individual workflows differently.

What part of this process requires analysis? What is repetitive? Where does someone spend time gathering or interpreting information? Where might an AI system assist without removing the human judgment the process depends on?

Those questions are much closer to real work than a generic demonstration of what a model can generate.

Putting the learning to work

On August 25, 15 Nubankers from PJ Operations and leadership participated in an applied AI workshop built around this approach. After exploring the technological foundations, the session moved into a hands-on challenge focused on the Performance Management Lifecycle. 

The choice of problem was important. Instead of experimenting with an artificial exercise created to showcase AI capabilities, participants worked on an existing challenge within PJ Operations: the quantitative and qualitative analyses involved in performance management.

This includes activities such as root-cause analysis and evaluating BPO performance, a work that requires teams to interpret information and turn it into insights that can support operational decisions.

The workshop received an overall satisfaction score of 4.8 out of 5, but the more important outcome of the session was the opportunity to connect the concepts participants had just explored with a problem they already understood deeply.

Starting with a real operational problem

Performance analysis is a significant workload within Ops PJ. Today, the quantitative and qualitative analysis involved in this workflow consumes approximately 100 hours per month, and that number provides a useful starting point for thinking about AI.

The opportunity is not simply to “add AI” to performance management. It is to examine the workflow and identify where AI capabilities could reduce manual effort and make the analysis process more agile, and this is where operational knowledge becomes fundamental.

People who work closely with the process can distinguish between tasks that are repetitive and tasks that require contextual judgment. They can identify where information must be gathered, compared, interpreted, or investigated. And they can help determine where AI assistance might be valuable without assuming that the entire workflow should be automated.

Technical understanding makes it easier to imagine what is possible. Operational understanding helps determine what is useful.

From ~100 hours to a potential ~25

For the Performance Management Lifecycle use case explored in the workshop, the expected opportunity is significant. The current workload is approximately 100 hours per month. With AI applied to parts of the workflow, the target is to reduce that effort to approximately 25 hours per month.

It’s important to mention that this figure is a projection, not a measured result. The workflow has not yet demonstrated a reduction from 100 to 25 hours in production. Instead, the estimate represents the potential efficiency identified for the use case and provides a concrete target against which future experimentation can be evaluated.

AI enablement, at the end of the day, is about helping teams identify opportunities grounded in real work, formulate hypotheses about where AI can help, and create a clearer path toward testing those hypotheses.

In this case, the projected reduction provides something concrete to investigate. The value of the exercise lies not only in the number itself, but also in the process that produced the opportunity: beginning with an operational problem, understanding the work behind it, and then examining where AI capabilities could fit.

Enablement as an ongoing capability

A workshop can introduce concepts and a tool can make experimentation easier, but long-term AI enablement depends on something broader: giving people enough understanding and practical experience to recognize opportunities for themselves.

That means moving beyond teaching people how to use a particular model or interface: tools will change, models will evolve, and new ways of connecting AI systems to data, applications, and workflows will continue to emerge.

The more durable capability is knowing how to look at a real problem and ask the right questions. Where is time being spent today? Which parts of the work depend on information synthesis or analysis? Which constraints matter? Where could technology assist? And how would we know whether an experiment actually improved the process?

For Operations, this makes technical knowledge actionable rather than abstract. For Engineering, operational context helps anchor technological possibilities in problems that matter.

Bringing those perspectives closer together is what turns AI enablement from learning about a technology into learning how to apply it, and that may be one of the most important steps in moving from an impressive AI demo to a meaningful improvement in the way work gets done.

Check our job opportunies