What It Takes to Move AI from Experiment to Operations

AI experiments are now easy to start but what happens after the experiment works?

That was one of the main themes of a roundtable at the Avalia Tech Forum 2026 in São Paulo, where leaders from DOT Digital Group and Avalia discussed what it takes to bring AI into real business operations.

DOT is a Brazilian EdTech company working with corporate and institutional clients across sectors including industry, agribusiness and entrepreneurship. Its educational operations work at significant scale, with more than 400 clients and over two million enrolments in a year.

The discussion showed that AI adoption at this scale is less about finding another tool and more about deciding where AI makes sense, how its impact will be measured, and how new ways of working can be adopted across an organisation.

Experimentation was not the problem

DOT had already been encouraging people to explore AI.

Teams were testing tools and finding ways to apply AI to their daily work and employees had space to experiment and develop their own knowledge.

But this also created a familiar problem: different teams could be working on similar solutions without knowing it. An agent could exist in one area while another team was building something almost identical.

So people were experiementing, but there was no way to decide which initiatives should move forward.

As Ana Paula Gutierrez from DOT Digital Group explained during the discussion, the challenge became one of governance: understanding whether an initiative was actually useful, whether it should be scaled and whether AI was making a process better or simply adding another layer of work. This led the teams to start looking at AI initiatives as a portfolio rather than a collection of isolated experiments.

Existing use cases were mapped and assessed according to factors such as the number of people they could affect, their expected value, implementation complexity and investment required. A catalogue was created to give the organisation a clearer view of what already existed and what was worth developing further.

Measuring AI means looking at the process

Educational content is central to DOT's business, but producing specialised material can take considerable time and often requires external subject-matter experts. The team redesigned part of this process using AI.

According to DOT, a content-production activity that previously took around 224 hours was reduced to approximately 48 hours. Output increased from around 1.3 pages per hour to 3.2, while the lead time for that stage of the process fell significantly.

But the interesting part was not simply the use of AI. The process itself changed.

Rather than asking a specialist to create the entire content from the beginning, DOT introduced a content producer role and used AI to support the initial production. A technical reviewer then validated the material afterwards.

What they got was a different operating model, not just a faster version of the old one.

This distinction came up repeatedly during the conversation. Applying AI to an existing process without questioning the process itself can produce very different results from redesigning the workflow around the capabilities of the technology, and sometimes the result is not positive.

DOT found, for example, that AI was improving efficiency in some types of scripting, while in another video-related workflow the person using AI was actually taking longer than before.

Without measurement, both could easily have been described as successful AI initiatives.

The expected ROI is only the beginning

This was also where Avalia's role became particularly relevant. Ana explained that Avalia helped DOT create what the team called calculation models for its AI initiatives: a structured way of defining what an initiative was expected to deliver before scaling it.

That included expected ROI, costs, efficiency gains and the operational benefit associated with each use case.

The important part, however, was continuing to measure the initiative after implementation.

An estimated saving made at the beginning of a project is only a hypothesis. Once AI becomes part of the actual workflow, the organisation needs to see whether those savings appear in practice.

The roundtable highlighted a practical problem here: once a new process is working, people quickly adapt to it. Three hours saved today may become invisible six months later because the employee is already using that time for something else.

Capturing the impact at the right moment therefore becomes part of AI governance itself.

Scaling changes the nature of the problem

A successful pilot can involve a few motivated people who are curious about the technology but scaling the same change across dozens or hundreds of employees is different.

DOT described this as one of its main challenges. People involved directly in the project often understood the new process and became comfortable with it. The next step was getting many more people, with different levels of digital maturity, to work in the same way.

That requires more than training. Processes need to be documented. Knowledge cannot remain with the person who created an agent or designed a workflow. Teams need standards, playbooks and a shared understanding of which tools are approved and how they should be used.

HR also became part of the work. During the project, DOT involved HR alongside technology and education teams, particularly around training, communication and concerns employees had about what AI might mean for their roles.

The discussion made an important distinction: technology can provide the infrastructure and guardrails, but adoption happens through people.

Governance does not have to mean centralising everything in IT

Another recurring question was who should own AI inside an organisation.

DOT's experience suggests that giving technology teams complete control over every experiment can create a bottleneck. Instead, the organisation began working towards a model where technology establishes the boundaries. Within those boundaries, business teams can experiment and build.

The team also created a catalogue of AI agents and a process for reviewing new requests. Before creating another agent, teams could ask whether something similar already existed, whether an agent was really necessary, or whether a simpler assistant or skill could solve the same problem.

The objective was not to restrict experimentation. It was to avoid turning experimentation into unnecessary duplication.

The next AI challenge is operational

The conversation at Avalia Tech Forum reflected a shift that is becoming increasingly important for organisations working seriously with AI.

The first question was whether AI could help. For many use cases, that question has already been answered.

The questions now are more operational: Where does it genuinely improve the process? How should the impact be measured? Who needs to be involved before something moves into production? And how can a successful experiment become a normal way of working without creating new complexity?

DOT's experience shows that the answers do not come from the technology alone.

They come from testing ideas against real work, measuring what changes and building enough structure around AI for useful experiments to become part of the organisation.