A research mentor for iFood's designers, created when the Research team had three people to support more than 180 designers.
The Research team had three people to support more than 180 designers. Part of that support happened in one-on-one conversations about questions that kept coming up: which method to use, how to structure a study, how to put together an interview guide or review questions.
When I measured how much time these conversations took, each person on the team was spending more than 7 hours a week on this kind of support.
Much of the guidance given in these conversations was already documented in guides, best practices and templates. Even so, to get to that content, people still turned to the Research team.
I started to see this knowledge as something people could access without depending on our availability. That's where the idea for the agent came from: bring together what we had already built and turn that material into guidance closer to the conversation people had with someone from Research.
I created the agent in Toqan, the AI platform iFood uses internally, based on the materials the Research team already had: guides, methods, interview guide templates, best practices and guidance for running studies.
I wanted it to work the way we guided a study day to day. When someone came asking for help putting together an interview guide, for example, the conversation started with context: what the problem was, what decision the study needed to support and who the person wanted to talk to.
I brought that logic into the agent and structured its instructions using the OCANES framework: Objective, to define what the agent should do; Context, with the information it needed to consider; Actions, with what it should do and in what order; Norms, with the rules and limits; Examples, with references for the expected behavior; and Output, defining how each answer should be delivered.
Within this structure, I defined seven pieces of information that guide the conversation: problem, decision, audience, method, objectives, research questions and hypotheses. The person doesn't need to arrive with all of this defined. The agent leads the conversation with questions and prompts that help build these definitions along the way.
The structure also defined what the agent should deliver for each type of support. A research plan, for example, needed to include problem, objectives, method, sample and deliverables. For qualitative interview guides and analyses, I created specific structures following the Research team's practices.
I also set limits for the answers. The agent couldn't fill gaps by inferring information. It had to work with what the person brought to the conversation and with the materials used to train it.
Since launch in August 2025, 121 people have used the agent, adding up to 4,289 interactions. Although it was built for designers, 28% of users came from other areas.
From October to April, usage stayed between 23 and 45 people a month. In that period, on average, 46% of those who used the agent in a given month came back the next month.
During the test phase, 23 people from different BUs and seniority levels rated their experience with the agent.
From the start, the agent advised people to check whether studies on the topic already existed before starting a new one. That search, however, still happened separately in Atena.
In April 2026, we integrated the two solutions. Looking up existing knowledge became part of the research process itself: first understand what iFood already knows, identify the gaps, and then use the Research Agent to structure what still needs to be investigated.
After the integration, the Research Agent started running inside Atena. That's why the data in this case goes up to April 2026. From then on, its interactions became part of Atena's usage metrics.