This post is the first of our new series of General articles where students, researchers and development stakeholders can share their experiences, insights and ideas around development and research. If you want to contribute, please contact us.
Francisco Flores is the co-editor of Economics That Really Matters blog and a PhD candidate at Cornell University’s Dyson School. His interests lie in agricultural economics and risk management. In this article he shares how AI agents assisted him in a recent fieldwork project.
Leading fieldwork for the first time is challenging for many economics students. The processes and skills it require are not usually taught in a standard program. Still, a fieldwork project is a great opportunity to answer a relevant research question when no adequate data exist to address it. In addition, many researchers value the qualitative information gained by observing the dynamics of the field site and talking with stakeholders.
A common way to acquire fieldwork skills is to work as an assistant on projects led by others. Other people start their own projects with no prior experience, guided by their advisors. That was my case, and I want to share the challenges I faced and the lessons I learned along the way. I will do this in 2 parts. This first part will describe how I used AI to customize and enhance my data collection. I hope it is useful for anyone thinking about starting a project who wants to map the obstacles they might meet and learn how others handled them.
The research question
To give context for the fieldwork, I start with the goal of the research project. The project had two motivations: (i) the limited use of agricultural insurance in developing countries, and (ii) anecdotal evidence that farmers do not buy insurance because they see the premium as a loss when no disaster occurs and they receive nothing in return. With these motivations, the research question was: Is insurance adoption limited, in part, by the perception of the premium as a loss when no disaster occurs?
The instrument
The method chosen to address this question required generating data through a lab experiment. In the experiment, participants made decisions about adopting risk-management tools under different conditions. In the first stage, they were given the context of the decision: they were told they were in charge of running a farm and were shown its production conditions. The only random element in these conditions was the weather, which could be good or bad with certain probabilities, and each state translated into a given profit.
In the second stage, the participant answered a series of questions. In each question, they were told the probability of the good state and were shown a risk-management tool. The participant had a few moments to review the tool’s features and the risk and then decided whether to take it. After the decision, a state of nature was drawn at random according to that question’s probability. Once the state was revealed, the profit was determined. In the third stage, one question from the second stage was drawn at random, and the profit from that question determined the participant’s compensation. This design should align incentives by encouraging participants to answer every question trying to earn the highest possible profit.
Building a survey in the age of AI agents
The lab experiment was applied with a survey in tablets and the user experience (UX) was highly appreciated by the enumerators and participants. This would not have been possible without the advances in AI in early 2026. For the first pilot, in late 2025, I used a survey programmed in oTree, a Python package for building behavioral experiments. One of its advantages is that it is free, but at the time I did not like its look or its workflows. Improving both would probably have required more programming, which would have taken more time to learn and implement. At that point, with the help of ChatGPT, I put together a version that was acceptable for the pilot.
Then, in early 2026, I started using Claude Code. Since one lesson from the first pilot was the need for a design that held the user’s attention, I set out to build a survey that was visually appealing, had a good balance of text, images and diagrams, and reflected as well as possible the context and nature of the decisions participants had to make.
To request a survey with these features, I first had to share with Claude Code a document describing the experimental design and what I wanted on each page of the survey. My prompts can be summarized as follows: “Read the experimental design document. Then review survey manuals for developing countries from institutions such as the World Bank, the IDB, universities, and any other institutions you consider relevant. From these manuals, identify the features a survey conducted with farmers in rural Peru should have. With those features in mind, check whether the survey plan I gave you is adequate and improve it accordingly.”
After several iterations, I was pleased with the result. I had a survey that was visually appealing and easy to understand. Claude suggested:
- illustrations to reinforce understanding of the experiment
- intuitive visual designs to represent probabilities
- tables and colors to show the trade-offs of each decision
- a balanced amount of text on each page
- aesthetic choices such as typography, colors and a progress bar
A second important contribution from AI agents was an app for tablets and phones. The app stored the survey, could be installed by scanning a QR code, and synced the data to a cloud server in a few simple steps. Claude Code also guided me through setting up that cloud server, importing the data into the local server I used, and converting the data into a format more familiar to economists.
I should clarify that this process did not take a few hours. It took several days of iterations and reviews on my part until I was satisfied with the survey experience and had confirmed that the whole process worked without failures, from taking the survey to receiving the data on the final server. But I am sure that if I had tried to learn how to do these steps from scratch, it would have taken weeks, if not months. I could have hired well-known data-collection firms, but I don’t think I would have had the same flexibility to get a survey with the user-experience features I wanted.
Final reflection
Carrying out a fieldwork project with no prior experience is challenging, but it doesn’t have to hold you back. You can start by making efficient use of the resources and tools you have at hand. For tasks outside our expertise that are still essential to the quality of the experiment, we can lean on the AI tools available today. AI agents can now cut the time and cost of tasks such as designing a survey or running data-management protocols, while also making those tasks more tailored to our needs.
