How an Emory economist turned a parenting problem into a global AI playbook

When Pedro Sant’Anna, an associate professor of economics at Emory University, describes his typical workday, he starts with the parts that have nothing to do with economics. He has five children under the age of 6, including an 8-month-old. He begins work around 10 a.m., gets in a few focused hours, then breaks for lunch with the kids and works again while they nap.

For years, the schedule had a stubborn flaw.

“I have to be in front of the computer all the time,” Sant’Anna says. “Everything is very intense. I have to type.”

Sant’Anna is one of the most cited econometricians of his generation, known for methods widely used across the social sciences that help researchers measure cause and effect in real-world data. Before joining Emory, he spent six years on the faculty at Vanderbilt University, then worked in the tech industry as a full-time economist at Microsoft and Amazon.

These days, much of his research and teaching gets built while he is nowhere near a keyboard. 

He speaks into his phone (sometimes through the microphone of his smart glasses) while feeding the baby or handling other tasks. His instructions are sent to a computer where a team of AI agents writes code, gathers data and drafts course materials. When the machines finish, or get stuck, they flag him for review. Santa’Anna still sets the goals, makes key decisions and verifies the work; AI handles much of the execution in between.

“The beauty of this system is that it doesn’t require me to be in front of the computer all the time,” he says.

A council of research assistants

The first thing Sant’Anna explains about his workflow is that it is not a chatbot conversation.

“Imagine that every time my research assistant has a blocker, he has to come back to my office. That’s ChatGPT,” he says. “It’s never going to work for me.”

Instead, the system works the way he works with graduate research assistants (RAs). He writes a two-page brief describing the goal, then sends it to AI agents that operate autonomously, trying approaches, checking one another’s work and returning only with a solution or a precise account of where they got stuck.

“It’s almost like I have a council of RAs,” he says. “They talk among themselves, they bounce what works, what doesn’t, and once they all agree, they come to me.”

He came to the technology as a skeptic. 

Late last year, a journal asked him to replicate an economics paper from the early 1990s. No data files or code survived, only the well-documented published article. He estimated the task would take a month or two of work for himself, a co-author and an assistant. At a friend’s urging, he wrote up instructions and gave the job to AI agents.

“Three hours later, the results came out, and I’m like, this is very powerful,” he recalls.

What convinced him was not the speed but the answer key. Because the original paper existed, every output could be checked. “If I didn’t have a way to verify the output, I wouldn’t have bought into it,” he says.

Trust nothing, verify everything

That principle anchors everything Sant’Anna has built since. The machines, he is quick to say, fail constantly.

“This happens every day, every hour,” he says. “This is not bulletproof. It is not meant to replace our expertise or be trusted blindly.”

So he sets traps. When his AI agents insisted that two versions of his statistical software matched perfectly, the claim did not pass his smell test. He planted 20 bugs in the code and told the agents to check again. They reported everything was fine. Caught, the system got an overhaul of its verification process.

He compares the practice to a vaccine: a small, controlled dose of error that strengthens the whole system. The deeper shift, he argues, is in where a researcher’s effort now goes. The slowest part of the job is no longer producing the work. It is checking it.

However, when it comes to brainstorming and ideation, Sant’Anna refuses to hand over that task. From choosing what research he will pursue to the content that he will teach in his course, he remains adamant that, “I don’t trust AI’s judgment for that.”

From an Emory classroom to WeChat

The workflow got its first real test at Emory. 

Sant’Anna was preparing a new doctoral course, “ECON 720: Causal Inference with Longitudinal Data,” with a class size three times his usual enrollment and drawing students from economics, business and public health. AI helped him rebuild years of dense technical material into something a mixed audience could use. 

The same rule Sant’Anna applies to his own research carries to the classroom: The person using AI remains responsible for the result.

That principle has also shaped how he teaches. In live demonstrations, he prompts the machines in front of the class, including when they give wrong answers, to emphasize the importance of verification. The data can be found. The code can be written. Every result students submit must be one they have personally verified and signed their name to.

Colleagues kept asking how he did it. Sant’Anna had no time to teach them, so he pulled his favorite move: He had the AI study its own usage logs and write the manual. In February, he published a guide to help people use the technology in their daily lives.

It took off in ways he never planned. The playbook has been copied more than 2,800 times on GitHub and drawn tens of thousands of readers on LinkedIn. Sant’Anna knew it had escaped his orbit when co-authors began texting from China to say it had gone viral on WeChat. “That’s when it hit me,” he says. 

The readers who surprise him most are people with no technical background. That, he says, is the point. 

The guide is free, published under the most permissive license available, and written so that no coding experience is required to use it. In the early days of AI, users had to master the art of crafting prompts; now the systems handle that structure on their own. Getting started with Sant’Anna’s playbook is as simple as copying one prompt from his guide and pasting it into an AI app on a computer or phone.

“I think that’s part of the success,” he says. “The entry point is very low.” The demand, he adds, is not coming from people with deep technical skills; people who have never opened a command line are reading the guide and using it.

Not about doing more

Sant’Anna is optimistic about what the technology means for academia, describing it as an expansion of the “possibility frontier” that lets researchers automate what they enjoy least and protect what they value most. 

Industry is well ahead of universities in adopting these tools, he notes, and he thinks academia should embrace the change deliberately, keeping in mind that AI is not always 100% factual — and guarding against distortions that can occur.

For colleagues wondering whether to follow his path, his advice has less to do with tools than with intention.

“Put your heart into what you care about,” Sant’Anna says. “If you put in the effort, the passion and the work, there’s always someone watching you. And once the opportunity comes, if it is the time, it will take off.”

As for what the machines have bought him, he does not talk about output. He talks about a trade. The hours he once spent chained to a keyboard now belong to his family, and neither his research nor his teaching paid the price.

“For me, it’s not about doing more,” he says. “I can be mentally available and physically available with my family. I can enjoy my small kids growing up. At the same time, I don’t have to compromise my academic and research life.”