Research

Advice only helps if people use it. We ask how algorithms should guide multi-step decisions so the advice helps now and still leaves people better on their own.

Our Management Science paper uses reinforcement learning to read logs of people's step-by-step choices and turn the costliest gap from a strong policy into one short tip, which we then tested in a virtual-kitchen game. Another paper asks why people in the same game pass over tips that are known to be good. In experiments on electric-vehicle charging routes, we compare an exact next action with a broader rule of thumb and ask what people do once the advice is gone.

We are also developing compliance-aware reinforcement learning for managing human agents.

We study how professionals actually use generative AI and AI assistants, what it changes about their work, and how much authority AI agents should have.

Much of this work starts with teachers. We followed K-12 teachers as they used ChatGPT on their own tasks and sorted the kinds of help they asked for, and a year-long case study looks at how generative AI changes their productivity and the tasks they choose. We also model how a professional advisor responds when clients can consult a personal AI.

Newer work looks at where an AI assistant's memory of its user breaks down, when an AI agent in a service workflow should act, ask or defer, and how AI changes which tasks people plan and which they carry out.

We study how workers build skill alongside app recommendations, and how platforms can assign work so people improve rather than just cover today's demand.

So far most of this work comes from gig and delivery platforms. Using more than a million delivery orders, we follow how couriers' speed and punctuality change with experience and how their use of the app's job recommendations shifts. We also ask how a platform should offer jobs when workers can turn offers down and improve at rates the platform can't see in advance, and we test learning algorithms for this on delivery-platform data.

Dispatching to Develop (SSRN, 2026) is our newest paper on this question, and a new project asks whether skill learned in one job carries over to another.

We study when and where flexible workers choose to work, and how platforms can pay, motivate and staff them in competitive labor markets.

Much of this line is about ride-hailing. We model when drivers start and stop working, looking at pay alongside habits such as earnings targets and the momentum of a long shift. With one platform's records and public New York City trip data, we also estimate how drivers split their time between competing apps and compare pay designs such as hourly guarantees, streak bonuses and wage floors.

Current projects look at how to design and sequence incentives for platform workers and how to staff fluid teams in a fragmented labor market. The line goes back to Park Sinchaisri's 2021 doctoral dissertation at the University of Pennsylvania, on managing the gig economy.