Frontier research on humans and AI
The Berkeley Operations and Behavioral Analytics Lab (BOBALAB) at UC Berkeley Haas runs experiments with real workers and companies to see how people use AI on the job, and builds tools that make them better at it.
Our alumni have gone on to graduate programs and firms like these
Alumni
From field data to the decisions people make every day
Research
We work with data from gig and delivery platforms, study how people use AI assistants and agents, and run experiments to see what actually changes behavior.
Partnerships
We run pilots with platforms, employers and public agencies, starting from one question and ending with tools they keep using.
Apps assign gig work, algorithms give advice and AI agents take on service tasks. Using field data and experiments, we study how people learn alongside these systems and how to design them better.
Platforms, employers and public agencies work with us on decisions people and algorithms make together. We scope one decision with your team, test changes on your own data, and leave you with tools you keep.
How a pilot starts
One decision, one question, tested where the work happens.
A partnership, step by step
Scope the decisionPick one recurring decision and what a better outcome would mean.
Pilot on your dataModel the decision with the data you already collect.
Field experimentRandomize the change and measure it in the field.
Tools you keepGuidance, dashboards or policies your team runs after we leave.
What working with us looks like
Four steps, from a question to tools your team keeps.
AmazonRouting, delivery and warehouse work, where algorithms plan the day
Publications
Recent papers from our lab
Proceedings of the ACM on Human-Computer Interaction, 9(7), 1–35, 2025. Draws on more than a million delivery orders to show couriers get faster and more punctual with experience. Newcomers tend to follow the app's job recommendations, while seasoned workers increasingly go their own way.
Learning on the Job
Learning on the Go
Submitted.
Can you move my delivery to Friday?
Checking what I can change...
[ask] Friday has a 2 to 6 pm window. Should I book it?
Major revision, SSRN working paper 5387320, 2025. Asks how a platform should offer jobs when workers may decline and improve at different, unknown rates. Proposes a tractable benchmark and learning algorithms, including a variant that caps the risk of market failure, and tests them on delivery-platform data.
Research in the field
Two papers on human-AI decision‑making presented at the MIT Initiative on the Digital Economy
Under reviewLearning to Gig: Workforce Development under Worker AgencyLearning to Gig
PreprintStale, Misattributed, or Late: Where Personal Memory Fails Before GenerationStale, Misattributed, or Late
A talk on skills in the AI era at the UC Berkeley-NASA symposium at NASA Ames
Research that makes algorithms easier to work with
Our goal is to make the algorithms people work with easier to learn from and to trust. We started with gig workers taking jobs from apps, and now study AI agents and the assistants people rely on every day.
A research grant from the Fisher Center for Business Analytics at Berkeley Haas
Grant, Jul 2023A talk on the future of skills in the AI era, at NASA Ames
Talk, Sep 2023Two papers on human-AI decision‑making, presented at MIT
Talk, Nov 2023Two papers on human-AI decision‑making, presented at Stanford
Talk, Mar 2024For Park's operations course in the Evening MBA program
Award, May 2024How K-12 teachers use ChatGPT on their own work
Accepted to CSCW, Sep 2024Runner-up for the INFORMS TIMES Best Working Paper award
Award, Oct 2024Two papers on how people and algorithms make decisions in sequence
Talk, Nov 2024Managing Multihoming Workers, selected for the 2025 symposium in Boston
Talk, May 2025Reinforcement learning that turns decision data into tips people can use
Published in Management Science, May 2025How gig workers learn alongside recommendation algorithms
Accepted to CSCW, Jul 2025Learning on the Go and Making ChatGPT Work for Me, presented in Bergen
Talks, Oct 2025How advisors respond when their clients also consult a personal AI
Preprint on arXiv, Mar 2026For teaching in the Evening & Weekend MBA program
Award, Apr 2026Whether algorithmic advice should name the next step or a broader rule
Major revision at M&SOM, Apr 2026Dispatching jobs to gig workers who learn and can turn jobs down
Major revision at M&SOM, May 2026Three papers presented at Harvard Business School
Talks, Jul 2026When an AI agent in service work should act, ask or defer
Submitted to POMS, Aug 2026Dispatch policies that build gig workers' skills
Posted on SSRN, Aug 2026Where an AI assistant's memory of its user goes wrong
Preprint on arXiv, Oct 2026How firms can learn about pricing without running costly live experiments
In progressCompliance-aware reinforcement learning for managing human agents
In progressHow to staff fluid teams in a fragmented labor market
In progressDoes skill learned in one job carry over to another?
In progressHow to design and sequence incentives for platform workers
In progressHow K-12 teachers use ChatGPT, and how it changes their productivity and the tasks they choose
In progressLooking for a specific paper? You'll find all of them on our publications page.
All publications