Many companies are currently introducing GenAI tools into everyday work: writing assistants, coding copilots, customer-service copilots, meeting summarizers, knowledge-search assistants, and AI tools for sales, HR, or consulting. However, companies often struggle to evaluate whether these tools actually improve work. They may track usage, adoption, or employee satisfaction, but these indicators do not show whether GenAI causally improves performance, quality, speed, learning, or decision-making.

A simple A/B test is often not enough. In organizational workflows, employees learn from AI, share tips with colleagues, and change their routines over time. This creates carryover effects: even after AI assistance is removed, employees may still behave differently because they have learned from prior AI exposure. In addition, teams often interact with each other, which creates interference: one employee’s access to AI may affect the work of others. The switchback experiment paper addresses exactly these kinds of settings by studying experiments where the same unit is repeatedly switched between treatment and control over time, while explicitly considering carryover effects and valid causal inference.

The thesis is supposed to answer the following research question: "How can switchback experiments be designed to evaluate the causal effects of GenAI assistance in organizational workflows where learning, carryover, and interference make simple A/B tests problematic?"

If you are interested in this master's thesis, feel free to send an email to Marc Grau (marcchristopher.grau@unisg.ch).