Where AI ‘digital twins’ pay off
Can an AI “digital twin” stand in for a real customer? A team of 17 Columbia Business School faculty led by Professor Olivier Toubia, the Glaubinger Professor of Business, put the idea to the test across 19 studies and 164 survey outcomes. Twins built from about 500 survey answers per person barely beat a simple demographic profile at predicting how people would respond.
The results looked different in the field. Working with Oliver Wyman, the team built twins of about 330,000 customers of a global bank from account and spending history. Used to personalize how offers were framed, the twins lifted credit card sign-ups by 6% and payroll transfers by 25% in an early test.
For alumni who buy or build these tools, the lesson is practical: feed twins behavioral data that matches the task, ask vendors to compare results against simpler baselines, and test predictions before the behavior happens. The School’s free ExploraTwin platform lets you compare twin answers with real survey data.