DoorDash Gives 10,000 Employees an AI Data Analyst

DoorDash has introduced an artificial intelligence agent named Vera to assist its employees in analyzing data. Employees can ask Vera questions about the company's data, which is stored in over 200,000 datasets totaling 350 petabytes. For example, employees can ask Vera about the impact of a new product launch, which merchants are at risk of leaving DoorDash's marketing products, and how many orders were canceled by hour at Arby's over the last 90 days.

The hard part is knowing what data to use, DoorDash said, as their earlier tools were either specialized or broad and shallow. DoorDash built Vera to answer these questions in conversation using company data. Vera passed 73% of 100 sample questions against an off-the-shelf AI coding agent, which passed 48%. Vera's pass rate improved from 43% to 90% over the project, with better AI models and organizing data around the model contributing to the gains.

Vera sorts questions into three levels, with the hardest being cause-and-effect analysis, experiment results, forecasts, and high-stakes calls. Vera tends to stop early and return shallow answers at the top level. DoorDash added a planning step, which improved results but swung between 64% and 100% on a sample of 14 questions. The project fits a wider AI push at the company, with an ordering agent building a grocery cart in under 2 minutes.

However, finance chiefs have been cautious with AI agents, with nearly 7% of enterprise CFOs in the U.S. having put them into live finance workflows. The company now focuses on whether faster answers move its core business numbers.