DoorDash built the agent for its own staff. It answers questions in conversation using company data, the company said in a blog post. More than 10,000 people at DoorDash rely on that data, per the blog post. The data sits in more than 200,000 datasets and totals 350 petabytes.
The questions are everyday ones, according to the company. How has a new product paced since launch? Which merchants are at risk of leaving DoorDash’s marketing products? How many orders were canceled by hour at Arby’s over the last 90 days?
With that much data, the hard part is knowing what to use, DoorDash said in the post. The company’s earlier tools were either deep in one business area or broad and shallow. “The problem is actually that our upstream data sources are super-fragmented,” a senior analytics manager said in the launch post.
Off-the-shelf AI tools often fall short on data this large and messy, the post said. DoorDash tested an off-the-shelf AI coding agent against Vera on 100 sample questions. Both ran on the same AI model, which is not the one Vera uses day to day. Vera passed 73%. The outside tool passed 48%. The outside tool had trouble picking the right data tables, DoorDash said. Vera is blocked from using data that hasn’t been verified.
DoorDash grades Vera like a student. The exam holds more than 1,900 questions across 19 areas including sales, finance, product and operations. Some came from employees and the experts who own each business area. Others were drafted by AI from old Slack threads and work tickets. An expert approves each one before it counts. A second AI model does the grading.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
Vera’s pass rate rose from 43% to 90% over the project, by the company’s own testing. The test set doubled in size twice along the way. Better AI models helped. The bigger gains came from organizing the data around the model, DoorDash said.
The payoff showed up in finance. “It has worked well for strategic finance questions across difficulty levels,” a finance director said after several weeks of testing.
Harder Questions Test the AI Agent’s Limits
DoorDash sorts questions into three levels. The first is lookups such as metrics and trends. The second is analysis that can sway a decision, including targeting, sizing an opportunity, setting priorities and allocating resources. The third is the hardest, covering cause-and-effect analysis, experiment results, forecasts and high-stakes calls.
Vera stumbles at the top level. It tended to stop early and return answers that sounded right but stayed shallow, according to the post. DoorDash added a planning step. Vera now lays out its plan and runs nothing until the employee approves it. Results improved but swung between 64% and 100% on a sample of 14 questions.
The work fits a wider AI push at the company. An ordering agent now builds a grocery cart in under 2 minutes, PYMNTS reported after DoorDash’s second-quarter earnings call. DoorDash CEO Tony Xu said on that call that order volume from outside AI agents remains low.
Finance chiefs have been cautious with AI agents. Nearly 7% of enterprise CFOs in the U.S. had put them into live finance workflows as of September 2025 and another 5% were running pilots, PYMNTS Intelligence found.
The hardest questions are DoorDash’s next focus. The company said it still has to show whether faster answers move its core business numbers.
Add as Preferred Source
