DoorDash
DoorDash builds a voice support assistant for Dashers with Claude on Amazon Bedrock
DoorDash worked with the AWS Generative AI Innovation Center to build a voice self-service assistant for Dashers using Anthropic's Claude 3 Haiku on Amazon Bedrock and Amazon Connect. AWS reports it fields hundreds of thousands of Dasher calls per day with responses in 2.5 seconds or less.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
About 2 months
Time from start to a solution ready for live testing1
50%
Reduction in development time using Amazon Bedrock1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
DoorDash receives hundreds of thousands of support requests per day from consumers, merchants and Dashers (its delivery drivers). The case study says Dashers prefer to call while on the road, so they need quick answers with little delay. DoorDash already had an IVR phone system built on Amazon Connect and Amazon Lex.
What they built
A voice self-service assistant that answers Dasher questions by drawing on DoorDash's public help center content. It reduces how often calls need a live agent.
How they did it
DoorDash worked about eight weeks with the AWS Generative AI Innovation Center. Amazon Connect handles calls. Claude models on Amazon Bedrock generate answers, and Knowledge Bases for Amazon Bedrock retrieves help center content to ground them. DoorDash adopted Claude 3 Haiku for its speed. An Amazon SageMaker framework runs A/B tests, tracks success metrics and checks responses against ground-truth data. The case study also says Claude helps guard against hallucination, prompt injection and abusive language.
Results
AWS reports the solution was ready for live testing in about two months, 50% less development time on Bedrock, and a 50x rise in automated testing capacity. Responses arrive in 2.5 seconds or less. The solution fields hundreds of thousands of Dasher calls per day and has cut escalations to live agents by thousands per day.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Response latency | 2.5 seconds or less1 | |
| Time from start to a solution ready for live testing | About 2 months1 | |
| Reduction in development time using Amazon Bedrock | 50%1 | |
| Increase in automated testing capacity | 50x1 | |
| Dasher calls fielded per day | Hundreds of thousands1 |
Lessons for your project
For voice, latency is a requirement. DoorDash picked a small, fast model (Claude 3 Haiku) to stay at or under 2.5 seconds.
Building an automated evaluation harness early let the team run thousands of tests per hour and compare against ground-truth answers.
Ground answers in the public help center through retrieval, so no custom data integrations are needed to launch.
Caveats
- The only source is an AWS case study written with DoorDash; figures are vendor-reported and not independently verified.
- The case study page has no publication date. The deployment month is approximate: the page says the test finished in early 2024.
- The assistant answers questions; actions through DoorDash's logistics workflow service are described as planned, not live.
Sources
- Building a Generative AI Contact Center Solution for DoorDash Using Amazon Bedrock, Amazon Connect Customer, and Anthropic's Claude
Amazon Web Services · Undated · Vendor case study · aws.amazon.com
Summarized in our own words from the sources above. Spotted an error? See corrections.