DoorDash Built a Gen AI Voice Assistant in Two Months

Many companies trial generative AI in customer service as a side project, walled off from their production systems. DoorDash went the other way.

Working with AWS's Generative AI Innovation Center program, the delivery platform built a self-service voice assistant powered by generative AI. According to the case study published by AWS^[1], the solution was ready for live testing within two months. AWS also reports that Amazon Bedrock and Anthropic's Claude model cut development time by 50%.

The project was built on top of Amazon Connect, the cloud contact center DoorDash already had in place.

The starting point was the phone system already in place

DoorDash fields hundreds of thousands of support requests every day. They come from three groups: consumers, partner merchants and Dashers, the independent couriers who deliver through the platform.

AWS reports that at the end of 2023 DoorDash had more than 37 million monthly active consumers and more than 2 million monthly active Dashers.

Before the project, DoorDash already had an automated IVR built with Amazon Connect and Amazon Lex. According to AWS, that IVR had cut transfers to live agents by 49%. It had also lifted first-contact resolution by 12% and delivered $3 million in annual savings.

Even so, most calls still ended up with a human agent. That gap is where the new idea came from.

In the AWS write-up, DoorDash says it wanted to help Dashers get quick answers to the most common questions. Typical requests involve app issues, sign-up and payment methods.

The decisive choice was a technical one. The new assistant plugs into the same contact center that was already handling the calls, which let the team focus on the model and the quality of its answers.

The four numbers AWS attributes to the project

The AWS case study lists these results.

• Development time: cut by 50% with Amazon Bedrock.
• Testing capacity: increased 50x.
• Response time: 2.5 seconds or less with Claude 3 Haiku.
• Volume: hundreds of thousands of calls a day handled by the solution.

All four figures come from AWS, which sells the technology involved. Independent verification is still missing.

Response time is the most useful figure for anyone designing a similar system. On a phone call, any pause before an answer is immediately noticeable. Two and a half seconds is a concrete benchmark to work from.

What the story leaves out

Three of the four numbers measure the build. They tell you how fast the system came together and how fast it responds.

The fourth measures volume. The published text says nothing about how many of those calls end with the problem actually solved.

The data that anyone weighing a similar investment needs is missing.

• The containment rate (share of calls closed by the assistant), before and after.
• Transfers to a live agent compared with the old IVR.
• Satisfaction among Dashers, consumers and partner merchants.
• Cost per call.

The obvious comparison is with the old IVR. That system came with clear outcome numbers: 49% fewer transfers and 12% more first-contact resolutions. The new assistant enters the story with a different kind of number altogether.

The account also lacks any mistake corrected along the way.

In the verified cases I have followed, the most instructive data point was always a course correction. Here the public account leaves it out.

For now, DoorDash's result is an engineering win. The customer service win has yet to be proven.

What to take away from DoorDash

The useful lesson is about where to start.

1. Start from the contact center you already run. DoorDash layered generative AI on top of Amazon Connect, which was already handling its calls.
2. Measure the old system before you change it. The IVR numbers become the baseline.
3. Set a response-time threshold. DoorDash reports 2.5 seconds or less with a small, fast model.
4. Look at the features already built into the platform. The Amazon Connect documentation^[2], for example, describes AI-generated contact summaries.
5. Pick your outcome metric up front. The number to publish is the share of calls resolved, before and after.

The question to bring to your next meeting

Two months to build is a good result. What is the number that proves your assistant actually solves the problem, not just that it answers quickly?