DoorDash2025

EnvisioningAI-poweredordermanagement

Hundreds of thousands of Merchants on DoorDash use the Order Manager Tablet to fulfill their daily orders, but for many of them, contacting support is the first instinct when something goes wrong.

Support ends up taking 10–15 min on average per issue. Some Merchants would rather cancel an order than deal with the hassle.

Our team hypothesized that AI-powered experiences could potentially shift Merchant behavior and help them save precious time, while leading to higher quality order outcomes and fewer cancellations.

My role & impact

As product design lead, I led a 4-week sprint to envision what deep AI integration might look like, ultimately leading to months of dedicated roadmap time for features and three shipped sub-agents that have so far collectively saved about $10M in order cancellations and quality-induced refunds. More sub-agents are currently in development.

Research told us that the biggest blockers to self-serve issue resolution were 1) poor discoverability of existing self-serve features, 2) friction-filled workflows, and 3) a lack of proactive guidance for Merchants.

Ihadnoideayoucouldtakesomethingoffofanorderwithoutcallingintosupport.

Restaurant Staff in Seattle, WA

My design explorations focused on addressing these key issues across a series of our most painful Merchant tasks, while considering how they might exist within a coherent and scalable system inside the Merchant tablet.

Example of using the AI agent to manage bulk out of stock operations, an otherwise extremely time consuming task.

After better understanding merchant’s relationship with AI, I developed two strongly held principles for a great experience:

  • AI suggests → Human decides → AI executes.

    Merchants are highly sensitive about providing great service, and therefore need control over what an AI agent does for them. Asking for permission before acting provides a level of trust that is critical to feature adoption.

  • Enable magic moments by automating painful tasks.

    Any time a merchant needs to start a chat or pick up a phone to call Support is an opportunity to enable a magical automation that takes the painful task off their plate.

Example of an outbound voice AI agent, which replaces the need for the merchant to call the Dasher or Support.

In testing, Merchants loved the idea of not having to pick up the phone, and were receptive to using AI in their operations.

The one recurring caveat: given the relatively high stakes nature of some actions (customer loyalty and sales are on the line!), Merchants expressed that even small missteps on the AI’s part could degrade their trust in it. We later used this insight as input to descope a set of potentially more sensitive subagents until the following year.

Thethingthattakesthemostiscommunicatingandwaitingfortheanswerfromsupport.IfAIcouldmakethisfasterthatwouldbeamazing

Restaurant manager in Philadelphia, PA
In the future, sufficiently mature AI agents have the opportunity to catch and resolve issues before the merchant even notices, like when Dashers are stuck in one spot for too long.

Designing AI-based interactions for the DoorDash tablet came with a unique set of constraints, with few established patterns to lean on.

  • Tapping, not typing

    Nobody likes typing on a tablet, and that’s especially true for restaurant staff workers with greasy kitchen gloves and very little time on their hands. Through testing, I found that it would be incredibly important to support single-tap interactions, and never rely on a typed response to continue. Every interaction ends with some sort of CTA, or set of possible CTAs.

  • Assessing voice as an input

    Voice was another interesting avenue, though it was plagued with its own set of constraints: noisy kitchens, language and accent barriers, and store ambiance were all found to be potential blockers. In the end, a CTA-driven model, while supporting other input modalities as secondary, felt to be the best combination of solutions.

  • Built for fluid multi-tasking

    Through rounds of testing, I landed on a side panel pattern that pushes over content on the screen rather than overlaying it. This allows the merchant to navigate and still have visibility and control over their orders instead of being locked into a modal view just to interact with the AI experience.

The AI assistant panel aims to be contextually helpful depending on what the merchant is currently looking at. It also doesn’t block interaction with the rest of the screen.
It’s easy to fall behind on the management work of optimizing a store’s operations. Merchants loved the idea that AI could continuously suggest easy-to-accept changes to keep things running smoothly.

As a result of this work, and a series of reviews and meetings with merchant team leadership, we landed on three quarters worth of investment to realize this vision.

Working with my PM and a data scientist, we also narrowed down on the highest potential ROI sub-agents to develop first, which were largely related to Dasher communication and handoff issues (about 26.4% of our total tablet support contact cases).

As of June 2026, the team has released 3 sub-agents to encouraging results. A Dasher pre-arrival subagent reduced late arrivals by 5%, equating to around 2M in annualized savings per year, and a post-Dasher handoff sub-agent reduced missing and incorrect items by 3%, equating to about 3.5M in annualized savings. We currently have an order adjustments subagent in experiment with promising results, showing a directional lift in order quality with an annualized 6M savings impact.

I can’t take all the credit for these metrics, though! After landing the vision work, the next phase didn’t start until early 2026 – at which point I acted as an advisor to another designer who completed the executional design work.

  • Interested in learning more about this project?

    This is just a part of this project and the process behind it. I’m happy to share more about it in a conversation or portfolio presentation.

    Email copied!

Next project