Why Handing Out Tools Isn't the End of It
After introducing Claude Code in 2025, DoorDash noticed people outside of engineering shipping directly from the terminal, and decided to give every employee a coding agent. The goal was to raise the company's baseline AI fluency — its 'floor.' Many employees still used AI only as a chat, but the moment it connects to Gmail, calendar, and Slack, the efficiency of knowledge work changes.
But as throughput rose, hidden bottlenecks surfaced. With code pouring in, merge queues and code review backed up, and security issues increased across the industry. DoorDash responded by automating this in turn with AI code review agents. Unblock one layer of speed and the next wall appears — automating that wall again is what AX actually looks like.
How Does Learning Spread Across the Whole Organization?
The most scalable way to spread learning is through written artifacts. Have your best people and success cases documented, and the whole company can read them — with the byproduct that agents gain material to reference too. Assigning someone to teach each person one by one does not scale at the size of an organization.
You can't let people share only the wins. There has to be a culture where failures — 'this workflow didn't work,' 'this MCP integration just wasted tokens' — are shared just as comfortably. Learning accumulates only when each team can surface both its results and its limits in its own channels.
What Do Experimental Teams Need to Actually Move?
When you give a team the goal of 'ship this 3-5x faster with AI,' asking them to do it within the confines of their normal job makes them give up quickly. The barriers outside of code generation — cross-functional alignment, review processes, approval steps — are still there.
So DoorDash attaches VP- and executive-level sponsorship to experimental teams. Beyond a token budget, they give teams the authority to escalate whatever is blocking them outside of writing code. Teams experiment safely when a leader personally leans in and clears the blockers.
How Do You Find AI Champions?
If you appoint 'automate this workflow' from the top, you usually pick the wrong workflow or the wrong person. Only the people who do the work every day actually know where the bottleneck is.
So you give people freedom and look for the champions who emerge on their own. Across sales, marketing, operations, and support, the people who dig into the tools even in their spare time reveal themselves — because their enthusiasm makes them unable to resist sharing results in Slack. These people find the real bottleneck in their domain and automate it, or turn it into a shared skill their coworkers can use.
So What Does the Organization Have to Redesign?
The most striking redesign was the review process. Engineers used to spend time getting a design pixel-perfect; now the engineer only gets it to a working state and the designer finishes the polish self-sufficiently. That made product review and design review the bottleneck instead, and DoorDash asked whether those reviews were still needed in the same form.
The core principle is self-sufficiency. The more self-sufficient individuals and teams are, the faster they move, so you keep only a few domain-expert gatekeepers and grow the rest into generalists who move fluidly across the codebase. When coding speed was the bottleneck, that speed dictated team structure, process, and review cadence — but now that engineering is fast, all of it has to be redesigned.
ROI is viewed two ways. One is raising the 'baseline' by cutting time on tasks everyone does; the other is automating specific workflows on a per-department, per-domain basis. The latter can have the bigger impact. Even a 10% productivity lift across every employee is an enormous difference at a large organization. In the end, the metric isn't how fast code gets merged — it's whether customer value is delivered faster.