Share with your CTO
Tony Xu is pushing back on the idea that AI coding tools are the whole engineering productivity story. In a recent podcast appearance, the DoorDash CEO noted that code-writing represents only 25 to 50 percent of an engineer’s day, with the remainder consumed by product reviews, design meetings, and cross-functional alignment. DoorDash reports that AI now writes roughly two-thirds of its code, yet Xu says that gain alone hasn’t justified restructuring teams or workflows. The real bottleneck is everything that surrounds the code, not the code itself.
What this means for your business
The shops that bought Copilot or Cursor licenses and declared victory on engineering productivity are measuring the wrong thing. If your engineers spend 60 minutes a day writing code and three hours in meetings, cutting the coding time in half saves you 30 minutes. The constraint was never the typing. It was the coordination overhead, and most AI coding tools don’t touch that at all.
Call this the workflow gap: the distance between where AI productivity actually lands (code generation) and where engineering throughput actually breaks down (decision latency across product, design, and business teams). Xu’s framing is correct, and it has a practical implication. The next round of engineering productivity investment shouldn’t be more coding assistants. It should be AI that compresses the review cycle, drafts the spec, summarizes the design debate, and surfaces the blocker before the standup. That’s a different tooling category entirely, and most enterprises haven’t started buying it yet.
Andrej Karpathy’s admission that AI-generated code is “bloaty” and “brittle” compounds this. CTOs who are counting AI-written code volume as a productivity metric are building on a shaky foundation. High code velocity plus low code quality plus unchanged coordination costs is a recipe for technical debt acceleration, not efficiency. The signal worth watching: which vendors start shipping AI tooling aimed at the meeting and review layer, not the editor layer.
Concept deep-dive: Workflow-native AI
Workflow-native AI means embedding AI across the full operational sequence of a function, not just its most visible task. In engineering, that’s the difference between autocompleting code and automating the entire path from customer feedback to shipped feature: requirements drafting, stakeholder alignment, code review, and deployment sign-off. It exists because point-tool AI hits diminishing returns fast once the surrounding process stays human-paced. Think of it like widening one lane of a four-lane highway. Traffic still jams at the other three. For CTOs, this reframes the build-vs-buy question around process redesign, not just tooling procurement.
Based on reporting from DoorDash CEO Says AI Coding Is Not Enough for Engineering Productivity, originally published 2026-07-29 00:19:00.
