How many designers have actually pushed AI-generated code to production?
The Craft chapter of State of AI Design 2026 says 50%, though only 20% identify as design engineers. By stage: 68% at early-stage companies, 50% at growth-stage, 33% at public companies.
This is a self-reported survey, not a census of all designers; the larger the company, the lower the share.
How does DoorDash's approach differ from Flux's?
DoorDash routes everything through GitHub PR approvals and adds a design review step, so humans gate before merge. Flux does post-facto PR review with AI risk scoring and multiple AI bots, and Brooks Solveig can merge his own PRs with no oversight in many cases.
The difference is whether the gate comes before or after merge, and who carries the responsibility: people or an automated system.
Which model is better?
The report doesn't say. It offers no comparison of reverts, rollbacks or incidents between human gates before merge and automated review after merge, and it doesn't disclose the error rate of Flux's risk scoring.
So we can say these models exist, but not which is better from this data; to judge, you need your own revert rate and incident record.
Can Anthropic's statement be treated as evidence?
Be careful. Anthropic is a partner of this report, and Joel Lewenstein, its head of product design, is quoted saying there are no excuses for not being deeply involved in final delivery, as a case.
That is a stated position, not neutral evidence. The other cases in the report (DoorDash, Watershed, Flux) are likewise self-described by respondents.
Once designers start pushing AI-generated code to production, the real question is no longer "can they" but "who stops a bad change." The Craft chapter of State of AI Design 2026 records how several companies handle it: DoorDash, Watershed, Flux, and a few unnamed firms. This piece lines them up on one spectrum. Two caveats first: Anthropic is a partner of the report, and Anthropic's head of product design, Joel Lewenstein, is quoted in it; and the content is self-reported by respondents, with no defect-rate or incident data.
The report finds that 50% of respondents have pushed AI-generated code to production, yet only 20% identify as design engineers. By company stage, it is 68% at early-stage companies, 50% at growth-stage and 33% at public companies; 70% of founders, and nearly half of executives, are shipping. On ownership, 40% say it fully belongs to them, 43% mostly, and 9% say it is hard to separate. Designers who code with AI were twice as likely to feel more creative and capable, and reported a higher quality bar. These are self-reported feelings, not measurements of product quality.
DoorDash puts designers through the same process as engineers. Shali Nguyen, Head of Consumer Experience Design, says "Everything goes through GitHub PR approvals," with designers held to the same production-quality bar as engineering counterparts; they also added a design review step in which tagged designers approve UI changes. Watershed's Head of Design, Hannah Hudson, says all of their designers have submitted production PRs, and designers review code alongside engineers. What these two share is a human gate before merge.
Flux sits at the other end of the spectrum. Staff Product Designer Brooks Solveig says they do post-facto PR reviews with an AI risk scoring system and reviews from multiple AI bots, and that he can merge his own PRs with no oversight in many cases. In other words, the gate moves from before merge to after, and automation carries part of the responsibility. The report also describes unnamed companies that use AI risk scoring to auto-push low-risk changes, set up sandboxes for engineering review, or have engineers rebuild prototypes. Stripe's Katie Dill offers a self-test: would I be happy with this if it took me a week to make?
Anthropic's Joel Lewenstein puts it more bluntly: there are no excuses anymore for not being deeply involved in the final delivery process. Because Anthropic is a report partner and also a cited case, treat that line as a stated position rather than neutral evidence.
These cases answer "how do they govern it," not "does it work well." The report offers no comparison of whether a human gate before merge or automated review after merge leads to fewer reverts, rollbacks or incidents. Flux's approach depends on how accurate its risk scoring is, and the report doesn't disclose that system's error rate. So we can say these three models exist, but not which is better.
If your designers are about to submit code, choose the kind of gate first and the tools second. A human gate costs engineers' review time; automation costs the work of building and maintaining risk scoring and bot review. Neither is free. There is no data showing post-merge review is cheaper, and none showing it is riskier.
A low-cost practical start is to begin with low-risk changes (copy, styling), keep human review, and track your revert rate. Your own accumulated data is more reliable than borrowing someone else's case.