Dropbox

Dropbox reported 90% AI adoption among its engineers and a 20% increase in merged pull requests, with a reduced change failure rate, figures Gergely Orosz and Laura Tacho used in September 2025 when surveying how 18 tech companies measure AI's impact on software development. The company sits in cloud storage and collaboration, but the reason it keeps surfacing in discussions of engineering productivity is that it publishes operator numbers rather than vendor claims.

The internal agent platform is Nova, which already accounts for roughly 1 pull request in 12 and has spread into migrations, flaky tests, bug investigations, and dependency updates. Kazuaki Okumura, writing on 28 May 2026 from a DX Annual 2026 talk, argues that Nova's value comes less from the model than from the systems around it: codebase context, internal practices, safe execution, workflow integration, and human review.

Okumura's thesis is that AI moves bottlenecks instead of removing them. Faster code generation pushed pressure downstream onto review, CI, validation, release coordination, and production operations. Dropbox's response was to stop treating PR throughput as sufficient and adopt a four-stage model running Fuel, Adoption, Output, Impact, paired with quality signals: code review turnaround time, first-run test pass rate, defect ratio, rework rate.

Okumura locates competitive advantage in the systems built around foundation models, since access to the models themselves is shared. Whether the four-stage measurement holds up as agent volume grows remains untested.