Intel / article

What Dropbox has learned from deploying AI at company scale

Source: dropbox.tech
Author: Dropbox Team
Date: Published Sep 23, 2026

Context

Producing code faster can move the bottleneck elsewhere. That is the central lesson in Dropbox's account of company-wide AI adoption, where Chief Technology Officer Ali Dasdan describes coding faster but then having too many code reviews, and generating more code that strains the tools and infrastructure supporting development. The company's leaders separate proxy metrics, such as pull request throughput, from the outcomes they treat as real return: revenue, cost, customer satisfaction and retention. Their argument is that extra output becomes valuable only when the whole workflow can turn it into useful software.

Dropbox says it uses Nova, its internal service for running coding agents, to connect usage to engineering workflows, validation and outcomes. The Nova platform description, covered in an earlier Intel item, shows what validation means there: a patch is checked against the real build and test environment, and failed checks go back to the agent for another attempt, up to a cap.

Key takeaways · AI-distilled

Dropbox CTO Ali Dasdan says giving people AI tools is only the first step: new bottlenecks emerge, such as too many code reviews or extra load on development infrastructure, so the workflow has to be re-examined end to end.
Dropbox says about 70% of its code is now AI-generated, close to a figure Uber shared, and it runs coding agents through an internal service called Nova.
Dropbox reports pull request throughput in the top 5% of a custom of peers with large, complex codebases, a change failure rate in line with peers at the 75th percentile, and usage among the lowest in that group.
Dasdan treats revenue, cost, customer satisfaction and retention as the real ROI, tracked through proxy metrics grouped under speed, effectiveness, quality and impact. Nova lets Dropbox tie token spend to the engineering work it produced.
Uma Namasivayam says leadership mandates alone do not drive adoption: Dropbox paired them with boot camps and show-and-tells, and aligned incentives so managers could see the return.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Read the source dropbox.tech