Dropbox's engineering organization now runs on AI-assisted workflows, with adoption exceeding 90% among developers this year. The company reports that figure surpasses adoption at many industry peers. Getting there took roughly two years, a first attempt that stalled in the low double digits, and a 2025 reset built around leadership alignment, internal tooling and structured change management. CTO Ali Dasdan and Uma Namasivayam, Senior Director of Engineering Productivity, described the path and the open questions in a Q&A.
From Copilot pilot to stalled adoption
Dropbox introduced GitHub Copilot in late 2023, around the point where coding models began to become powerful. Engagement stayed limited. Adoption held in the low double digits through 2024. Engineers who were enthusiastic about the tool's potential were outnumbered by those who saw too little value to fold it into everyday work. Namasivayam describes the feedback as a mix of curiosity, uncertainty and friction rather than active rejection.
Dasdan frames the productivity argument simply: time shifted away from meetings and toward coding, with AI taking over writing code, running code reviews, generating tests, debugging and incident resolution. It also lowers the cost of starting a task, such as writing a test, which opens room for exploration and ideation. The company started with coding and has since widened its tooling to more areas.
Three workstreams and a turning point
Early 2025 marked the shift, driven primarily by leadership alignment: AI tooling became a company-level engineering priority. Vision alone did not fix the friction, so Dropbox split the rollout into three workstreams:
- Evaluation and adoption of external AI tools, continually refreshed given the pace of model improvements;
- Exploration and construction of internal AI tools tuned to Dropbox's needs, which also builds engineering depth in AI development;
- A company-wide change management plan covering tracked and shared metrics, training and enablement, and communications that set clear expectations.
Namasivayam credits culture with scaling the effort: engineers learning from each other and sharing successes is, in her view, the only way adoption spreads. The metrics work was designed to show progress in data rather than anecdote, and the change plan was built around listening to developer feedback.
Build versus buy
The split is pragmatic. If a third-party tool works well, the priority is access and adoption. Where external offerings fall short, notably in code reviews and incident debugging, Dropbox builds its own tools on top of foundation models. It has already produced custom tooling for migrations, testing and debugging, and these builds fill gaps that shaped the overall adoption number.
Evaluation applies the same standard to both internal and vendor tools. Usage is not the test; measured developer productivity counts, including development velocity and developer experience sentiment. Tools that move those numbers get scaled, the rest get dropped, and outside partners are expected to demonstrate value and business impact before scaling.
Beyond code completion
Coding assistants were chosen for the fastest wins, but Dasdan says the roadmap now spans the full software development lifecycle. AI supports code migrations and automated code fixes, and the company is exploring developer experience pain points such as documentation and automated testing.
The expansion reaches past engineering. Product managers use AI to analyze customer feedback when shaping roadmaps, designers use it to generate and iterate prototypes, and go-to-market teams get faster reads on market trends.
Where the industry misjudges AI
Dasdan points to overestimation in treating low-level metrics, such as the volume of AI-generated code, as a proxy for business value; that link does not hold yet, and acting on it quickly carries risk. Underestimation, meanwhile, shows up as large companies underestimating the urgency and falling behind. He argues for staying in sync with development pace and being deliberate about it.
On the next year, he expects these tools to improve rapidly and cover an ever-wider range of engineering tasks, while insisting that excellent, specialized engineers remain essential. The connection to real business outcomes, he says, is not fully realized yet.
Namasivayam reports one early signal in Dropbox's data: developers shipping more PRs. Much of the tedious, repetitious work, such as tests and boilerplate, can now be automated, but the tools are not there yet for deeper infrastructure or production-grade systems. The company is actively testing hypotheses and measuring impact with concrete outcomes.
The adoption playbook
Namasivayam's advice to other teams rests on three moves: remove friction so teams can try tools and get support easily; run AI enablement like a product rollout, with a clear strategy, onboarding investment and continuous feedback; and meet engineers where they are, since workflows and needs vary. Flexibility through tool choice and targeted support is what made adoption stick at Dropbox.
For its own customer-facing AI, embodied by Dropbox Dash, Dasdan's target is concrete before-and-after value: doing work better, faster and more affordably. He cites customers struggling to locate a single video among thousands, whose real pain is wasted time under pressure. The outcome he hopes to hear a year from now is that Dash helped customers deliver more value to their own customers.



