Dropbox uses GrowthBook to safely drive AI product development

Executive Summary

With over 3 billion feature evaluations and 1 billion logs processed daily, Dropbox is operating at a scale where the infrastructure team needs to be world-class.

As the company grew through acquisitions and launched new AI-driven products, Dropbox found itself managing a fragmented ecosystem of internal tools for feature flagging, experimentation, and analysis. The result was rising complexity, slower decision-making, and growing maintenance costs.

By adopting GrowthBook as a self-hosted experimentation and release platform, Dropbox consolidated tooling, standardized release practices, and enabled data science teams to run analysis directly on their existing Databricks data lake. Today, GrowthBook supports everything from high-velocity AI feature rollouts to large-scale experimentation, with plans to deepen experimentation even further in 2026.

‍

The Challenge: Fragmentation at Scale

Dropbox’s legacy experimentation and feature flagging systems were built in-house and served the company well early on. But over time, cracks started to show.

Key challenges included:

• Tool sprawl and rising complexity
  Multiple internal and third party tools handled feature flags, experimentation, and analysis, each with different workflows and assumptions.
  
  
• Slow analysis cycles
  Experiment analysis often took days and required custom notebooks or bespoke pipelines.
  
  
• Limited scalability across teams and stacks
  Dropbox supports multiple programming languages and platforms across web, desktop, and mobile, complicating standardization.
  
  
• Growing pressure from AI product development
  New AI-driven products required rapid iteration, staged rollouts, and fast rollback capabilities without risking stability.

‍

Why GrowthBook

After evaluating alternatives, Dropbox selected GrowthBook based on three non-negotiables:

1. Self-Hosted, Warehouse-Native Analytics

Dropbox runs GrowthBook self-hosted on AWS and analyzes experiments directly in its Databricks data lake. This avoided costly and risky data duplication to third-party vendors while preserving full control over sensitive data.

‍

2. Scales Across Teams and Languages

GrowthBook supports Dropbox’s diverse technical ecosystem, including Python, PHP, and other languages inherited through acquisitions, without requiring custom rebuilds.

‍

3. Modern UX and Rapid Product Evolution

Compared to Dropbox’s legacy tools, GrowthBook’s UI and analytics dramatically lowered friction for engineers and data scientists. Dropbox also values GrowthBook’s fast pace of improvement and responsiveness to feature requests. 

‍

From Feature Flags to Safe, Standardized Releases

One of the most immediate wins came from standardizing feature releases.

Dropbox introduced a “feature ring” system using GrowthBook, creating a single source of truth for rollout stages:

• Internal teams
• All Dropbox employees
• External alpha users
• Self-serve customers

This eliminated overlapping audience definitions and confusion around targeting, while enabling precise, staged rollouts across desktop, web, and mobile platforms. 

For emerging AI products like Dash, GrowthBook became the backbone for safe, high-velocity releases in a fast-moving market. 

‍

Experimentation at Dropbox Today

While GrowthBook is heavily used for release management today, experimentation remains a core capability and strategic priority.

• 30–50 experiments run monthly across Dropbox
• Engineers and data scientists build experiments directly
• Product managers primarily consume results and make decisions
• Analysis is performed directly in GrowthBook and Databricks 

GrowthBook’s analytics significantly outperform Dropbox’s legacy experimentation tools, especially for statistical analysis and visualization. An unexpected bonus: GrowthBook can analyze experiments configured in legacy systems as long as exposure data is consistent, allowing Dropbox to modernize analytics without forcing immediate migration.

‍

Real Experiment Learnings

Dropbox uses experimentation not just to validate wins, but to avoid costly mistakes and guide product direction with statistical rigor.

“Experimentation is how we decide where to invest. It prevents us from shipping things that feel right but don’t actually move metrics. Neutral or negative results save us from scaling the wrong ideas.”
— Fernando Nogueira, Staff Data Scientist, Dropbox

Positive impact example

A cancellation flow experiment personalized messaging by showing users how many collaborators would lose access to shared files if they canceled. By grounding the message in real user data, Dropbox reduced cancellations and later replicated the strategy across multiple product surfaces.

This wasn’t just a copy change. It demonstrated how targeted, data-backed messaging can drive measurable retention gains.

Learning from losses

In another test, an onboarding redesign replaced an image with an educational video. The team expected improved understanding and conversion. Instead, conversion rates dropped due to added load time and friction.

The experiment prevented a broad rollout of a well-intentioned but harmful change.

Fernando notes that this distribution of outcomes is normal at scale, out of every 10 experiments::

• 2–3 winning experiments
• 6 neutral outcomes
• 2 losses 

For Dropbox, that math reinforces a simple truth: experimentation isn’t about chasing wins. It’s about systematically reducing risk while uncovering incremental gains across the product surface.

‍

Operational Efficiency and Tool Consolidation

GrowthBook also played a key role in reducing operational overhead.

• Two internal tools are being retired
• Experiment analysis is centralized in GrowthBook
• Engineers spend less time maintaining bespoke systems
• Data science teams gain faster access to results 

For Dropbox, this isn’t just about speed. It’s about sustainability as experimentation volume grows.

‍

What’s Next: Scaling Experimentation

Looking ahead, Dropbox plans to deepen its experimentation practice with:

• Global holdouts for large-scale controlled experiments
• Standardized metrics and governance across teams
• Quarterly and annual experimentation reporting

With new data science leadership driving renewed focus, GrowthBook is positioned as the foundation for Dropbox’s next phase of experimentation maturity.

‍