What Dropbox has learned from deploying AI at company scale

Scaling AI means reworking the workflow, not just handing out tools

Dropbox CTO Ali Dasdan and Senior Director of Engineering Productivity Uma Namasivayam have spent the past year deploying AI across the company and now describe what that experience taught them about adoption, measurement, and where human skills still matter.

The starting point, in Dasdan's view, is the outcome you want. Distributing AI tools is only step one; as people use them, assumptions get tested and new bottlenecks appear. A team may code faster but then face too many code reviews, or generate significantly more code and strain the tools and infrastructure supporting development. Meaningful impact requires examining the workflow end to end and adapting it continually. "There's no silver bullet," Dasdan said.

Namasivayam adds that the effort needs a product mindset. A leadership mandate to adopt AI tools is useful but insufficient on its own — ground-up awareness through boot camps and show-and-tells drives deeper involvement, and incentives must be aligned so managers understand the ROI they're getting.

Measuring productivity: outcomes first, proxies second

Attributing business results to individual code changes is close to impossible, Dasdan argues: ten lines pushed to production can't readily be tied to a specific amount of customer satisfaction or revenue. Nor does volume settle the question — a million lines a day is worthless if the product is wrong and customers don't want it.

The real ROI metrics, then, are outcomes such as revenue, cost, customer satisfaction, and retention. Dropbox layers proxy metrics on top to connect output to those outcomes: whether teams move faster, produce more code, run more experiments, and ship more features, and whether the site is faster and more reliable. There is no single metric. The framework covers speed, effectiveness, quality, and impact, with multiple metrics under each — and, Dasdan notes, the industry has yet to find the full solution.

How Dropbox stacks up against peers

Defining peers as the top technology companies Dropbox hires from and competes with for talent, Dasdan says the problems and metrics look very similar. Roughly 70% of Dropbox's code is now AI-generated, a figure comparable to the ~70% Uber recently shared. Both companies see a need for internal agentic coding solutions — Uber has its own tools, Dropbox has Nova, its internal service for running coding agents. Larger companies have devoted more resources to these problems, while Dropbox is "super lean," yet Dasdan considers the company comparable to that peer group across multiple metrics, including how quickly it adopted AI and hit its targets, and ahead of the broader industry.

Namasivayam points to pull request throughput — the rate at which teams complete code changes — as one industry proxy for productivity. It isn't perfect, but Dropbox's throughput ranks in the top 5% of a custom benchmark of companies with similarly large, complex codebases and engineering systems. Quality is tracked alongside speed: its change failure percentage, which measures how often a deployment leads to degraded performance or failure, sits in line with peers at the 75th percentile. Token usage is also among the lowest in the peer group, which may indicate efficient teams or systems. Together, she says, these measures give a fuller picture against similar companies.

Deciding where the next AI dollar goes

AI funding spans a large portfolio at Dropbox: products used by individual functions, capabilities embedded in tools like Zoom and Slack, company-wide deployments such as ChatGPT Enterprise, automation tools, and coding models. Benefits are visible — people are fixing tech debt, long-running migrations are getting finished, and most of the roadmap is delivered with fewer resources.

Organizations still have to choose which ideas to pursue and when to invest more, often before a direct line to ROI can be drawn, which demands domain expertise, the right incentives, and plenty of judgment. Dasdan places Dropbox alongside the rest of the industry in still working through that question. One shift is underway in how spend is evaluated: less attention to raw token consumption, more to the engineering value those tokens create. With Nova, for instance, Dropbox can connect agent usage to actual engineering workflows, validation, and outcomes, making it possible to judge AI spend by the engineering work produced rather than by token counts.

Skills that appreciate as the cost of software falls

For Dasdan, the core of engineering and computer science — knowing how to solve a problem — isn't going away. AI may enable bigger or faster solutions, but engineers still must supply better context, ask the right questions, judge whether answers are valid, know how to iterate, and produce a good spec and problem definition. They also need to see how a solution fits the broader system and adapt as tools change. Problem solving, communication, systems thinking, and judgment grow more important precisely because agents depend on people to guide and evaluate their work.

Namasivayam names three skills. Problem solving comes first, then judgment — if the next phase of AI drives the cost of engineering toward zero, choosing the right problems and evaluating AI output become critical. Leadership is the third: understanding people, navigating ambiguity as work changes, bringing teams together, communicating clearly, building trust, and keeping incentives aligned. These skills, she says, won't disappear; their value will only grow.

Context and continuity as differentiators

However capable models and agents become, Dasdan notes they don't automatically know which information an organization trusts, what decisions have already been made, who has access, or what needs to happen next. Without that context, even a strong model can return generic or incomplete results. Dropbox's role, as he describes it, is to connect supported AI experiences to customer-owned content and context, then give AI-generated work a place to be saved, shared, reviewed, approved, and continued — with people making the consequential decisions while the work keeps moving, retaining its sources, history, and collaborators.

Consistent with that, Dropbox aims to let people bring organizational context into the supported AI tools they choose while preserving existing permissions and controls, reducing time spent finding, uploading, and re-explaining what already exists.

The year ahead

This year's focus has been heavily on engineering. Namasivayam says Dropbox is now beginning to understand workflows across product, design, and other non-engineering teams. A year out, the goal is for people on those teams to come up with ideas, test them with customers, and potentially ship some of those features — for example, a customer experience team spotting an issue and using agents to develop a solution and see whether it works for the customer. If the company proves the model internally, it could potentially enable external customers to build on Dropbox platforms and use AI with its building blocks.

Human judgment remains central throughout: people still need to pick the right problems, give agents the context they need, evaluate their output, and stay in control of important decisions and actions. That is what turns AI-enabled output into real value.