Saturday, 5 September 2026
Artificial intelligence is rapidly increasing the amount of work companies can produce, but Dropbox CTO Ali Dasdan argues that output alone is the wrong measure of whether AI is creating value. The more important test is whether greater speed translates into better products, stronger decisions, higher customer satisfaction and measurable business results.
Speaking at Ai4 2026 in Las Vegas, Dasdan described how Dropbox is applying AI across engineering, collaboration and product development. His central argument was that companies should distinguish between inputs, outputs and outcomes. AI may generate more code, documents, summaries or decisions, but those outputs only matter when they eventually improve something important to the customer or the business.
That distinction is becoming increasingly relevant as companies report large increases in AI-assisted work. Dropbox says adoption among its software engineers rose rapidly from around 40% in March 2025 to effectively universal use by the end of that year. Dasdan said AI use has since become widespread across the wider company, while many engineers now work with several AI tools rather than relying on a single system.
Dropbox has also developed Nova, its internal platform for running coding agents. The system allows engineers to assign work to AI agents operating within Dropbox’s engineering environment, while retaining the context, validation processes and human oversight required to move changes into production. The experience illustrates one of the emerging problems with AI productivity. Faster coding does not automatically result in faster product delivery. As developers generate more code, pressure moves further down the development chain to code review, continuous integration, testing, validation, security and deployment.
Dropbox has acknowledged this effect in its own engineering research. Its engineers have found that increasing AI-assisted coding throughput can expose capacity constraints elsewhere in the software development lifecycle. The implication is that companies cannot simply introduce coding tools and expect overall productivity to rise at the same rate.
Measurement therefore becomes critical. Dropbox is increasingly evaluating productivity through a combination of speed, effectiveness, quality and business impact rather than relying on a single metric such as lines of code or the number of pull requests completed. Quality remains particularly important. AI-assisted development has to operate within security, privacy, reliability, performance and data-quality requirements. Measuring a large increase in code production without examining what happens to testing, vulnerabilities or customer experience could create the appearance of productivity while shifting additional costs elsewhere.
Dasdan also warned against using productivity dashboards as employee surveillance systems. Individual metrics can easily create undesirable incentives and can often be manipulated. A more useful approach is to combine different indicators and use them to identify where processes are improving or where new bottlenecks are developing.
One potentially more meaningful indicator is what employees do with the capacity that AI releases. Dasdan said Dropbox has seen engineers use additional time for work that sits beyond formal product-roadmap commitments, including security improvements, technical-debt reduction and infrastructure migrations. One example presented involved the migration of approximately 1,900 Python packages, which Dasdan said one employee was able to automate using AI over a period of roughly two weeks. Such examples are internal Dropbox results rather than independent productivity benchmarks, but they demonstrate the type of work companies may increasingly automate as AI becomes integrated into engineering systems.
This raises a broader organisational question. Companies originally expected AI productivity primarily to mean completing the existing workload with fewer hours. A potentially more important effect may be that employees use the released capacity to address problems that organisations previously lacked the resources to prioritise.
Collaboration presents a different challenge. AI systems become substantially more valuable when they understand the context surrounding the work rather than responding only to individual prompts. For businesses, that context may exist across files, emails, meetings, messaging systems, project-management platforms and customer databases. Connecting those sources allows an AI system to answer questions based on an organisation’s actual knowledge rather than relying primarily on the general information contained in its underlying model.
Dropbox is pursuing this idea through Dropbox Dash, its AI-powered search and knowledge platform. Dash can connect information from different workplace applications, allowing authorised employees to search across multiple sources from a common environment. Dasdan described situations where an executive entering a board meeting could receive an unexpected question and retrieve relevant internal information through AI without having to locate a colleague who already knows the answer.
The important element is not simply search. It is the combination of search with organisational context and access controls. This creates a major security issue for enterprise AI. Information from different systems cannot simply be indexed and exposed to an AI layer without maintaining the permission structures surrounding it. AI systems are capable of combining information from multiple sources, making access boundaries considerably more complicated than conventional document search.
Dropbox says Dash respects the permissions established within connected applications so users should only receive information they are authorised to access. Dasdan argued that this type of control must be incorporated into the architecture from the beginning rather than added after an AI system has already been created.
Organisational adoption also depends on more than providing employees with software. Dropbox has experimented with AI champions, internal training, hack events and collaboration between technical and non-technical teams. This is becoming increasingly important because generative AI is allowing employees without conventional programming experience to build applications and automate workflows themselves. That democratises software development, but it introduces another type of risk.
A non-technical employee may be able to create a functioning application without understanding monitoring, security, scalability or the downstream consequences of operating it. Dropbox’s approach therefore includes pairing employees with people who understand the technical environment and encouraging AI champions within different business functions.
An AI specialist working inside human resources, for example, may be more effective at demonstrating relevant applications to colleagues than an engineer explaining the same technology from outside the department. The language, problems and workflow are already familiar. The result can become an organisational network effect. Once employees see colleagues performing work that previously required specialist technical knowledge, experimentation spreads more quickly across the business.
The third major area of change is product development itself. Generative AI has dramatically reduced the time required to create software, but the bottleneck is increasingly shifting from writing code to deciding what should be built. If companies can produce new features far more quickly, understanding customer problems becomes more valuable rather than less.
Traditional product development usually involves product managers interviewing customers, conducting research and translating what they learn into requirements for engineers. Generative AI creates the possibility of widening that feedback loop substantially. Customer-support tickets, calls, product feedback, transcripts and other unstructured information can increasingly be analysed continuously. Instead of forming a product hypothesis after speaking with a relatively small sample of customers, companies can potentially identify recurring problems across a much larger proportion of their customer base.
Those insights can then be supplied directly to developers and AI agents. This begins to blur the traditional boundaries between product management, engineering and design. If an AI development agent can access the original customer problem, supporting evidence and criteria defining a successful solution, there may be less need to translate information repeatedly through several organisational layers before development begins.
However, faster product development introduces another constraint that AI cannot easily remove: the speed at which customers themselves can absorb change. Technically, software companies may increasingly be able to release major changes daily or even several times per day. Customers may not want interfaces and workflows to change at anything approaching the same rate.
Companies therefore face an emerging mismatch between machine development speed and human adoption speed. The ability to produce more features will not necessarily justify releasing them all immediately. This reinforces Dasdan’s original argument that productivity cannot be judged by output alone. Generating more software is not useful if customers cannot understand it, do not need it or experience constant disruption as products change.
Economics will also increasingly influence how companies design AI systems. The most powerful model will not necessarily be the appropriate choice for every problem. Businesses are likely to combine different models, using relatively inexpensive systems for routine work and more capable models for complex reasoning. They must also decide which parts of the AI stack should be developed internally and which should be purchased from external providers.
Data and organisational context may ultimately become one of the most defensible advantages. As access to powerful AI models becomes increasingly widespread, competitors can often obtain similar underlying technology. What differs is the information surrounding the model: internal documents, customer knowledge, operational history, workflows, permissions and proprietary data. That context can determine whether AI produces a generic response or a genuinely useful business answer.
Dropbox’s experience therefore points towards a broader transition in corporate AI strategy. The first stage was experimentation. The second was adoption. The next phase is likely to involve rebuilding workflows, development systems and information architecture around a workplace in which AI-generated output is abundant.
The limiting factor may no longer be how quickly employees can produce work. It may instead be whether organisations can identify the right problems, provide AI with the right context, maintain quality and security, redesign processes to absorb much greater output and ultimately convert that additional capacity into something customers actually value. That is the difference between using AI to perform tasks faster and using it to create business leverage.
Source: CIJ.World Research & Analysis Team
