Dropbox Hits 100% AI Adoption as Code Surge Strains Delivery Systems

AI-Driven Productivity, Collaboration, and Product Development with Dropbox — Ai4

Ai4

With AI generating up to 80% of codebase changes, software delivery bottlenecks have shifted from raw developer output to continuous integration pipelines and customer feature absorption limits.

AI-assisted summary of AI-Driven Productivity, Collaboration, and Product Development with Dropbox — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 23, 2026, this session features Dropbox CTO Ali Dasdan discussing how artificial intelligence is transforming knowledge worker collaboration, developer workflows, and product development into proactive, AI-native workspaces. The presentation outlines the technical architectures, retrieval systems, governance considerations, and operational strategies required to responsibly deploy enterprise AI systems at scale.

## TL;DR Ali Dasdan details Dropbox's journey to achieving universal, company-wide AI adoption and explains how true organizational leverage requires tying AI outputs directly to high-order business outcomes like security, quality, and customer satisfaction. He breaks down how Dropbox evaluates engineering velocity using the DX Core 4 framework augmented with AI-specific telemetry, demonstrating how freed developer bandwidth drives autonomous tech debt remediation. The talk explores the architectural design of contextual platforms like Dropbox Dash, highlighting the necessity of day-one permission boundaries and strategic model orchestration. Dasdan also analyzes downstream system bottlenecks in CI/CD pipelines and explains how piping unstructured customer feedback directly into internal agentic frameworks eliminates traditional product development friction.

## Atmosphere & style The presentation maintains a pragmatic, highly authoritative, and engineering-grounded tone, avoiding tech hype in favor of concrete operational data and structural frameworks. Dasdan speaks with the measured clarity of a seasoned technology executive, structuring the narrative around clear organizational maturity stages and architectural diagrams. Watching the talk feels like an insider masterclass on enterprise execution, balancing ambitious visions of agentic development with hardheaded realities around security, infrastructure strain, and human absorption capacity.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=oZiXWnHSjko&t=4s) AI creates organizational leverage only when advancing concrete business outcomes (quality, security, margins, cycle time, CSAT) rather than merely speeding up isolated task outputs. - [02:41](https://www.youtube.com/watch?v=oZiXWnHSjko&t=161s) Scaling AI adoption from 40% to 100% across an enterprise requires a structured four-stage playbook: explicit leadership signaling, protected exploration time, lateral peer enablement via AI Champions, and continuous reinforcement. - [02:41](https://www.youtube.com/watch?v=oZiXWnHSjko&t=161s) Dropbox developed 'Nova,' a proprietary internal infrastructure, to orchestrate agentic development workflows where 70% to 80% of engineers actively use multiple AI tools concurrently. - [04:31](https://www.youtube.com/watch?v=oZiXWnHSjko&t=271s) Measuring engineering productivity in the AI era requires balancing the DX Core 4 dimensions (Speed, Effectiveness, Quality, Impact) with AI-specific telemetry, such as PR throughput for codebases with 70% or more AI-generated logic. - [06:59](https://www.youtube.com/watch?v=oZiXWnHSjko&t=419s) Productivity dashboards should serve as collaborative conversation starters to diagnose systemic planning friction rather than surveillance mechanisms that incentivize metric gaming. - [06:59](https://www.youtube.com/watch?v=oZiXWnHSjko&t=419s) Productivity gains from AI empower engineers to autonomously reinvest reclaimed bandwidth into technical debt, security remediation, and massive migrations—such as updating 1,900 Python packages in a two-week sprint. - [09:51](https://www.youtube.com/watch?v=oZiXWnHSjko&t=591s) AI collaboration platforms like Dropbox Dash require a unified context layer with deeply integrated permission boundaries from day one to prevent unauthorized enterprise data leakage during synthesis. - [11:56](https://www.youtube.com/watch?v=oZiXWnHSjko&t=716s) Democratizing AI for non-technical employees creates powerful network effects but requires technical pairing and domain-specific champions to enforce production rigor, monitoring, and prompt evaluation. - [14:12](https://www.youtube.com/watch?v=oZiXWnHSjko&t=852s) With 70% to 80% of code generated by AI across top tech companies, the primary engineering bottleneck has shifted downstream to continuous integration (CI) infrastructure and human review capacity. - [14:12](https://www.youtube.com/watch?v=oZiXWnHSjko&t=852s) Product delivery speed must be balanced against human 'absorption cost,' acknowledging that end users have a finite capacity to adapt to rapid interface and workflow changes. - [18:09](https://www.youtube.com/watch?v=oZiXWnHSjko&t=1089s) Streaming real-time, unstructured customer signals directly into agentic development frameworks bypasses traditional requirement handoff bottlenecks between product managers and engineers. - [18:09](https://www.youtube.com/watch?v=oZiXWnHSjko&t=1089s) Proprietary enterprise data and contextual knowledge form the primary defensible moat against AI commoditization, requiring strategic model orchestration across open-weight and commercial LLMs. - [19:44](https://www.youtube.com/watch?v=oZiXWnHSjko&t=1184s) Moving past basic drafting and summarization to deep AI value requires embedding autonomous agents into core engineering operations (SRE, security) and deploying large-scale semantic indexing over billions of assets.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=oZiXWnHSjko&t=4s) The Core Thesis: Inputs, Outputs, and Outcomes

AI creates true organizational leverage only when it advances a concrete business outcome rather than merely speeding up an isolated task. While outputs are necessary stepping stones toward outcomes, simply producing things faster does not guarantee value. An organization can have an extraordinarily fast software developer writing massive amounts of code, but if they are building the wrong product, the raw speed is useless. True outcomes are defined by what genuinely matters to the business and its customers: baseline quality, robust security, healthy margins, faster time to market to test new hypotheses, and elevated customer satisfaction.

Productivity serves to systematically generate these higher-order outcomes by increasing organizational velocity. As an operational principle, "speed is the only advantage of any company"—it may not guarantee long-term survival on its own, but lacking speed is invariably fatal. Beyond velocity alone, the ultimate goal of AI-driven productivity is enabling better decisions that systematically compound into superior business results.

```mermaid flowchart LR accTitle: Progression from AI Inputs to Business Outcomes A["Input Layer<br/>(Prompts, training data, context, learnings)"] --> B["Output Layer<br/>(Drafts, decisions, code, tickets, summaries)"] B --> C["Business Outcomes<br/>(Quality, security, margins, time to market, CSAT)"] ```

- **Input Layer:** Prompts, specialized training, enterprise data access, organizational learnings. - **Output Layer:** First drafts, synthesized decisions, generated code, resolved tickets, summaries. - **Outcome Layer:** Core business metrics including customer satisfaction, security baselines, product margins, and reduced cycle time for experimentation.

---

[02:41](https://www.youtube.com/watch?v=oZiXWnHSjko&t=161s) Driving Company-Wide AI Adoption

Scaling AI across an enterprise requires a deliberate, multi-phase operational strategy. In March 2024, software engineering adoption of AI tools at Dropbox hovered around 40%. By the end of the year, engineering reached 100% adoption, and AI usage subsequently reached 100% across the entire company. Today, between 70% and 80% of Dropbox engineers actively utilize multiple AI tools simultaneously, supported by Dropbox's proprietary internal infrastructure, "Nova," built to orchestrate agentic software development workflows.

```mermaid flowchart TD accTitle: Four Stages of Enterprise AI Adoption S1["1. Leadership Signal<br/>(Budget, strategic priority, executive conviction)"] --> S2["2. Dedicated Exploration<br/>(Protected time to learn and experiment)"] S2 --> S3["3. Peer Enablement<br/>(AI Champions program, peer influence, use-case sharing)"] S3 --> S4["4. Continuous Reinforcement<br/>(Moving beyond initial adoption to maximum utilization)"] ```

Achieving universal adoption relies on four foundational stages: - **Leadership Signal:** Executives must provide explicit backing through dedicated budgets, concrete organizational goals, and genuine conviction so employees recognize AI adoption as an institutional priority. - **Exploration Time:** Leadership must grant teams explicit bandwidth and protected time to learn, experiment, and integrate new AI tools into their day-to-day routines. - **Peer Enablement:** Adoption scales through social proof and lateral learning rather than top-down mandates. Programs like the internal "AI Champions" network allow teams to showcase proven workflows and inspire organic peer adoption. - **Continuous Reinforcement:** Initial wins must be continuously reinforced and expanded upon, shifting the operational focus from baseline adoption to extracting maximum leverage and value.

---

[04:31](https://www.youtube.com/watch?v=oZiXWnHSjko&t=271s) Measuring Developer Productivity with the DX Core 4 Framework

To rigorously assess engineering productivity in the AI era, Dropbox moved beyond standard DORA metrics and implemented the DX Core 4 framework (developed by DX, acquired by Atlassian). The framework balances four interrelated dimensions: speed, effectiveness, quality, and impact. Speed tracks pull request (PR) cycle time and throughput; effectiveness captures qualitative developer sentiment and tool satisfaction via developer experience surveys; quality enforces service-level objectives (SLOs) spanning security, privacy, system stability, performance, and data integrity; and impact measures top-line returns such as revenue per engineer and operational cost efficiency.

```mermaid flowchart TD accTitle: DX Core 4 Framework for Engineering Productivity subgraph DX["DX Core 4 Framework"] S["Speed<br/>(PR cycle time, throughput)"] E["Effectiveness<br/>(Developer satisfaction, tool sentiment)"] Q["Quality<br/>(Security, privacy, stability, performance SLOs)"] I["Impact<br/>(Revenue per engineer, cost efficiency)"] end DX --> AI["AI-Augmented Metrics<br/>(AI PR throughput at ≥70% AI code, AI-generated code coverage)"] ```

Dropbox is actively augmenting this measurement system with AI-specific telemetry. Because massive influxes of generated code can easily obscure underlying engineering trends, Dropbox isolates metrics specifically for AI-generated assets—such as PR throughput for pull requests containing 70% or more AI-generated code, as well as test coverage tracked specifically across AI-generated logic. Tracking AI impact at this granular level ensures that velocity gains do not cause silent quality degradations across the codebase.

---

[06:59](https://www.youtube.com/watch?v=oZiXWnHSjko&t=419s) Dashboards as Conversation Starters and Autonomous Tech Debt Elimination

Engineering telemetry and productivity dashboards must never be weaponized as surveillance systems. Tracking individual output metrics like raw PR volume creates perverse incentives that encourage engineers to game the system. Instead, dashboards should serve as balanced "conversation starters" to diagnose system health—monitoring leading indicators such as roadmap health, quarterly commitment delivery rates, and scope churn. When roadmap churn exceeds 15%, for example, leadership can intervene to address underlying planning friction rather than penalizing individuals.

```mermaid flowchart LR accTitle: Autonomous Reinvestment of AI Productivity Gains A["AI Productivity Gains<br/>(Core roadmap completed faster)"] --> B["Freed Developer Bandwidth"] B --> C["Bottom-Up Technical Ownership"] C --> D["Non-Critical Security Remediation"] C --> E["Massive Framework Migrations<br/>(1,900 Python packages in 2 weeks)"] C --> F["System Resiliency & SLO Improvements"] ```

At Dropbox, measuring capacity over time reveals an upward trajectory of discretionary engineering investments completed outside baseline roadmap deliverables. Because customer trust (encompassing security, privacy, compliance, stability, performance, and data quality) is the company's highest strategic priority, engineers are autonomously redirecting their newly gained bandwidth into proactive platform health:

- **Autonomous Security Remediation:** Engineers are actively resolving non-critical security tickets that previously languished behind core roadmap commitments. - **Massive Automated Migrations:** In a standout initiative, a single software engineer leveraged AI tooling to fully automate the migration of 1,900 Python packages within a two-week sprint. - **Bottom-Up Agency:** High-velocity teams use their reclaimed hours to eliminate domain-specific technical debt and resolve pain points that might otherwise fall outside top-down corporate priorities.

---

[09:51](https://www.youtube.com/watch?v=oZiXWnHSjko&t=591s) Contextual AI Collaboration and the Dropbox Dash Architecture

True collaboration is inherently human and distributed across fragmented communication channels, documents, and meetings. AI cannot enhance collaboration simply by acting as an isolated generic chatbot; it must operate on top of a comprehensive "context layer." Dropbox Dash was designed around this principle as a unified contextual AI platform capable of federating and synthesizing disparate organizational knowledge into real-time answers.

```mermaid flowchart TD accTitle: Dropbox Dash Contextual AI Architecture Agent["Agent Platform<br/>(Intelligent assistants, cross-application orchestrators)"] Context["Context Platform<br/>(Real-time data ingestion, unified semantic understanding)"] Integrations["Enterprise Connectors<br/>(Salesforce, Slack, Google Workspace, Third-Party Tools)"] Trust["Native Governance & Trust Layer<br/>(Real-time access control, permission boundary enforcement)"]

Integrations --> Context Trust -.-> Context Trust -.-> Agent Context --> Agent ```

During high-stakes scenarios—such as unexpected executive inquiries during board meetings—contextual AI allows leaders to retrieve precise, verified answers across organizational repositories in real time without needing to pull engineers away from active work. Crucially, enterprise governance and permissions cannot be casually "bolted on" after indexing. Because generative models blend contextual data during synthesis, permission boundaries must be structurally embedded into the core architecture from day one to prevent unauthorized information leakage across user roles.

---

[11:56](https://www.youtube.com/watch?v=oZiXWnHSjko&t=716s) Driving Organizational Network Effects and Cross-Functional Enablement

Widespread AI enablement generates powerful organizational network effects, but it also introduces new operational risks. Today's generative tools empower non-technical personnel across business units to independently build functional, complex applications. While this unlocks unprecedented employee agency, non-technical creators frequently overlook critical downstream production requirements, occasionally prompting confusion when asked basic infrastructure questions such as "are you monitoring this?"

```mermaid flowchart LR accTitle: Cross-Functional AI Network Effects A["Domain-Specific AI Champions<br/>(HR, Operations, Finance)"] --> B["Workflow Pairing<br/>(Forward-Deployment Engineering)"] B --> C["Targeted Hackathons & AI Showcases"] C --> D["Broad Organizational Capability Uplift"] ```

To bridge this operational gap and maintain system integrity, organizations must build intentional cross-functional scaffolding: - **Internal Workflow Pairing:** Technical staff partner directly with non-technical builders—analogous to forward deployment engineers—to establish essential monitoring, prompt evaluation rigor, and operational lifecycle practices. - **Domain-Specific AI Champions:** Rather than relying exclusively on central software engineers, functional units (such as HR or Finance) deploy dedicated champions who understand and communicate within the native terminology of their respective disciplines. - **Targeted Hackathons and Recognition:** Dedicated, AI-specific hack events and public showcase competitions demystify technical barriers, allowing employees to see peers with no formal programming background building powerful workflows.

---

[14:12](https://www.youtube.com/watch?v=oZiXWnHSjko&t=852s) Resolving System Bottlenecks in AI Product Development

While generative AI has largely solved raw code generation, end-to-end system delivery has emerged as the industry's critical bottleneck. At a recent industry AI summit hosted by Dropbox alongside engineering peers from Uber, Netflix, and other major technology firms, shared patterns surfaced: top tech organizations have roughly doubled engineering productivity, with 70% to 80% of all committed code being AI-generated. However, this sheer volume of code puts unprecedented strain on continuous integration (CI) environments and code review infrastructure, pushing platforms like GitHub to their operational limits and prompting tools like Cursor to re-architect downstream delivery pipelines.

```mermaid flowchart LR accTitle: Modern AI Software Delivery Pipeline and Verification Gen["AI Code Generation<br/>(70-80% of codebase)"] --> Vuln["Automated Vulnerability Hunter<br/>(Security scanning, fault injection)"] Vuln --> Review["Mandatory Human Code Review<br/>(Final approval gate)"] Review --> CI["CI/CD Pipeline Verification<br/>(System delivery bottleneck)"] ```

To maintain code health under high throughput, Dropbox balances automated security tooling with mandatory human oversight: - **Human-in-the-Loop Reviews:** While automated review tools accelerate pull request evaluations, human developers maintain final approval authority over production changes. - **Automated Vulnerability Hunting:** Security engineers created an internal "vulnerability hunter" that automatically flags vulnerabilities and performs continuous fault-injection tests to verify system resilience without incurring massive external token overhead. - **Customer Absorption Constraints:** Even when teams can deploy production features multiple times per day, product teams must account for human "absorption cost"—the rate at which end users can realistically adapt to interface updates and workflow modifications. - **Role Blending and Unstructured Customer Data:** Modern agentic workflows are dissolving the traditional boundary between Product Managers, UX researchers, and developers. By leveraging generative models to extract structured insights from unstructured customer call transcripts and support tickets in real time, organizations can feed verified customer pain points directly into agentic development workflows.

[18:09](https://www.youtube.com/watch?v=oZiXWnHSjko&t=1089s) Connecting Customer Needs to Autonomous Agents and the Emerging AI Architecture

Directly connecting real customer pain points to engineering teams and autonomous systems fundamentally transforms how software gets built. Rather than relying on the traditional, imperfect telephone game where engineers build features based on what they thought they understood from a product manager who only spoke with ten people, organizations can pipe live user signals directly into their engineering workflows. When paired with an internal agentic AI development framework, engineers and autonomous agents are equipped to solve authentic user frustrations far more comprehensively and rapidly than legacy product development cycles ever allowed.

At the same time, the architectural and economic realities of AI are shifting how products are constructed and valued. Interfaces are becoming dynamically adaptive; when built upon a robust underlying platform, AI allows users and developers to spin up personalized, custom interfaces on demand. However, this flexibility introduces strict economic constraints because running advanced AI models is far from cheap. To succeed, companies must recognize that their proprietary data and unique contextual knowledge represent their true competitive moat. Moreover, organizations cannot rely on a single model to solve every task. Technology leaders must actively navigate model orchestration—deciding how to blend open-weight models with proprietary ones, and evaluating whether to build customized AI interfaces internally or purchase them from the growing ecosystem of specialized startups.

```mermaid flowchart LR accTitle: End-to-End Agentic Product Development Flow A["Real Customer Signals & Context"] --> B["Internal Agentic AI Framework"] B --> C["Engineers & Specialized Agents"] C --> D["Adaptive Custom Interfaces"] C --> E["Domain-Specific Models (Open-Weight & Proprietary)"] ```

- **Bridging the requirement gap:** Direct integration of customer feedback with agentic frameworks bypasses the bottleneck of relying solely on small PM interview samples. - **Data as a moat:** Proprietary enterprise data and context are the foundational elements that make AI genuinely powerful and defensible against commoditization. - **Model orchestration:** No single model fits every use case; teams must strategically balance performance, cost, and open-weight versus closed-source architectures.

---

[19:44](https://www.youtube.com/watch?v=oZiXWnHSjko&t=1184s) Q&A: Moving Beyond Remedial Tasks to Deep Value with Dropbox Dash

During the brief audience Q&A, an attendee asked how Dropbox managed to move beyond low-level, remedial applications like drafting emails to achieve substantial productivity gains. Ali Dasdan explained that while standard knowledge worker capabilities like writing assistance and text summarization are widely adopted across the company, Dropbox focuses on deploying AI where the technical leverage and organizational scale are most profound.

```mermaid flowchart TD accTitle: Dropbox AI Application Ecosystem A["Dropbox AI Adoption"] --> B["Internal Engineering Lifecycle"] A --> C["Product Capabilities (Dropbox Dash)"]

B --> B1["Code Generation & Review"] B --> B2["Custom SRE Agents"] B --> B3["Internal Security Agents"]

C --> C1["Semantic Indexing Engine"] C --> C2["Search Across Millions of Files"] C --> C3["Retrieval Over Hundreds of Billions of Images & Videos"] ```

Internally, Dropbox has embedded AI across every phase of software development—ranging from automated coding and rigorous code reviews to purpose-built, internal site reliability engineering (SRE) agents and dedicated security agents. Externally, this philosophy powers products like Dropbox Dash. With enterprise users managing individual environments containing millions of files, alongside a global platform corpus spanning hundreds of billions of images and videos, traditional keyword search breaks down. By implementing advanced semantic indexing, the platform enables users to surface precisely what they are searching for based on intent and deep context. Rather than settling for superficial chatbots, AI is deeply integrated into operational systems and large-scale data retrieval to solve concrete platform-scale challenges.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=oZiXWnHSjko&t=4s) speed is the only advantage of any company — Ali Dasdan

> [00:04](https://www.youtube.com/watch?v=oZiXWnHSjko&t=4s) AI creates true organizational leverage only when it advances a concrete business outcome rather than merely speeding up an isolated task. — Ali Dasdan

> [02:41](https://www.youtube.com/watch?v=oZiXWnHSjko&t=161s) In March 2024, software engineering adoption of AI tools at Dropbox hovered around 40%. By the end of the year, engineering reached 100% adoption, and AI usage subsequently reached 100% across the entire company. — Ali Dasdan

> [06:59](https://www.youtube.com/watch?v=oZiXWnHSjko&t=419s) Engineering telemetry and productivity dashboards must never be weaponized as surveillance systems. Tracking individual output metrics like raw PR volume creates perverse incentives that encourage engineers to game the system. — Ali Dasdan

> [06:59](https://www.youtube.com/watch?v=oZiXWnHSjko&t=419s) A single software engineer leveraged AI tooling to fully automate the migration of 1,900 Python packages within a two-week sprint. — Ali Dasdan

> [09:51](https://www.youtube.com/watch?v=oZiXWnHSjko&t=591s) Because generative models blend contextual data during synthesis, permission boundaries must be structurally embedded into the core architecture from day one to prevent unauthorized information leakage across user roles. — Ali Dasdan

> [11:56](https://www.youtube.com/watch?v=oZiXWnHSjko&t=716s) are you monitoring this? — Ali Dasdan

> [14:12](https://www.youtube.com/watch?v=oZiXWnHSjko&t=852s) Top tech organizations have roughly doubled engineering productivity, with 70% to 80% of all committed code being AI-generated. However, this sheer volume of code puts unprecedented strain on continuous integration (CI) environments and code review infrastructure — Ali Dasdan

> [18:09](https://www.youtube.com/watch?v=oZiXWnHSjko&t=1089s) Directly connecting real customer pain points to engineering teams and autonomous systems fundamentally transforms how software gets built. — Ali Dasdan

> [18:09](https://www.youtube.com/watch?v=oZiXWnHSjko&t=1089s) To succeed, companies must recognize that their proprietary data and unique contextual knowledge represent their true competitive moat. — Ali Dasdan

> [19:44](https://www.youtube.com/watch?v=oZiXWnHSjko&t=1184s) Rather than settling for superficial chatbots, AI is deeply integrated into operational systems and large-scale data retrieval to solve concrete platform-scale challenges. — Ali Dasdan

## Mentioned Ali Dasdan, Dropbox, Ai4, Nova, DX Core 4, DX, Atlassian, DORA, Python, Dropbox Dash, Salesforce, Slack, Google Workspace, Uber, Netflix, GitHub, Cursor

## Video description Ali Dasdan - CTO, Dropbox

Artificial intelligence is fundamentally reshaping how knowledge workers collaborate, create, and build software. Productivity applications are evolving into AI-native workspaces where intelligent assistants can summarize meetings, generate content, automate workflows, retrieve organizational knowledge, and help developers write, review, debug, and understand code. Rather than simply responding to commands, modern AI systems are becoming proactive collaborators that anticipate user needs, personalize experiences, and orchestrate work across applications.

This transformation extends well beyond chatbots. AI is changing how organizations capture institutional knowledge, coordinate projects, manage information, and develop software. As these capabilities become embedded across productivity suites, communication platforms, developer tools, and enterprise applications, organizations face new challenges around context management, accuracy, security, governance, privacy, and trust.

This session explores the technologies, architectures, and product strategies powering the next generation of AI-driven productivity, collaboration, and product development. We’ll examine how retrieval, enterprise knowledge, personalization, coding and operations copilots, workflow automation, and organizational memory combine to deliver intelligent experiences that are accurate, secure, and deeply integrated into everyday work. We’ll also discuss how vendors are rethinking user experience, pricing, and business models as AI becomes a core capability rather than an add-on.

Attendees will leave with practical insights into designing AI experiences that improve productivity, accelerate product development, enhance collaboration, and responsibly deploy AI systems that users can trust.

[Source](https://www.youtube.com/watch?v=oZiXWnHSjko)

中文简报

Dropbox 实现 100% AI 普及率，代码量激增对交付系统构成压力 随着 AI 生成的代码变更占比高达 80%，软件交付瓶颈已从开发者的原始产出转移到了持续集成流水线以及客户对功能吸收能力的极限。 Ali Dasdan 详细介绍了 Dropbox 实现全公司全面采用 AI 的历程，并解释了真正的组织杠杆需要将 AI 产出直接与安全、质量和客户满意度等高阶业务成果挂钩。他剖析了 Dropbox 如何使用结合了 AI 专用遥测指标的 DX Core 4 框架来评估工程速度，展示了解放出的开发者精力如何推动自主的技术债务修复。演讲探讨了 Dropbox Dash 等上下文平台的架构设计，强调了在项目初期确立权限边界和战略性模型编排的必要性。Dasdan 还分析了 CI/CD 流水线中的下游系统瓶颈，并阐释了将非结构化客户反馈直接接入内部 Agent 框架如何消除传统产品开发中的摩擦。 只有在推动具体的业务成果（质量、安全、利润率、周期时间、客户满意度）而非仅仅加速孤立的任务产出时，AI 才能创造组织杠杆。 (0:04) 在整个企业中将 AI 普及率从 40% 扩大到 100% 需要一套结构化的四阶段指南：明确的领导层信号、受保护的探索时间、通过 AI Champions 进行横向同侪赋能，以及持续强化。 (2:41) Dropbox 开发了专有内部基础设施“Nova”来编排 Agent 开发工作流，目前有 70% 至 80% 的工程师在积极同时使用多种 AI 工具。 (2:41) 在 AI 时代衡量工程生产力，需要兼顾 DX Core 4 维度（速度、效能、质量、影响力）与 AI 专用遥测指标，例如针对 AI 生成逻辑占比达到或超过 70% 的代码库的 PR 吞吐量。 (4:31) 生产力仪表盘应当用作诊断系统性规划摩擦的协作对话契机，而不是诱发钻指标空子行为的监控机制。 (6:59)

Reading summary

AI creates organizational leverage only when task-level speed translates into concrete business outcomes such as baseline quality, security, and customer satisfaction. According to Dropbox engineering leadership, capturing this value requires structured enterprise adoption, balanced productivity telemetry, and context-aware platforms that direct reclaimed bandwidth toward platform resilience and real customer needs. Scaling AI Adoption Toward Business Outcomes (0:04) Operational velocity matters only when it serves higher-order outcomes, as accelerating raw software output yields little value if engineers build the wrong products. To systematically realize business gains, Dropbox expanded software engineering adoption of AI from 40% in March 2024 to 100% by year-end, with 70% to 80% of engineers now running multiple AI tools concurrently through its proprietary 'Nova' agentic development framework. Leadership outlines four stages for achieving enterprise-wide adoption: establishing explicit executive backing and dedicated budgets, granting teams protected exploration time, enabling lateral learning through an internal 'AI Champions' network, and continuously reinforcing utilization. Furthermore, as non-technical employees build functional applications, organizations must provide cross-functional scaffolding—pairing technical engineers with business units to implement essential monitoring, prompt evaluation rigor, and operational lifecycles. Telemetry and Autonomous Tech Debt Remediation (4:31) To evaluate engineering performance in the AI era without creating perverse incentives, Dropbox implemented the DX Core 4 framework, balancing speed, effectiveness, quality, and impact. Because large volumes of generated code can mask underlying quality issues, the company tracks specific AI telemetry, such as pull request throughput for submissions containing 70% or more AI-generated code and test coverage across AI-authored logic. Dashboards function as diagnostic conversation starters rather than surveillance systems, prompting managerial intervention only when leading indicators—like roadmap churn exceeding 15%—reveal planning friction. Critically, productivity gains are tracked by measuring engineering capacity reinvested outside core roadmap commitments. With customer trust designated as Dropbox's highest strategic priority, developers autonomously redirect freed hours into platform health. Standout results include engineers independently clearing non-critical security backlogs and a single developer using AI tools to migrate 1,900 Python packages in a two-week sprint. Managing Delivery Bottlenecks and Review Constraints (14:12) While generative tools have largely solved initial code synthesis, downstream delivery has emerged as the critical system bottleneck. At an industry summit hosted by Dropbox alongside peers such as Uber and Netflix, engineering leaders noted that top tech firms have roughly doubled productivity, with 70% to 80% of committed code being AI-generated. This volume strains continuous integration pipelines and review infrastructure, pushing tools like GitHub and Cursor to their operational limits. To safeguard production stability under high throughput, Dropbox maintains mandatory human approvals for code reviews while using internal automated 'vulnerability hunters' to execute security scans and fault injections without massive token overhead. Concurrently, product organizations must pace deployment frequency against customer absorption costs—the rate at which users can realistically adapt to workflow and interface modifications. Contextual Platforms and Direct Customer Feedback Loops (9:51) Generic chatbots fail in enterprise environments because collaboration requires federated organizational context. Platforms like Dropbox Dash address this by operating on a unified context layer that integrates disparate tools, with permission boundaries structurally embedded from inception to prevent unauthorized information leakage across user roles. At platform scale, semantic indexing replaces traditional keyword search across corporate repositories containing millions of files and hundreds of billions of images and videos. Directly feeding unstructured customer signals—such as support tickets and call transcripts—into internal agentic frameworks eliminates the traditional communication breakdown between product managers and engineers. Navigating this emerging architecture requires technology leaders to orchestrate mixtures of proprietary and open-weight models, recognizing that proprietary enterprise data remains the primary defensible competitive moat. While AI tooling resolves the bottleneck of initial content and code generation, enterprises face ongoing challenges in managing downstream continuous integration limits, enforcing data governance across synthesized contexts, and pacing product releases against user absorption capacity.

精炼阅读摘要

Dropbox指出，人工智能只有在推动具体业务成果而非单纯加速孤立任务时，才能产生真正的组织杠杆。企业若要将AI生产力转化为持久竞争优势，关键在于将速度转化为更好的决策、建立多维效能度量，并将专有上下文内嵌于系统架构之中。 业务成果导向与企业级AI采纳路径 (0:04) Dropbox的核心论点认为，AI必须服务于具体的业务成果，包括基线质量、安全性、利润率、上市实验速度以及客户满意度。虽然“速度是企业唯一的优势”，缺乏速度必定致命，但若缺乏明确业务方向，单纯依靠AI高速编写代码或生成内容并无实际价值。生产力的终极目标是通过提升组织周转速率，赋能更优决策并沉淀为商业成果。 在组织推广方面，Dropbox软件工程团队的AI采纳率从2024年3月的约40%迅速增长至年底的100%，随后实现全公司100%覆盖。目前70%至80%的工程师依托内部专有基础设施“Nova”同时运行多个AI工具进行智能体开发。该公司将全员采纳归纳为四个阶段：高管通过明确预算与战略优先级释放明确信号、保障团队探索与试错时间、设立“AI冠军”网络进行同级经验推广，以及持续强化以从基线使用迈向价值最大化。 DX四维效能度量与自主技术债治理 (4:31) 为了在AI时代精准评估工程生产力，Dropbox超越传统DORA指标，引入了DX Core 4框架，从速度（PR周期与吞吐量）、有效性（开发者满意度与工具体验）、质量（安全、隐私、系统稳定性与数据完整性SLO）以及业务影响（人均创收与成本效率）四个维度进行综合平衡。针对海量生成代码对工程趋势的遮蔽，Dropbox设立了AI专项遥测，独立追踪AI代码占比达到70%以上的PR吞吐量，并对AI生成逻辑的测试覆盖率实施严格监控，防止研发速度提升隐蔽地破坏代码质量。 Dropbox明确将监控仪表板定位为诊断系统健康度的“对话启动器”，而非针对个人的绩效监视系统，以避免诱发刷量作弊行为。当路线图变更率超过15%等前置指标异常时，管理层介入解决规划摩擦。数据表明，效率提升释放的工程容量正自发流向底层架构维护：工程师主动修复长期搁置的非关键安全工单，并在两周内由单名工程师借助AI全自动完成了1900个Python依赖包的框架迁移。 全流程交付瓶颈与系统级质量防御 (14:12) Dropbox在与Uber、Netflix等科技企业的行业交流中发现，AI代码生成已推动工程生产力近乎翻倍，提交的代码中有70%至80%由AI生成。然而，这一趋势导致行业瓶颈从代码编写转移到了下游的持续集成（CI）与代码审查基础设施，迫使平台重构交付管线。同时，即便工程团队能够每天多次部署功能，产品发布仍必须考量最终用户的“吸收成本”，即用户适应界面与工作流调整的心理承受速率。 为确保高通量交付下的系统质量，Dropbox采取了双重防御策略：一方面依靠内部安全团队研发的“漏洞猎手”自动扫描漏洞并进行持续故障注入测试；另一方面严格坚持人类在环（Human-in-the-Loop）机制，保留人工对生产环境变更的最终审批权。此外，通过生成式AI将客户通话记录和支持工单等非结构化数据实时转化为结构化洞察，产品经理、UX研究员与开发人员的角色边界正在消融，真实痛点得以直接输入智能体开发管线。 上下文协同架构与企业数据壁垒 (9:51) 协同AI无法通过通用聊天机器人实现，必须建立在统一的上下文层之上。Dropbox Dash产品架构将上下文平台、代理平台与底层企业连接器打通。由于生成式模型在合成信息时会打通上下文数据，权限边界与治理规则必须结构化内嵌于核心架构中，杜绝事后修补带来的跨角色越权泄露隐患。在底层检索能力上，针对用户个人数百万文件以及全平台数千亿图片和视频，Dash通过语义索引实现基于意图与深度上下文的高精度检索。 在跨职能推广中，非技术人员利用AI构建复杂应用时常忽略生产环境监控。对此，Dropbox推行技术人员与业务人员配对机制以补齐生命周期管理，并在各职能部门部署专属AI冠军。面对高级模型的高昂运行成本，企业必须清醒认识到专有数据与业务上下文才是真正的防御壁垒，技术决策者需要在开源权重模型与商业专有模型之间进行理性编排与取舍。 Dropbox的实践表明，AI为企业带来的核心优势并非单纯的代码生成速度，而是专有上下文的整合深度与系统交付韧性。然而，当下游CI/CD管线负荷加剧、高级模型推理成本居高不下，以及终端用户对频繁迭代的功能吸收能力存在天然上限时，如何平衡快速交付与长期系统架构及用户承受力，仍是企业深化AI应用过程中必须持续应对的结构性挑战。

Detailed notes — zh

## Background 本内容由 Ai4 于 2026 年 8 月 23 日发布在 YouTube 平台上。视频记录了 Dropbox 首席技术官 Ali Dasdan 的主题演讲，主要探讨在知识工作者协作、企业知识检索与软件研发全生命周期中，AI 如何从被动响应工具蜕变为主动协作的智能工作空间，以及企业在上下文管理、安全信任与架构演进上面临的战略抉择。 ## TL;DR 演讲开宗明义地指出，AI 的核心价值在于推动实际业务成果而非单纯完成任务产出。Ali Dasdan 详细分享了 Dropbox 工程师实现 100% 内部 AI 采纳的四步推广路径，以及基于 DX Core 4 平衡速度、有效性、质量与影响力的度量体系。他进一步剖析了当前研发瓶颈从代码生成向持续集成与系统交付的转移，并介绍了企业级上下文检索平台 Dropbox Dash 与 Agentic AI 如何颠覆传统 PRD、直连海量客户一手痛点的工程实践。 ## Atmosphere & style 整场演讲呈现出典型硅谷顶尖工程掌舵者的沉稳、务实与前瞻气质，语言精炼透彻，毫无空洞的概念炒作。Dasdan 穿插着诸如非技术员工反问“什么是监控”等生动风趣的实战掌故，既有对“指标被投机取巧”与 CI 管道过载等现实痛点的敏锐洞察，又洋溢着对 Agentic 研发范式演进的笃定与激情，让听众在紧凑的节奏中深切感受到一线科技巨头在 AI 落地风暴眼中的架构思考与组织变革魄力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=oZiXWnHSjko&t=4s) AI 只有在推动实际业务成果（Outcome）而非单纯完成某项具体任务（Task）时，才能真正释放商业杠杆价值。 - [00:04](https://www.youtube.com/watch?v=oZiXWnHSjko&t=4s) 研发速度是企业唯一的竞争壁垒，但提速的核心目的在于借助更短的反馈周期达成更优的决策与质量，而非盲目追求代码产出量。 - [02:41](https://www.youtube.com/watch?v=oZiXWnHSjko&t=161s) 全员 AI 普及需依托“明确管理信号、留出探索时间、跨部门全方位赋能、持续强化应用”这四步推广框架。 - [04:31](https://www.youtube.com/watch?v=oZiXWnHSjko&t=271s) 科学衡量研发效能须在速度、有效性、质量与商业影响（Core 4 框架）之间维持平衡，警惕效能看板异化为监控工具导致员工钻营指标。 - [07:07](https://www.youtube.com/watch?v=oZiXWnHSjko&t=427s) AI 带来的研发产能盈余赋能工程师自下而上主导技术攻坚，如自主修复次级安全漏洞及两周内完成 1,900 个 Python 依赖包的自动迁移。 - [09:51](https://www.youtube.com/watch?v=oZiXWnHSjko&t=591s) 企业协同的核心瓶颈在于信息割裂，Dropbox Dash 依托上下文平台打通 Slack、Salesforce 等异构数据源并严格继承原始权限。 - [11:05](https://www.youtube.com/watch?v=oZiXWnHSjko&t=665s) 企业级 AI 检索系统的安全与信任机制必须深植于底层架构，事后通过顶层打补丁或叠加权限过滤无法阻断 AI 综合推理导致的越权泄漏。 - [11:56](https://www.youtube.com/watch?v=oZiXWnHSjko&t=716s) 推行“前向部署工程师”工作流配对与业务部门 AI 拥趸制度，能够化解非技术人员开发应用时忽视系统运维与监控的脱节隐患。 - [14:12](https://www.youtube.com/watch?v=oZiXWnHSjko&t=852s) 软件研发的瓶颈已从代码生成转移至持续集成（CI）吞吐与交付基础设施，海量 AI 生成代码给自动化测试与安全审查带来巨大压力。 - [15:54](https://www.youtube.com/watch?v=oZiXWnHSjko&t=954s) Agentic AI 能够直接将海量客服工单与通话数据提炼为全局痛点看板，颠覆传统层层转述的 PRD 模式，让研发与智能体直面真实客户诉求。 - [15:54](https://www.youtube.com/watch?v=oZiXWnHSjko&t=954s) 高频迭代必须考量客户吸收成本，若用户界面变动过于频繁人类将无法适应，除非交互演化为智能体对智能体的自主对接。 - [19:51](https://www.youtube.com/watch?v=oZiXWnHSjko&t=1191s) 超越日常琐碎任务的关键在于构建深度语义索引以及全研发生命周期的专用智能体（如 SRE Agent 和 Security Agent）。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=oZiXWnHSjko&t=4s) 核心论点：AI 的价值在于推进业务成果，而非单纯完成任务 Dropbox 首席技术官 Ali Dasdan 开门见山地提出了整场演讲的核心论题：“AI 只有在推进实际成果（Outcome）而非仅仅完成某项具体任务（Task）时，才能真正产生杠杆效应（leverage）”。他幽默地表示希望大家还没有听烦 AI，并强调尽管任务的输出（Output）是产生成果的必经阶段，但单纯追求产出很容易陷入误区。比如一位软件开发者凭借 AI 获得了极高的编码速度，但如果写出来的是错误的产品，这种快速产出就毫无商业价值。 ```mermaid flowchart LR accTitle: 从输入到成果的AI杠杆演化路径 subgraph S1["输入层 (Input)"] I1["Prompt 提示词"] I2["训练与数据"] I3["知识沉淀"] end subgraph S2["产出层 (Output)"] O1["草稿与决策"] O2["代码与工单"] O3["会议摘要"] end subgraph S3["成果层 (Outcome)"] R1["产品质量与安全"] R2["利润率与上市时间"] R3["客户满意度"] end S1 --> S2 S2 --> S3 ``` 真正的生产力必须落脚在业务与客户真正关心的核心成果上。质量和安全性是业务的准入门槛（Table stakes），在此之上则是利润率、上市时间（Time to Market，决定了团队能尝试多少新想法）以及客户满意度。Ali Dasdan 向团队灌输的座右铭是：“速度是任何公司唯一的优势；它或许不是决定你能否存活的唯一决策指标，但如果速度不够快，那就必然会有大麻烦。”然而，提速的终极诉求绝不仅仅是快，而是通过更快的反馈机制带来更好的决策与更优的结果。 ### [02:41](https://www.youtube.com/watch?v=oZiXWnHSjko&t=161s) 从 40% 到 100%：Dropbox 的 AI 内部普及与推广四步法 在落地实施方面，Dropbox 的 AI 普及经历了一场彻底的变革。以软件研发团队为例，2025 年 3 月时使用 AI 的工程师比例仅为 40% 左右；到了年底，该比例迅速攀升至 100%，目前全公司已实现 100% 的 AI 使用覆盖。在工程团队中，有 70% 至 80% 的人日常同时使用多个 AI 工具。此外，Dropbox 还自主研发了名为“Nova”的底层基础设施，用于支持 Agent 级别的自主开发（Agentic Development）。 驱动这种全员转型的关键在于明确的四步采纳路径： - **明确信号与领导层支持（Signal）**：领导层必须展示坚定的意愿，提供预算、将其确立为明确的组织目标，并以身作则，让员工看到管理层认真的态度。 - **给予充足的探索时间（Time for Exploration）**：如果只提要求却不给员工尝试工具、学习摸索的时间，推广就会沦为空谈。 - **全方位赋能与同伴示范（Enablement）**：大多数学习源于同事之间的观察、相互影响甚至适度的同伴压力。Dropbox 设立了“AI 冠军计划”（AI Champions Program），展示优秀用例供大家复用和借鉴。 - **持续强化与迭代（Reinforcement）**：初期成功之后不能止步，必须持续强化应用习惯，从“如何开始用”进阶到“如何最大化发挥 AI 价值”。 ### [04:31](https://www.youtube.com/watch?v=oZiXWnHSjko&t=271s) 平衡研发度量：基于 DX Core 4 的多维生产力评估 为了科学评估生产力，Dropbox 引入了 DX（被 Atlassian 收购的研发效能度量机构）的 Core 4 框架，而非单纯依赖传统的 DORA 指标。Ali Dasdan 强调必须在这四个维度之间寻求动态平衡，防止单一指标导致团队行为变形： | 评估维度 | 核心定义与指标范畴 | 结合 AI 的延伸与细化度量 | |:--- |:--- |:--- | | **速度 (Speed)** | PR 流转周期（Cycle Time）、PR 吞吐量（Throughput）等研发速度指标 | 专门追踪“AI 生成为主的 PR”吞吐量（例如 AI 生成代码占比超 70% 的 PR 流转表现） | | **有效性 (Effectiveness)** | 开发者体验（DevEx），主要基于问卷调查获取质性反馈，衡量工具满意度 | 评估开发者在 AI 协同工作流程中的主观体验与摩擦力 | | **质量 (Quality)** | 全面的服务等级目标（SLO），覆盖安全、隐私、系统稳定性、性能及数据质量 | 单独评估由 AI 生成的代码的测试覆盖率（Code Coverage），防止大批量代码稀释质量缺陷 | | **影响 (Impact)** | 商业效益指标，如每位工程师创造的收入（Revenue per Engineer）、成本效率等 | 衡量 AI 带来的复合研发产出与实际商业价值增益 | 在组织推进路径上，Dropbox 内部通常以“EPD”（即工程 Engineering、产品 Product、设计 Design）为核心发力点。工程部门最先接入并跑通流程与质量机制，随后产品和设计团队迅速跟进，最后扩展到业务运营（BizOps）及全公司其他职能部门。 ### [07:07](https://www.youtube.com/watch?v=oZiXWnHSjko&t=427s) 避免监控陷阱与自下而上的效能释放：1900 个 Python 包的迁移奇迹 针对效能度量看板，Ali Dasdan 发出了严肃警告：如果度量仪表盘被员工视为“监视系统”（例如紧盯谁提交的 PR 更大），必然会催生投机取巧的行为（Gaming the metrics）。看板应该作为复盘与改进的“对话切入点”（Conversation Starter），结合路线图健康度（如季度交付承诺达成率、需求变更率，若需求变动或取消超过 15% 则需警惕规划动荡）来全面观察，而非用于对员工进行优劣排队。 随着 AI 带来显著的生产力红利，Dropbox 观察到一个极具价值的现象：在不增加资源投入的情况下，工程师在完成预定路线图承诺之余，自发投入到了更多高价值领域： - **非关键安全漏洞修复**：尽管客户信任（涵盖安全、隐私、合规、稳定性等）已被列为公司第一优先级，但工程师利用节省下来的时间，主动清除了大量原本未被列为紧急事项的次级安全工单。 - **技术债清理与自动化迁移**：一名工程师借助 AI 工具，仅用两周时间就彻底完成了整个代码库中 1,900 个 Python 依赖包的自动化迁移工作。 - **自下而上的主观能动性（Agency）**：Dropbox 的路线图规划采取自下而上的方式，生产力盈余让工程师能够在对齐公司战略的同时，有充裕的自主时间去攻克那些自己认为至关重要但未列入公司最高优先级的工程痛点。 ### [09:51](https://www.youtube.com/watch?v=oZiXWnHSjko&t=591s) 协同的核心是上下文：Dropbox Dash 的架构与实战价值 转向协作（Collaboration）领域时，Ali Dasdan 认为最核心的基础设施是“上下文层”（Context Layer）。协同本质上是高度复杂的组织活动，分散在聊天记录、文档和会议等大量载体中。单纯给员工分发 AI 工具无法解决协同痛点，必须构建一个能够将组织信息实时汇聚、提供情境化理解的 AI 体系，才能显著提升协同效能。 ```mermaid flowchart TD accTitle: Dropbox Dash 上下文AI架构 subgraph Agents["智能体平台 (Agent Platform)"] A1["任务规划与执行"] A2["企业级应用交互"] end subgraph Context["上下文平台 (Context Platform)"] C1["跨系统实时数据流"] C2["严格权限控制与映射"] end subgraph Sources["数据源接入"] S1["Salesforce"] S2["Slack"] S3["第三方系统 / 网盘文档"] end Sources --> Context Context --> Agents ``` Dropbox 为此打造了企业级上下文 AI 平台——Dropbox Dash。该平台顶层为 Agent 平台，中层为上下文平台，能够实时集成 Slack、Salesforce 以及各类第三方系统中的数据，并在全流程中严格遵循用户的原有权限设置。Ali Dasdan 分享了一个切身的使用场景：在参加董事会会议时，他偶尔会遇到无法当场解答、也无法即时联系团队成员解答的突发问题；此时他直接在 Dropbox Dash 中发起搜索，系统便能基于企业上下文立刻为他整理出高度精准的答案。 ### [11:05](https://www.youtube.com/watch?v=oZiXWnHSjko&t=665s) 构建兼具上下文与信任的企业级 AI 检索架构 Ali Dasdan 首先探讨了 AI 搜索的底层架构本质。他强调，现代企业级 AI 搜索绝不仅仅是语言模型凭空捏造答案，而是必须通过丰富、精准的上下文进行强化（Context-Enhanced Search）。整个检索系统的架构支柱清晰地分为四个核心环节：连接（Connecting）、理解（Understanding）、编排（Orchestrating）以及控制（Control）。 在系统架构中，安全与信任（Trust）必须是内生构件，正如软件工程中一句古老的格言所说："你无法把信任事后螺栓式硬拧进系统"（You cannot bolt it in）。例如，当系统从第三方平台拉取数据时，绝不能采取"先索引所有数据、再在顶层叠加权限过滤"的偷懒做法。由于 AI 模型具有极强的混合与综合能力，在生成回答时会把各类数据碎片融合在一起，导致权限边界变得极为模糊甚至失效。因此，数据访问控制与合规边界必须从系统设计的第一天起就深植于底层架构之中。 ```mermaid flowchart LR accTitle: 企业级 AI 检索核心流程 subgraph S1["数据层"] A["连接第三方数据"] --> B["理解与语义解析"] end subgraph S2["核心处理层"] B --> C["上下文编排"] C --> D["生成与检索控制"] end subgraph S3["底层基石"] E["原生内置权限与信任机制"] -.-> A E -.-> B E -.-> C E -.-> D end ``` --- ### [11:56](https://www.youtube.com/watch?v=oZiXWnHSjko&t=716s) 组织网络效应：跨部门工作流配对与赋能生态 在推动全员 AI 赋能（AI Enablement）的过程中，核心在于构建一种"组织网络效应"（Organizational Network Effect）。当前的技术环境对非技术人员而言堪称天堂：借助强大的生成式 AI，非技术员工能够独立构建出功能丰富、外观专业的完整应用，这给他们带来了前所未有的自信心。然而，这种繁荣背后潜藏着显著隐患——由于缺乏系统工程与底层架构训练，他们往往完全忽略下游运维风险。Dasdan 幽默地分享了一个常见场景：当他询问一位刚写出应用的非技术员工"你做监控了吗？"时，对方却一脸茫然地反问："什么是监控？"（"What's monitoring?"）。 为了化解这种脱节，Dropbox 大力推行"工作流配对"（Workflow Pairing），在内部通常被称为前向部署工程师（Forward Deployment Engineer）机制。通过技术专家与业务人员结对，非技术人员能迅速理解提示词工程不仅关乎自然语言编写，更关乎对结果的即时评估与后续维护流程。与此同时，企业还需要设立跨部门的"价值转译者"（Value Translators / AI Champions）——一个技术专家向非技术团队宣讲，其效果远远不如来自 HR 部门自身的 AI 拥趸，因为后者才真正掌握该业务领域的行业术语与实际痛点。 此外，Dropbox 结合专项激励机制深化这种网络效应： - **AI 专项黑客马拉松**：在一年一度的传统 Hack Week 之外，专门增设针对 AI 的专项 Hack 活动，打破岗位壁垒，彻底解放全员的创意灵感。 - **系统性表彰竞赛**：评选最具创意的落地成果。通过展示"一个毫无编程背景的普通员工做出了令人惊叹的应用"，在全公司范围内引发示范效应，激励更多人投身 AI 创新。 --- ### [14:12](https://www.youtube.com/watch?v=oZiXWnHSjko&t=852s) 研发瓶颈转移：从代码生成到持续集成与安全防御 谈及产品研发领域的变革，Dasdan 直言：代码生成（Code Generation）本身在业界几乎已被视作一个"已解决的问题"，但系统交付（System Delivery）正成为全新的关键瓶颈。在 Dropbox 举办的生产力 AI 峰会上，Uber、Netflix 等一线科技巨头齐聚一堂，大家面对的现实高度一致：各家公司的整体研发效率近乎翻倍，代码库中 70% 至 80% 的代码都由 AI 生成，但这股巨大的代码洪流给下游的持续集成（CI）系统带来了极度严峻的压力，甚至连 GitHub 等主流托管平台的吞吐能力都开始捉襟见肘，促使 Cursor 等工具厂商着手自研更高效的基础设施。 在代码审查层面，尽管自动化程度不断攀升，但 Dropbox 依然坚持在最终环节保留人工审核（Human-in-the-loop）。此外，为了应对 AI 生成代码可能带来的漏洞隐患，红队与安全团队自研了一款名为"漏洞猎手"（Vulnerability Hunter）的内部工具。该系统能自动化扫描内部代码漏洞，其检测能力堪比前沿的 Mythos 模型，却完全无需支付高昂的外部 Token 费用；更重要的是，它具备故障注入（Fault Injection）与交叉验证能力，能够模拟外部恶意注入并自我检验整个防御流程的有效性。 --- ### [15:54](https://www.youtube.com/watch?v=oZiXWnHSjko&t=954s) 告别传统 PRD：基于 Agentic AI 重构客户痛点直连链路 在超越代码自动补全之后，研发必须回归客户问题的本质。然而，开发团队在高速迭代中必须清醒认识到"客户吸收成本"（Customer Absorption Cost）的存在：即使团队具备每天发布多次新功能的能力，如果界面交互和工作流发生剧烈变动，人类用户的适应节奏也无法跟上这种频繁变更——除非未来的交互模式彻底演变为"Agent 对 Agent"的静默对接。 ```mermaid flowchart TD accTitle: 传统研发与 Agentic AI 驱动研发模式对比 subgraph Traditional["传统软件研发流水线"] T1["PM 访谈个别客户 / 搜集 UX 研究"] --> T2["撰写冗长需求文档 PRD"] T2 --> T3["工程师阅读理解需求并写代码"] end subgraph Agentic["Agentic AI 实时研发模式"] A1["海量工单 / 通话录音 / 客服数据"] --> A2["GenAI 结构化非结构化数据"] A2 --> A3["实时全景痛点看板"] A3 --> A4["内部 Agentic 框架 + 工程师"] A4 --> A5["高契合度精准交付"] end ``` 更根本的颠覆在于产品经理的角色与需求传递链条。过去数十年间，产品经理调研 10 位客户、撰写需求文档（PRD）再转交工程师的瀑布式闭环虽行之有效，但在智能体 AI（Agentic AI）时代已显滞后。大语言模型最擅长从非结构化数据中提炼结构化洞察。Dropbox 与行业前沿正在实现的是：利用 AI 实时解析所有客服工单、客户通话录音等全量交互数据，将整个客户群体的真实焦虑以全景视图书卷般实时投喂给工程师及内部 Agentic AI 研发框架。让代码与智能体直接对接活生生的第一手痛点，而非经过层层转述的二手理解，彻底消除了需求失真。 关于产业影响与底层技术选型，Dasdan 指出： - **自适应界面（Adaptive UI）**：随着底层平台的完善，AI 赋予了按需实时生成个性化交互界面的能力。 - **构建私有护城河**：模型推理成本依然高昂，企业独有的上下文与业务专有数据将是打造竞争壁垒的关键。 - **多元模型策略**：没有任何单一模型可以通吃所有场景，合理搭配开源权重模型（Open-weight models）与商用闭源模型、权衡自建与外购网关平台，是企业架构师的长期必修课。 --- ### [19:51](https://www.youtube.com/watch?v=oZiXWnHSjko&t=1191s) 现场问答：从日常杂务迈向 Dash 语义索引与全流程自治 在演说的尾声，现场观众提出了一个核心问题：企业究竟该如何跨越"仅用 AI 草拟邮件或处理琐碎杂务"的初级阶段，真正释放深层生产力？ Dasdan 以 Dropbox 的实践进行了有力回应： - **产品端（Dropbox Dash 的演进）**：面对拥有数百万文件的大型企业用户，以及平台上数百亿量级的视频与图片资产，传统的关键词搜索完全失效。Dropbox 引入深度语义索引（Semantic Indexing），使用户能够凭借模糊意图和自然语言直接跨系统秒级找回所需资产。 - **研发运维端（自研全流程智能体）**：Dropbox 将 AI 纵深打入全研发生命周期。除了前述的代码补全与自动化代码审查之外，公司在运维与安全领域构建了专有的 SRE Agent 与 Security Agent，让智能体跨越简单的文本生成，深入核心业务逻辑与全天候系统运维，实现了全技术栈的智能化升级。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=oZiXWnHSjko&t=4s) AI 只有在推进实际成果（Outcome）而非仅仅完成某项具体任务（Task）时，才能真正产生杠杆效应（leverage） — Ali Dasdan > [00:04](https://www.youtube.com/watch?v=oZiXWnHSjko&t=4s) 速度是任何公司唯一的优势；它或许不是决定你能否存活的唯一决策指标，但如果速度不够快，那就必然会有大麻烦。 — Ali Dasdan > [11:05](https://www.youtube.com/watch?v=oZiXWnHSjko&t=665s) 你无法把信任事后螺栓式硬拧进系统 — Ali Dasdan > [11:56](https://www.youtube.com/watch?v=oZiXWnHSjko&t=716s) 什么是监控？ — 某非技术员工 > [14:12](https://www.youtube.com/watch?v=oZiXWnHSjko&t=852s) 代码生成（Code Generation）本身在业界几乎已被视作一个“已解决的问题”，但系统交付（System Delivery）正成为全新的关键瓶颈。 — Ali Dasdan ## Mentioned Dropbox, Ali Dasdan, Nova, DX, Atlassian, DORA, Dropbox Dash, Slack, Salesforce, Uber, Netflix, GitHub, Cursor, Vulnerability Hunter, Mythos, SRE Agent, Security Agent ## Video description Ali Dasdan - CTO, Dropbox Artificial intelligence is fundamentally reshaping how knowledge workers collaborate, create, and build software. Productivity applications are evolving into AI-native workspaces where intelligent assistants can summarize meetings, generate content, automate workflows, retrieve organizational knowledge, and help developers write, review, debug, and understand code. Rather than simply responding to commands, modern AI systems are becoming proactive collaborators that anticipate user needs, personalize experiences, and orchestrate work across applications. This transformation extends well beyond chatbots. AI is changing how organizations capture institutional knowledge, coordinate projects, manage information, and develop software. As these capabilities become embedded across productivity suites, communication platforms, developer tools, and enterprise applications, organizations face new challenges around context management, accuracy, security, governance, privacy, and trust. This session explores the technologies, architectures, and product strategies powering the next generation of AI-driven productivity, collaboration, and product development. We’ll examine how retrieval, enterprise knowledge, personalization, coding and operations copilots, workflow automation, and organizational memory combine to deliver intelligent experiences that are accurate, secure, and deeply integrated into everyday work. We’ll also discuss how vendors are rethinking user experience, pricing, and business models as AI becomes a core capability rather than an add-on. Attendees will leave with practical insights into designing AI experiences that improve productivity, accelerate product development, enhance collaboration, and responsibly deploy AI systems that users can trust. [Source](https://www.youtube.com/watch?v=oZiXWnHSjko)

Source quotes

[0:04] Ali Dasdan: AI 只有在推进实际成果（Outcome）而非仅仅完成某项具体任务（Task）时，才能真正产生杠杆效应（leverage） [0:04] Ali Dasdan: 速度是任何公司唯一的优势；它或许不是决定你能否存活的唯一决策指标，但如果速度不够快，那就必然会有大麻烦。 [11:05] Ali Dasdan: 你无法把信任事后螺栓式硬拧进系统 [11:56] 某非技术员工: 什么是监控？ [14:12] Ali Dasdan: 代码生成（Code Generation）本身在业界几乎已被视作一个“已解决的问题”，但系统交付（System Delivery）正成为全新的关键瓶颈。

Source quotes

[0:04] Ali Dasdan: speed is the only advantage of any company [0:04] Ali Dasdan: AI creates true organizational leverage only when it advances a concrete business outcome rather than merely speeding up an isolated task. [2:41] Ali Dasdan: In March 2024, software engineering adoption of AI tools at Dropbox hovered around 40%. By the end of the year, engineering reached 100% adoption, and AI usage subsequently reached 100% across the entire company. [6:59] Ali Dasdan: Engineering telemetry and productivity dashboards must never be weaponized as surveillance systems. Tracking individual output metrics like raw PR volume creates perverse incentives that encourage engineers to game the system. [6:59] Ali Dasdan: A single software engineer leveraged AI tooling to fully automate the migration of 1,900 Python packages within a two-week sprint. [9:51] Ali Dasdan: Because generative models blend contextual data during synthesis, permission boundaries must be structurally embedded into the core architecture from day one to prevent unauthorized information leakage across user roles. [11:56] Ali Dasdan: are you monitoring this? [14:12] Ali Dasdan: Top tech organizations have roughly doubled engineering productivity, with 70% to 80% of all committed code being AI-generated. However, this sheer volume of code puts unprecedented strain on continuous integration (CI) environments and code review infrastructure [18:09] Ali Dasdan: Directly connecting real customer pain points to engineering teams and autonomous systems fundamentally transforms how software gets built. [18:09] Ali Dasdan: To succeed, companies must recognize that their proprietary data and unique contextual knowledge represent their true competitive moat. [19:44] Ali Dasdan: Rather than settling for superficial chatbots, AI is deeply integrated into operational systems and large-scale data retrieval to solve concrete platform-scale challenges.

Original transcript

[0:04] Hello everybody. Hopefully you're not tired of  hearing about AI. I'm going to talk about AI [0:11] more. Um, we're going to have three sections. Um,  actually, sorry, let's start with the thesis. Um, [0:20] hopefully it's a no-brainer. Uh, AI creates  leverage when it advances an outcome, [0:25] not just a task. um definitely without an output  outcome doesn't come. So there's a progression [0:32] from there but it's important to keep this in mind  again even though it's a simple lesson I'll try to [0:38] emphasize it multiple times along the way. Um  so I'm going to talk about three things. Um I [0:45] speak fast and I have an accent so um if I if you  guys don't get it so I'll try to leave some time [0:50] at the end to ask some questions. So first things  is about productivity then collaboration and and [0:57] product development they're kind of infused with  each other embedded with each other there are [1:01] connection across across all three but I'll try  to um provide some details along the way and give [1:08] examples from Dropbox uh uh so for productivity  um you know start with the input layer uh [1:18] whatever goes into that you know could be prompts  training access data data learnings and all that. [1:24] Next one is definitely output layer. Whatever  comes out of your process, right? Could be drafts, [1:29] decisions, code, tickets, summaries, etc. Then  hopefully some of these will become outcomes, [1:33] right? Not everything become as an outcome.  Uh you might be the amazingly fast, you know, [1:38] um software developer, but if it's not the  right product, doesn't matter basically. Um [1:43] so outcome is in ultimate is definitely whatever  matters for the business and customer. You know, [1:49] quality has to be let's say as a table  steak and security things like that. [1:52] than the margin you know time to market so you can  actually try more things customer satisfaction the [1:57] way that we are measuring it um and this is  the purpose of productivity basically to to [2:01] hopefully generate more outcome uh for you um it  it enables you to move faster I mean so fast so [2:09] much so that I have a motto with my team that you  know speed is the only advantage of any company [2:14] uh it's not the uh final decision maker for  your survival but if you're not fast enough [2:21] uh we have a problem. Um but beyond fast in  the end you want it's okay beyond you want [2:28] uh better results better decisions right  so whatever you are trying to improve [2:32] um hopefully these decisions will lead to  even better results as a result of that um [2:41] um so with respect to AI it starts with  adoption so last year around March March 2025 [2:49] uh with respect to software engineers using  AI we were around 40%. Uh so by the end of the [2:54] year we we reached 100%. Um now it is across the  company 100% usage. Uh if you look at the software [3:03] engineers what they use almost 70 80% are using  multiple at the same time. We have built our own [3:08] infrastructure uh called Nova you know to to to do  agentic development. But coming back to adoption [3:15] uh first it starts with the signal. There has to  be support from the leadership. Um uh it could [3:21] be budget, could be the willingness, could be uh  setting it as a goal. Um uh and you have to be you [3:27] have to mean it. Uh you have to uh so that people  can look up to you and adop adopt it as the step. [3:34] Um next one is uh you know go adapted do this  but at the same time if you're not giving enough [3:42] um time to people to try or invest in the  in these tools and learning then we have a [3:48] problem. So second one is that definitely give  that time for exploration. Third one is enable [3:53] people as much as possible. I mean uh most of  the adoption comes because people learn from [3:57] each other. They look at each other influence  each other you know sometimes peer pressure. [4:02] uh but seeing these examples we have this thing  called AI champions program for example that you [4:07] can see the best use use use cases and as a result  you try to copy that um so that is the third phase [4:13] and the last one is definitely you have to keep  reinforcing it so the first success is not enough [4:18] you just have to go uh further than that and u  and then repeat this and go beyond now we are [4:23] at a stage that adoption is not the problem then  you go now how can we get the most use out of this [4:31] um so with respect to productivity specifically uh  we we started measuring it's called DX core metric [4:37] I don't know if you guys know DX the company  that Atlas bought so uh originally Dora metrics [4:42] are mostly um um uh you know not as general as DX  core 4x core 4 basically try to for me is speed [4:53] effectiveness quality and impact tries to provide  a balance across all of these so speed metrics [4:58] are PR cycle time PR throughput and all that so  these are the productivity metrics applicable [5:01] the software developers. Um then effectiveness  you measure through developer experience. These [5:07] are survey based qualitative results, right? Are  they satisfied with their tools and the way that [5:12] they are working with AI? On the quality side, uh  across the board, we have everything as an service [5:17] level objective, security, privacy, stability, you  know, performance, product quality, data quality [5:22] and all that. On top of it, um you have to measure  um impact, right? What's the revenue per engineer? [5:28] You know, the cost efficiency and things like  that. Now with AI we are extending these more [5:33] and more with AI specific metrics. So if it's not  only PR throughput but PR throughput with respect [5:39] to the PRs that are mostly AI generated you know  what's an AIPR I don't know let's say 70% of the [5:44] code is AI generated and let's just also look into  that one if it is code coverage then you say code [5:50] coverage that is of the code that is gen generated  by AI unless you look at that detail sometimes you [5:56] might miss because amount of code is so large  right the effect uh you might miss the real [6:01] metrics or improvements or sometimes you know um  degradation Okay. And one nice thing about this [6:07] one is I think you know it's a compounding right  the more you improve the more it leads to more [6:12] developments. Um with respect to mapping you know  the AI value. There are some differences between [6:19] engineering product design and uh the rest of  the company. You know over here I'm just using [6:22] business operations as an as an example. At topox  we say engineering product and design. We refer [6:27] we use the acronym EPD. Um definitely engineering  was the first starting point and that led to you [6:35] know is the flow at the right rate you know the  quality is there and you know it is not fully sold [6:41] but it has improved drastically then product and  design again we started ne last year um with that [6:47] enabling them uh as much as possible and then  the rest of the company so many of these moved [6:52] parallel but the speed uh changed as well as what  AI could do given the capabilities of the products [6:59] Okay. uh with respect to showing the results  right dashboards um if if it if it is seen [7:07] as a surveillance system whose PR is larger or  whatever the metric that you care then it creates [7:13] not the right incentives right I mean any any  measurement can be gamed that's why you have to [7:18] have multiple of them it has to be an across the  uh board balance but you know like I'm saying I'm [7:24] trying to say over here is a conversation starter  it should be there you're using it as an input to [7:28] do something better rather than again monitoring  um um exactly who's who's better uh versus worse [7:35] and you know it could be metrics across like  you [clears throat] know road map health let's [7:39] say you know how much of our commitments per  quarter we have delivered what's the change rate [7:45] right so some initiatives will be dropped or dep  prioritized or whatever if it is less let's say [7:50] more than 15% you know there could be a problem  like that kind of a churn is is is a problem so [7:55] it's a good idea to look at from different  metrics just to uh create an opinion again [7:59] this is basically inputs to Hey, are we on the  right path or not? Right? And you know, metrics [8:04] across the other areas too. Um, so this is real  statistics from Dropbox and you know on the x-axis [8:10] you see the time uh and y-axxis we don't tell you  the exact number but um so this is the work that [8:17] developers have been doing beyond their road map  commitments. So we already have customer trust as [8:23] the number one priority of the company. customers  trust, you know, within that we have security, [8:27] privacy, you know, compliance, uh, stability, um,  performance and product and data quality. So, they [8:34] are already there, commitments are already there.  But as productivity is increasing, uh, people are [8:39] finding extra time and they spend that time in  things that they feel valuable. In this case, for [8:45] example, they actually went ahead and, you know,  fixed additional security tickets even though they [8:50] were not critical security tickets. But then for  for them, they wanted to pay pay attention to that [8:56] um tech that reduction migrations, you know, one  single person person automated the entire Python [9:02] um um package migration like in two weeks, 1900  of them were fully automated and using AI. Um [9:11] and you see other capabilities too and it's it  the fact that you have this upward trajectory [9:16] even though our resources are not increasing  uh that also shows that productivity gains [9:21] are coming right through that because you see you  see the more and more and more usage and increase [9:26] um uh as you go um and this also gives the  incentive as well as agency to everybody [9:33] actually to focus on what they sometimes feel  um is important. Our road map process is bottom [9:39] up definitely what we try to prioritize what  they feel like is important you know aligning [9:43] with the company strategy and all that but on  top of it sometimes my burning point might not [9:47] be the top priority for the company but if I am  gaining extra productivity I can actually use it [9:51] for that purpose um then move to collaboration  for collaboration what matters is the context [9:57] layer um so just I gave you a tool and just  somehow you improve the collabor collaboration [10:04] you know in the end it's a human activity you  know we have lots of artifacts through chats [10:09] and documents and meetings etc. Um, and you want  to make sure that all this information comes back [10:15] uh through a context layer and that when you do  AI, it's a contextual AI. When you try to solve [10:20] a problem, it's it's with context. Actually,  results are far better as a result of that. Um, [10:25] this is a high level architecture of Dropbox  Dash, uh, which is our contextual AI platform. Um, [10:31] so you see at the top you have the agent platform.  Then in the middle is context platform. you know [10:36] if you if you willing to connect your data let's  say from Salesforce and Slack or what whatever [10:42] thirdparty systems that you have uh we can  actually enable it right it's fully uh honoring [10:47] the permissions that you have it's in real time  all that information will flow to the system and [10:52] then the result will be um super relevant what  you're looking for literally you know I sometimes [10:57] go to a board meeting and there's a question comes  and I do not know the answer and I cannot reach [11:01] somebody from my team at that moment but I do a  search over there and just gives me the answer [11:05] uh it's an AI search but AI search with the  enhance with the context it's not just model [11:10] is answering out of out of nowhere okay um and  uh on the left side you are seeing basically [11:16] the elements of that connection connection  connecting understanding orchestrating and control [11:23] um and trust has to be built into the system  architecture okay again it's a very old saying [11:29] everybody knows that you cannot bolt it in bolt  it in um for example if I'm pulling data from [11:35] third party. I cannot just say let me just index  it and then s on top of it I'm just going to add [11:41] um you know permissions let's say you got  to be careful with that because AI might [11:45] be blending it in such a way that you cannot  differentiate sometimes where is the boundary [11:49] for permissions so that's why it has to  be part of the system from the beginning [11:56] um now with respect to enablement AI enablement  uh you have to create this organizational network [12:01] effect uh I mentioned this in the adop adoption  too uh for example workflow pairing right so I [12:08] mean right now it's it's a heaven for nontechnical  people because first of all they are using AI [12:13] and creating these applications um everybody's  so happy because the power of with power with [12:19] with the power of AI the complexity that they can  create how it looks like a real application gives [12:24] unbelievable confidence but at the same time  if they do not have the technical background [12:29] then you they end up might sometimes ignoring  the downstream effects of that. So if somebody [12:37] develops an application sometimes I ask like  are you monitoring this? They go like what's [12:40] monitoring? Um so um so but then more and more  that language changes right because the more they [12:47] do the more the more they interact creating these  the more uh they actually work with the technical [12:53] people then they realize okay there are other  steps that they have to pay attention to right [12:57] um and if you know how to ask to the AI it's not  just like you know English you know you have to be [13:01] able to like evaluate the results really fast and  that that has to come with that and that pairing [13:05] helps a lot uh sometimes they call it like forward  develop deployment engineer internally within the [13:12] company. um value translation I mentioned like  the champions right from each team you know a [13:17] tech champion talking to non techch all the  time is does not have the same effect as a AI [13:22] champion that's let's say from the HR team right  because they know the language and terminology [13:27] of of that team and that organization um practical  experiments right hack events for example you know [13:34] um um we we have hackw week every year um  but on top of we created specific hack events [13:42] um just for AI purposes uh just to make sure  that we can actually open open things up and [13:47] people just let let everything ideas flow so they  can actually understand what's coming um and the [13:53] system incentives like rewarding awards you know  some kind of competition sometimes what's the most [13:58] creative thing that people have created uh just  the showcasing itself just opens up completely [14:03] people's minds and saying that oh you know I  didn't know that this person who has no technical [14:08] background can do that I can also do it right just  change the organization That's why it's network [14:12] effect basically. Now the last point is product  development. Um now with AI coding definitely [14:20] code generation is you know people sometimes  say it's solved right. Uh but system delivery [14:24] is still still the bottleneck. Um the more I mean  we had an AI summit on productivity. We invited [14:31] uh Uber and Netflix and all these companies. You  can see everybody's facing the same problems. [14:36] Metrics are similar. Everybody kind of doubled  the productivity all these top tech companies. [14:41] uh but then you can see that the amount of code is  so large they're around 70 80% just like us with [14:46] respect to AI generated code but then this creates  extra stress on their CI systems even GitHub is [14:53] having difficulty if if you know u that's why  curs and some other companies started creating [14:58] their own uh systems so those bottlenecks have  to be resolved um you still need humans I mean we [15:04] still use humans for code reviews even though we  are automating it as much as possible with AI but [15:09] the final step is through a human. I don't know  one day this might be completely automated but we [15:14] still we still keep it over there and we also had  to add extra uh checks for example um uh somebody [15:22] um you know a developer from the red team and from  the security created this vulnerability hunter [15:27] uh it actually finds vulnerabilities automatically  almost like mythos running internally you know [15:32] without paying the token price um u and also  it can do fault injection actually try to [15:37] find exactly what was inserted in injected from  another system so it can also do a verification [15:42] of its own process. So it's it works really  well. Um and all of these have to be so that [15:48] balance has to change and hopefully uh the entire  thing will be automated uh soon across the board. [15:54] Uh moving beyond simple code completion, right,  you still have to focus on the customer problem. [16:01] Um and at the end uh definitely you know you have  to define if I know the customer problem correctly [16:08] um if I know what good looks like uh if I have a  good process on building building and validation [16:15] um I am learning and iterating with the customers  but you should always remember that there is still [16:20] an absorption cost right you know just because you  changed and you are moving let's say we I push a [16:25] feature every day it is little possible today you  know multiple times a day uh but then customers [16:30] might not be able adapt to it especially if you  are drastically changing the interface and the way [16:34] that workflows are working. So there's still that  human speed unless agent to agent are talking. As [16:40] long as that speed is over there, nothing's bad  about it. It's just that that's how the humans [16:43] work. Uh as a result, basically this this whole  thing has to adapt to that. Um and you have to [16:50] close that distance and this is uh let me see this  is one uh one example. Um so if if if you follow [17:00] the current uh current way of doing work you know  uh software development where product managers [17:07] you know they think about uh customer problem they  talk to some customers maybe they use UX research [17:12] um and then they try to you know express this in  a requirements document talk to some engineers and [17:17] they get an understanding of the problem. This has  worked so so well so far. But now with agentic AI, [17:23] this this has to change, right? The problem  is agents can do a really good job if you can [17:28] define for them the problem the right way. Which  means developers have to have or agents have to [17:33] have access to the real problem definitions, real  pain points of the customer as well as how these [17:39] things should be validated, right? Uh that has to  be designed really well. So which means basically [17:43] this blending of the roles in a way is happening.  Um and today again using AI, right? One of the [17:49] best things about Genai is it can create structure  out of unstructured data. So all this customer [17:54] information in the customer support tickets um  customer calls recordings or whatever now that all [18:00] all that information there are startups doing that  definitely we we have done it ourselves to pulling [18:04] that information in a form that in real time you  are literally seeing the entire field of what [18:09] customers are worried about and if you tie that to  first of all provide that information to engineers [18:15] as well as through them to agents especially  if you have an internal agentic AI development [18:19] uh framework results are far because what you know  you are solving real problems not the problems [18:27] that you thought you understood from product  manager who talked to 10 10 people only to get [18:31] that one. So this this has to change and this  is changing basically and all these steps are [18:35] um um getting far more comprehensive and far  more powerful. Um with respect to the implication [18:44] of AI also sometimes this is used you know to  reduce the stock price of certain companies but [18:50] uh the inter interface become adaptive right if  you have a good platform then uh using AI I can [18:56] create my own interface on top of it right uh  economics is changing definitely now you have [19:00] to pay attention to because these things are not  cheap uh how you're going to blend it if you have [19:05] proper context proper data that becomes your mode  then that's make that's going to make your AI far [19:10] more powerful Um and also now you have to think  about how I'm going to you know what's the right [19:15] models right a single model cannot solve every  problem um again open weight models how I'm going [19:20] to manage this should I create that interface  myself should I buy it from somewhere else there [19:25] are again lots of start startup companies trying  to do that um so hopefully this was useful I went [19:30] super fast on very very complicated topics so  maybe we have 26 seconds for a question thank [19:43] [applause] [19:44] We can we can take one question if  somebody would like to right here. Great. [19:51] How how did your company move to make um  the the AI to be a little more productive [19:56] and move past just like drafting emails  or kind of more of the remedial tasks. [20:01] Um I mean with that option basically the we  apply AI to all the tasks that we are doing [20:08] right you know coding of course we are people are  also using it to create documents summarization [20:13] all that those use cases are already there  but as part of for example Dropbox dash right [20:18] uh we have customers with sometimes million  million files and they have huge number of we [20:22] have hundreds of billions of videos and images on  our platform people are trying to find something [20:26] right and we are using now semantic indexing for  example to find exactly what you're looking for [20:31] So all these are coming because of the customer  problems that we are facing as well as internally [20:35] the way that we are developing product from  coding to code reviews to operations to you [20:39] know we have created our own SR agent uh security  agent as I mentioned so across the board. Yeah.