Dataiku Urges CEOs to Manage AI Agents as Digital Workforce

Stop Governing Your AI. Start Hiring It with Dataiku — Ai4

Ai4

Treating AI agents as a managed digital workforce with specific budget owners prevents shadow deployment risks while enabling rapid business innovation.

AI-assisted summary of Stop Governing Your AI. Start Hiring It with Dataiku — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 23, 2026, this video features a presentation by Dataiku CMO Mark Abramowitz and SVP of AI & Platform Jed Dougherty. According to the page description, the session challenges traditional AI governance models by advocating that organizations stop merely governing AI and instead manage autonomous agents as a digital workforce tied directly to business unit budgets, drawing on insights from 900 CEOs.

## TL;DR The presentation argues that shadow AI usage is a legitimate user demand signal reminiscent of the 1980s enterprise spreadsheet boom, but with vastly higher operational risks as autonomous agents execute real-world business actions. Abramowitz and Dougherty assert that traditional centralized IT governance cannot keep pace with rapid AI adoption and propose treating agents as labor owned, funded, and evaluated directly by business units. Meanwhile, central IT shifts to providing safety, identity, and observability infrastructure, preventing chaotic spreadsheet-like sprawl through unified control planes such as Dataiku's offerings.

## Atmosphere & style The presentation carries a pragmatic, authoritative, and slightly provocative tone, opening with uncomfortable realities about enterprise risk that immediately hook executive viewers. The speakers deliver a tight, highly structured narrative that blends historical tech analogies with stark survey statistics, transitioning from alarming warnings about ungoverned agents to a clear, solution-oriented operational framework. Watching the session feels like attending a focused corporate keynote that challenges conventional IT management logic while maintaining an engaging, analytical pace.

## Key takeaways - [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) A Dataiku survey of 900 enterprise CEOs revealed that 96% believe employees use generative AI without explicit corporate approval, and 42% say over half their workforce does so. - [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) Shadow AI usage represents a legitimate business demand for speed and autonomy that central IT cannot suppress, echoing the historical enterprise adoption of electronic spreadsheets 40 years ago. - [04:29](https://www.youtube.com/watch?v=yntVXCUf6kw&t=269s) Unlike static spreadsheet errors, ungoverned AI agents introduce dynamic, real-time risks because they directly execute actions like emailing customers, moving enterprise funds, and accessing sensitive data. - [04:29](https://www.youtube.com/watch?v=yntVXCUf6kw&t=269s) Dataiku introduced its 'Co-Build' product to provide enterprise-grade controls, visibility, and governance over natural language and 'vibe-coded' applications built by non-technical business users. - [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) Traditional central IT gatekeeping built on technology scarcity is obsolete; organizations should manage AI agents as a digital workforce owned, funded, and evaluated directly by business units. - [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) In a digital workforce model, central IT shifts to a role similar to HR or Facilities, providing platform infrastructure, identity management, access control, and safety guardrails. - [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) Survey data shows intense executive pressure, with 80% of CIOs feeling their jobs are directly at risk if they fail to deliver measurable AI ROI within the year. - [11:51](https://www.youtube.com/watch?v=yntVXCUf6kw&t=711s) An audit of a global manufacturing firm uncovered over 400 active AI agents across multi-cloud environments—10 times the CIO's original estimate of 40—highlighting severe visibility gaps. - [11:51](https://www.youtube.com/watch?v=yntVXCUf6kw&t=711s) Dataiku developed an Agent Management control plane to track ownership, system permissions, execution costs, and operational performance across disparate multi-cloud ecosystems. - [11:51](https://www.youtube.com/watch?v=yntVXCUf6kw&t=711s) Enterprises face a critical choice between repeating the chaos of ungoverned spreadsheets at machine speed or establishing a governed, accountable agentic workforce.

## Detailed walkthrough Here is the detailed, faithful walkthrough of the transcript:

[00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) Shadow AI Usage and the Spreadsheet Precedent

Mark Abramowitz opens the presentation by pointing out an uncomfortable reality: across enterprises today, employees who would never be granted access to production systems are actively building AI solutions. These individuals are not rogue actors attempting to disrupt operations; rather, as the people closest to actual business problems, they are responding rationally to accessible tools because the cost of waiting for official approval is simply too high. A survey conducted by Dataiku across 900 CEOs of large companies (each generating $500M+ in revenue) across eight countries reveals the sheer scale of shadow AI: * **96% of CEOs** believe their employees are using Generative AI without explicit corporate approval. * **42% of CEOs** confirm that at least half of their entire workforce is doing so.

Rather than viewing this as a pure compliance failure to be clamped down on, Abramowitz argues it represents a strong demand signal: workers cannot wait for formal permission. The core thesis of the talk is that while AI demand is legitimate, the necessary enterprise infrastructure is missing. Drawing a historical parallel, Jed Dougherty takes the audience back 40 years to the dawn of personal computing. The most successful software deployment in enterprise history—the electronic spreadsheet—was never built, approved, or controlled by central IT. Despite IT's best efforts to eliminate it, Excel became the backbone of financial M&A modeling, drug trial logistics, and payroll execution because it placed power in the hands of business experts rather than system administrators. IT spent four decades trying to lock down spreadsheets, failing to realize that the legitimate demand for user-driven tools was never going to disappear.

``` LEGITIMATE BUSINESS DEMAND (Users want speed & autonomy) │ ┌───────────────────────┴───────────────────────┐ ▼ ▼ Spreadsheet Era (1980s+) AI Agent Era (Present) ┌───────────────────────────┐ ┌───────────────────────────┐ │ • Localized formula errors│ │ • Dynamic, execution-based│ │ • Static spreadsheet cells│ │ • Interacts with customers│ │ • Auditable upon inspection│ │ • Moves enterprise funds │ └───────────────────────────┘ └───────────────────────────┘ │ │ └───────────────────────┬───────────────────────┘ ▼ MISSING FOUNDATION & GOVERNANCE INFRASTRUCTURE ```

[04:29](https://www.youtube.com/watch?v=yntVXCUf6kw&t=269s) From Spreadsheet Errors to Unmanned Agents: The Risk Evolution

While spreadsheet risks were largely localized—a miscalculated formula sitting inside a single cell that a human might eventually catch—AI software and vibe-coded dashboards introduce entirely new failure modes. Dougherty illustrates this with a visual example of a sleek, vibe-coded sales performance dashboard. While its charts look convincing, one hidden line of code pulled last quarter's actual revenue instead of current quarter targets. Because incorrect data design presents visually clean results, identifying such flaws requires parsing thousands of lines of auto-generated logic.

The risk escalates exponentially when moving from vibe-coded code to autonomous AI agents. An ungoverned agent doesn't merely display incorrect numbers on a screen; it directly executes actions by emailing customers, moving enterprise funds, and training on proprietary corporate data. To harvest the benefits of vibe-coding—rapid development in natural language by domain experts—enterprises require continuous visibility and governance without needing massive IT overhead. To address this gap, Dataiku launched its "Co-Build" product in June, designed to provide enterprise controls over user-built AI applications.

- **Vibe-Coded Dashboard Example**: Demonstrates how a clean visual chart can mask severe underlying errors, such as computing revenue targets against previous quarterly actuals. - **Agent Governance Requirements**: Organizations must know when agents break, maintain full inspection capabilities, and handle remediation without large, dedicated engineering teams.

[07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) Treating AI Agents as a Digital Workforce

Addressing the central operating model of enterprise technology, Abramowitz highlights that traditional centralized IT governance was built for an era of technology scarcity, when hardware and software were rare, expensive, and fragile. In that era, annual procurement cycles were acceptable. Today, the speed of AI adoption makes waiting an existential threat to individual careers and business units alike.

Rather than treating AI management as a technological procurement challenge, organizations must address it as an organizational design problem. Executive keynotes frequently refer to AI agents as a "digital workforce," but few apply the literal operational logic of labor management:

``` ┌─────────────────────────────────────────────────────────────────┐ │ BUSINESS UNIT RESPONSIBILITY │ ├─────────────────────────────────────────────────────────────────┤ │ • Hires & deploys agents based on specific ROI goals │ │ • Funds agents directly out of the unit P&L budget │ │ • Conducts performance reviews, evaluations, and firings │ └────────────────────────────────────────┬────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────────────┐ │ IT AS THE FOUNDATION │ ├─────────────────────────────────────────────────────────────────┤ │ • Provides Identity, Access Control, and Safety Frameworks │ │ • Acts like HR / Facilities: Enables safe, fast deployment │ │ • Serves as the enablement platform rather than a gatekeeper │ └─────────────────────────────────────────────────────────────────┘ ```

When agents are assigned a specific budget owner with measurable revenue or efficiency targets, governance becomes self-enforcing. Business leaders will not maintain or pay for duplicate, ineffective, or underperforming AI agents. In this framework, central IT shifts its role from being a permission gatekeeper to functioning like HR, Facilities, or Finance: providing agent identity, access controls, system-of-record integration, and safety rules to make hiring agents fast, safe, and compliant.

- **Survey Insights on Executive Risk**: - **80% of CIOs** report that their own jobs are directly at risk if they fail to deliver measurable AI ROI this year. - **77% of CIOs** believe a peer at another company will be fired due to a failed AI strategy or operational crisis.

[11:51](https://www.youtube.com/watch?v=yntVXCUf6kw&t=711s) Agent Visibility, Observability, and Future Trajectories

To support this operating model, IT requires clear visibility into all running assets. Dougherty shares an anecdote from a conversation with a CIO at a major global manufacturing company. When asked how many AI agents were active across their enterprise, the CIO estimated roughly 40 in production. However, after Dataiku ran an audit across their diverse environment (spanning Databricks, Snowflake, AWS, Azure, and other platforms), they discovered over 400 active agents running across various business units. Most of these agents had no clear ownership, performance metrics, cost tracking, or active supervision.

Dougherty compares this 10x discrepancy to an organization losing track of 90% of its human headcount. To solve this, Dataiku built an Agent Management control plane to track agent ownership, system permissions, execution costs, and real-time performance across multi-cloud environments.

Concluding the session, Abramowitz contrasts two potential futures facing modern enterprises: 1. **The Unchecked Spreadsheet Repeat**: Companies recreate the chaotic spreadsheet era at machine speed, resulting in thousands of uninspectable, action-taking, vibe-coded agents running rampant. 2. **The Governed Agentic Workforce**: Companies build a standardized platform foundation, enabling business teams to build and manage accountable AI agents while central IT provides the visibility, safety, and operational infrastructure.

- **56% of surveyed CEOs** admit that their competitors currently possess an AI strategy superior to their own, highlighting declining executive confidence at a critical inflection point.

## Notable quotes > [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) 96% of CEOs believe their employees are using Generative AI without explicit corporate approval. — Mark Abramowitz

> [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) 42% of CEOs confirm that at least half of their entire workforce is doing so. — Mark Abramowitz

> [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) The most successful software deployment in enterprise history—the electronic spreadsheet—was never built, approved, or controlled by central IT. — Jed Dougherty

> [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) IT spent four decades trying to lock down spreadsheets, failing to realize that the legitimate demand for user-driven tools was never going to disappear. — Jed Dougherty

> [04:29](https://www.youtube.com/watch?v=yntVXCUf6kw&t=269s) An ungoverned agent doesn't merely display incorrect numbers on a screen; it directly executes actions by emailing customers, moving enterprise funds, and training on proprietary corporate data. — Jed Dougherty

> [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) 80% of CIOs report that their own jobs are directly at risk if they fail to deliver measurable AI ROI this year. — Mark Abramowitz

> [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) 77% of CIOs believe a peer at another company will be fired due to a failed AI strategy or operational crisis. — Mark Abramowitz

> [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) When agents are assigned a specific budget owner with measurable revenue or efficiency targets, governance becomes self-enforcing. — Mark Abramowitz

> [11:51](https://www.youtube.com/watch?v=yntVXCUf6kw&t=711s) 56% of surveyed CEOs admit that their competitors currently possess an AI strategy superior to their own — Mark Abramowitz

## Mentioned Mark Abramowitz, Jed Dougherty, Dataiku, Dataiku Co-Build, Excel, Databricks, Snowflake, AWS, Azure

## Video description Mark Abramowitz - Chief Marketing Officer, Dataiku Jed Dougherty - SVP of AI & Platform, Dataiku

Every session here tells you how to govern your AI better. This one tells you to stop. Everyone says agents are a digital workforce. Almost nobody follows that logic to its conclusion: if agents are labor, the business hires them, pays for them, and fires the ones that underperform. The best AI governance tool ever invented is a budget line with someone's name on it. Half your company is already building AI you never approved. Dataiku CMO Mark Abramowitz and Head of AI Platform Jed Dougherty bring data from 900 CEOs, a live demo, and the one decision that separates AI winners from casualties.

[Source](https://www.youtube.com/watch?v=yntVXCUf6kw)

中文简报

Dataiku呼吁CEO将AI智能体作为数字劳动力进行管理 将AI智能体视为拥有明确预算归属的受管数字劳动力，既能防范影子部署风险，又能推动快速的业务创新。 演讲指出，影子AI的使用是合理的用户需求信号，类似于20世纪80年代企业电子表格的普及热潮，但由于自主智能体会执行真实的业务操作，其运营风险要高得多。Abramowitz和Dougherty认为，传统的集中式IT治理无法跟上AI快速落地的步伐，并建议将智能体视为直接由业务部门拥有、资助和评估的劳动力。与此同时，中央IT部门应转向提供安全、身份识别和可观测性基础设施，借助Dataiku等产品提供的统一控制平台，防止出现类似电子表格失控蔓延的混乱局面。 Dataiku 对 900 位大型企业 CEO 的调查显示，96% 的受访者认为员工在未经公司明确批准的情况下使用生成式 AI，42% 的企业中已有超半数员工存在此类行为，这反映出业务端对效率与自主性的合理需求，类似于 40 年前电子表格的普及历程。 (0:08) 与静态电子表格错误不同，缺乏管控的 AI Agent 会直接执行外发动向（如向客户发送邮件、转移企业资金），带来实时且动态的风险；为此 Dataiku 推出 Co-Build 产品以应对非技术业务人员用自然语言构建应用的管理与审查难题。 (4:29) 企业应将 AI Agent 视为“数字劳动力”，由各业务部门直接负责预算投入、雇佣及绩效考核，而中央 IT 则应从传统把关人转变为类似 HR 或行政的角色，专注于提供身份认证、权限管控与安全护栏等基础设施。 (7:10) 调查显示 80% 的 CIO 感到无法交付可衡量的 AI 投资回报率将直接危及其职位；同时多云环境审计显示，一家全球制造企业的实际运行 Agent 数量（超 400 个）达到 CIO 预估的 10 倍，凸显出严重的资产可见性缺失。 (11:51)

Reading summary

Dataiku executives Mark Abramowitz and Jed Dougherty argue that enterprise AI should be managed as an accountable digital workforce rather than restricted through traditional IT gatekeeping, which risks repeating the governance failures of the spreadsheet era. The Rise of Shadow AI and Spreadsheet Parallels (0:08) A survey conducted by Dataiku across 900 CEOs of companies generating at least $500 million in revenue across eight countries indicates that 96% believe employees use generative AI without corporate approval, and 42% confirm that at least half of their workforce does so. Abramowitz characterizes this shadow AI usage as a legitimate response from workers seeking speed when formal approval processes take too long. Dougherty compares the current landscape to the emergence of personal computing 40 years ago. Central IT spent decades attempting to lock down electronic spreadsheets like Excel, yet business units adopted them anyway for critical functions like financial modeling and payroll because they empowered domain experts. Abramowitz and Dougherty contend that enterprise demand for user-driven automation will persist despite IT friction. Escalating Operational Risks from Vibe-Coding to Autonomous Agents (4:29) Dougherty points out that whereas historical spreadsheet errors were typically localized inside static formula cells, generative AI tools introduce more complex failure modes. In one example, a visually polished, vibe-coded sales dashboard obscured a flaw where generated logic pulled previous quarterly actuals instead of current targets, masking the discrepancy behind clean charts. The operational risk increases significantly when moving from dashboards to autonomous agents. Ungoverned agents do not merely display figures; they can execute actions such as contacting customers, transferring enterprise funds, and training on internal data. Dougherty argues that organizations need continuous inspection and remediation capabilities to capture the speed of natural-language development without relying on large, dedicated engineering teams. Reframing AI Governance Through Workforce Management Models (7:10) Traditional centralized IT governance was designed for technology scarcity and lengthy procurement cycles, which Abramowitz argues is unsuited to current AI adoption. Survey data highlighted in the talk reveals that 80% of CIOs feel their jobs are directly at risk if they fail to show measurable AI return on investment this year, and 77% believe a peer will be fired over a failed AI strategy or crisis. Rather than treating AI purely as a software procurement issue, Abramowitz proposes organizing it like a digital workforce. In this framework, business units act as hiring managers: they fund agents from departmental P&L budgets, monitor performance metrics, and decommission underperforming models. Central IT shifts to an enablement role analogous to HR or Facilities, establishing identity, access control, and safety guardrails across the organization. Multi-Cloud Blind Spots and Agent Control Planes (11:51) A major barrier to workforce-style governance is operational visibility. Dougherty cites an audit Dataiku conducted for a global manufacturing enterprise whose CIO estimated roughly 40 agents were operating in production. The multi-cloud audit across environments including Databricks, Snowflake, AWS, and Azure uncovered more than 400 active agents—a tenfold disparity where most assets lacked defined ownership, cost tracking, or performance oversight. Dougherty compares this visibility gap to losing track of 90% of an organization's human headcount, underscoring the need for centralized control planes that monitor agent permissions, infrastructure costs, and behavior across cloud platforms. Abramowitz presents two enterprise outcomes: organizations either recreate the spreadsheet era at machine scale with thousands of unmonitored, autonomous agents, or establish a standardized foundation where business teams deploy accountable agents within IT-managed safety frameworks. This choice arrives as 56% of surveyed CEOs admit their competitors currently possess a superior AI strategy.

精炼阅读摘要

Dataiku高管Mark Abramowitz与Jed Dougherty指出，影子AI的广泛蔓延实质上反映了业务端追求敏捷提效的合理诉求；企业应吸取过往封禁电子表格失败的历史教训，将AI代理转变为由业务部门直接出资、考核与雇佣的“数字劳动力”，并由IT部门提供底层的治理与安全支撑。 影子AI的蔓延与电子表格的历史镜像 (0:08) Abramowitz援引Dataiku针对全球8个国家900位大型企业CEO（所在公司年营收均超5亿美元）的调查指出，企业内部未受监管的AI使用已成普遍现象。调查显示，96%的受访CEO认为员工在未获得公司明确批准的情况下使用生成式AI，42%的CEO证实其所在企业至少有一半员工存在此类行为。Abramowitz认为，直接面对业务痛点的基层员工因难以承受官方漫长的审批成本而选择自行构建方案，这属于理性的需求反应，而非恶意违规。 Dougherty通过回顾40年前个人计算的发展历程指出，企业史上最成功的软件部署——电子表格（Excel）最初同样未经过中央IT部门的批准与构建。尽管IT部门花费数十年试图对其进行封禁和管控，但Excel最终仍凭借对业务专家的直接赋能，成为并购财务建模、临床试验物流以及薪酬核算的核心支柱。Dougherty强调，业务专家驱动的技术需求无法被行政禁令消除，当前AI领域正暴露出相同的业务需求与底层企业基础设施之间的脱节。 从单元格错误到自主代理的风险演进 (4:29) Dougherty对比了传统办公软件与AI代理的风险形态。电子表格的潜在隐患通常局限于静态单元格内的公式计算错误，在人工复核时仍有被发现的可能；然而，当前依靠自然语言快速生成的软件仪表板引入了更具隐蔽性的故障模式。例如，一个外观结构完备的销售绩效看板，其底层自动生成的代码可能错误提取了上一季度的实际收入而非当季目标，这类深嵌在数千行机器代码中的缺陷极难通过视觉排查识别。 当技术形态进一步转向具备自主执行能力的AI代理时，潜在威胁显著升级。未受治理的代理不仅会呈现错误数据，更会直接调用系统权限向客户发送邮件、转移企业资金或调用核心商业机密进行训练。为应对自然语言开发普及带来的风险，Dataiku在6月推出了Co-Build产品，旨在让企业在无需配备庞大专属工程团队的前提下，保持对用户自建AI应用的持续审查与故障修复能力。 将AI代理重塑为业务驱动的数字劳动力 (7:10) Abramowitz分析认为，传统集中式IT治理模式脱胎于硬件与软件稀缺且脆弱的历史时期，因而习惯于依赖漫长的年度采购与审批机制；但在AI技术快速迭代的当下，漫长的等待已对业务部门乃至个人的职业生涯构成直接威胁。企业应当将AI管理从单纯的技术采购转变为组织架构设计问题，真正将AI代理视作受劳动力机制约束的数字员工。 在该管理模型中，AI代理的部署权应归属于具体业务部门，由业务主管依据清晰的投资回报率（ROI）指标进行雇佣、直接从部门的损益（P&L）预算中承担运行成本，并负责后续的绩效考评与解雇裁撤。业务主管为避免浪费预算，会主动清理冗余或低效的代理。在此机制下，中央IT部门的角色转向类似于人力资源、财务或行政部门的赋能平台，专注于提供代理身份认证、系统访问控制与安全合规框架。调查显示，80%的受访CIO认为若今年无法产出可衡量的AI投资回报，自身职位将面临风险，且有77%的CIO认为同行可能会因AI战略失误或运营危机而遭解雇。 跨云环境下的资产可见性盲区与治理 (11:51) Dougherty强调，数字劳动力运营模式的落地必须以IT部门对运行中AI资产的全局可见性为前提。他以一家全球大型制造企业的实操审计为例：该企业CIO最初估计全公司在生产环境中仅运行着约40个AI代理；但在Dataiku对其跨越Databricks、Snowflake、AWS和Azure等复杂多云环境展开资产摸底后，实际排查出400多个处于活跃状态的AI代理，形成了十倍的认知偏差。 这些分散运行的代理大多缺乏明确的归属责任人、性能跟踪指标以及成本监视。Dougherty将这种盲区比喻为企业丢失了90%员工的人事档案。针对多云异构环境下的管理失控，Dataiku构建了AI代理管理控制平面，用以集中追踪代理的所属权、系统访问权限、执行成本及其实时运行指标。 Abramowitz最后指出，现代企业正处于分水岭：若缺乏统筹，企业将以机器速度重蹈电子表格时代的失控局面，导致大量不可审查且具备系统操作权限的AI代理在各部门肆意运行；若构建起统一的标准平台底座，则能在释放业务端敏捷构建能力的同时守住治理底线。伴随56%的受访CEO坦承竞争对手的AI战略优于自身，如何建立透明可控的AI运营框架已成为企业缓解高管问责焦虑的关键课题。

Detailed notes — zh

## Background 该内容由频道 Ai4 于 2026 年 8 月 23 日发布在 YouTube 平台上。页面简介显示，演讲者为 Dataiku 首席营销官 Mark Abramowitz 与 AI 及平台高级副总裁 Jed Dougherty，他们基于 900 位 CEO 的调研数据与现场演示，阐述了如果将 Agent 视作数字化劳动力，企业就应通过业务预算与雇佣机制进行闭环治理，而非单纯依赖传统 IT 审批。 ## TL;DR 演讲指出 96% 的 CEO 证实员工正在私下使用生成式 AI，这种现象反映出一线急迫的业务需求，传统 IT 审批已彻底失效。回顾 40 年前 Excel 突破 IT 封锁的历史，演讲强调自主 Agent 具备外部执行能力与系统级破坏力，但解决之道并非死板管控，而是将 Agent 真正当作劳动力对待——由业务部门自担预算（P&L）、评估绩效乃至解雇低效 Agent。IT 部门则需转型为类似 HR 与后勤的底层赋能平台，通过统一的控制平面实现资产观测与默认安全，彻底颠覆现有的治理逻辑。 ## Atmosphere & style 整场演讲语调犀利、直击痛点，充斥着对传统企业官僚体制的颠覆感与现实紧迫感。两位演讲者配合默契，从抛出惊人的调研数据引发全场共鸣，到引用真实制造企业将 40 个 Agent 错估为 400 个的鲜活案例，层层剥开企业管理者的盲区与焦虑。现场氛围既有撕开“影子 AI”假象的冷峻警示，又带有为一线业务正名的务实底气，给听众带来强烈的情感冲击与认知重构。 ## Key takeaways - [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) 绝大多数员工在未经IT批准的情况下私自使用生成式AI（96%的CEO对此认可），表明一线业务需求巨大且不可遏制，核心问题在于底层基础设施缺失而非合规漏洞。 - [02:44](https://www.youtube.com/watch?v=yntVXCUf6kw&t=164s) 过去40年电子表格（Excel）普及的历史证明，业务端自主掌控业务工具的需求终将战胜IT的层层管控，试图通过强行封锁阻挡技术演进注定会失败。 - [04:29](https://www.youtube.com/watch?v=yntVXCUf6kw&t=269s) “氛围感编码”（Vibe Coding）生成的自主Agent带来了全新的危险物理法则：静默逻辑错误极难人工排查，且Agent具备向客户发信、资金调拨等实际行动力，风险远超传统公式计算错误。 - [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) 中央集权式的IT审批是技术资源匮乏时代的产物，冗长的审批流程无法压制员工在AI时代关乎工作生存与业务交付的敏捷需求。 - [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) Agent的本质是数字劳动力：如果将Agent视为劳动力，就必须由最懂业务的部门主导构建、入职、绩效评估乃至解雇。 - [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) 最好的AI治理工具是一张写着业务负责人名字的预算单，将Agent成本直接移入业务部门损益表（P&L），能促使业务端自发进行ROI考核与风险约束。 - [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) 高管面临严酷的AI回报问责压力：80%的CIO表示今年若无法从AI投资中交付可衡量的ROI将面临职位被撤换的风险。 - [10:07](https://www.youtube.com/watch?v=yntVXCUf6kw&t=607s) IT部门需从“审批盖章的守门人”蜕变为AI时代的“HR与底层安全平台”，负责提供数字身份管理、权限控制和审计基线，而非成为业务创新的阻碍。 - [12:01](https://www.youtube.com/watch?v=yntVXCUf6kw&t=721s) 企业对底层AI资产普遍缺乏可观测性：有制造企业管理层预估公司仅有40个Agent，排查后却发现超过400个未受监控的影子Agent正在跨云运行。 - [14:13](https://www.youtube.com/watch?v=yntVXCUf6kw&t=853s) 企业正面临两种未来的抉择：要么以机器速度复制电子表格灾难，要么在坚实底座上由业务部门像管理真实员工一样正式“雇佣”并全生命周期管理AI。 ## Detailed walkthrough ### [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) 96%的员工在“暗度陈仓”：不可遏制的真实需求与缺失的基础设施 在大会开场，Dataiku 首席营销官 Mark Abramowitz 向台下抛出了一个直击企业痛点的开场白：此时此刻，就在大家坐在这里交流的同时，你公司内部某个你绝对不会允许其接触生产系统的人，实际上正在搭建一个生产系统。这群人并不是企业里的捣乱分子，他们没有恶意，反而是离具体业务问题最近的人。由于等待 IT 审批的代价过于高昂，面对能够让工作更轻松、更高效的新工具，他们做出了最理性的自发反应。而更令人坐立难安的事实是——他们的做法是对的。 为了印证这一点，Dataiku 对来自全球 8 个国家、年营收在 5 亿美元以上大型企业的 900 位 CEO 进行了深度调研。调研得出的核心数据令人震惊： - **96% 的 CEO** 坚信他们的员工正在未经企业批准的情况下私自使用生成式 AI（GenAI）。 - **42% 的 CEO** 明确证实，公司内部已有超过半数的员工正在这么做。 这种普遍现象绝不是少数几个边缘团队的个别行为，而是半个公司的常态。面对这一现状，大多数企业的本能反应是“进一步封锁管控”，但 Mark 强调这完全吸取了错误的教训。这种暗流涌动的现象根本不是合规漏洞，而是一个强烈的业务信号：员工之所以不等待许可，是因为业务节奏根本等不起。 整场演讲的核心论点由此明确立下：“业务需求是完全合情合理的，缺失的是底层基础设施。” 在 AI 时代能够最终胜出的企业，绝非那些 AI 采用速度最快或管控力度最严厉的公司，而是那些能够为员工已经自发开展的工作铺设好坚实底层基座的企业。 ```mermaid flowchart LR accTitle: AI需求与治理错位分析 subgraph S1["传统管控困境"] A["业务一线迫切需求"] --> B["IT漫长审批流程"] B --> C["员工绕道私建影子AI"] end subgraph S2["Dataiku破局模型"] D["承认正当业务需求"] --> E["构建受控底层基座"] E --> F["安全敏捷业务交付"] end ``` --- ### [02:44](https://www.youtube.com/watch?v=yntVXCUf6kw&t=164s) 40年前电子表格的战争：IT输了，业务赢了 为了阐明当下的 AI 治理困境，演讲将视线拉回到 40 年前个人计算机时代的黎明。企业软件历史上最成功、普及最广的部署始于 40 年前，而那款产品从诞生起就从未由 IT 部门构建，从未获得过 IT 的正式审批，而且在过去 40 年里，尽管 IT 部门竭尽全力试图扼杀它，却始终未能如愿——那就是电子表格（Excel）。 整整四十年来，企业的并购财务模型、临床药物试验的物流追踪、乃至日常薪酬核算全部跑在 Excel 之上。整个应用生命周期（包括数据采集、清洗、计算、报表和业务预测），完全由真正理解业务的人来掌控，而非理解系统底层架构的 IT 部门。IT 部门与财务业务部门在这种隐秘的拉锯战中痛苦纠缠了 40 年。那么最终谁赢了？IT 彻底输了，业务赢了。 电子表格无处不在的事实证明，将系统交给懂业务的人并不是一件坏事。IT 当年吸取的错误教训是“下次一定要管得更死”，但电子表格真正揭示的规律是：使用工具的业务需求是正当的，当年缺失的同样是基础设施。过去 40 年里，没有任何人真正为电子表格搭建起稳固的底层基座，整个企业界不过是在被迫与安全风险共存。 --- ### [04:29](https://www.youtube.com/watch?v=yntVXCUf6kw&t=269s) 从公式错误到失控Agent：“Vibe Coding”带来的全新危险物理法则 电子表格虽然存在隐患，但其风险上限基本局限于“算错数据”，而且公式错误通常停留在某个单元格内，在局部范围内肉眼尚可排查。然而，当今基于自然语言快速生成的“氛围感编码”（Vibe Coding）与自主智能体（Agent），其破坏力遵循着截然不同的全新物理法则。 演讲者在现场展示了一个外观精美、看似极其专业的销售业绩仪表盘。图表排布整齐，视觉效果极具说服力，完全符合向管理层汇报的标准。然而其底层代码仅有一行写错了日期范围，导致系统直接拿上个季度的实际业绩去对比本季度的目标配额。如果这是一个排版糟糕的图形，人类在两秒钟内就能发现颜色错乱或版面颠倒；但这种在设计与代码逻辑层面的静默错误，在成千上万行自动生成的代码中极难被审查出来。一旦企业基于这种错误数据向董事会汇报并做出战略决策，后果不堪设想。 而未经治理的自主 Agent 带来的危害甚至远甚于此： | 维度 | 传统电子表格 (Spreadsheet) | 未受控的自主 Agent (Ungoverned Agent) | |:--- |:--- |:--- | | **错误范围** | 单元格局部公式计算错误 | 跨系统业务逻辑与操作全链路失真 | | **执行能力** | 静态计算与数据呈现，无外部行动能力 | 自主向客户发送邮件、直接发起资金调拨与交易 | | **数据风险** | 局部文件拷贝或权限泄露 | 擅自抓取并使用企业核心私有数据进行训练 | | **排查难度** | 肉眼或审计公式尚可追踪定位 | 海量自然语言生成的逻辑链难以人工回溯 | 企业真正需要的，是既能保留“用自然语言极速构建”的敏捷性，让最懂业务的人亲手搭建工具，同时必须具备企业级底线控制。企业必须能够清晰透视被构建的内容、在系统崩溃时实时感知，并且无需庞大的 IT 运维和工程团队介入就能即时修复。为此，Dataiku 于 6 月正式推出了名为 **Co-Build** 的产品，旨在让业务人员在坚实的安全治理基座上实现自由构建。 --- ### [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) 最好的AI治理工具：写着负责人名字的预算表单 当前整个科技行业都在苦苦思索“如何严格控制 AI 的普及应用”，但这完全是一个错误的前提假设，因为它预设了企业的组织架构是不可变动的，技术必须削足适履。中央集权式的 IT 运营模式本质上是旧时代“技术资源匮乏”的产物——因为过去的软硬件昂贵、脆弱且罕见，企业才不得不将专业人员集中管理。当业务需求以“年”为单位递进时，这套模式尚能运转；但生成式 AI 彻底打破了这一平衡，因为“等待”所付出的时间成本对业务而言已经具有毁灭性。一线员工如今疯狂构建 AI 是因为他们的工作饭碗系于其上，“你永远无法通过一套冗长繁琐的采购审批流程，去压制员工关乎职业生存的本能。” 这一痛点并非纯粹的技术采纳问题，而是“披着技术外衣的组织架构设计问题”。业界普遍将 Agent 标榜为企业的“数字劳动力”，却极少有人顺着逻辑推导至终局：**如果 Agent 是劳动力，企业就必须将其完全视同劳动力来对待。** - **业务主导构建与雇佣**：由对业务产出结果直接负责的业务部门决定构建和部署何种 Agent，正如业务部门决定招聘哪些员工一样，因为业务最懂实际工作。 - **业务部门承担开销**：Agent 的算力与运行成本必须移出 IT 成本中心，直接计入业务部门的损益表（P&L），对其产出效率与投资回报率（ROI）全权负责。 - **自发约束的治理模型**：“有史以来发明过的最好的 AI 治理工具，就是一张写着负责人名字的预算单（a budget line with someone's name on it）。” 一旦某个 Agent 绑定了具体的业务预算负责人和业绩考核指标，治理问题就会形成自我约束。没有任何管理者愿意持续为冗余、低效的数字员工买单。业务部门将全权主导 Agent 的入职、绩效评估、定期考核乃至“解雇”——未通过考核评估的 Agent 就像通不过试用期的员工一样，坚决不予上线。 ```mermaid flowchart TD accTitle: 业务主导的Agent劳动力生命周期管理 subgraph Dept["业务部门 P&L 责任闭环"] H["提出业务需求"] --> I["自主构建 / 雇佣 Agent"] I --> J["承担预算成本 (P&L)"] J --> K["定期绩效与 ROI 考核"] K -- "表现合格" --> L["持续生产运行"] K -- "表现低下" --> M["直接解雇 / 下线下架"] end ``` 这种责任绑定在当前尤为残酷。针对这 900 位 CEO 的进一步调研显示： - **80% 的 CIO** 直言，如果今年无法从 AI 投资中交付可衡量的明确 ROI，自己的职位将面临被撤换的风险。 - **77% 的高管** 确信，一旦企业遭遇 AI 战略彻底失败或引发重大 AI 危机，同僚中必有人会被直接开除。 --- ### [10:07](https://www.youtube.com/watch?v=yntVXCUf6kw&t=607s) IT角色的蜕变：从审批盖章的守门人，到AI时代的HR与安全底座 权力与责任从来都是对等的：业务部门想要获得 AI 的自主构建权，就必须承担起最终的 ROI 和产出风险。在这种新型劳动力范式下，IT 部门的角色必须发生根本性蜕变——它不再是层层阻拦的“关卡守门人（Gatekeeper）”，而是转变为类似于人力资源（HR）、后勤保障或财务部门的角色。 HR 部门之所以不会被视为业务招聘的绊脚石，是因为它负责为人类员工提供工牌发放、薪酬合规、法务保障等基础服务，让用人部门能够安全合规且高效地招募人才。在充满自主智能体的企业中，IT 部门正是 Agent 的“HR 与后勤保障”：负责提供统一的数字身份管理、严格的权限访问控制、底层审计记录系统以及核心安全合规基线。 “IT 应当成为底层赋能平台，而不是一张通行审批单（the platform, not the permission slip）。” 这一理念让会场两端的参会者都感到不适：它明确警告业务领导者“没有绝对的自主权而不伴随严格的问责”；同时告诫 IT 领导者“你们的工作不再是审批盖章，而是构建能够让审批流程变得完全多余的底层平台”。 --- ### [12:01](https://www.youtube.com/watch?v=yntVXCUf6kw&t=721s) 真实案例冲击：从预估的40个到查出的400个Agent 为了让 IT 顺利转型为平台支撑角色，企业必须首先解决一个致命短板：对全盘 AI 资产的可观测性（Observability）。Dataiku 平台高级副总裁 Jed Dougherty 在现场分享了一个近期发生在全球大型制造企业的真实案例。 在与该企业 CIO 交流时，Jed 询问了一个看似非常基础的问题：“你们公司目前在生产环境中运行了多少个 Agent？” 该 CIO 颇具信心地回答：“大概在 40 个左右。” 然而，当技术团队对其底层分散运行的 Databricks、Snowflake、AWS、Azure 等跨云与数据平台进行全面深度排查后，实际结果令人瞠目结舌： - 该公司实际在生产环境中运行的 Agent 数量**超过了 400 个**，与管理层预估的数字相差整整 10 倍。 - 企业对这些 Agent 处于完全失控状态：不知道是谁构建的、性能与稳定性如何、由谁负责日常维护、消耗了多少云算力成本，甚至连它们是否还在正常运行都一无所知。 Jed 犀利地指出，试想如果一家企业的员工编制人数出现了 10 倍的统计偏差，完全没有人知道这些员工是谁招进来的、归谁管理、在做什么、是否在发薪水，这家企业将沦为一个彻底的笑话。“一个你看不到的底座根本就不是底座，那只是纯粹的侥幸与幻想。” 这也是 Dataiku 重点构建的第二大基座能力：**Agent Management（智能体全生命周期管理）**。企业必须建立统一的控制平面（Control Plane），清晰掌握全公司存在哪些 Agent、所有权归属、系统访问权限、运行成本以及实际业务绩效表现。 --- ### [14:13](https://www.youtube.com/watch?v=yntVXCUf6kw&t=853s) 两种未来的分野：在毁灭边缘重蹈覆辙，还是开始正式“雇佣”AI 演讲尾声指出，基于当前的产业数据，企业面前正在清晰地展开两种截然不同的未来： ```mermaid flowchart TD accTitle: 企业的两种AI未来路径对比 A["企业面对AI浪潮的抉择"] --> B["未来一：重蹈覆辙"] A --> C["未来二：底座赋能"] subgraph Future1["未来一：重蹈覆辙 (失控扩散)"] B --> B1["以机器速度复制电子表格灾难"] B1 --> B2["数以千计不可检视、粗制滥造的影子Agent"] B2 --> B3["具备自主执行权却无人负责，引发系统性崩溃"] end subgraph Future2["未来二：底座赋能 (正规雇佣)"] C --> C1["最懂业务的人使用受控平台快速构建"] C1 --> C2["平台保障默认安全 (Safe by Default)"] C2 --> C3["业务部门像管理员工一样对Agent与ROI全权负责"] end ``` 在调研中，已有 **56% 的 CEO 坦承竞争对手部署的 AI 战略显著优于自身**，这表明在 AI 赌注空前高昂的当下，企业高层的信心正在发生严重动摇。 40 年前，由于缺乏基础设施，企业与电子表格的风险共存了数十年；但在技术以秒级迭代的今天，企业根本没有另一个 40 年的时间去试错。胜出的关键从来不是比拼谁上线的速度最快，也不是比拼谁的封锁手段最严厉，而是在于谁能最先在员工日常的实际工作之下，搭建起坚固、透明、安全的基础设施底座。真正的工作属于最理解业务的人，其余一切则是底层平台的使命。企业必须从今天起彻底转变思维方式：“停止试图去死板管控你的 AI，开始以正式的态度去雇佣它。” ## Notable quotes > [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) 此时此刻，就在大家坐在这里交流的同时，你公司内部某个你绝对不会允许其接触生产系统的人，实际上正在搭建一个生产系统。 — Mark Abramowitz > [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) 业务需求是完全合情合理的，缺失的是底层基础设施。 — Mark Abramowitz > [00:08](https://www.youtube.com/watch?v=yntVXCUf6kw&t=8s) 在 AI 时代能够最终胜出的企业，绝非那些 AI 采用速度最快或管控力度最严厉的公司，而是那些能够为员工已经自发开展的工作铺设好坚实底层基座的企业。 — Mark Abramowitz > [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) 你永远无法通过一套冗长繁琐的采购审批流程，去压制员工关乎职业生存的本能。 — Mark Abramowitz > [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) 如果 Agent 是劳动力，企业就必须将其完全视同劳动力来对待。 — Mark Abramowitz > [07:10](https://www.youtube.com/watch?v=yntVXCUf6kw&t=430s) 有史以来发明过的最好的 AI 治理工具，就是一张写着负责人名字的预算单。 — Mark Abramowitz > [10:07](https://www.youtube.com/watch?v=yntVXCUf6kw&t=607s) IT 应当成为底层赋能平台，而不是一张通行审批单。 — Mark Abramowitz > [12:01](https://www.youtube.com/watch?v=yntVXCUf6kw&t=721s) 一个你看不到的底座根本就不是底座，那只是纯粹的侥幸与幻想。 — Jed Dougherty > [14:13](https://www.youtube.com/watch?v=yntVXCUf6kw&t=853s) 停止试图去死板管控你的 AI，开始以正式的态度去雇佣它。 — Mark Abramowitz ## Mentioned Mark Abramowitz, Jed Dougherty, Dataiku, Ai4, Co-Build, Agent Management, Excel, Databricks, Snowflake, AWS, Azure ## Video description Mark Abramowitz - Chief Marketing Officer, Dataiku Jed Dougherty - SVP of AI & Platform, Dataiku Every session here tells you how to govern your AI better. This one tells you to stop. Everyone says agents are a digital workforce. Almost nobody follows that logic to its conclusion: if agents are labor, the business hires them, pays for them, and fires the ones that underperform. The best AI governance tool ever invented is a budget line with someone's name on it. Half your company is already building AI you never approved. Dataiku CMO Mark Abramowitz and Head of AI Platform Jed Dougherty bring data from 900 CEOs, a live demo, and the one decision that separates AI winners from casualties. [Source](https://www.youtube.com/watch?v=yntVXCUf6kw)

Source quotes

[0:08] Mark Abramowitz: 此时此刻，就在大家坐在这里交流的同时，你公司内部某个你绝对不会允许其接触生产系统的人，实际上正在搭建一个生产系统。 [0:08] Mark Abramowitz: 业务需求是完全合情合理的，缺失的是底层基础设施。 [0:08] Mark Abramowitz: 在 AI 时代能够最终胜出的企业，绝非那些 AI 采用速度最快或管控力度最严厉的公司，而是那些能够为员工已经自发开展的工作铺设好坚实底层基座的企业。 [7:10] Mark Abramowitz: 你永远无法通过一套冗长繁琐的采购审批流程，去压制员工关乎职业生存的本能。 [7:10] Mark Abramowitz: 如果 Agent 是劳动力，企业就必须将其完全视同劳动力来对待。 [7:10] Mark Abramowitz: 有史以来发明过的最好的 AI 治理工具，就是一张写着负责人名字的预算单。 [10:07] Mark Abramowitz: IT 应当成为底层赋能平台，而不是一张通行审批单。 [12:01] Jed Dougherty: 一个你看不到的底座根本就不是底座，那只是纯粹的侥幸与幻想。 [14:13] Mark Abramowitz: 停止试图去死板管控你的 AI，开始以正式的态度去雇佣它。

Source quotes

[0:08] Mark Abramowitz: 96% of CEOs believe their employees are using Generative AI without explicit corporate approval. [0:08] Mark Abramowitz: 42% of CEOs confirm that at least half of their entire workforce is doing so. [0:08] Jed Dougherty: The most successful software deployment in enterprise history—the electronic spreadsheet—was never built, approved, or controlled by central IT. [0:08] Jed Dougherty: IT spent four decades trying to lock down spreadsheets, failing to realize that the legitimate demand for user-driven tools was never going to disappear. [4:29] Jed Dougherty: An ungoverned agent doesn't merely display incorrect numbers on a screen; it directly executes actions by emailing customers, moving enterprise funds, and training on proprietary corporate data. [7:10] Mark Abramowitz: 80% of CIOs report that their own jobs are directly at risk if they fail to deliver measurable AI ROI this year. [7:10] Mark Abramowitz: 77% of CIOs believe a peer at another company will be fired due to a failed AI strategy or operational crisis. [7:10] Mark Abramowitz: When agents are assigned a specific budget owner with measurable revenue or efficiency targets, governance becomes self-enforcing. [11:51] Mark Abramowitz: 56% of surveyed CEOs admit that their competitors currently possess an AI strategy superior to their own

Original transcript

[0:08] morning. Wow. So, somewhere in your company right  now, and I mean right now while we're having this [0:16] conversation here today, there's someone you would  never let go near a production system, is actually [0:23] building one. They're not a rogue. They're not  trying to do cause problems. They're the person [0:29] closest to the problem. Responding rationally to a  system that made it easier, better the uh and the [0:38] consequences of waiting too high. And here's what  I what's a little uncomfortable. They're right. [0:48] So we surveyed 900 CEOs from eight  different countries around the world and [0:54] uh from large companies, companies that are  half a billion plus in revenue. And here's [0:59] what they told us. 96% of them believe their  employees are using Genai without approval. [1:06] That's almost all of them. 42% confirm that  half their employees are already doing this. [1:18] So, not a few rogue, not a small team,  but half the company. And now the reflex [1:23] here is to lock this down, to control it even  more. But that's the wrong lesson. The energy [1:30] in that stat is not a compliance problem.  It's a signal. These people aren't waiting [1:37] for permission because they [clears throat]  can't. So, here's the uncomfortable um sorry, [1:43] here's the argument that we're going  to make over the next 15 minutes here, [1:46] and I'm going to say it plainly now, so  you can decide if you agree or disagree. [1:54] The demand is legitimate and the infrastructure is  missing. This has happened before, but we learned [2:02] the wrong lesson. The companies that are going to  win the AI era aren't those that going to adopt AI [2:09] the fastest or control it the hardest. They'll be  the ones that build the missing foundation under [2:16] the work that their people are already doing.  That's the whole talk. Everything from here on [2:22] out will be uh what that means, what the costs  are if you don't do that and what we actually [2:29] mean by this foundation. And here's the thing,  we're likely to challenge both sides of the room, [2:37] both the business folks in the room and  the IT folks in the room. I'd like to take [2:44] you back 40 years, 40 years to the dawn of  the personal computing age. We've run this [2:51] experiment before 40 years ago. And the results  are in. The most successful enterprise software [2:58] deployment in history started 40 years ago. And  it was one that it never built, never approved, [3:06] and has never ever, despite their very best  efforts, been able to kill the spreadsheet. [3:17] For four decades, finance has run M&A models  in Excel. For four decades, drug trials have [3:23] been tracked their logistics in Excel. Payroll  runs in Excel. An entire application life cycle, [3:30] data sourcing, manipulation, reporting,  projections, all owned not by it, [3:38] but by the people who understand the business,  not the people who understand the system. It [3:44] finance spent 40 years in quiet agony over  this tension. So, who won? It lost. They did. [3:57] Spreadsheets are everywhere and I would argue  that's not a bad thing. The lesson it took was [4:03] lock it down harder next time. But the lesson  the spreadsheet was actually teaching was [4:08] one that Mark just gave you. The demand is  legitimate. The demand to use spreadsheets, [4:14] the demand to use agents, the demand to use AI.  The infrastructure was what was missing. No one [4:20] ever built the foundation under the spreadsheet.  We just lived with the risk for 40 years. [4:29] Now that saved Amanda's back, except the  spreadsheet could only miscalculate. And [4:34] this is the part I really want you  to sit with because this is where it [4:37] gets genuinely dangerous. A formula error in a  spreadsheet sits in a cell. You could see it. [4:45] It's [clears throat] largely localized. Someone  might even catch it. Let me show you what I mean. [4:54] This is a vibe coded sales performance  dashboard. It looks great. It looks very [4:59] believable. I believe it, I guess. Clean layout,  nice charts. But let me show you what's going [5:07] on under the hood. One line, one line of code  pulled the wrong date range. Revenue is being [5:16] calculated against last quarter's actuals  instead of this quarter's targets. The table [5:22] looks right. The chart looks believable. It looks  like something you can put in front of your boss, [5:28] but the inside is completely wrong. If this were  badly designed graphic, you'd know immediately if [5:36] it rendered sideways or the colors were broken.  A human being would be able to catch this in two [5:41] seconds. A data table that's wrong by design,  though, it's incredibly difficult to check. [5:51] You present this to your board. You make  decisions on it. And the moment you try to [5:58] trace through thousands and thousands of fivecoded  logic, well, good luck. An agent is even worse. [6:09] An ungiven agent doesn't just miscalculate. It  emails your customers. It moves your money. It [6:15] trains on your proprietary data. the same demand  as the spreadsheet but with the brand new physics. [6:23] So what do we actually need? We need the  magic of five coding. The ability to build [6:30] fast in plain language by the people who  understand the work with the governance [6:35] and controls enterprises actually require.  You need to see what is being built. You [6:42] need to know when it breaks and you need to  be able to fix it without a fragance team and [6:48] without a massive IT overhead. That's the  first piece of the foundation we're going [6:54] to be talking about today. We've actually  invested in solving this problem. In fact, [6:58] we built a product to do it and we launched it  in June. We call it co-uild. Come find me at the [7:05] booth after this session and I'll give you a live  demo. You can even try to break it if you want. [7:10] So the entire industry is asking how do we control  AI adoption and that's the wrong question because [7:20] this question assumes the org chart is fixed and  that technology must conform to it. We need to [7:27] flip this thinking. [snorts] The central IT  operating model was an artifact of scarcity. [7:35] Technology was expensive, fragile and rare. And  so we centralized the people who could handle [7:41] it. That that made sense and it sort of worked  for a while in a world where business demand [7:47] came on annual increments. But AI broke that  model not because it got worse but because the [7:54] value of not waiting became existential because  people are building as if their jobs depend on [8:02] it. because they do because you cannot procurement  process your way out of an existential incentive. [8:12] [snorts] This is not a technology adoption  problem. It's an organizational design problem [8:20] wearing a technology costume. Everyone in this  room says agents are a digital workforce. I hear [8:30] it constantly. keynotes, panels, podcasts,  but almost nobody takes that to its natural [8:36] conclusion. If agents are labor, we should treat  them like labor. The business unit that owns the [8:43] outcome decides what agents to build and deploy  the same way it decides who to hire because the [8:51] business understands the work. the business pays  for them which means they move off the IT cons [8:57] uh cost center and move into the business P&L  accountable to productivity and ROI uh just [9:05] like any other resource in the business. And so  the best um AI governance model ever invented [9:18] is a budget line with someone's name on it. [9:24] The moment an agent has a budget owner with  a number to hit, your governor's problem [9:30] um becomes self- enforcing. Nobody keeps  paying for a duplicative uh underperforming [9:36] hire. The business manages performance,  onboarding, evaluation, reviews, and yes, [9:44] firing. An agent that can't uh pass review doesn't  ship just like a hireer who can't pass probation. [9:55] And so this is personal. Now we went back and  asked those same 900 CEOs a few more questions. [10:07] 80% of those CIOS say that their role is  at risk if they fail to deliver measurable [10:14] ROI from AI this year. And 77% actually  believe someone's going to get fired, [10:21] one of their peers is going to get fired if they  uh have a failed AI strategy or an AI crisis. [10:33] And so this has always been true that the deal  with labor has always come with you have to out [10:41] own the outcomes. You have to own the ROI and  it well it stops being so much of a gatekeeper [10:50] and it becomes something closer to what HR  facilities or finance are for human labor. [10:58] They become the foundation that makes hiring  safe, legal, and fast. Identity for agents, [11:04] access controls, systems of records, safety codes.  Nobody calls HR a bottleneck just for making sure [11:11] that new hires get badges and payroll. That can  be the destination for it in an agentic company. [11:21] It can be the platform, not the permission slip.  This is the part that no one on the floor is [11:30] saying because it's uncomfortable for both sides  of the room. It tells business leaders you don't [11:36] get autonomy without accountability. It tells  it leaders your job is changing from approving [11:43] things to building the underlying foundation that  makes approval unnecessary. Both sides are getting [11:51] challenged and that's how we know it's true. But  here's the catch. To play that role, it needs [12:01] visibility over what's being created. If those  numbers Mark just showed you don't scare you, [12:07] uh maybe this will. I was recently speaking  with a customer, a major global manufacturer, [12:13] and our conversation landed on, unsurprisingly, I  suppose, agent observability. I asked the CIO what [12:22] seemed like a really simple question. I asked,  "How many agents do you have running inside [12:27] your company right now?" They were like, "Ah,  I think something around 40 40 in production." [12:35] After we did a full analysis and investigated  across their different underlying platforms, [12:40] they had data bricks, snowflake, AWS, Azour,  etc. [snorts] uh turned out to be more than [12:46] 400 in production running across the business. No  visibility into who made them. [clears throat] No [12:53] visibility into how they're performing, who's  managing them, how much they cost, if they're [12:58] even running. Imagine if your employee headcount  was off by 10x. No visibility into who hired them, [13:05] who's managing them, how they're performing,  and if they're actually working at all. That [13:10] would be chaos. That'd be a joke of a company. A  foundation you can't see is no foundation at all. [13:19] It's just a hope. [snorts] It's just a hope. If  the business is hiring agents on five platforms [13:28] and deploying them across six units, somebody  has to know what's running, who built it, [13:34] what it's doing, and what it costs. And somebody  should be it. A foundation you can't see is not a [13:40] foundation. It's a hope. And that's why the second  piece of the foundation is agent management. [13:49] You need a control plane for the workforce the  business is already beginning to hire here in this [13:55] new age. You need to know what agents exist, who  owns them, what they can access, what they cost, [14:01] and whether they're actually performing. That's  the other problem we're building for here at Data [14:06] IQ. I can demo data IQ attitude management at  the booth also. Please come find me afterwards. [14:13] Two futures are unfolding both visibly visible  in the data. In one company's the spreadsheet [14:21] er the the companies repeat the spreadsheet era at  machine speed. Thousands of vibecoded duplicative [14:27] uninspectable agents the new spreadsheets except  these take action. 56% of the CEOs we surveyed [14:38] now admit that their competitors have deployed  an AI strategy that is superior to their own. So [14:46] confidence is eroding just when the AI stakes are  the highest and the other future the same energy [14:53] runs on this foundation. The people who understand  the work are the ones that do the building. The [15:01] platform makes building safe by default and the  businesses the business is accountable for ROI and [15:08] what its agents produce the way it's always been  accountable for its people and what its people [15:15] produce. And so we started with this. The demand  is legitimate. The infrastructure is missing. [15:26] And that was true for the spreadsheet 40 years  ago, and we never fixed it. But now, we don't [15:33] have 40 years. The companies that win won't be the  companies that deploy AI the fastest or govern it [15:41] the hardest. The companies that are going to win  are the ones that build this foundation under [15:46] the work that people are already doing. The  work belongs to the people who understand it. [15:56] everything else is the foundation.  So, it's time to shift your mindset [16:03] and think about how your company  will start to hire AI. Thank you.