Oximy Finds 4% of Enterprise Workers Drive 40% of Shadow AI Output

The Control Tower for AI at Work: Seeing the Work Nobody Can Currently See With Oximy — Ai4

Ai4

Corporate bans fail to curb unapproved tool sprawl because top performers bypass rigid interfaces, trapping proprietary context in vendor silos unless companies adopt unified workbenches with dynamic model routing.

AI-assisted summary of The Control Tower for AI at Work: Seeing the Work Nobody Can Currently See With Oximy — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 17, 2026, this presentation features Naman Ambavi, CEO of Oximy, addressing how enterprises can gain visibility into untracked AI spending and workflows. The description highlights that AI work is scattered across unauthorized tools and personal accounts, framing the session as a guide to establishing a system of record that routes requests to the right models and tools.

## TL;DR Traditional enterprise methods for tracking AI usage rely on fragmented logs and surveys that fail to capture reality, functioning like pre-radar air traffic control. Telemetry reveals that 71% of AI requests bypass official channels, while a core 4% of power users generate 40% of all AI output, leading to severe loss of proprietary context in external silos. To solve this, organizations must control two critical choke points: providing a high-quality internal workbench that employees prefer over consumer tools, and deploying a dynamic model router that optimizes cost, latency, and reasoning. Oximy closes this loop by turning AI visibility into a compounding governance flywheel that can reduce token costs by up to 95%.

## Atmosphere & style The presentation carries a sharp, analytical, and authoritative tone anchored in real-world enterprise data and systems architecture. Ambavi speaks with the pragmatic clarity of an experienced builder, eschewing top-down corporate bans in favor of frictionless product design and automated routing. The delivery feels urgent yet calculated, transforming an overwhelming narrative of enterprise sprawl into an organized, step-by-step engineering roadmap.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=4s) Official enterprise lists of approved AI tools are fundamentally incomplete because current reconciliation methods—finance invoices, IT logs, employee surveys, and vendor dashboards—provide only a fragmented, manual view akin to pre-radar aviation paper strips. - [02:17](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=137s) Telemetry reveals a massive shadow AI footprint: while organizations average contracts for around nine approved tools, employees actively use over 90 different tools, with 38% bypassing procurement entirely. - [02:17](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=137s) Roughly 71% of all enterprise AI requests occur through unapproved shadow channels, with only 29% traveling through designated corporate AI platforms. - [02:17](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=137s) Enterprise AI output is heavily concentrated among a top 4% of power users who produce ~40% of total AI work, yet almost none of these power users operate within official corporate channels. - [02:17](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=137s) Using consumer AI web apps creates a 'lost context' problem where companies pay twice: once on tool subscriptions and again through the permanent loss of institutional knowledge trapped inside third-party vendor silos. - [05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) Governing enterprise AI requires controlling two critical choke points in the request lifecycle: the interaction surface where work happens (the Workbench) and the execution layer where models are selected (the Model Router). - [05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) To prevent workers from reverting to shadow AI, internal workbenches must offer enterprise-grade capabilities—including sandboxing, MCP integrations, and collaborative artifacts—that match or exceed consumer tool UX. - [05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) Dynamic model routers can deliver 60% to 95% inference cost savings by dynamically triaging requests across small, open-source, and frontier models based on cost, reasoning needs, and latency. - [08:48](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=528s) Blanket bans and forced adoption of basic internal chat UIs fail; effective governance requires an attract-and-optimize flywheel that maps power users, draws them into a superior workbench, and routes requests efficiently. - [08:48](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=528s) A complete enterprise AI control tower must go beyond passive traffic monitoring to actively close the loop between visibility, unified workbenches, and dynamic model routing.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=4s) The Flawed "Approved AI List" and the Paper Strip Dilemma

Every enterprise maintains an official list of approved AI tools deemed safe and compliant, yet in practice, that list is fundamentally incorrect and incomplete. Organizations attempt to reconcile their AI footprint through four distinct methods, none of which provide the true picture:

```mermaid flowchart TD accTitle: Four Methods of Enterprise AI Reconciliation subgraph Traditional["Current Reconciliation Methods"] F["1. Finance & Procurement<br/>Expense reports & invoices"] IT["2. IT & Security<br/>Auth logs & network traffic"] S["3. Internal Surveys<br/>Slack / Teams self-reporting"] V["4. Vendor Dashboards<br/>Selective vendor metrics"] end Traditional --> Blindspot["Incomplete Visibility<br/>'Paper Strip' View"] ```

1. **Finance and Procurement**: Reviewing expense reports and purchase orders to catch billed tools. 2. **IT and Security Logs**: Scanning authentication records and network logs for AI platform access. 3. **Internal Surveys**: Polling staff via Slack or Microsoft Teams—often yielding biased responses where employees simply report what management wants to hear. 4. **Vendor Dashboards**: Relying on metrics provided by the tool vendors themselves, which only highlight favorable active user counts and retention numbers.

To capture the inadequacy of these fragmented methods, the state of AI management can be compared to early aviation before radar existed. Air traffic controllers relied on pilots radioing in coordinates, which controllers manually transcribed onto paper strips arrayed across a physical map table. Today's enterprise AI governance similarly depends on self-declared "paper strips," leaving organizations blind to what is actually happening in real time. True oversight requires an automated "radar" capable of continuously identifying what is being used, why it is being used, what it costs, and how to govern it effectively.

---

[02:17](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=137s) Enterprise Realities: Shadow AI, The Top 4%, and the Cost of Lost Context

Large-scale telemetry gathered across tens of thousands of employees reveals a dramatic gap between corporate policy and actual workplace behavior:

```mermaid flowchart LR accTitle: Enterprise AI Reality vs Policy subgraph ToolSpread["AI Tool Adoption"] direction TB A1["9 Enterprise Contracts"] A2["90 Actual Tools in Use"] A3["38% Unprocured Shadow AI"] end subgraph TrafficRoute["Request Routing"] direction TB B1["29% Approved Safe Paths"] B2["71% Unapproved Shadow Paths"] end subgraph WorkloadSplit["Workload Distribution"] direction TB C1["Top 4% Power Users"] C2["~40% of All AI Output"] C3["0% on Standard Corporate Path"] end ```

- **Tool Discrepancy**: While organizations typically reported having enterprise contracts for approximately nine approved tools (such as ChatGPT, GitHub Copilot, or Cursor), employees actively utilized over 90 different tools, including SaaS platforms with embedded co-pilots and third-party extensions. - **Unprocured Spend**: Roughly 38% of all AI tools in active use bypassed procurement entirely. - **Unsafe Request Routing**: A staggering 71% of all AI requests occurred outside official corporate channels, while only 29% traveled through designated enterprise AI platforms. - **The Pareto Effect**: A hyper-concentrated 4% of power users generated nearly 40% of all AI output across an organization. Almost none of these high-volume users relied on approved tools, having bypassed internal red tape to optimize their own workflows using personal accounts and unauthorized interfaces.

This unmanaged sprawl leads directly to the problem of "lost context." When employees feed organizational knowledge, project context, messages, file attachments, and memory into disparate third-party web apps like ChatGPT or Claude, the company receives only a one-time text output in return. As a result, the enterprise ends up "paying twice"—first on the tool subscription invoice, and second through the permanent loss of proprietary context trapped inside external vendor silos.

---

[05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) Anatomy of an AI Request and the Two Critical Choke Points

Taking control of enterprise AI requires a first-principles breakdown of the anatomy of an AI interaction. Every prompt follows a standard lifecycle: an employee opens an application interface, a specific underlying model is selected (often steered toward expensive frontier models by vendors seeking higher margins), and the model returns an answer. Within this workflow lie two primary control choke points.

```mermaid flowchart LR accTitle: Two Choke Points in the Enterprise AI Request Lifecycle User(["Employee"]) --> CP1

subgraph CP1["Choke Point 1: The Workbench"] App["Unified Work Surface<br/>Web, Desktop, Slack"] end

CP1 --> CP2

subgraph CP2["Choke Point 2: The Model Router"] Router{"Dynamic Router"} Router -->|"Cost Optimization"| M1["Open-Source / SLM"] Router -->|"Reasoning / Intelligence"| M2["Frontier Model"] Router -->|"Lowest Latency"| M3["Fast Inference Engine"] end

CP2 --> Out(["Returned Answer"]) ```

#### Choke Point 1: Where the Work Happens (The Workbench) Tech-forward organizations like Stripe, Ramp, Coinbase, and Sierra build or adopt unified internal workbenches across desktop, web, or Slack to own the interaction layer between employees and AI. While building a basic chat interface is trivial, maintaining an enterprise workbench quickly becomes unsustainable when scaling features like sandboxing, credential management, Model Context Protocol (MCP) integrations, skills, collaborative artifacts, and shared canvas sites. When internal tools feel clunky, workers revert to shadow AI. Companies must provide a modern, enterprise-grade workbench that users genuinely prefer over consumer apps.

#### Choke Point 2: Where the Model Gets Chosen (The Model Router) Open-source and specialized models now rival frontier models across many enterprise tasks while delivering 60% to 95% cost reductions at equivalent quality and latency. Instead of locking into static enterprise model agreements, organizations need dynamic model routers that evaluate each request and optimize for one of three core dimensions: - **Cost**: Routing routine tasks (such as simple support ticket categorizations) to the most cost-efficient small or open-source model. - **Quality & Reasoning**: Directing complex legal contract reviews and strategic analysis to high-capability frontier models. - **Latency**: Prioritizing ultra-low latency models for interactive voice agents and customer-facing assistants to avoid robotic pauses.

---

[08:48](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=528s) The Complete Control Tower Loop and Oximy

A true AI control tower cannot stop at passive visibility. Merely graphing traffic solves only the diagnostic problem; complete governance requires closing the loop across visibility, work surfaces, and backend routing.

```mermaid flowchart TD accTitle: The Compounding Control Tower Flywheel M["1. Map & Observe<br/>Identify top 4% power users & shadow workflows"] --> W["2. Provide Unified Workbench<br/>Deliver superior UX without punitive bans"] W --> R["3. Route Backend Traffic<br/>Dynamically optimize cost, quality & latency"] R --> M ```

Rather than issuing blanket bans or forcing employees onto rigid internal chat UIs ("Hey, you guys got to come here from tomorrow with your own internal GPT platform. That's not going to work"), leaders should deploy a compounding flywheel: 1. **Identify**: Map actual AI activity and discover the power users driving 40% of the company's output. 2. **Attract**: Invite those users to a feature-rich, integrated workbench that removes friction. 3. **Optimize**: Place an intelligent model router behind the workbench to silently drive down token costs while safeguarding enterprise context.

Drawing from experience leading a prior Sam Altman-backed enterprise AI transformation company that served multi-billion-dollar organizations, Naman Ambavi founded Oximy to serve as this comprehensive control tower and dynamic router. Organizations can pilot the router across millions of tokens to validate that routing reduces inference costs by up to 95% without compromising output quality.

## Notable quotes > What if the single fastest-growing line in your budget is also the one you can see the least? — Naman Ambavi

> AI at work no longer lives inside one approved tool. It's scattered across ChatGPT, Claude, Copilot, Cursor, personal accounts, shadow tools, and model traffic nobody is routing — Naman Ambavi

> If you own AI strategy, spend, or governance and you're tired of guessing, come see what visibility actually looks like. — Naman Ambavi

> [00:04](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=4s) Every enterprise maintains an official list of approved AI tools deemed safe and compliant, yet in practice, that list is fundamentally incorrect and incomplete. — Naman Ambavi

> [00:04](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=4s) Today's enterprise AI governance similarly depends on self-declared 'paper strips,' leaving organizations blind to what is actually happening in real time. — Naman Ambavi

> [02:17](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=137s) Roughly 38% of all AI tools in active use bypassed procurement entirely. — Naman Ambavi

> [02:17](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=137s) A staggering 71% of all AI requests occurred outside official corporate channels, while only 29% traveled through designated enterprise AI platforms. — Naman Ambavi

> [02:17](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=137s) A hyper-concentrated 4% of power users generated nearly 40% of all AI output across an organization. Almost none of these high-volume users relied on approved tools — Naman Ambavi

> [08:48](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=528s) Hey, you guys got to come here from tomorrow with your own internal GPT platform. That's not going to work — Naman Ambavi

## Mentioned Naman Ambavi, Oximy, Ai4, ChatGPT, Claude, Copilot, Cursor, Stripe, Ramp, Coinbase, Sierra, Slack, Microsoft Teams, Model Context Protocol (MCP), Sam Altman

## Video description Naman Ambavi - CEO, Oximy

What if the single fastest-growing line in your budget is also the one you can see the least?

AI at work no longer lives inside one approved tool. It's scattered across ChatGPT, Claude, Copilot, Cursor, personal accounts, shadow tools, and model traffic nobody is routing, which means most enterprises can't answer basic questions about what they're paying for, what's actually working, or where AI work should happen.

In this session I'll show what enterprise AI usage really looks like underneath the dashboards, drawn from deploying across companies from Series A startups to global enterprises. I'll walk through how leaders are turning that chaos into a system of record: seeing every AI task, routing work to the right tool, and running requests on the right model.

If you own AI strategy, spend, or governance and you're tired of guessing, come see what visibility actually looks like.

[Source](https://www.youtube.com/watch?v=Lj-Ngz1acsk)

中文简报

Oximy发现4%的企业员工贡献了40%的影子AI产出 企业禁令未能遏制未经批准工具的蔓延，因为绩优员工会绕开僵化界面，导致专有上下文被困在供应商的数据孤岛中，除非企业采用具备动态模型路由功能的统一工作台。 传统企业跟踪AI使用情况的方法依赖于零散的日志和问卷调查，无法反映真实情况，运作方式如同雷达出现前的空中交通管制。遥测数据显示，71%的AI请求绕过了官方渠道，而仅占4%的核心超级用户就产生了40%的AI总产出，致使专有上下文严重流失于外部孤岛中。为解决这一问题，机构必须把控两个关键关口：提供员工偏好度高于消费级工具的高质量内部工作台，以及部署能够优化成本、延迟和推理能力的动态模型路由器。Oximy通过将AI可见性转化为可产生复利效应的治理飞轮来完成这一闭环，最高可将Token成本降低95%。 企业现有的批准 AI 工具清单在本质上是不完整的，财务发票、IT 日志、员工问卷与供应商仪表板等传统对账方式仅能提供碎片化视角，犹如雷达发明前依赖纸质便签条的空中交通管制。 (0:04) 遥测数据显示企业影子 AI 泛滥：虽然企业平均仅签署约 9 款工具的采购合同，但员工实际使用的 AI 工具超过 90 款，且约 71% 的 AI 请求走的是未经批准的影子渠道。 (2:17) 企业 AI 产出呈现高度集中，仅占 4% 的超级用户贡献了全公司约 40% 的 AI 输出，但这批高频用户几乎全都在官方审批渠道之外使用个人账户和影子工具。 (2:17) 员工将专有上下文输入第三方消费级 Web 应用会导致组织知识永久流失，形成企业既付订阅费又损失上下文的“双重付费”困境。 (2:17) 治理企业 AI 必须把控两大关键卡点：一是提供体验匹敌消费级工具的企业工作台（Workbench），二是通过动态模型路由器（Model Router）按成本、推理与延迟智能分流，可降低 60% 至 95% 的推理开销。 (5:00)

Reading summary

Enterprises systematically underestimate employee reliance on unauthorized artificial intelligence tools, exposing proprietary context and inflating vendor costs. According to Oximy founder Naman Ambavi, addressing this blind spot requires replacing static audit lists with an operational control architecture that combines an attractive user workbench with dynamic model routing. Flaws in Conventional AI Oversight (0:04) Organizations typically attempt to track AI usage through four retrospective mechanisms: procurement expense reports, IT security logs, internal self-reporting surveys, and vendor dashboards. Ambavi likens these fragmented methods to early aviation before radar, when air traffic controllers relied on pilots radioing coordinates that were manually transcribed onto paper strips across a map table. Relying on self-declared logs and selective vendor metrics similarly leaves enterprise security and finance teams without real-time visibility into what tools are used, why they are chosen, and what they actually cost. Telemetry Findings and the Lost Context Penalty (2:17) Telemetry collected across tens of thousands of corporate employees highlights a stark divergence between corporate policy and actual workflow. While enterprises averaged contracts for about nine approved tools, staff routinely utilized more than 90 distinct tools, with 38% bypassing procurement entirely and 71% of AI requests traveling through unapproved shadow paths. Activity was heavily skewed: a core 4% of power users generated roughly 40% of total AI output, with almost none utilizing corporate-designated tools due to workflow friction. This unmanaged sprawl forces enterprises to 'pay twice.' Beyond paying tool subscriptions, organizations suffer permanent context loss. When employees input proprietary files, messages, and project histories into third-party web interfaces like ChatGPT or Claude, the company receives only an isolated text response, while proprietary institutional memory remains locked inside external vendor silos. Governing the Two AI Interaction Choke Points (5:00) Managing enterprise AI effectively requires controlling two specific choke points in the lifecycle of an AI request: the application interface where work occurs, and the backend layer where models are selected. Forward-looking tech firms such as Stripe, Ramp, and Coinbase address the first choke point by deploying unified internal workbenches across desktop, web, and Slack. Ambavi emphasizes that basic chat wrappers are easy to create but hard to maintain once organizations require advanced capabilities like credential management, sandboxing, collaborative artifacts, and Model Context Protocol (MCP) integrations. If internal workbenches feel clunky, workers will inevitably defect to consumer shadow applications. The second choke point resides at model selection. Static enterprise agreements frequently lock companies into costly frontier models for simple tasks. By introducing dynamic model routers, enterprises can route queries based on cost, quality, and latency. Simpler tasks like ticket triage can be shifted to open-source or small language models that deliver 60% to 95% cost reductions at comparable quality, reserving expensive frontier models for intricate legal or strategic analysis. Deploying a Closed-Loop Governance Flywheel (8:48) Ambavi argues that sustainable governance cannot rely on punitive bans or mandatory internal clones, which consistently encounter user resistance. Instead, an effective AI control tower deploys a three-part compounding flywheel: mapping actual user telemetry to uncover shadow workflows and power users, attracting those users into an integrated workbench with superior usability, and routing backend traffic dynamically. Oximy was designed around this closed-loop model, enabling enterprises to pilot dynamic routing across millions of tokens to reduce inference expenses by up to 95% while retaining institutional knowledge. The proposed control tower framework hinges on the enterprise's ability to provide internal tools compelling enough to attract power users away from cutting-edge consumer applications, making sustained user experience parity the central dependency for enterprise security and cost control.

精炼阅读摘要

Oximy创始人Naman Ambavi指出，企业现行的AI安全核算与批准工具清单严重脱离实际，绝大多数AI生成与交互均发生在IT监控之外的影子路径中。他主张企业必须摒弃被动监测与强制封禁的旧模式，在员工工作台与底层模型路由两个关键卡点建立自动化闭环控制塔，以挽回企业流失的业务上下文并大幅降低推理成本。 传统AI核算方式的盲区与纸条式治理困境 (0:04) Ambavi指出，几乎所有企业都会维护一份经过安全合规审查的官方批准AI工具清单，但这类清单在实践中往往彻底失真。企业目前主要依赖四种传统手段来统计内部AI使用情况：一是财务与采购部门报销和发票审计，仅能捕捉付费工具；二是IT与安全部门对网络流量和身份验证日志的静态扫描；三是通过Slack或Microsoft Teams等渠道进行员工问卷自评，此类调查通常充斥着员工迎合管理层预期的偏差回复；四是依赖AI软件供应商提供的后台看板，而这些看板通常只展示对供应商有利的活跃度与留存数据。 Ambavi将当前的治理模式比作雷达诞生之前的早期民航调度。当时的空管人员完全依赖飞行员通过无线电口头报送坐标，调度员再将信息手写在纸条上并摆放在物理沙盘中，这使得管理层对真实的空中态势处于半盲状态。现今企业的AI管控同样建立在各类自我声明的“纸条”之上，无法实时感知真实世界发生的风险与成本，因此企业急需一种类似雷达的自动化控制系统，持续识别具体工具的使用场景、成因、成本及治理路径。 影子AI泛滥、上下文永久流失与4%超级用户 (2:17) 根据涵盖数万名企业员工的大规模遥测数据显示，企业名义上的AI政策与员工实际操作之间存在巨大的断层。调查显示，一家企业平均签订正式采购合同的官方AI工具约为9种（例如ChatGPT、GitHub Copilot或Cursor），但在组织内部被实际调用的AI工具却超过90种，其中约38%的活跃工具完全绕过了企业采购流程。在请求路由层面，高达71%的AI调用发生在官方企业通道之外，仅有29%的流量通过了指定的安全合规企业平台。 数据还揭示了极度倾斜的二八法则（Pareto效应）：组织内仅占4%的超级用户（power users）贡献了全公司近40%的AI生成输出量。更为严峻的是，这批高产出的核心员工几乎没有一个人使用公司官方批准的标准工具，而是为了摆脱繁琐的内部流程，选择利用个人账户和第三方非授权界面推进日常工作。这种混乱蔓延带来了一个深层代价——业务上下文的流失。当员工将公司的内部知识、项目背景、沟通信息、附件与长期记忆频繁输入外部独立Web应用（如消费级ChatGPT或Claude）时，企业仅能换取一次性的文本产出；这导致企业在支付工具订阅费用的同时，还永久性地将核心业务上下文锁死在第三方供应商的数据孤岛中，承受着双重损失。 请求链路剖析：工作台与动态路由两大核心卡点 (5:00) 要从第一性原理重构企业AI管控，必须对单次AI请求的完整生命周期进行拆解：员工在前端界面发起提示词，系统选择底层模型并执行计算，模型随后返回结果。Ambavi强调，这一链路中存在两个必须牢牢掌握的控制卡点。第一个卡点是承载工作的交互界面（统一工作台）。诸如Stripe、Ramp、Coinbase和Sierra等技术领先企业，倾向于在桌面端、网页端或Slack中自研或部署统一工作台，以掌握员工与AI之间的交互入口。搭建简易的聊天界面虽然门槛较低，但一旦涉及沙箱隔离、凭证管理、模型上下文协议（MCP）集成、外部技能库以及团队协同画布等进阶能力，自建系统的维护成本将呈指数级上升；如果内部工具体验粗糙，员工便会再次流向影子AI，因此企业必须提供用户体验优于消费级产品的企业级工作台。 第二个卡点则是模型选择环节（动态模型路由）。目前开源模型与专用模型在诸多企业级任务中的表现已足以媲美头部前沿模型，且在保持同等响应质量与延迟的前提下，能够实现60%至95%的成本降幅。Ambavi认为，企业不应受制于与单一模型厂商的静态长期协议，而应建立动态模型路由器，根据三个核心维度对每笔请求实施精细分流：对于日常支持工单分类等基础任务，路由至成本最优的小型或开源模型；对于复杂的法律合同审查与战略分析，导流至高智能的前沿大模型；而对于交互式语音智能体等对卡顿极为敏感的前端场景，则优先调度至超低延迟推理引擎。 打造闭环控制塔：从被动阻断走向飞轮治理 (8:48) Ambavi指出，真正的AI控制塔绝不能停留在被动的数据抓取与可视化图表层面。如果仅仅绘制流量走势图，企业只能诊断问题而无法解决问题；完整的治理必须在可见性、工作表面与后端路由三者之间形成闭环运作。企业领导者不应急于发布一刀切的封禁禁令，或者强迫全体员工在次日统一迁移至生硬笨拙的内部自建聊天平台上，这种高压推行方式往往以失败告终。 可行的治理飞轮应包含三个步骤：首先是“发现与测绘”，通过监控精准掌握真实的影子AI工作流，并找出驱动全公司40%产出的前4%核心超级用户；其次是“吸引与聚合”，通过无摩擦、功能完备且体验卓越的统一工作台将这些核心用户自然吸引至受控环境中；最后是“优化与路由”，在工作台后端接入智能模型路由器，在留存组织业务上下文与核心资产的同时，悄然压缩推理Token成本。Ambavi结合自身此前在Sam Altman投资的企业AI转型机构中服务大型客户的经验创立了Oximy，旨在将这一控制塔理念与动态路由方案落地；他在演讲中表示，企业通过试运行数百万Token的实际流量，即可验证该路由方案能在确保输出质量的前提下降低最高达95%的推理成本。 该演讲的核心结论在于，企业治理影子AI的有效途径并非压制性禁令，而是以高体验的工作台收拢用户，再辅以智能路由降低成本并保留数据资产；但演讲内容主要基于Oximy自身产品架构与技术理念，如何在多变的业务场景中建立客观量化的输出质量评测标准，以及企业在跨模型分流时如何兼顾不同供应商的数据合规与隐私界限，依然是该方案在实际工程落地中需要进一步解决的问题。

Detailed notes — zh

## Background 该视频由 YouTube 频道 Ai4 于 2026 年 8 月 17 日发布，主讲人为 Oximy 首席执行官 Naman Ambavi。视频围绕企业内部 AI 预算与工具失控的痛点展开，旨在揭示常规仪表盘之下的真实 AI 使用全貌，并分享如何通过控制塔（Control Tower）建立记录系统、实现任务路由与模型优化。 ## TL;DR 演讲指出企业传统的 AI 审批清单与四种追踪手段（财务报销、IT 日志、全员问卷、供应商后台）存在严重盲区，实测显示实际使用的 AI 工具数量往往是官方名单的 10 倍，且存在 4% 核心用户完成 40% 载荷的极度不均现象。为解决数据与上下文流失问题，企业必须掌控两大咽喉点：一是提供体验媲美商用软件、能沉淀内部上下文的统一 Workbench，二是建立在成本、质量与延迟间动态权衡的智能模型路由（Model Router）。通过雷达观测、专属工作台与模型路由器构成的自动化闭环，企业能够在保障安全与质量的同时降低 60% 至 95% 的模型调用成本。 ## Atmosphere & style 演讲整体节奏紧凑、直击痛点且极具实战说服力。主讲人以民航早期纸条空管的生动比喻切入，随后抛出颠覆认知的实测统计数据，迅速打破了管理层对合规现状的虚假安全感。整场分享风格务实清晰，既有第一性原理的架构拆解，又充满落地操作指南，带给观众一种从混乱混沌中理清脉络、掌握主动权的笃定感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=4s) 企业传统的“已审批 AI 工具清单”多为谎言，实际使用情况远远超出管理层的掌控范围。 - [00:04](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=4s) 依赖财务报销、IT 日志、全员问卷和供应商后台的追踪方式存在严重盲区，犹如依靠飞行员无线电口报坐标的原始纸条时代，企业迫切需要一套现代“雷达系统”。 - [02:53](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=173s) 实测数据显示，企业实际使用的 AI 工具数量往往是官方签署清单的 10 倍，且 38% 的工具从未走过正规采购流程。 - [02:53](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=173s) 高达 71% 的 AI 请求完全绕过了企业批准的安全通道，仅 29% 走官方路径。 - [02:53](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=173s) 企业 AI 使用存在极端不均的“4/40 幂律分布”：仅 4% 的重度用户完成了约 40% 的 AI 工作量，且这批核心高产员工大多因审批繁琐而使用未经批准的外部工具。 - [02:53](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=173s) 员工将商业上下文和内部知识持续喂给第三方黑盒工具，导致企业陷入“既付订阅费、又损失不可逆上下文资产”的双重锁定代价。 - [05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) 企业治理 AI 必须掌控两大核心咽喉（Choke Points）：员工交互的工作台（Workbench）和模型选择路由（Model Selection）。 - [05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) 企业需要完全拥有并掌控交互工作台，自研简易网页难以应对沙箱隔离、MCP、Skills 扩展等复杂需求，明智之举是采用可扩展的成熟工作台底座。 - [05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) 开源模型的崛起使得成本可骤降 60% 至 95%，企业应根据业务场景在成本、质量和延迟之间动态调度模型，而非盲目绑定最贵的前沿模型。 - [08:48](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=528s) 真正的企业 AI 控制塔（Control Tower）不能停留在静态监控，必须形成“雷达观测识别核心用户 → 引导至专属 Workbench → 底层智能模型路由分流优化”的自动化正向飞轮。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=4s) 企业 AI 审批清单的谎言与观测盲区 在任何大型组织中，管理者通常都会拿出一份“已获审批的 AI 工具清单”（Approved AI List），声称这是企业内部用于保证安全合规的 AI 工具集。然而，现实情况往往大相径庭——这份清单既不准确，也远远谈不上详尽，员工实际使用的工具根本不受其约束。 目前企业通常依赖四种传统途径来核对和追踪员工使用的 AI 工具，但这四种方式均存在致命短板： 1. **财务部门报销与采购记录**：拉取发票、报销单和采购审批流，但只能覆盖走正规预算报销的商业工具。 2. **IT 与安全团队日志**：通过网络防火墙、安全网关及 SSO 认证日志排查 AI 流量，但极易遗漏非受控设备和个人账户。 3. **全员问卷调查**：在 Microsoft Teams 或 Slack 上向成千上万名员工推送问卷。这种方式看似直接，但“员工往往只会回答你希望听到的答案”，没人会主动承认自己在违规使用未受控工具。 4. **供应商后台仪表盘**：各大 AI 供应商（如 OpenAI 等）提供的控制台，只展示它们想让你看的数据（如活跃用户数、留存率、文件访问情况），无法提供全局视角。 这就好比早期民航领域的空中交通管制——在没有雷达的年代，飞行员需要主动通过无线电报告自身坐标，空管员把坐标抄在纸条（Paper Strips）上，桌面上铺满的纸条就构成了所谓的“航线图”。如今企业管理 AI 依然停留在这种依赖员工“自我申报”的原始纸条阶段。企业真正迫切需要的是一套现代“雷达系统”，能够穿透表象，清晰洞察：究竟哪些工具在被使用、为什么被使用、产生了多少开销，以及掌握全局后该如何干预与治理。 ```mermaid flowchart TD accTitle: 传统企业 AI 追踪方式与盲区 subgraph Traditional["传统碎片化追踪（纸条时代）"] Finance["财务与采购报销<br/>（仅限报销项目）"] ITLogs["IT/安全/SSO 日志<br/>（盲区严重）"] Surveys["全员问卷调查<br/>（充满虚假汇报）"] VendorDash["单一供应商仪表盘<br/>（数据割裂孤立）"] end Radar["雷达观测层 (Oximy Control Tower)<br/>全景透视：工具类型、真实使用动机、端到端成本"] Traditional --> Radar ``` --- ### [02:53](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=173s) 大规模实测揭露的残酷真相：影子 AI 与 4/40 幂律法则 基于在从数万人规模企业到高成长初创公司中的实际部署与数万名员工的观测数据，隐藏在常规仪表盘之下的真实图景令人震惊： - **工具数量暴增 10 倍**：某企业官方声称并签署企业合同的 AI 工具仅有 9 种（包括 Cursor、GitHub Copilot、ChatGPT 等），但监控雷达实际检测到的 AI 工具竟多达 90 种（包含内嵌 AI 功能的各类 SaaS、垂直场景协作者等）。 - **38% 未经采购审核**：检测到的工具中有高达 38% 从未走过正规采购流程，属于典型的“影子 IT”（Shadow AI）。 - **71% 请求绕过安全路径**：高达 71% 的 AI 请求完全未走企业合规的安全路径，仅有 29% 的请求通过了企业批准的官方通道。 - **4/40 帕累托法则（极度不均的使用分布）**：头部仅 4% 的重度用户消耗并完成了约 40% 的全量 AI 工作载荷。尤为讽刺的是，这群最能利用 AI 创造价值的核心员工，几乎无一例外都不使用公司审批的官方工具；因为大企业的官方审批流程极其繁琐冗长，他们为了提高交付效率，直接绕过红线使用自己习惯的外部工具。 这种碎片化使用带来了巨大的隐形代价：企业的商业上下文、日常沟通消息、内部知识资产、附件以及长期协作记忆，源源不断地流向第三方的黑盒应用（如 ChatGPT 或 Claude）。企业换回来的仅仅是一个扁平的答案，实质上是在“支付双重代价”——不仅要为工具订阅发票付费，更在为供应商锁定（Vendor Lock-in）导致的不可逆企业上下文流失付费。 --- ### [05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) 请求解构：企业必须掌控的两大核心咽喉（Choke Points） 要彻底重构治理体系，必须运用第一性原理去解构一个 AI 请求的完整生命周期。一次完整的交互链条包含清晰的三步：用户访问工作应用 $\to$ 模型路由选择 $\to$ 接收输出结果。在这条链路上，存在两个至关重要的卡点（Choke Points）： ```mermaid flowchart LR accTitle: AI 请求生命周期的两大咽喉点 subgraph Flow["AI 请求解构"] User["业务人员<br/>（市场/财务/通用）"] --> Choke1["咽喉点 1：工作台<br/>(Workbench)<br/>控制员工交互与上下文"] Choke1 --> Choke2["咽喉点 2：模型路由<br/>(Model Router)<br/>控制成本/质量/延迟"] Choke2 --> Provider["模型供应商<br/>(开源/闭源模型)"] Provider --> Answer["返回结果"] end ``` #### 咽喉点 1：工作实际发生的地方（The Workbench） 这是属于“人”的入口。无论是市场人员使用的专属工具，还是财务使用的分析副驾，企业必须掌控员工与 AI 发生首次交互的界面。类似 Stripe、Ramp、Coinbase、Sierra 等技术领先的 AI 原生企业，都在不遗余力地打造统一的内部 Workbench。这个工作台的呈现形态并不受限（可以是桌面客户端、Slack 应用或网页端），核心在于**企业必须完全拥有并掌控交互界面**。 很多企业在此容易犯错：起初觉得自研一个内部 ChatGPT 网页极其简单，但随着场景深入，需要集成沙箱隔离（Sandboxing）、凭据管理（Credentialing）、MCP（Model Context Protocol）、Skills 扩展、协作画布（Artifacts）等复杂功能时，自研工具很快沦为体验平庸的鸡肋，员工随即迅速流失并重回外部商业工具。明智的做法是采购成熟可扩展的工作台底座，底层接入中立解耦的工具体系。 #### 咽喉点 2：模型被选择的地方（The Model Selection） 这是属于 IT 与财务成本控制的核心枢纽。绝大多数 SaaS 工具会诱导用户选择最贵的前沿模型，鼓吹其“智商最高、幻觉最少”，但核心驱动力在于供应商能赚取更多差价。在未来 3 到 9 个月内，开源模型已全面崛起并在规模化能力上正面对抗前沿模型，且成本直接骤降 60% 至 95%，同时保证同等水准的质量与响应延迟。 企业无法在前期单向采购阶段就把模型选型定死，而是需要根据动态业务场景在“成本（Cost）”、“质量（Quality）”、“延迟（Latency）”三者之间寻求最佳平衡： - **客户工单处理**：优先追求成本，选择能满足要求的最低成本模型极速处理。 - **合同与法务审查**：优先追求质量，必须调用顶级前沿模型提供极高水准的推理与智能支持。 - **语音客服智能体**：优先追求极低延迟，避免由于模型停顿让客户明显察觉是在与机器人对话。 --- ### [08:48](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=528s) 闭环控制塔与 Oximy 解决方案 真正的“企业 AI 控制塔”（Control Tower）绝不仅仅是一个静态的可视化看板或合规监控层。单纯给出一张“现状地图”只能解决最初级的感知问题。一个真正生效的控制塔必须形成自动化正向飞轮： ```mermaid flowchart TD accTitle: 控制塔闭环运作机制 A["1. 观测雷达<br/>(识别全貌与 Top 4% 高价值员工)"] --> B["2. 引入专属 Workbench<br/>(提供沉淀企业上下文的统一交互入口)"] B --> C["3. 底层接入智能模型路由<br/>(按需动态调度成本、质量与延迟)"] C --> A ``` 通过雷达精准识别出创造 40% 业务价值的 Top 4% 核心用户，将他们平滑引导至体验优秀的企业工作台；工作台底层由智能模型路由器（Model Router）接管请求流量，动态分流优化开销。这种方式远胜于简单粗暴地封禁所有外部网站、强推难用的内部替代品。 演讲者 Naman Ambavi 介绍，其此前创立的公司曾获得 Sam Altman 的投资，专门为营收数十亿美元的巨头企业提供 AI 转型架构服务。如今创立的 Oximy 便是上述高并发实战经验的体系化结晶。企业可以通过其开放的免费测试额度接入数百万 Token，直接验证在生成质量毫不受损的前提下大幅压降推理成本的效果。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=4s) 已获审批的 AI 工具清单既不准确，也远远谈不上详尽，员工实际使用的工具根本不受其约束。 — Naman Ambavi > [00:04](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=4s) 在没有雷达的年代，飞行员需要主动通过无线电报告自身坐标，空管员把坐标抄在纸条上，桌面上铺满的纸条就构成了所谓的‘航线图’。如今企业管理 AI 依然停留在这种依赖员工‘自我申报’的原始纸条阶段。 — Naman Ambavi > [00:04](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=4s) 企业真正迫切需要的是一套现代‘雷达系统’，能够穿透表象，清晰洞察：究竟哪些工具在被使用、为什么被使用、产生了多少开销，以及掌握全局后该如何干预与治理。 — Naman Ambavi > [02:53](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=173s) 某企业官方声称并签署企业合同的 AI 工具仅有 9 种，但监控雷达实际检测到的 AI 工具竟多达 90 种。 — Naman Ambavi > [02:53](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=173s) 高达 71% 的 AI 请求完全未走企业合规的安全路径，仅有 29% 的请求通过了企业批准的官方通道。 — Naman Ambavi > [02:53](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=173s) 头部仅 4% 的重度用户消耗并完成了约 40% 的全量 AI 工作载荷。尤为讽刺的是，这群最能利用 AI 创造价值的核心员工，几乎无一例外都不使用公司审批的官方工具。 — Naman Ambavi > [02:53](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=173s) 企业换回来的仅仅是一个扁平的答案，实质上是在‘支付双重代价’——不仅要为工具订阅发票付费，更在为供应商锁定导致的不可逆企业上下文流失付费。 — Naman Ambavi > [05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) 一次完整的交互链条包含清晰的三步：用户访问工作应用 → 模型路由选择 → 接收输出结果。在这条链路上，存在两个至关重要的卡点。 — Naman Ambavi > [05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) 起初觉得自研一个内部 ChatGPT 网页极其简单，但随着场景深入，需要集成沙箱隔离、凭据管理、MCP、Skills 扩展、协作画布等复杂功能时，自研工具很快沦为体验平庸的鸡肋，员工随即迅速流失并重回外部商业工具。 — Naman Ambavi > [05:00](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=300s) 在未来 3 到 9 个月内，开源模型已全面崛起并在规模化能力上正面对抗前沿模型，且成本直接骤降 60% 至 95%，同时保证同等水准的质量与响应延迟。 — Naman Ambavi > [08:48](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=528s) 真正的‘企业 AI 控制塔’绝不仅仅是一个静态的可视化看板或合规监控层。单纯给出一张‘现状地图’只能解决最初级的感知问题。一个真正生效的控制塔必须形成自动化正向飞轮。 — Naman Ambavi > [08:48](https://www.youtube.com/watch?v=Lj-Ngz1acsk&t=528s) 通过雷达精准识别出创造 40% 业务价值的 Top 4% 核心用户，将他们平滑引导至体验优秀的企业工作台；工作台底层由智能模型路由器接管请求流量，动态分流优化开销。 — Naman Ambavi ## Mentioned Naman Ambavi, Oximy, Ai4, ChatGPT, Claude, GitHub Copilot, Cursor, OpenAI, Microsoft Teams, Slack, Stripe, Ramp, Coinbase, Sierra, Sam Altman, MCP (Model Context Protocol) ## Video description Naman Ambavi - CEO, Oximy What if the single fastest-growing line in your budget is also the one you can see the least? AI at work no longer lives inside one approved tool. It's scattered across ChatGPT, Claude, Copilot, Cursor, personal accounts, shadow tools, and model traffic nobody is routing, which means most enterprises can't answer basic questions about what they're paying for, what's actually working, or where AI work should happen. In this session I'll show what enterprise AI usage really looks like underneath the dashboards, drawn from deploying across companies from Series A startups to global enterprises. I'll walk through how leaders are turning that chaos into a system of record: seeing every AI task, routing work to the right tool, and running requests on the right model. If you own AI strategy, spend, or governance and you're tired of guessing, come see what visibility actually looks like. [Source](https://www.youtube.com/watch?v=Lj-Ngz1acsk)

Source quotes

[0:04] Naman Ambavi: 已获审批的 AI 工具清单既不准确，也远远谈不上详尽，员工实际使用的工具根本不受其约束。 [0:04] Naman Ambavi: 在没有雷达的年代，飞行员需要主动通过无线电报告自身坐标，空管员把坐标抄在纸条上，桌面上铺满的纸条就构成了所谓的‘航线图’。如今企业管理 AI 依然停留在这种依赖员工‘自我申报’的原始纸条阶段。 [0:04] Naman Ambavi: 企业真正迫切需要的是一套现代‘雷达系统’，能够穿透表象，清晰洞察：究竟哪些工具在被使用、为什么被使用、产生了多少开销，以及掌握全局后该如何干预与治理。 [2:53] Naman Ambavi: 某企业官方声称并签署企业合同的 AI 工具仅有 9 种，但监控雷达实际检测到的 AI 工具竟多达 90 种。 [2:53] Naman Ambavi: 高达 71% 的 AI 请求完全未走企业合规的安全路径，仅有 29% 的请求通过了企业批准的官方通道。 [2:53] Naman Ambavi: 头部仅 4% 的重度用户消耗并完成了约 40% 的全量 AI 工作载荷。尤为讽刺的是，这群最能利用 AI 创造价值的核心员工，几乎无一例外都不使用公司审批的官方工具。 [2:53] Naman Ambavi: 企业换回来的仅仅是一个扁平的答案，实质上是在‘支付双重代价’——不仅要为工具订阅发票付费，更在为供应商锁定导致的不可逆企业上下文流失付费。 [5:00] Naman Ambavi: 一次完整的交互链条包含清晰的三步：用户访问工作应用 → 模型路由选择 → 接收输出结果。在这条链路上，存在两个至关重要的卡点。 [5:00] Naman Ambavi: 起初觉得自研一个内部 ChatGPT 网页极其简单，但随着场景深入，需要集成沙箱隔离、凭据管理、MCP、Skills 扩展、协作画布等复杂功能时，自研工具很快沦为体验平庸的鸡肋，员工随即迅速流失并重回外部商业工具。 [5:00] Naman Ambavi: 在未来 3 到 9 个月内，开源模型已全面崛起并在规模化能力上正面对抗前沿模型，且成本直接骤降 60% 至 95%，同时保证同等水准的质量与响应延迟。 [8:48] Naman Ambavi: 真正的‘企业 AI 控制塔’绝不仅仅是一个静态的可视化看板或合规监控层。单纯给出一张‘现状地图’只能解决最初级的感知问题。一个真正生效的控制塔必须形成自动化正向飞轮。 [8:48] Naman Ambavi: 通过雷达精准识别出创造 40% 业务价值的 Top 4% 核心用户，将他们平滑引导至体验优秀的企业工作台；工作台底层由智能模型路由器接管请求流量，动态分流优化开销。

Source quotes

Naman Ambavi: What if the single fastest-growing line in your budget is also the one you can see the least? Naman Ambavi: AI at work no longer lives inside one approved tool. It's scattered across ChatGPT, Claude, Copilot, Cursor, personal accounts, shadow tools, and model traffic nobody is routing Naman Ambavi: If you own AI strategy, spend, or governance and you're tired of guessing, come see what visibility actually looks like. [0:04] Naman Ambavi: Every enterprise maintains an official list of approved AI tools deemed safe and compliant, yet in practice, that list is fundamentally incorrect and incomplete. [0:04] Naman Ambavi: Today's enterprise AI governance similarly depends on self-declared 'paper strips,' leaving organizations blind to what is actually happening in real time. [2:17] Naman Ambavi: Roughly 38% of all AI tools in active use bypassed procurement entirely. [2:17] Naman Ambavi: A staggering 71% of all AI requests occurred outside official corporate channels, while only 29% traveled through designated enterprise AI platforms. [2:17] Naman Ambavi: A hyper-concentrated 4% of power users generated nearly 40% of all AI output across an organization. Almost none of these high-volume users relied on approved tools [8:48] Naman Ambavi: Hey, you guys got to come here from tomorrow with your own internal GPT platform. That's not going to work

Original transcript

[0:04] Thank you so much for having me.  While we get the slides up over here, [0:09] something I want to put across is every single  one of you in your organizations would have [0:15] an approved AI list. A list that you guys have  been approved to use for safe AI, for secure AI. [0:22] That's wonderful. But the matter  of fact is that list is incorrect. [0:27] That list is not the exhaustive list that we  are using to get the work going. There are [0:33] four different ways in which someone reconciles  your AI list in terms of what they're using. [0:39] The first one is from the finance department  where you kind of pull in your expense reports, [0:44] you put in your procurement reports and you  understand here are all the AI tools being used. [0:50] The second is basically going to be with your  IT or your security team where you're kind of [0:54] explaining here's my security logs, here's my  authentication logs, and here's how we are able [1:00] to kind of get the people that are using AI.  The third is my most favorite. It's a survey. [1:06] You get out on Teams, you get out on Slack, you  send a survey to thousands of peoples over there [1:11] and you're kind of able to understand what these  people might be using, but the matter of fact is [1:16] they might just be telling you what you want to  hear. And the fourth is the vendor dashboard. [1:21] The vendor dashboards are going to be the one  that will tell you exactly what they would like [1:26] for you to see. Your active users, your retention,  people trying to access your kind of files, etc. [1:32] The part that is most interesting is that none  of them actually give you the whole picture, [1:38] which is fair. None of them will actually tell you  what is really going on. How do you get to that? [1:45] To explain that, let me give you a very simple  scenario. I'm assuming all of you have come in via [1:50] flights and aviation etc. Um, before there were  like radars in air traffic control, people and [1:58] pilots used to tell the air traffic controls that  hey here's my coordinate, here's where I'm at, [2:04] write it on a paper strip and the table where the  paper strips are was the map. This is the current [2:10] situation for AI. Those are the paper strips of  like the tools that you have been selfdeclared. [2:17] So I propose you guys should have something like  a radar. What does a radar do at a high level? [2:25] A radar can exactly tell you what's being used,  why is it being used, how much is it costing you, [2:32] and once you have an entire picture, what can  you really do about it? Right? We ran this across [2:41] hundreds and thousands and tens of thousands of  employees. We have multiple deployments at scale [2:48] and here are some findings I want to share with  you and then what you can do about it as well. [2:53] The first finding was that they named  nine different AI tools that they used [2:58] which are enterprise AI contracts. They had  cursor co-pilots. You have chat GPT etc. [3:04] The matter of fact were there were 90 different  tools being used. They could be tools with AI [3:10] features, SAS platforms with AI co-pilots and so  much more. And 38% of what they saw actually never [3:19] went through procurement. This is a real fact. The  second finding we actually found was out of these [3:25] different tools 71% of the requests that came  in actually never went in through the safe path [3:33] and the rest of the 29% went through the approved  path through the enterprise AI platforms that [3:38] they had access to which is quite shocking.  The third one is again my most favorite one [3:46] out of this this was a very interesting stat. This  is a paritive effect which is very very common. [3:51] The top 4% of the people that we have accumulated  to roughly 40% of all AI work which is [3:59] but obvious. And at the same time almost every  single one of them didn't use an approved AI tool. [4:06] It just came in from various tools that they were  comfortable using. The corporates had a lot of [4:11] red tapes to go across it. So they just  kind of like optimized their work around it. [4:16] Very interesting fact. And once you have this,  what are you going to do? You have a company. [4:22] The company has context, messages,  knowledge layers, attachments, memory, [4:29] collaboration materials, all going to somebody  else's app, which is going to be either a chat GPT [4:35] or a cloud or something like that, and all you get  back is an answer. The funny part is that you're [4:41] actually paying twice now. You're paying one for  the invoice and second is what you're paying for [4:46] lost context because you can't get that context  back from the vendor lockins that you guys have. [4:51] So you kind of think okay never mind this  is great you gave me a picture in terms of [4:55] what's wrong what's going on. What the hell do  we do about it? How do we kind of take forward [5:00] to take control of it? So from a first principles  thinking which is what really matters in this era [5:06] is we break down the anatomy of what a request  looks like. When a request looks like there's a [5:13] person who goes to an app where the work actually  happens. If it's a marketing guy, he goes to a [5:19] marketing app. If it's a finance guy, he goes to a  financial AI co-pilot of sorts. If it's a general [5:24] purpose tool, that's where they go. Once they  go, the next thing is the model that they choose. [5:31] Probably the apps are going to try to lucrate you  to use the most expensive model because that's [5:37] apparently the most smart like the smartest model.  Um, it's the least hallucinating model, so on and [5:43] so forth. But a secret is they probably want you  to choose that model because they make more money. [5:50] Wow. Okay. So, and then you get back an answer.  The answer is just what they were looking for. [5:57] These are two choke points that I think are very,  very, very, very, very fundamentally important. [6:02] The first choke point is where the work happens.  This is for the people. The second choke point [6:08] is where the model gets chosen. And this is  mainly for finance. This is mainly for IT. [6:14] Going deeper into this for the first choke  point. You essentially want to come with your [6:20] own workbench which is what most folks are getting  into. A lot of AI native companies, a lot of tech [6:26] forward companies. You have the examples of the  stripes, the ramps, the coinbases of the world, [6:32] the Sierras of the world who are building their  own workbench. The surface doesn't matter. It [6:36] could be a desktop app. It could be a Slack app.  It could be a a web app. It doesn't really matter. [6:42] What matters is that they own the interaction of  the employee interacting with AI. And that's very [6:49] very important. The common mistakes again we  see is people try to get excited and make it [6:55] themselves. And a first version of a chat app is  actually very very easy. But the minute you kind [7:00] of add things like sandboxing, credentiing, MCP,  skills, collaborative materials, artifacts, sites, [7:06] hundreds of different things, you lose touch.  The app becomes mediocre and people go back [7:12] to what they preferred, which is what you don't  want. So you want to come with a way that you can [7:15] kind of have a workbench that you can buy and then  kind of have your own agnostic tools beneath it. [7:21] That's the first layer, the first choke point.  You want to be able to control what your employee [7:27] does when the first interaction of AI happens. The  second choke point is where the model gets chosen. [7:33] This is fundamentally going to be the most  important topic. If not now, then maybe 3 to 6 to [7:40] 9 months from now because one, open-source models  are now taking charge and competing with frontier [7:47] models at scale. There is roughly a 60 to 95% cost  deduction when it comes to it with the same amount [7:54] of quality and the same amount of latency. What  you really want to do over here is that if it's [8:00] you want to optimize for three things.  You either want to optimize for cost, [8:04] you want to optimize for quality or you want to  optimize for latency. And this depends on the use [8:09] case. You can't necessarily decide that when you  procure a model or an enterprise agreement. If [8:14] it's a support ticket, you want to be quick. you  want to get it done as the cheapest model possible [8:17] to get the job done. If it's a contract and a  legal review, you want a frontier model to have [8:22] reasoning and intelligence on top. If you have a  customer support wise agent, for example, you want [8:27] to have the fastest model so that the customer  doesn't feel it's a robot talking to them. [8:31] Routers make this possible. Routers can actually  analyze the traffic and optimize for one of these [8:37] three things. And I think this is very important  when you want to optimize for one of these three [8:42] things, which is which is what everyone is  going to be look for. This is at large of [8:48] how the control tower should be. Typically when we  see control tower, it always looks like a layer of [8:54] visibility in terms of you're controlling, you're  governing, you're securing, which is great. Like [8:58] I completely understand that. But giving a map  and an observability layer of what's going on [9:04] only solves problem number one. Problem number  two is can you control where the work happens? [9:10] And problem number three is can you also control  which models are picked for that exact same work. [9:16] And it's a loop. It kind of compounds over time.  You see the work happening which is your top 4% [9:21] of people. You invite them to a workbench  that you've made. behind the workbench, [9:25] you can have a model router that can optimize your  costs and you can keep running this loop which is [9:30] a phenomenal way of getting people to adopt  your own company's workspace without forcing [9:37] them to just block everything else and be like,  "Hey, you guys got to come here from tomorrow [9:41] with your own internal GBD platform. That's not  going to work. This is typically how you should [9:46] be doing it. I run a company called Oxy where  we focus on all of this. Before this, I ran [9:52] another company which is backed by Sam Altman  where we did AI transformations for companies [9:56] with billions of dollars of revenue. OxyMe is the  version that my learnings have compounded into. [10:03] If you ever would like to get your picture of what  is really going on in your company, you should [10:09] just come visit us at our booth at K921. I'm going  to be around in this room around as well. Um, [10:15] and if you would like to try out our model  router, that is also free. Try it with like a [10:20] few million tokens. You can really see the quality  not deteriorating, but your cost going all the way [10:25] down. That's kind of the point. Um, thank you so  much for your time. I enjoy the rest of the day.