Vibrant Capital CEO Says AI Fails to Boost Corporate EBIT

Scaling AI on Main Street: From Demos to Bottom Line With Vibrant Capital — Ai4

Ai4

Shadman Zafar argues non-tech enterprises miss out on AI profits because they deploy lazy architectures and fail to redesign workflows around small language models.

AI-assisted summary of Scaling AI on Main Street: From Demos to Bottom Line With Vibrant Capital — original source. Check the original for full context.

Source published: 2026-08-22

Detailed notes

## Background Published on YouTube by Ai4 on August 22, 2026, this session features Shadman Zafar, CEO of Vibrant Capital. Based on the page description, Zafar draws on two decades of CIO experience to present an operator-led playbook for Main Street industries looking to transition from hype and pilots to deep, repeatable AI deployment that impacts bottom-line business results.

## TL;DR Despite massive civilizational hype surrounding AI, most enterprise operators struggle to realize actual bottom-line financial gains due to structural hurdles like tool rush fatigue, lazy tech stacks, and outdated governance. To fix this gap, companies must shift from chasing every new model release to implementing durable architectural patterns, such as using dynamic prompt routers to send routine tasks to lower-cost Small Language Models (SLMs) and building unified enterprise data fabrics. Furthermore, organizations need to modernize governance as automated code and restructure into an 'hourglass' organizational model that retains entry-level talent for evaluation and data custodianship while scaling leadership through agent orchestration.

## Atmosphere & style The presentation carries an authoritative, pragmatic, and slightly contrarian tone, characteristic of an experienced enterprise operator speaking candidly to peers. Zafar delivers uncomfortable truths with calm clarity, blending analytical data points with relatable anecdotes about corporate AI panics. The session feels intensely practical and grounded, cutting through Silicon Valley grandiosity to offer a realistic, operator-focused reality check.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) A major AI paradox exists between civilizational hype from tech leaders and the lack of visible EBIT impact on non-tech Main Street balance sheets. - [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) A PwC study of 4,454 CEOs revealed that 56% report zero financial return from AI, while 12% report minor improvements under 5% of EBIT. - [02:23](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=143s) The primary cause of poor AI return on investment lies in four key operational hurdles: lazy architecture, legacy governance, outdated org structures, and tool rush fatigue. - [05:02](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=302s) Technology operators frequently suffer from 'tool rush fatigue', repeatedly switching software tools without allowing teams to extract actual workflow value. - [05:02](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=302s) Relying on 'lazy architecture'—a thin app layer over a single frontier model with basic RAG—creates high token costs and fails at enterprise scale. - [08:41](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=521s) Deploying a dynamic prompt router with a model portfolio strategy can cut token expenditures by 40% to 50% by delegating simple workflows to Small Language Models. - [08:41](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=521s) Outdated or poisoned enterprise context is 600% more likely to yield dangerous or out-of-policy responses than raw model hallucinations. - [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) Enterprise AI governance must shift from manual, slow approval committees to automated 'governance as code' and dedicated lifecycle management for autonomous agents. - [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) Organizing enterprises into an 'hourglass structure' retains essential junior roles for evaluations and data stewardship while leveraging AI to compress middle management. - [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) Eliminating entry-level pipelines to rely entirely on external models risks severe 'intelligence leakage' and mortgages the long-term future of the enterprise.

## Detailed walkthrough Here is a detailed, faithful walkthrough of the transcript.

---

[00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) The Great AI Paradox: Civilizational Hype vs. Main Street P&L

Shadman Zafar opens the session by framing what he considers one of the biggest paradoxes happening in artificial intelligence today. On one side, frontier lab creators, tech CEOs, CFOs, media gurus, and visionaries speak of AI in dramatic, "civilizational terms." They claim AI will rewrite history, upend every industry, reshape civilization, and perhaps alter humanity as we know it. However, if you step just a few miles outside Silicon Valley and walk into a Main Street enterprise—a bank, a hospital, a logistics firm, or an insurance company—and inspect their quarterly P&L statements, you encounter an entirely opposite reality.

Most non-tech companies are simply not seeing any meaningful EBIT impact from AI yet. To back this up, Zafar references a recent PwC study surveying 4,454 CEOs. When asked whether AI has delivered tangible cost reductions or revenue growth, 56% of CEOs reported seeing zero impact—not even 1% or 2%. Meanwhile, only 12% reported seeing "some improvement," which amounted to less than a 5% shift in EBIT. For financial operators, a 5% improvement in EBIT typically translates to a mere five basis points on the overall bottom line. This disconnect does not mean AI technology is ineffective—it clearly works in daily life—but its value is failing to land on corporate income statements.

- **PwC Survey Data**: 4,454 CEOs surveyed regarding AI financial returns. - **56% of CEOs**: Reported zero cost reduction or revenue uplift from AI. - **12% of CEOs**: Reported minor improvements (less than 5% of EBIT, often working out to roughly ~5 basis points). - **Core Paradox**: Unprecedented civilizational expectations on one side, versus a few basis points in financial filings on the other.

---

[02:23](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=143s) Technology Creators vs. Technology Operators

Zafar draws a clear line between two distinct groups in the ecosystem: **technology creators** (the engineers, labs, and startups building new frontier models and tools) and **technology operators** (the CIOs, CTOs, CISOs, and COOs tasked with running real-world enterprises). The operators find themselves trapped right in the middle of this paradox. They face immense pressure from CEOs and CFOs demanding 50% cost cuts or 200% revenue surges, while struggling to deliver more than a couple of basis points in savings.

After polling the room to gauge how many attendees identify as creators versus operators, Zafar outlines why current implementations are falling short from the operator's perspective. He identifies five key structural hurdles, highlighting four primary drivers: 1. **Lazy Architecture**: Relying on quick-fix tech stacks rather than durable design patterns. 2. **Outdated Governance**: Using legacy policy checks instead of modern automated guardrails. 3. **Organizational Inertia**: Retaining legacy team structures instead of redesigning the workforce around AI. 4. **Tool Rush Fatigue**: Constantly swapping software tools before realizing actual workflow value.

He briefly notes a fifth factor often cited by observers—that "technology diffusion takes time"—but dismisses it as a scapegoat. In Vibrant Capital's studies, diffusion lag accounts for less than 6% to 10% of the delay. The real bottlenecks lie in how technology operators design, govern, and execute. He warns that enterprises face a "Borders vs. Barnes & Noble" tipping point: those who integrate AI deeply into workflows will survive and thrive, while those who fail will be left behind.

---

[05:02](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=302s) Tool Rush Fatigue and "Lazy Architecture"

Tracing the problem back to the post-November 2022 explosion of generative AI, Zafar points out that enterprise CIOs spent decades operating in deterministic, declarative programming environments. When generative tools began evolving on a weekly basis, many CIOs panicked about how to handle the onslaught and responded by repeatedly deploying new tools without understanding the underlying architectural paradigms.

To illustrate this "tool rush fatigue," Zafar shares an anecdote about a CIO he coaches. A year and a half ago, the CIO secured a budget to deploy GitHub Copilot across his software organization, meticulously training leaders and developers. Just as the rollout finished, peers told him Copilot was already outdated and he was "stupid" for not using Cursor. Without updating his board, he quietly requested more money and spent four to five months rolling out Cursor. Frustratingly, 45 minutes after completing the Cursor deployment and training, colleagues informed him Cursor was obsolete and he needed to switch to Claude Code. He was embarking on his third tool migration without his engineering team ever deriving sustained business value from any of them.

``` +-----------------------------------------------------------------------+ | TYPICAL "LAZY" ARCHITECTURE | | | | [ Thin Custom App Layer ] --> [ Prompt Firewall / RAG Infrastructure ]| | | | | v | | [ Single Frontier Model ] | | | | | v | | [ MCP Server / Legacy DBs ] | +-----------------------------------------------------------------------+ ```

Zafar contrasts this reactionary approach with "lazy architecture"—the widespread pattern of slapping a thin application layer over a single frontier model, hooking up basic Retrieval-Augmented Generation (RAG), adding a prompt firewall, and deploying a Model Context Protocol (MCP) server to access legacy databases. While simple to set up for a demo, this setup fails to scale or adapt.

---

[08:41](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=521s) Durable Architectural Patterns: Model Portfolios, Routers, and Data Fabrics

Instead of chasing every model release or overspending on frontier tokens, Zafar advises CIOs to build durable architectural patterns. A prime example is implementing a **model portfolio strategy** driven by a dynamic prompt router (such as the open-source OpenRouter framework). By categorizing prompts by workflow complexity, enterprises can route lower-complexity tasks to Small Language Models (SLMs) and reserve frontier models only for multi-step, high-reasoning tasks.

To demonstrate the power of SLMs, Zafar cites a study analyzing 594 distinct operational workflows within a consumer bank: all but 177 of those workflows could be executed by small language models at the exact same capability level as a high-cost frontier model. Because high-volume customer transactions fall into these SLM-capable buckets, deploying a router immediately slashes enterprise token expenditure by 40% to 50%. Token cost explosions are usually not an inherent AI problem, but rather a direct symptom of lazy architecture.

``` +-----------------------------------------------------------------------+ | DURABLE ROUTER ARCHITECTURE | | | | [ Inbound Prompts ] | | | | | v | | [ Dynamic Prompt Router ] | | / | \ | | / | \ | | v v v | | [ SLMs (70%+) ] [ Mid Models ] [ Frontier Models ] | +-----------------------------------------------------------------------+ ```

A second critical pattern is establishing an enterprise **data fabric**. Historically, companies maintained separate silos for analytical data, operational data, and OLTP databases, later wedging in a vector database or graph for RAG. However, in decoupled systems, data drifts daily, leading to "poisoned context."

- **Context Poisoning vs. Hallucinations**: In an enterprise setting, poisoned or outdated context is **6 times (600%) more likely** to produce a dangerous, out-of-policy response than a model hallucination. - **Resilience Tools**: Technologies like Couchbase provide high-performance, real-time synchronization across analytical and non-analytical data stores, ensuring AI models receive live, un-drifted data.

---

[13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) Modernizing AI Governance and Redesigning the Hourglass Organization

Moving to governance, Zafar notes that traditional committee approvals and manual sign-offs drag cycle times down to a crawl. Modern enterprise governance must be expressed directly as code—using automated frameworks and specialized solutions (such as Autopill or Trustwise). He poses a fundamental question for agentic deployments: "If agents are becoming a huge chunk of your workforce, where is the HR department for your agents?" Just as human employees require onboarding, monitoring, and compliance reviews, AI agents require programmatic governance.

Equally important is restructuring the enterprise workforce. While many industry experts argue that AI transforms corporate pyramids into "inverted pyramids" (few leaders, zero junior staff, and heavy automation), Zafar warns that this is a dangerous error. Instead, progressive enterprises are adopting an **hourglass organizational model**:

- **Top Layer**: Senior leaders whose expertise is scaled significantly through AI agent orchestration. - **Middle Layer**: A compressed middle management layer, as routine oversight and execution are handled by autonomous agents. - **Bottom Layer**: A robust entry-level pipeline of graduates, interns, and junior staff focused on model evaluations, reinforcement learning, data custodianship, and lifecycle management.

Zafar cautions that eliminating entry-level roles "mortgages the future" of the company and triggers severe "intelligence leakage." By default, enterprise contracts forbid frontier labs from training public models on private corporate data. If a company routes all cognitive tasks to external frontier models without capturing local feedback loops, fine-tuning, or reinforcement learning internally, its organizational intelligence freezes in stasis while employees fail to develop deep expertise over time.

---

[17:11](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=1031s) Operator Playbook and Vibrant Capital

In closing, Zafar urges technology leaders to take full responsibility for mastering internal AI lifecycle management rather than blindly outsourcing core cognitive architecture to third-party providers.

Reflecting on his own career, Zafar shares that after serving nearly 30 years as a Fortune 50 CIO and CTO across major institutions—including Citi, JPMorgan Chase, and Verizon—he now focuses on guiding the next generation of technology leaders. He highlights two key initiatives: - **CIO Fellows Society (`ciofellows.org`)**: A non-profit fellowship program built to give back to the tech community, elevate the craft of the CIO, and enable tech executives to learn operator-led best practices from one another. - **Vibrant Capital (`vibrantcapital.ai`)**: An operator-led growth network and investment firm where he serves as CEO, helping enterprise leaders cut through the noise of thousands of AI point-solutions to implement durable tools that deliver measurable bottom-line value.

---

## Notable quotes > [00:00](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=0s) What they're actually scaling is experiments. — Shadman Zafar

> [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) On one side, frontier lab creators, tech CEOs, CFOs, media gurus, and visionaries speak of AI in dramatic, 'civilizational terms.' — Shadman Zafar

> [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) Most non-tech companies are simply not seeing any meaningful EBIT impact from AI yet. — Shadman Zafar

> [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) 56% of CEOs reported seeing zero impact—not even 1% or 2%. — Shadman Zafar

> [02:23](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=143s) The real bottlenecks lie in how technology operators design, govern, and execute. — Shadman Zafar

> [05:02](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=302s) He was embarking on his third tool migration without his engineering team ever deriving sustained business value from any of them. — Shadman Zafar

> [08:41](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=521s) Token cost explosions are usually not an inherent AI problem, but rather a direct symptom of lazy architecture. — Shadman Zafar

> [08:41](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=521s) In an enterprise setting, poisoned or outdated context is 6 times (600%) more likely to produce a dangerous, out-of-policy response than a model hallucination. — Shadman Zafar

> [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) If agents are becoming a huge chunk of your workforce, where is the HR department for your agents? — Shadman Zafar

> [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) Eliminating entry-level roles 'mortgages the future' of the company and triggers severe 'intelligence leakage.' — Shadman Zafar

## Mentioned Shadman Zafar, Vibrant Capital, PwC, GitHub Copilot, Cursor, Claude Code, OpenRouter, Couchbase, Autopill, Trustwise, Citi, JPMorgan Chase, Verizon, CIO Fellows Society

## Video description Shadman Zafar — CEO, Vibrant Capital

Most companies think they're scaling AI. What they're actually scaling is experiments. In this session, I will argue that Main Street industries, from banking and insurance to healthcare, manufacturing, logistics, retail, and communications, have seen enough demos. What they're missing is AI built into the everyday workflows where cycle time, cost, risk, and revenue actually move. Drawing on two decades as a CIO and what we are building at Vibrant Capital through our Growth Partners Network and CIO Fellows Society, I will share a more operator-led playbook for moving from pilots and hype to repeatable deployment. My view is a bit contrarian: the winners will not be the companies with the flashiest models, but the ones that redesign work around AI, govern it well, and make it accountable to business results.

If you are a tech operator trying to make AI stick inside a real enterprise, you will leave with a sharper lens, a few uncomfortable truths, and a practical way to act.

[Source](https://www.youtube.com/watch?v=L6AoBsRRl7s)

中文简报

Vibrant Capital CEO称AI未能提振企业EBIT Shadman Zafar认为，非科技企业之所以错失AI利润，是因为采用了懒惰架构，且未能围绕小语言模型重新设计工作流。尽管围绕AI存在巨大的文明级炒作，但由于工具追逐疲劳、懒惰技术栈以及过时治理等结构性障碍，大多数企业运营者仍难以实现切实的底线财务收益。为了弥合这一差距，企业必须从盲目追逐每一次新模型发布，转向落实持久的架构模式，例如利用动态提示词路由将常规任务分派给低成本的小语言模型（SLM），并构建统一的企业数据网络。此外，企业还需将治理现代化为自动化代码，并重组为“沙漏型”组织模式，在保留初级人才以承担评估与数据保管职责的同时，通过智能体编排来扩展领导力。普华永道对 4,454 位 CEO 的调研显示，56% 的受访者表示 AI 未带来任何财务回报，12% 仅看到低于 5% 的息税前利润（EBIT）小幅改善，凸显出行业宏大叙事与实体企业财报间的巨大脱节。 (0:04) 企业 AI 投资回报率低迷并非源于技术普及滞后，而是受困于四大运营障碍：懒惰架构（Lazy Architecture）、陈旧治理、组织惯性以及频繁更换软件工具引发的“工具狂热疲劳”（Tool rush fatigue）。 (2:23) 依赖单一前沿模型加基础 RAG 的“懒惰架构”成本高昂且难以规模化；采用动态 Prompt 路由器结合小语言模型（SLM）的组合策略，可直接削减 40% 至 50% 的 Token 支出。 (8:41) 在企业实际场景中，过时或受污染的上下文导致危险或违规响应的概率比模型幻觉高出 600%，因此需建立统一的实时数据编织（Data Fabric）与代码化治理（Governance as Code）。 (11:20) 企业组织应重构为“沙漏型模型”，由 AI 智能体压缩中层管理并放大高管能力，同时保留初级人才负责模型评估、强化学习与数据管理，以防出现智力资产流失与人才断层。 (13:28)

Reading summary

Former enterprise technology executive Shadman Zafar argues that artificial intelligence has failed to generate meaningful corporate earnings because enterprise operators rely on brittle architectures, rapid tool churn, and outdated organizational structures rather than durable engineering patterns. The P&L Paradox and Operator Hurdles (0:04) While technology creators promote artificial intelligence in sweeping civilizational terms, Zafar points out that enterprise income statements show virtually no bottom-line benefit. He cites a PwC survey of 4,454 corporate chief executives where 56 percent reported zero cost reductions or revenue growth from AI, and only 12 percent observed minor gains under 5 percent of earnings before interest and taxes (EBIT)—a shift that typically yields around five basis points to the bottom line. Zafar identifies enterprise technology operators, such as CIOs and CTOs, as caught between executive demands for massive margin shifts and structural deployment bottlenecks. Rejecting the common defense that technological diffusion simply requires extended time, Zafar claims Vibrant Capital research attributes less than 6 to 10 percent of adoption delays to diffusion lag. Instead, he attributes the gap directly to architectural design failures, legacy governance, and organizational resistance. Tool Rush Fatigue and Brittle Stacks (5:02) Enterprise leaders accustomed to deterministic programming environments frequently panic amid rapid generative AI releases, falling into what Zafar calls 'tool rush fatigue.' To illustrate, he recounts an enterprise CIO who consecutively funded and rolled out GitHub Copilot, Cursor, and Claude Code over 18 months, repeatedly replacing systems immediately after developer onboarding without ever measuring sustained operational value from any single platform. This continuous tool migration is compounded by 'lazy architecture.' Zafar describes this pattern as wrapping a thin application layer around an external frontier model, basic retrieval-augmented generation (RAG), a prompt firewall, and a model context protocol server connecting to legacy databases. While sufficient for isolated demonstrations, Zafar warns that these setups lack resilience, fail to adapt to workflow variance, and rapidly inflate operational costs. Dynamic Routing and Data Fabric Integration (8:41) To build durable systems, Zafar advises deploying dynamic prompt routers that triage enterprise tasks across a portfolio of models. Citing an analysis of 594 operational workflows at a consumer bank, he notes that 417 of those workflows could be handled by small language models (SLMs) at performance parity with top-tier models. Directing standard, high-volume transactions to SLMs while reserving complex reasoning for frontier models immediately cut enterprise token expenditures by 40 to 50 percent. Equally critical is resolving data fragmentation through unified data fabrics. Decoupling analytical data, operational databases, and vector stores leads to data drift and outdated context. Zafar states that in enterprise workflows, poisoned or stale context is six times (600 percent) more likely to cause a dangerous, out-of-policy output than an internal model hallucination, requiring continuous synchronization layers to feed models live operational ground truth. Governance as Code and Hourglass Staffing (13:28) Traditional governance committees and manual sign-offs introduce severe operational latency. Zafar contends that deploying autonomous agent fleets requires 'governance as code'—automated compliance, monitoring, and policy guardrails that function as programmatic human resources departments for software agents. Structurally, Zafar challenges the prediction that AI turns enterprises into inverted pyramids with few managers and no junior workers. Instead, he advocates an hourglass organizational structure: senior leaders amplified by agent orchestration at the top, a compressed middle-management layer where routine oversight is automated, and a preserved junior workforce at the base. Eliminating entry-level roles, Zafar warns, causes 'intelligence leakage' and freezes organizational expertise, as junior teams are needed to handle evaluation benchmarks, reinforcement learning, and data custodianship while training future leadership. Zafar concludes that enterprises must retain internal mastery over AI lifecycle management and architectural routing rather than outsourcing cognitive processes entirely to frontier labs, leaving unanswered how non-technical firms can build such sophisticated engineering capacity without inflating initial overhead.

精炼阅读摘要

Vibrant Capital 首席执行官 Shadman Zafar 指出，尽管前沿科技实验室与媒体对人工智能抱有文明层面的宏大叙事，但多数实体企业尚未在利润表中看到实质性财务回报。他认为这一脱节的主因并非技术扩散滞后，而是企业运营者在架构设计、工具选型、治理模式和组织架构上陷入了结构性误区。宏大叙事与财务表现的脱节 (0:04) Zafar 强调，在硅谷以外的银行、医院、物流与保险等实体企业中，AI 的应用尚未对息税前利润（EBIT）产生普遍的拉动作用。PwC 对 4,454 位首席执行官（CEO）展开的一项调查显示，56% 的受访 CEO 表示 AI 带来了零成本削减或零收入增长；仅有 12% 报告了轻微改善，但其对 EBIT 的贡献低于 5%，反映到综合利润表底线上通常仅约 5 个基点。技术创造者持续宣称 AI 将重塑历史与文明，而企业的技术运营者（如 CIO 与 CTO）却在管理层削减 50% 成本或激增 200% 收入的重压下举步维艰。Zafar 认为技术并非无效，而是企业无法将其变现。Vibrant Capital 的研究显示，常被用作借口的技术扩散滞后在整体延迟因素中仅占不到 6% 至 10%，真正的阻碍在于技术架构和运营落地的执行偏差。工具追逐疲劳与偷懒式架构 (5:02) 自 2022 年底生成式 AI 爆发以来，习惯于确定性、声明式编程环境的 CIO 们陷入了频繁更换工具的恐慌性恶性循环。Zafar 举例说明，一位企业 CIO 在花费数月部署 GitHub Copilot 后，被同行指责落后而转向 Cursor；而在 Cursor 部署培训完成仅 45 分钟后，他又被劝说更换为 Claude Code。这种频繁迁移导致工程团队尚未从任何工具中获得持续业务价值，就已陷入严重的工具疲劳。与追逐新工具并存的是“偷懒式架构”（Lazy Architecture）：企业往往仅在前沿模型上方搭建一层薄弱的应用层，简单接入基础检索增强生成（RAG）、提示词防火墙，并配合模型上下文协议（MCP）服务访问遗留数据库。这种简易堆叠能快速实现演示原型，但缺乏应对实际复杂业务的可扩展性与适应力，最终沦为无法产生效益的技术负债。模型路由策略与数据编织防线 (8:41) 为替代单调依赖高成本前沿模型的做法，Zafar 建议部署基于动态提示词路由器（如 OpenRouter 框架）的模型组合策略。系统根据业务工作流的复杂度进行任务分流：将多步骤、高推理要求的任务交给前沿大模型，而将常规任务分流至小型语言模型（SLM）。在针对一家零售银行 594 个运营工作流的分析中，研究发现多达 417 个流程（即除 177 个流程外）由 SLM 处理即可达到与高成本前沿模型完全相同的能力水平。由于高频客户交易大多落在该区间，引入路由器可直接降低企业 40% 至 50% 的 Token 开销。与此同时，企业必须构建实时的数据编织（Data Fabric）来解决数据漂移引发的“上下文投毒”（Poisoned Context）。企业过去长期割裂分析型、运营型及 OLTP 数据库，并在事后生硬接入向量库用于 RAG。数据显示，在企业级生产环境中，过时或被污染的上下文导致危险或违规响应的概率，是模型自身产生幻觉的 6 倍（600%）。借助如 Couchbase 等能够实现跨存储实时同步的工具，才能保障 AI 模型获得未经漂移的活数据。代码化治理与沙漏型组织重构 (13:28) 传统的委员会审核与人工审批流程已无法适配 Agent（智能体）的部署节奏。Zafar 主张将治理逻辑直接转化为代码，利用自动化框架（如 Autopill 或 Trustwise）对 Agent 进行类似于人力资源部门对员工进行的入职、持续监控与合规核查。缺乏程序化治理的自主智能体将给企业带来无法承受的运营风险。针对行业内“组织将变为倒金字塔型”（即完全取消基层、全面依靠自动化）的论调，Zafar 提出反驳，认为这会导致“智力泄漏”并抵押企业的未来。领先企业正在转向“沙漏型组织”：顶层是借助 Agent 放大能力的资深领导与业务专家；中层管理层因常规执行与监督被 Agent 接管而大幅压缩；底层则保留活跃的毕业生与实习生梯队，专门负责模型评估、强化学习、数据管护与生命周期管理。由于主流商业合同禁止前沿实验室利用私有数据训练公共模型，企业若不保留初级梯队在内部沉淀反馈闭环与微调能力，其组织智力将被冻结停滞。 Zafar 呼吁技术高管停止盲目将认知架构外包给第三方，必须亲自主导企业内部的 AI 生命周期管理。他正通过非营利组织 CIO Fellows Society 和投资机构 Vibrant Capital 推广这些基于一线运营经验的最佳实践。然而，在组织层面上，企业能否在压缩中层管理的同时成功建立精通模型评估与数据治理的初级梯队，仍是该模式在多行业铺开时需要检验的实际挑战。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 22 日发布在 YouTube 平台上。内容基于 Vibrant Capital 首席执行官 Shadman Zafar 的主题演讲，聚焦于传统实体经济（Main Street）企业如何穿透技术炒作，将 AI 从概念验证与 Demo 演示真正转化为对企业损益表（P&L）有实质贡献的可持续工作流。 ## TL;DR 演讲揭示了当前 AI 的文明级叙事与实体企业财报回报几乎为零之间的刺眼断层，指出盲目堆砌工具与简陋脆弱的“懒惰架构”是落地的主要障碍。技术运营者必须转向“耐用型架构”，通过小型领域模型（SLM）与智能路由大幅压缩 Token 成本，并借助企业数据织锦（Data Fabric）消除比模型幻觉危害大 6 倍的数据漂移隐患。同时，企业应推行代码化治理（Governance as Code）并将组织重构为沙漏型架构，防止核心业务认知全盘外包导致企业智力停滞。 ## Atmosphere & style 整场演讲风格沉稳、老练且极具逆向思辨色彩，充满了资深技术操盘手的务实底蕴。讲者以毫不留情的“不适真相”和令人捧腹却又无比真实的行业踩坑故事，直击当前企业盲目追逐工具的焦虑与虚火。叙事从对泡沫乱象的犀利解构平稳过渡到清晰落地的工程与治理范式，带给观众强烈的共鸣感、清醒感与重归专业匠心的行动指引。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) 宏大叙事与实体企业财务表现断层：尽管 AI 被冠以文明重构级别的叙事，但实体经济企业的财报 EBIT 几乎未见改变；普华永道调研显示 56% 的全球 CEO 表示 AI 带来的降本或增收改善完全为零。 - [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) 断层根源于技术创造者与技术运营者的鸿沟：硅谷模型创造者专注于研发前沿模型，而传统企业的 CIO 与 CTO 等运营者则被夹在董事会的不切实际预期与微弱现实回报之间。 - [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) 阻碍价值转化的核心是非时间因素：导致实体企业无法兑现 AI 价值的关键是架构粗糙、治理过时、组织形态脱节及工具泛滥，单纯的技术扩散时间延迟在归因中仅占不足 10%。 - [05:12](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=312s) 盲目追逐新工具引发团队疲态且无法沉淀价值：缺乏底层认知的企业陷入工具轮换的怪圈，在频繁采购与培训新工具中空耗生产力，却未能积累出实际的业务资产。 - [05:12](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=312s) 警惕拼凑式“懒惰架构”的脆弱性：仅在大模型外围简单拼凑外围薄应用、简易 RAG、提示词防火墙和 MCP 接口的懒惰架构，无法支撑复杂的企业级业务场景。 - [08:41](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=521s) 践行耐用架构需要构建模型组合策略：企业无需执着于单一终极模型，而应通过智能路由将绝大多数高频业务交由小型领域模型（SLM）处理，以削减 40% 至 50% 的 Token 成本。 - [08:41](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=521s) 数据漂移造成的“脏上下文”危害远超模型幻觉：若缺乏打通事务与分析的数据织锦，生产环境中过时与失效上下文引发合规事故的概率比模型自身幻觉高出 600%。 - [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) AI 治理必须彻底转向代码化治理（Governance as Code）：依赖人工签字审批的静态传统治理难以应对自主智能体，必须引入运行时动态审计与代码化边界约束。 - [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) 组织形态应重塑为“沙漏型”而非取消底层的倒金字塔：顶部专家借助 AI 获得杠杆、腰部常规流程流转被压缩，同时必须保持底部吸纳年轻人才以持续培养未来的技术梯队。 - [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) 警惕外包导致的“企业智能外泄”与组织认知停滞：企业若将核心认知完全外包给外部大模型且缺乏内部强化学习（RL）闭环，将永远无法沉淀组织智慧，陷入认知停滞。 - [17:11](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=1031s) 重拾技术运营者的专业底蕴：AI 时代最终的赢家不是盲目跟风模型的企业，而是能够重塑工作流、掌控全生命周期治理并对最终损益表负责的技术运营者。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) AI 时代的宏大叙事与实体企业财报的巨大落差在当前围绕人工智能的讨论中，存在着一个极其刺眼的悖论。前沿实验室、科技创造者、主流媒体和科技巨头的高管们，都在用“文明级别”的宏大语境畅谈 AI——它将改写历史、颠覆一切行业、重构人类文明，甚至让人类现存的社会形态不复存在。然而，只要离开硅谷几英里，走进任何一家传统银行、医院或其他非纯科技行业，要求查看他们上一季度的损益表（P&L），看到的却是截然相反的景象：大多数实体经济（Main Street）企业根本没有在财报的息税前利润（EBIT）上看到 AI 带来的任何实质性改变。普华永道（PwC）近期对 4,454 位全球 CEO 进行的一项大规模调研印证了这一点。当被问及企业是否通过 AI 实现了成本削减或收入提升时，高达 56% 的 CEO 回答完全为零（连 1% 或 2% 的提升都没有）；仅有 12% 的受访者表示看到了一些改善，但这部分改善在 EBIT 上的体现普遍低于 5%。对于精通财务的人而言，这实际上只相当于 5 个基点（basis points）的微弱波动。AI 显然不是无效的，人们在个人日常生活中都能体会到它的强大，但这种能力却始终无法转化为企业损益表上的真金白银。这种断层根源于技术创造者（Technology Creators）与技术运营者（Technology Operators）之间的鸿沟。前者专注于研发前沿大模型和全新工具，而后者则是传统实体企业的 CIO、CTO、CISO 和 COO。当下的技术运营者正被夹在不切实际的董事会预期（“我的 50% 成本削减和 200% 收入增长在哪里？”）与仅有几个基点的现实回报之间，承受着巨大的身心煎熬。实体企业未能提炼出 AI 价值的核心症结，主要集中在以下五个方面： - **粗糙懒惰的系统架构（Lazy Architecture）**：缺乏可持久复用的架构范式。 - **与 AI 特性脱节的治理体系（Legacy Governance）**：依旧依赖传统静态策略与审批流。 - **未随技术重塑的组织形态（Outdated Organization）**：缺乏匹配 AI 工作流的人才梯队。 - **工具泛滥引发的疲态（Tool Rush Fatigue）**：在盲目追逐新工具的过程中消耗了生产力。 - **技术扩散周期（Technology Diffusion）**：虽然技术渗透需要时间，但在归因模型中，纯粹的时间延迟在成因中占比不足 6% 到 10%，前四项才是主要症结。 --- ### [05:12](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=312s) 工具狂潮与“懒惰架构”的死循环自 2022 年底生成式 AI 爆发以来，传统 CIO 们面临着前所未有的知识断层。过去几十年间，他们习惯于确定性、声明式的编程环境，拥有成熟的技术直觉；但在 AI 浪潮下，每周都有层出不穷的新工具席卷而来。由于缺乏对底层技术的深度理解，许多企业领导者在应对焦虑时做出了最糟糕的反应——盲目堆砌和轮换工具。有位在大型实体企业任职的 CIO 朋友曾陷入过极具讽刺意味的追逐怪圈：一年半前，他申请了预算采购并推广 GitHub Copilot，对全公司研发团队完成了部署和培训。然而刚部署完毕，就有人嘲讽他“还在用 Copilot 简直是生活在旧石器时代”。于是，他悄悄申请了第二笔预算，耗费四五个月采购并部署了 Cursor；结果在培训刚完成 45 分钟、正准备向上汇报时，又有人告诉他“不用 Claude Code 就彻底落伍了”。团队被这种频繁折腾搞得疲惫不堪，而在三轮工具更迭后，企业根本没有沉淀下任何实际业务价值。无论是智能体开发（Agentic Software）还是大模型选型，类似的盲目折腾都在频繁上演。技术运营者必须承担起构建可持续、耐用架构范式的责任，而不是依赖当前企业中普遍存在的“懒惰架构”。目前，许多市值超过 250 亿美元的大型企业所搭建的 AI 系统简陋得令人发指： ```mermaid flowchart TD accTitle: 典型懒惰 AI 架构与耐用型架构对比 subgraph Lazy["典型懒惰架构 (脆弱且易受污染)"] direction TB L1["前沿大模型 (如 GPT-4 / Claude)"] --> L2["外围薄应用层"] L3["简易 RAG / 向量库"] -.-> L2 L4["MCP 服务 / 工具接口"] -.-> L2 L5["提示词防火墙"] -.-> L2 L6[("静态遗留数据库")] --> L4 end subgraph Durable["耐用型架构范式 (高韧性与低成本)"] direction TB D1["业务工作流 Prompt"] --> D2{"智能路由层 (如 OpenRouter)"} D2 -- "90%+ 任务" --> D3["小型领域模型 (SLMs)"] D2 -- "特定高认知任务" --> D4["前沿大模型 (Frontier Models)"] D5[("实时业务 OLTP")] <--> D6["企业数据织锦 (Data Fabric / 如 Couchbase)"] D7[("分析型数据仓库")] <--> D6 D6 ==>|消除数据漂移| D3 D6 ==>|消除数据漂移| D4 end ``` 这种在前沿模型外部生硬拼凑 RAG、提示词防火墙和 MCP 协议的简陋系统，既无法承载复杂的企业业务，也让企业在日新月异的技术迭代面前无所适从。 --- ### [08:41](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=521s) 耐用架构实践：模型组合路由与企业数据织锦优秀的 CIO 和 CTO 应当在企业内部落地经得起考验的“耐用型模式”（Durable Patterns）。以模型选型为例，企业无需在“选用哪个终极模型”上浪费时间，而应建立“模型组合策略”（Model Portfolio Strategy），根据工作流的复杂度将各类提示词打标签并分流： - **小模型（SLM）承载核心高频流量**：在一项针对零售银行业务的研究中，银行日常处理的 594 种细分业务流中，除 177 种极其复杂的长尾逻辑外，其余绝大多数场景均可由微调的小型语言模型（SLM）以与前沿模型相当的准确率完成。 - **遏制 Token 成本爆炸**：盲目调用昂贵的前沿模型必然导致 Token 成本失控，引入智能路由（例如开源的 `openrouter.ai` 等多模型路由机制）能直接使企业 Token 开支削减 40% 到 50%。另一个致命缺陷是缺乏**企业数据织锦（Data Fabric）**。在传统系统架构中，分析数据、操作数据与联机交易（OLTP）数据长期彼此割裂；许多企业在接入 AI 时，仅仅搭建了一套简易的 RAG 或图数据库便以为万事大吉。然而，各系统间缺乏实时同步会导致数据自上线第一天起就开始产生“漂移”（Drift）。企业对于模型“幻觉”的过度担忧往往掩盖了更严重的隐患——在实际的企业生产环境中，因数据漂移导致的**“脏上下文”（Poisoned Context）所引发的合规事故与错误决策概率，是模型自身幻觉的 6 倍（高出 600%）**。前沿模型本身表现优异，但如果喂入的业务上下文已经过时或失效，输出的结果必然不可接受。通过引入如 Couchbase 这类能够打通分析型与非分析型数据、实现极低延迟抗漂移的成熟基础设施，企业才能为 AI 工作流筑牢数据根基。 --- ### [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) 代码化治理与沙漏型组织重构除架构之外，AI 时代的**治理体系（Governance）**同样需要彻底重写。过去那种依靠纸质审批、合规策略委员会长达数月层层签字的传统治理模式已经无法应对智能体系统。现代 AI 治理必须直接转化为代码和运行时监控（Governance as Code）。随着自主智能体（Agent）逐步承担关键业务操作，企业面临着一个全新的命题：“既然企业为人类员工设立了 HR 部门，那么智能体员工的 HR 部门在哪里？”利用如 Autopilots、Trustwise 等专为智能体生命周期设计的代码化治理工具，才能在运行时对 AI 行为进行边界约束与动态审计。与此同时，企业的**组织架构（Organizational Design）**必须经历一场彻底的形态重塑。传统的企业组织呈金字塔型，部分观点甚至误认为 AI 时代应转向“高管居多、几乎没有底层执行人员”的倒金字塔型： | 组织形态 | 结构特征 | AI 时代的适应性评估 | |:--- |:--- |:--- | | **传统金字塔** (Pyramid) | 顶层高管极少，中层管理者较多，底层包含大量基础业务执行人员。 | 随着低门槛认知工作被 AI 承接，中间与底层的人力冗余将成为成本重担。 | | **倒金字塔** (Inverted Pyramid) | 极少数核心资深专家主导，底层彻底取消，所有执行完全外包给前沿 AI。 | **巨大失误**。由于缺少底层梯队培养，将导致企业未来领袖断层并失去自主学习机制。 | | **沙漏型** (Hourglass) | **顶部扩张**：资深专家通过智能体获得极大杠杆；<br>**腰部收窄**：中层管理与常规流转被高度自动化压缩；<br>**底部宽阔**：大量吸纳应届生、实习生与年轻人才。 | **推荐范式**。在确保降本增效的同时，培育下一代模型评测、强化学习及数据治理人才。 | 重塑组织架构的核心在于防止**“企业智能外泄”（Leakage of Intelligence）**。在目前的商业合作中，企业普遍要求外部前沿大模型不得使用企业私有数据进行模型训练。这导致了一个荒谬的闭环： - 企业将高价值认知工作全部外包给第三方前沿模型； - 模型不会记住任何错误，企业内部也没有建立专属的强化学习（RL）闭环； - 系统在反复运行中无法沉淀任何组织智慧，组织陷入长期的“认知停滞状态”（Stasis）。人类新员工在犯错和纠错中会在三年内成长为业务骨干，但完全依赖外部模型且不进行企业内部强化学习的组织，其智力水平将永远停留在第一天。这无异于为了短期的单点自动化而抵押了企业的长期未来。 --- ### [17:11](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=1031s) 结语：重拾技术运营者的专业底蕴技术运营者必须扛起深度理解并掌管全生命周期 AI 管理的职责，停止将企业核心认知能力盲目外包给前沿模型厂商。分享者回顾了自己长达三十年、横跨花旗银行（Citi）、摩根大通（JPMorgan Chase）和威瑞森（Verizon）等财富 50 强企业的 CIO 与 CTO 生涯。退休后，他创办了非营利组织 CIO Fellows（`ciofellows.org`），致力于向技术运营者传承 CIO 的工程专业匠心；同时作为 Vibrant Capital（`vibrantcapital.ai`）的 CEO，他也带领团队帮助技术创造者穿透上千种工具泡沫的杂音，打造出真正能为实体企业带来业务实效、扎根于真实工作流中的耐用解决方案。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) 大多数实体经济（Main Street）企业根本没有在财报的息税前利润（EBIT）上看到 AI 带来的任何实质性改变。 — Shadman Zafar > [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) 当被问及企业是否通过 AI 实现了成本削减或收入提升时，高达 56% 的 CEO 回答完全为零（连 1% 或 2% 的提升都没有） — Shadman Zafar > [00:04](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=4s) 我的 50% 成本削减和 200% 收入增长在哪里？ — 董事会 > [05:12](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=312s) 团队被这种频繁折腾搞得疲惫不堪，而在三轮工具更迭后，企业根本没有沉淀下任何实际业务价值。 — Shadman Zafar > [08:41](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=521s) 盲目调用昂贵的前沿模型必然导致 Token 成本失控，引入智能路由（例如开源的 openrouter.ai 等多模型路由机制）能直接使企业 Token 开支削减 40% 到 50%。 — Shadman Zafar > [08:41](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=521s) 因数据漂移导致的“脏上下文”（Poisoned Context）所引发的合规事故与错误决策概率，是模型自身幻觉的 6 倍（高出 600%）。 — Shadman Zafar > [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) 既然企业为人类员工设立了 HR 部门，那么智能体员工的 HR 部门在哪里？ — Shadman Zafar > [13:28](https://www.youtube.com/watch?v=L6AoBsRRl7s&t=808s) 人类新员工在犯错和纠错中会在三年内成长为业务骨干，但完全依赖外部模型且不进行企业内部强化学习的组织，其智力水平将永远停留在第一天。这无异于为了短期的单点自动化而抵押了企业的长期未来。 — Shadman Zafar ## Mentioned Shadman Zafar, Vibrant Capital, CIO Fellows, 普华永道 (PwC), GitHub Copilot, Cursor, Claude Code, OpenRouter, Couchbase, Autopilots, Trustwise, 花旗银行 (Citi), 摩根大通 (JPMorgan Chase), 威瑞森 (Verizon), GPT-4, Claude ## Video description Shadman Zafar — CEO, Vibrant Capital Most companies think they're scaling AI. What they're actually scaling is experiments. In this session, I will argue that Main Street industries, from banking and insurance to healthcare, manufacturing, logistics, retail, and communications, have seen enough demos. What they're missing is AI built into the everyday workflows where cycle time, cost, risk, and revenue actually move. Drawing on two decades as a CIO and what we are building at Vibrant Capital through our Growth Partners Network and CIO Fellows Society, I will share a more operator-led playbook for moving from pilots and hype to repeatable deployment. My view is a bit contrarian: the winners will not be the companies with the flashiest models, but the ones that redesign work around AI, govern it well, and make it accountable to business results. If you are a tech operator trying to make AI stick inside a real enterprise, you will leave with a sharper lens, a few uncomfortable truths, and a practical way to act. [Source](https://www.youtube.com/watch?v=L6AoBsRRl7s)

Source quotes

[0:04] Shadman Zafar: 大多数实体经济（Main Street）企业根本没有在财报的息税前利润（EBIT）上看到 AI 带来的任何实质性改变。 [0:04] Shadman Zafar: 当被问及企业是否通过 AI 实现了成本削减或收入提升时，高达 56% 的 CEO 回答完全为零（连 1% 或 2% 的提升都没有） [0:04] 董事会: 我的 50% 成本削减和 200% 收入增长在哪里？ [5:12] Shadman Zafar: 团队被这种频繁折腾搞得疲惫不堪，而在三轮工具更迭后，企业根本没有沉淀下任何实际业务价值。 [8:41] Shadman Zafar: 盲目调用昂贵的前沿模型必然导致 Token 成本失控，引入智能路由（例如开源的 openrouter.ai 等多模型路由机制）能直接使企业 Token 开支削减 40% 到 50%。 [8:41] Shadman Zafar: 因数据漂移导致的“脏上下文”（Poisoned Context）所引发的合规事故与错误决策概率，是模型自身幻觉的 6 倍（高出 600%）。 [13:28] Shadman Zafar: 既然企业为人类员工设立了 HR 部门，那么智能体员工的 HR 部门在哪里？ [13:28] Shadman Zafar: 人类新员工在犯错和纠错中会在三年内成长为业务骨干，但完全依赖外部模型且不进行企业内部强化学习的组织，其智力水平将永远停留在第一天。这无异于为了短期的单点自动化而抵押了企业的长期未来。

Source quotes

[0:00] Shadman Zafar: What they're actually scaling is experiments. [0:04] Shadman Zafar: On one side, frontier lab creators, tech CEOs, CFOs, media gurus, and visionaries speak of AI in dramatic, 'civilizational terms.' [0:04] Shadman Zafar: Most non-tech companies are simply not seeing any meaningful EBIT impact from AI yet. [0:04] Shadman Zafar: 56% of CEOs reported seeing zero impact—not even 1% or 2%. [2:23] Shadman Zafar: The real bottlenecks lie in how technology operators design, govern, and execute. [5:02] Shadman Zafar: He was embarking on his third tool migration without his engineering team ever deriving sustained business value from any of them. [8:41] Shadman Zafar: Token cost explosions are usually not an inherent AI problem, but rather a direct symptom of lazy architecture. [8:41] Shadman Zafar: In an enterprise setting, poisoned or outdated context is 6 times (600%) more likely to produce a dangerous, out-of-policy response than a model hallucination. [13:28] Shadman Zafar: If agents are becoming a huge chunk of your workforce, where is the HR department for your agents? [13:28] Shadman Zafar: Eliminating entry-level roles 'mortgages the future' of the company and triggers severe 'intelligence leakage.'

Original transcript

[0:04] So today we're going to talk about uh what I consider one of the biggest paradox happening [0:10] in AI. And that is is that most of the frontier labs, creators, CEOs, CFOs, media gurus, they're [0:19] talking in civilizational terms about AI, how AI is going to rewrite the history, rewrite the every [0:25] industry, civilization will change. Maybe humanity as we know it may not exist. However, if you walk [0:34] few miles out of Silicon Valley and walk into a bank or a hospital or any other industry other [0:42] than tech and ask them for their P&L report for the last quarter, it is an exact opposite story. [0:53] Most of the main street companies are not seeing a meaningful EBIT impact of AI yet [1:00] anywhere. And I and I kid you not is like PWC quite recently did a did a study of 4,454 [1:10] CEOs. That's not a small number. And majority of and they asked them you know if they if you have [1:16] seen AI based cost reduction or revenue upload and the majority of the answer was zero 56% of [1:26] the CO said nothing like 2% 1% nothing 12% say it saw some improvement and the some improvement [1:35] was roughly within le less than 5% of the EBIT and less than 5% of the EBIT just I know they're [1:42] all finan financially savvy people in this room effectively doesn't mean 5% of the company it [1:47] means five basis points typically so while we are having civilizational change on one hand [1:54] we are talking on the P&L and by the way this is like for this quarter if you go and look at [1:58] the main street and you get their actual reports from um from their filings it's really that it's [2:05] in few basis points and that's that doesn't mean like AI doesn't work it's obviously works I mean [2:10] uh it it works we see it in our personal life but somehow it's not getting captured in the P&L and [2:17] uh we feel there is a there is a real uh reason for that and it really comes down [2:23] to the operators of technology. So I separate technology into technology creators, the people [2:29] who are sort of creating new technology, new tools, new capabilities, new frontier models [2:34] and everything else and the technology operators, the CIOS, the CTO's, the CISOs, the cos of main [2:43] street companies. And in my opinion, those CIOS, those technology operators are in the midst of [2:50] this paradox. incredibly unbelievable expectations and few basis points in the P&L statement and the [2:58] CI and the CEOs and CFOs are like where is my 50% cost reduction or 200% revenue uplift and [3:07] I do believe while I feel really sad and bad for them how many how many technology operators in the [3:12] room or how many versus how many techn operator okay very few technology creators few technology [3:18] creators people who like work for a company that develop technology ology that yourself. Okay. So, [3:24] got it. Okay. So, we we got both sides. So, um here's what I would say. We going to talk about [3:31] it from a technology operator's perspective of why it has not been working. And uh it really [3:38] comes down to in my view five key elements. It's an element around lazy architecture. Number one. [3:49] Number two, um, a governance that was not designed for AI. Number three, [3:58] the organization itself. Number four, tool rush fatigue. That is a real thing that is happening [4:08] um, and it's painful and it's actually slowing down uh, the industry. And also the fourth which I [4:14] think so most of the people use to actually write off that why you haven't seen the impact which [4:18] is like technology diffusion takes time I mean honestly speaking I'm not going to talk about that [4:23] uh but I do believe that that is less than 10% of the cause it's less than 6% of the cause in some [4:29] of our studies we have seen diffusion takes time but it has taken enough time it's really the first [4:33] four that actually causes the issue and impact okay so what those are and then I'll tell you [4:38] a little bit about what people can do to actually actually accelerate that because I do believe the [4:45] companies that will actually implement and extract value quickly will really survive and the ones who [4:51] won't will go it's like a borders versus Barnes & Noble time uh where you'll see some that will sort [4:57] of adapt technology in a deep basis and and they will carry on and thrive as well as few that may [5:02] you know be has spins so um I'll say the biggest issue that happened as the as November 22nd moment [5:12] happened. Um the CIOS have been trained over the years in in a deterministic declarative [5:22] programming environment. They knew they have built certain kind of instincts and the rest [5:27] of the technology came with a big onslaught with tools and capabilities changing on a weekly basis. [5:34] First problem that that people actually approached was that I have so much technology coming my way. [5:40] How do I handle it? And they handled it not the way I think so a technology operator typically [5:45] should handle without understanding the deep technology. They start implementing tools. I'll [5:50] give you an example of a a friend CI so I coach um I should say I I coach a lot of CIOS of main [5:57] street companies. uh today I was talking to one of the CIOS and he comes and you know pulling [6:02] his hair out and he's like shan you know uh a year and a half ago I got some budget to sort of deploy [6:09] uh software development AI tools and I went and I did a big deployment and um trained all of my [6:17] leaders and all of my software developers and he deployed GitHub Copilot and by the time he [6:24] deployed somebody came to him and says GitHub Copilot you're stupid. If you're doing that, [6:30] you are living in the past. And he's like without going and giving my update that I have [6:34] successfully did this deployment. I quietly went and asked for more money and I start deploying [6:38] cursor and I did cursor deployment for next four five months or so. I completed it. I trained [6:43] everybody. My people were getting frustrated that before they start to utilize the tool, [6:48] I put something new in and 45 minutes after I finish deployment all of that and training and [6:53] I was going to go and uh update my leaders and they says cursor you are just stupid. If you have [7:00] not used cloud code yet, you're you know living in the past and now I'm doing my third round without [7:05] anybody actually getting any uh value out of it. This is a story that's being repeated for multiple [7:12] things for how you do agentic software, how do you do model selection, how do you do anything [7:18] else and the problem yes there is a lot of tool rush and there's a crazy amount of technology [7:26] tools that are coming at the operator's level but there's a responsibility that sits with the [7:30] operator and that is to actually implement uh sustainable and durable architectural patterns [7:39] an example example I would give you and this is what I mean by lazy architecture right so most of [7:43] I'm sure many of you will recognize this many of the enterprises that I talk to I usually talk to [7:49] enterprises that are 25 billion in market cap and above so maybe slightly different in a market down [7:57] market but in the up market this architecture very similar you have put like a thin layer of like [8:03] apps on top of a frontier model you have created some kind of like a rag infrastructure on the side [8:08] you have had like some kind of prompt firewall and whatnot in place as well and more recently [8:14] we have set up like an MCP server with some tools behind it so you can access legacy databases more [8:20] or less I'm oversimplifying but honestly not by a lot that that's what's happening that's how [8:26] the architecture has been set up and I feel that is a big mistake the the the technology operators [8:33] are doing that they're doing a disservice to their enterprise and they themselves doing a disservice [8:36] to themselves because they can't keep up with the change that's happening on a daily basis. [8:41] The kind of durable patterns I tell them to implement are patterns like so so when the [8:47] when the model selection was a big issue I says don't waste time in figuring out the model make [8:52] sure you put like a m model portfolio strategy in place you categorize every single prompt and [9:00] you categorize those prompts by workflow and you know what level of intelligence you need [9:05] you'll send certain things to SLMs you'll send certain things to regular sized models you send [9:09] some things to Frontier model depend and most of your high traffic will actually come into the [9:14] SLMs. Actually, in in a banking study, 594 things that a consumer bank does, every single one of the [9:27] workflows with an exception of 177 out of 594 were executable at the same level of capability as a [9:36] frontier model by SLMs. That is like an incredible number. And most of those are the hightraic [9:43] environment. People are talking about token token economics explosion happening but that's because [9:48] of lazy architecture. You didn't implement rest of the stuff happening on the small language [9:53] models the right way and have a router in place. So people implemented that that became a dur that [9:58] is a durable pattern. A CIO worth his salt or a CTO worth is salt will actually implement that. [10:04] Same thing goes for developers. Same thing goes for governance. Governance is a big topic right [10:08] now as well. So we say build build build these frameworks or durable patterns. Um and and by the [10:16] way talk to other operators. I got to tell you one thing that there are these incredible technology [10:21] creators and and and no disrespect to anybody who's in the room who is that they they are sharp [10:27] as hell in trying to figure out how to create new capabilities with AI. However, they have not run [10:32] large operational main street companies. They don't know what are the problems you run into. [10:39] That's where a CIO and CTO comes into play in their judgment of how to utilize the technology [10:45] with durable patterns that don't change very quickly. Um I I'll give two more examples of the [10:52] type of patterns that are important. One I call data fabric. Any anybody who has not implemented a [10:57] data fabric for the enterprise and the enterprise is sizable. They are they are missing the ball. [11:02] uh there used to be a time where it was okay to have your analytical data separate from your [11:07] operational data separate even from your OOLTP and then most of the time people came in and put like [11:13] a fourth leg of rag or some kind of graph database where you put everything in there and say that [11:18] that would work because now I can sort of take that data and put it into AI doesn't the reason [11:25] it doesn't is because your data starts to drift almost by the day you put that in place so you [11:31] need to have some kind of a data fabric which uh which can keep things live between your analytical [11:37] versus operational data. So your your context is not um poisoned. And now another interesting [11:44] statistic people worry about about hallucinations. Your poisoned context in an enterprise is 6x 600% [11:55] more likely to give a bad unacceptable out of policy answer than a hallucination. I honestly [12:02] don't even care about hallucination that they're that low. I probably shouldn't. Don't rec quote [12:06] me. I don't care. I care. But it's like it's really low compared to context that is poisoned [12:12] because of a lazy architecture. and and and there are tools out there same technology creators that [12:18] are creating these cool demos. They are they are others that are creating these durable patterns [12:22] for CIOS who have learn taken time to learn about AI and core capability. There's a even open source [12:30] there is something called open router.ai. How many of you have heard of open router? Few. Okay, [12:35] great. Right. I mean you don't need to even go buy like a massive new capability. It's it's a highly [12:40] scalable model that you can uh router that can actually go across multiple models and you don't [12:45] have to worry about your your token cost goes down roughly anybody who implements that by 40 [12:50] 50% just by having that implementation in place or or commercial companies like there's a company [12:56] called Couchbase. Anybody have heard of Couchbase? Um it it's a fantastic company that actually [13:03] provides if you need extremely high resilience and high performance tracking undrifted data [13:11] of your both your analytical data non-analytical data you implement couch as a as a tool by itself [13:18] and you will have that in play. You have now undrifted data that you can provide to your AI. [13:22] So there it's not like cutting edge a lot of these durable capabilities are actually in place today. [13:28] The second thing uh other than the architecture is governance. So in governance the same question [13:34] is happening. CIOS talk about it's taking a long time to actually governance to work. But [13:40] um if you implement governance the right way which is forget about the days when you have governance [13:46] done through policy and approvals and this and that and then you go through execution. Governance [13:51] now is code. You have to utilize tools like autopill or there's a company called trustwise [13:58] which which basically asks the question that if agents are big chunk of your workforce where is [14:05] the HR department for your agents you have them for humans you don't have them for agent that's [14:11] where your governance needs to come play again there's the standard capabilities and tools that [14:15] are available to uh implement that uh and the last part comes in is the organization itself [14:22] I think most of the folks have not taken the hard action of redesigning the organization. [14:29] Organizations used to be a pyramid. Senior leaders at the top more middle management [14:34] and a lot of people at the bottom. Uh some people have been talking about that the new model is like [14:40] inverted pyramid. You know some senior people less number overall but more senior people who [14:46] can really automate and rest of the stuff is being done by the agents. You don't need a lot of people [14:50] at the bottom. I think that's a huge mistake. The the emerging model for organizations is actually [14:56] an hourglass. You have senior people at the top who can actually multiply themselves significantly [15:02] through agent. middle actually goes really thin because most of the work can get executed by [15:07] the agents and the bottom you got to be hiring actively significant amount of new hires interns [15:14] college students graduates and whatnot because otherwise you're mortgaging your future and they [15:19] need to get trained on the next generation of capabilities model evals uh reinforcement [15:25] learning data custodianship uh management of the data for verification versus non-verification life [15:31] cycle management of AI that is an organization ational structure that has to get redesigned. [15:35] that provides benefits but it also provides an ability for the organization to learn as you move [15:40] forward because if the organization does not have that then even if you're automating something you [15:46] are actively making your organization a stupider or or leakage of intelligence as I like to call it [15:54] because if you have by contract we have told most of the frontier models you cannot use my data to [16:01] train your model right I mean everybody familiar with that right and and and the models actually [16:06] do that uh dutifully. So, but by doing that and using them for every single cognitive processing, [16:15] no reinforcement learning yourself in the organization, what have you done? They [16:19] are throwing away any learnings they could do. You are not using any of the learnings in your [16:25] own enterprise loop of execution, finding the errors, using those as a reinforcement learning [16:31] to fine-tune the organization better. That happens today because each time human makes [16:36] a small mistake they get corrected. They get smarter. The person you hire first day is not [16:41] at the same level three years later as they were on the first day. But if you're using only the [16:46] frontier models, no reinforcement learning inside the company, then you are effectively [16:51] making your organization in stasis forever. If you succeed, by the way, so your success, [16:58] your best case scenario is a horrible answer. your worst case scenario is obviously that [17:04] doesn't happen plus you also don't automate. So uh important to have that implementation and that I [17:11] will leave with the last answer is most important for the uh operators to honestly take on the [17:17] responsibility of deeply learning what it means to have an AI life cycle management in an enterprise. [17:26] It's easy to outsource that cognitive capability to frontier models and you will mortgage your [17:31] future and do the same thing for the uh for the organization organizational design as well. Just [17:38] as a uh at the end kind of an introduction for myself uh I was a former Fortune50 CIO and CTO [17:46] for 30 years or so between like City Bank, JP Morgan, Chase, Verizon, whatnot. I am retired. [17:53] I have a nonprofit organization uh primarily designed to give back to the society on the [18:01] craftsmanship of being a CIO. Uh those of you who are or are no good CIOS who would like to [18:08] join again as a nonprofit to contribute to the community, we have something called CIO [18:13] fellows. It's a fellowship ciph.org. Uh you can go join there if you're a technology creator. I [18:20] also guide and advise the companies of how to become meaningful and matter for the operators [18:28] rather than throwing yet another tool out into the enterprise. We have the company as says I'm a CEO [18:34] of a vibrant capital. So we run that actually as a profit if you are interested to join a vibrant [18:39] network to tell you of how to not be sifted from the noise and the AI the thousands of tools that [18:46] are being created for the ones that actually matter in the main street. So um with that I am [18:53] at the end. Thank you very much. Thanks for your attention and if anybody wants to reach out once [18:56] again cofellows.org or vibrant capitalap.ai uh on either side. Thank you very much.