S&P AI Chief Warns Point Solutions Fail Without Workflow Rewiring

Enterprise AI Transformation: Turning Strategy into Business Value With S&P Global — Ai4

Ai4

Enterprise AI investments stall because bolt-on tools cannot automate the 80% of operational processes locked in unwritten tribal knowledge without federated agent governance.

AI-assisted summary of Enterprise AI Transformation: Turning Strategy into Business Value With S&P Global — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by Ai4 on YouTube on August 19, 2026, this keynote features Ashit Talukder, Global Head of AI and Automation at S&P Global. According to the video description, the session explores how organizations can transition from fragmented experimentation to scaled enterprise capabilities, aligning AI vision, governance, operating models, and agent architectures with measurable business value.

## TL;DR Ashit Talukder diagnoses why enterprise AI initiatives frequently stall, cautioning leaders against the dual traps of ungoverned shadow 'vibe coding' and isolated bolt-on point solutions. Drawing lessons from industrial electrification, he details a four-pillar operating model across Data, Platforms, People, and digitized Processes required to structurally rewire organizations for AI. The talk outlines practical architectures for long-running autonomous agents—incorporating governance control planes, identity management, and disciplined tokenomics—before examining how process digital twins, embodied robotics, and edge AI will fundamentally reshape corporate structures and emerging industries.

## Atmosphere & style Talukder delivers a grounded, authoritative, and analytical presentation that blends deep scientific research pedigree with hard-won commercial realism. The mood is refreshingly candid and clear-eyed, directly confronting internal corporate friction, employee skepticism, and the hidden risks of unmanaged prototype deployments. Watching the session feels like an executive strategy briefing, methodically stripping away market hype to focus on architectural discipline, engineering economics, and the tangible realities of enterprise transformation.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=1htxOFU_bfI&t=4s) Artificial intelligence is a powerful transformative tool rather than an end in itself; organizations must move beyond research and experimentation to produce measurable business outcomes. - [02:37](https://www.youtube.com/watch?v=1htxOFU_bfI&t=157s) Enterprise leadership urgently demands rapid AI deployment, but the workforce is fragmented into three conflicting camps: indifferent employees, enthusiastic vibe coders, and fearful skeptics or saboteurs. - [04:13](https://www.youtube.com/watch?v=1htxOFU_bfI&t=253s) Ungoverned vibe coding represents shadow IT with a new hat, introducing severe operational risks, lack of identity management, brittle deployments, and potential rogue agent threats. - [04:13](https://www.youtube.com/watch?v=1htxOFU_bfI&t=253s) Single-point bolt-on AI solutions optimize narrow tasks while moving the operational bottleneck upstream or downstream, failing to deliver end-to-end enterprise value. - [07:45](https://www.youtube.com/watch?v=1htxOFU_bfI&t=465s) Similar to how electric motors failed to boost industrial productivity until factories rewired their physical layouts, enterprise AI requires a fundamental redesign of the operating model. - [07:45](https://www.youtube.com/watch?v=1htxOFU_bfI&t=465s) A scalable enterprise AI operating model rests on four pillars: semantically searchable governed data, federated secure platforms, workforce AI fluency, and the standardization of tribal processes. - [07:45](https://www.youtube.com/watch?v=1htxOFU_bfI&t=465s) Up to 80% of enterprise processes exist as opaque, unwritten tribal knowledge that must be discovered, standardized, and digitized before AI can automate them. - [11:05](https://www.youtube.com/watch?v=1htxOFU_bfI&t=665s) Organizations must advance beyond single-turn reactive chatbots to autonomous agents equipped with execution harnesses, digital identities, auditability, and context memory. - [11:05](https://www.youtube.com/watch?v=1htxOFU_bfI&t=665s) Agentic production architectures require strict governance, digital identities, role-based access control, and confidence thresholds that trigger human-in-the-loop escalation. - [14:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=884s) Fiscal discipline and escape from 'pilot hell' require ruthlessly prioritizing high-impact business KPIs and implementing intelligent model routing to handle 80% of routine workloads on compact models. - [18:10](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1090s) Building process digital twins over the next 6 to 18 months enables digital simulation and bottleneck diagnosis, drastically reducing integration friction and change management costs. - [20:05](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1205s) Over the 2- to 3-year horizon, the convergence of edge AI chips, embodied robotics, and autonomous agent ecosystems will enable billion-dollar solopreneurs and flatten corporate enterprise structures.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=1htxOFU_bfI&t=4s) From Rocket Science to Enterprise AI: Setting the Stage

Ashit Talukder opens with an introduction to his unique background, noting with a smile that having started his career doing machine learning research at NASA JPL, he is "one of the few rocket scientists at S&P." Back then, artificial intelligence was strictly a research topic, and the federal government was virtually the only entity with pockets deep enough to fund machine learning investigations. As deep learning matured over the years, he transitioned to NIST (National Institute of Standards and Technology) to launch their data science program initiatives, before moving into the commercial sector to lead AI strategy and product functions at Moody's Analytics and now S&P Global Ratings.

Drawing from this span of foundational research and enterprise-scale commercial execution, Talukder sets the agenda for the talk: breaking down the recurring hurdles leaders face when driving AI transformations and illustrating how organizations can move past experimentation to produce measurable business value. He emphasizes that AI is fundamentally a powerful tool, not an end in itself, and must be harnessed to generate durable business outcomes.

To structure the discussion, Talukder outlines a three-part journey: - **Learning from the Past**: Examining prior industrial and technological revolutions to understand why structural transformation is essential. - **Navigating the Present**: Addressing the operational realities, governance risks, and operating model redesigns required to scale AI across modern enterprise workflows. - **Anticipating the Future**: Preparing enterprise systems and agent architectures to adapt to rapidly evolving technologies without becoming obsolete.

---

[02:37](https://www.youtube.com/watch?v=1htxOFU_bfI&t=157s) The 2026 AI Landscape: Executive Demands vs. A Divided Workforce

Reflecting on the historical perception of AI, Talukder notes that from its inception at Stanford in the 1950s all the way through 2022, organizational attitudes barely shifted beyond a niche group of ML researchers. The public debut of ChatGPT in late 2022 disrupted that equilibrium. Fast-forwarding to 2026, the pendulum has swung completely: enterprise leadership now urgently wants "AI solutions yesterday," creating intense organizational pressure.

However, the workforce on the ground is deeply fragmented into three distinct camps: - **The Indifferent**: Employees who remain disengaged and do not actively incorporate AI into their daily work. - **The Vibe Coders**: Self-taught practitioners who enthusiastically leverage generative coding tools to spin up their own AI prototypes. - **The Skeptics and Saboteurs**: Workers who fear AI will eliminate their jobs, leading to real instances where staff actively push back against or sabotage workplace AI initiatives.

---

[04:13](https://www.youtube.com/watch?v=1htxOFU_bfI&t=253s) The Dual Trap: Shadow AI "Vibe Coding" and Single-Point Bolt-On Solutions

In the rush to deliver value, enterprise AI efforts typically fall into two problematic patterns. The first is active "vibe coding"—individual employees building ad-hoc solutions with generative coding tools. Talukder characterizes this as essentially "shadow IT with a new hat." While these prototypes look shiny and quickly impress executives who immediately ask to deploy them into production, they carry severe operational risks. They lack identity management, audit trails, governance, and compatibility with core enterprise technology stacks. When rolled out beyond pilot datasets, they become brittle and fail, leaving dedicated AI teams to inherit the maintenance burden. In an agentic environment, an ungoverned vibe-coded agent could easily go rogue, penetrate internal networks, introduce cybersecurity vulnerabilities, or delete critical data, potentially taking down a multi-billion-dollar business.

```mermaid flowchart TD accTitle: Shadow AI vs Federated Platform Architecture

subgraph Ungoverned["The Shadow AI Risk (Vibe Coding)"] A["Vibe-Coded Prototype"] --> B["Bypasses Security & Identity"] B --> C["Brittle Deployment / Production Failures"] C --> D["Rogue Agent / Security & Data Breach"] end

subgraph Federated["The Governed Enterprise Architecture"] E["Governed AI Application"] --> F["Federated AI Platform"] F --> G["Central Identity, Observability & Guardrails"] G --> H["Stable, Resilient Business Value"] end ```

The second trap is the "bolt-on" single-point solution. AI teams identify low-hanging fruit in a complex workflow and attach an AI point feature (such as an isolated chatbot or summarizer). Talukder compares this to taking a sluggish two-lane highway and widening a single stretch into a five-lane highway: while traffic accelerates through that one segment, the broader process immediately bottlenecks before and after it. Point solutions fail to move the needle on overall business outcomes, forcing AI leaders into uncomfortable conversations with CTOs and CFOs about why large investments have not produced measurable enterprise returns.

- **Mitigating Shadow AI**: Enforce a mandatory federated AI platform equipped with centralized identity management, role-based access control, policy governance, and real-time observability across all business units. - **Moving Beyond Bolt-Ons**: Stop optimizing isolated tasks and instead evaluate and re-engineer end-to-end workflow topologies.

---

[07:45](https://www.youtube.com/watch?v=1htxOFU_bfI&t=465s) Rewiring the Enterprise: Lessons from the Electric Motor

To explain why bolt-on tools fall short, Talukder draws a historical analogy to the industrial transition from steam power to electricity. When factories initially adopted electricity, they merely replaced central steam engines with electric motors while keeping the old mechanical drive shafts and floor layouts intact. As a result, industrial productivity remained flat for nearly 40 years. It was only in the 1920s—when factory floors were completely redesigned and rewired to distribute power directly to individual workstations—that the full productivity explosion of electric power was realized. Enterprise AI faces an identical inflection point: organizations have deployed hundreds of disparate chatbots and co-pilot tools, but until they fundamentally rewire their operating models, substantial gains will remain out of reach.

```mermaid flowchart LR accTitle: Four Pillars of Enterprise AI Operating Model

subgraph OperatingModel["The Enterprise Operating Model"] direction TB P1["1. Data\n(Semantic search, lineage, AI-ready)"] P2["2. Platforms\n(Federated governance, identity, security)"] P3["3. People\n(Fluency, upskilling, champion networks)"] P4["4. Processes\n(Digitizing opaque tribal workflows)"] end

OperatingModel --> Val["Scalable Business Transformation"] ```

Talukder outlines the four essential pillars required to rewire an enterprise operating model: - **Data**: Must serve as clean fuel for the AI engine—semantically searchable, discoverable, and governed with end-to-end lineage so models have accurate business context. - **Platforms**: Robust, federated infrastructure with unified security, identity protocols, and system observability. - **People**: Active workforce upskilling and building "AI fluency" through champion networks that evangelize practical adoption across business lines. - **Processes**: The most critical and challenging pillar. In typical enterprises, only 20% to 30% of business processes are standardized, digitized, and observable. The remaining 70% to 80% exist as opaque, unwritten tribal knowledge that AI cannot access or automate. Rewiring requires exposing, standardizing, and digitizing these hidden workflows.

---

[11:05](https://www.youtube.com/watch?v=1htxOFU_bfI&t=665s) Transitioning to Actionable Agents: Harnesses, Control Planes, and Context

Enterprises must transition away from single-turn, reactive Q&A chatbots and move toward autonomous agents capable of executing complex, multi-step tasks end-to-end. A well-constructed agent does not simply rely on a base LLM; it actively plans, retrieves necessary information, manages execution context, and completes actions 24/7. This transition automates roughly 80% of repetitive manual effort, freeing human employees to focus on high-value creative problem-solving and client interactions.

```mermaid flowchart TD accTitle: Agentic AI Production Stack

subgraph Execution["Agent Core & Tooling"] A["LLM Base Engine"] B["Harness & Prompt Instructions"] C["MCP Tools / Enterprise Data Connectors"] A --- B B --- C end

subgraph Governance["Governance & Control Plane"] D["Identity, RBAC & Permissions"] E["Auditability & Observability Layer"] end

subgraph Runtime["Long-Running Execution"] F["Context Management & Memory"] G["Confidence Thresholds & Human-in-the-Loop Escalation"] end

Execution --> Governance Governance --> Runtime Runtime --> Output["Durable Enterprise Value"] ```

Deploying production-grade agents requires a disciplined architectural stack: - **The Harness and Tooling Layer**: Context frameworks, precise instruction sets, and Model Context Protocol (MCP) integrations that allow agents to interact securely with internal databases and APIs. - **The Control and Governance Plane**: Agents must be provisioned with dedicated digital identities, distinct role-based permissions, and strict audit logs, mirroring the onboarding and boundaries assigned to junior human analysts. - **Long-Running Execution & Context Management**: Architectures that maintain state over hours or days, equipped with confidence-scoring mechanisms that proactively pause and alert human supervisors when ambiguity arises. - **Vendor Management vs. In-House Builds**: Avoiding tool sprawl across proliferating commercial vendors by strategically deciding where to build custom in-house agent tooling versus leveraging standardized platforms.

---

[14:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=884s) Financial Discipline: Managing Tokenomics, Prioritization, and the AI Flywheel

As autonomous agents proliferate, CFOs inevitably raise concerns over runaway computational and inference costs. Talukder stresses that leaders cannot "let a thousand flowers bloom" unchecked; they must enforce ruthless prioritization by selecting the top two or three initiatives that tie directly to measurable business KPIs—such as cycle-time reduction, operational throughput, revenue growth, or client retention—and escape "pilot hell" where projects spend six months chasing unattainable perfection.

To build leadership alignment and maintain fiscal control, Talukder leverages two core concepts: - **The AI Flywheel**: AI systems improve as they consume operational telemetry and user interaction logs. Deploying an imperfect—yet strictly governed—system into production allows models to learn from human feedback loops, establishing a self-reinforcing flywheel of continuous improvement. - **Engineering Economics & Model Routing**: Avoid defaulting to expensive frontier models for basic queries. By implementing intelligent model routers and managing context window sizes, organizations can direct roughly 80% of routine workloads to compact, cost-effective models, reserving high-cost frontier models for the 20% of tasks requiring deep reasoning.

---

[17:04](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1024s) Looking Ahead: Digital Twins, Self-Learning Loops, and Computer Use

Concluding with a look toward emerging frontiers, Talukder warns that AI architectures must remain modular and future-ready rather than over-indexed on legacy patterns. Near-term breakthroughs will expand agent reasoning capabilities over much longer planning horizons.

Key emerging capabilities that enterprise leaders must prepare for include: - **Process Digital Twins**: Creating living digital models of end-to-end enterprise workflows, enabling agents to understand upstream and downstream dependencies across complex operational ecosystems. - **Self-Learning Agent Loops**: Autonomous systems that iteratively evaluate their own output, learn from errors, and integrate human corrections dynamically. - **Computer-Use Agents**: Agents capable of directly interacting with user interfaces and operating systems, which will demand new paradigms for governance, safety perimeters, and access controls.

[18:10](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1090s) Near-Term Innovations: Digital Twins and Process Simulation

In the immediate 6- to 18-month horizon, the emergence of self-evolving and self-learning AI solutions stands out as the next major transformation for enterprise workflows. Recalling the metaphor of an iceberg concealing hidden organizational processes, building an enterprise digital twin allows organizations to construct a digital replica of their internal operations. Instead of conducting tedious manual interviews with analysts and staff members to pinpoint inefficiencies, leaders can use digital models to automatically diagnose bottlenecks, simulate diverse operating conditions, and test new AI workflows before live operational rollout.

```mermaid flowchart LR accTitle: Enterprise Digital Twin Simulation Workflow subgraph Capture["1. Signal Ingestion"] A["Enterprise Signals<br/>(Multiple Channels)"] B["Omni-Modal Models"] end subgraph Modeling["2. Digital Twin"] C["Digital Models of<br/>Process Components"] D["Bottleneck Detection &<br/>Process Orchestration"] end subgraph Deployment["3. Operational Rollout"] E["Digital Simulation & Testing"] F["Low-Cost Integration &<br/>Reduced Change Friction"] end

A --> B B --> C C --> D D --> E E --> F ```

The greatest expense in enterprise AI is not the initial model development, but rather the integration into existing systems and the friction of change management. Digital twins dramatically slash these integration and change management expenses by enabling organizations to experiment and iterate entirely within a digital environment. As omni-modal models mature and capture multi-channel enterprise signals simultaneously, creating comprehensive, robust digital models will allow enterprises to resolve operational bottlenecks seamlessly and deploy new solutions with minimal friction.

- **Near-Term Horizon:** 6 to 18 months for self-learning systems and digital twin enterprise adoption. - **Cost Reduction:** Drastically reduces change management overhead by testing workflow modifications digitally before production. - **Bottleneck Resolution:** Replaces manual staff interviews with automated, digital process diagnostics.

---

[20:05](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1205s) Long-Term Horizon: Embodied AI, New Industries, and Organizational Shifts

Looking further down the road into the 2- to 3-year horizon, several foundational technologies will converge to reshape both commerce and corporate structure. Specialized, low-power AI chips will execute inference locally without requiring constant cloud connectivity, significantly lowering enterprise deployment costs. Concurrently, embodied intelligence—AI integrated directly into physical robotics that interact with and learn continuously from the physical environment—will expand alongside world models and continuous lifelong learning. Autonomous agents operating across both digital and physical domains will catalyze brand-new commercial sectors, including Robotics as a Service (RaaS) and digitally native ventures across pharmaceuticals, advanced manufacturing, and materials science.

The discovery of novel materials and autonomous physical capabilities will create a compounding flywheel effect, driving breakthroughs across adjacent frontiers such as quantum computing and space exploration. Beyond macroeconomic shifts, the internal structure of organizations will undergo a radical evolution:

```mermaid flowchart TD accTitle: Future AI Ecosystem and Organizational Transformation subgraph Frontiers["Foundational AI Capabilities (2-3 Years)"] A["Edge AI Chips<br/>(Low-Power, Offline)"] B["Embodied Robotics &<br/>Physical Agents"] C["Omni-Modal World Models &<br/>Continual Learning"] end subgraph EconomicImpact["Industry Flywheel"] D["Robotics as a Service (RaaS)"] E["Digitally Native Pharma & Materials"] F["Quantum Tech & Space Exploration"] end subgraph OrgStructures["Organizational Shifts"] G["Single-Person Billion-Dollar Solopreneurs"] H["Small Teams (10 people) with Multi-Billion Valuations"] I["Flatter, Outcome-Driven Enterprise Teams"] end

A & B & C --> D & E E --> F D & E --> G & H & I ```

This structural transformation will usher in the era of the solopreneur—giving rise to single-person billion-dollar enterprises and nimble teams of around ten people commanding tens of billions in valuation, powered by agents handling administrative overhead. Within traditional enterprises, rigid, hierarchical divisions will dissolve in favor of flatter, agile, outcome-driven teams designed to harness the full potential of autonomous agents.

- **Core Hardware & Architecture Shifts:** Low-power edge AI chips operating independently of the cloud, embodied physical robotics, and real-time continual learning models. - **Emerging Industries:** Robotics as a Service, digitally native materials and drug discovery pipelines, and accelerated quantum and aerospace advancements. - **Workforce & Enterprise Evolution:** Emergence of billion-dollar single-operator businesses, ultra-high-leverage small teams, and flatter, outcome-oriented enterprise organizational structures.

## Notable quotes > AI is no longer just a technology initiative; it is becoming a strategic transformation agenda for the enterprise. — Ashit Talukder

> [00:04](https://www.youtube.com/watch?v=1htxOFU_bfI&t=4s) one of the few rocket scientists at S&P. — Ashit Talukder

> [02:37](https://www.youtube.com/watch?v=1htxOFU_bfI&t=157s) AI solutions yesterday — Ashit Talukder

> [04:13](https://www.youtube.com/watch?v=1htxOFU_bfI&t=253s) shadow IT with a new hat — Ashit Talukder

> [04:13](https://www.youtube.com/watch?v=1htxOFU_bfI&t=253s) vibe coding — Ashit Talukder

> [07:45](https://www.youtube.com/watch?v=1htxOFU_bfI&t=465s) AI fluency — Ashit Talukder

> [14:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=884s) let a thousand flowers bloom — Ashit Talukder

> [14:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=884s) pilot hell — Ashit Talukder

> [18:10](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1090s) The greatest expense in enterprise AI is not the initial model development, but rather the integration into existing systems and the friction of change management. — Ashit Talukder

## Mentioned Ashit Talukder, S&P Global, S&P Global Ratings, NASA JPL, NIST (National Institute of Standards and Technology), Moody's Analytics, Stanford University, ChatGPT, Model Context Protocol (MCP), Ai4

## Video description Ashit Talukder - Global Head of AI and Automation, S&P Global

AI is no longer just a technology initiative; it is becoming a strategic transformation agenda for the enterprise. This talk will explore how organizations can move from AI ambition and experimentation to scaled enterprise capabilities that deliver measurable business value. The talk will include a practical strategy for aligning AI vision, business priorities, data transformation, AI-native workflows, production platforms, governance, and adoption into a cohesive enterprise operating model. I will explain why many AI and agentic AI initiatives stall after pilots, and how enterprises can overcome these barriers by redesigning workflows, transforming enterprise assets into AI-consumable services, and managing context, memory, and agents at scale. The session will illustrate how leaders can connect AI investments directly to ROI, productivity, innovation, risk reduction, and competitive advantage.

Attend this talk to learn how to turn enterprise AI strategy into scalable transformation and durable business value.

[Source](https://www.youtube.com/watch?v=1htxOFU_bfI)

中文简报

标普AI主管警告：不重构工作流程，单点解决方案终将失效 企业AI投资陷入停滞，是因为在缺乏联邦智能体治理的情况下，外挂式工具无法将深锁于未成文团队经验中的80%运营流程实现自动化。 Ashit Talukder剖析了企业AI项目频频停滞的原因，警示领导层防范失控的影子“氛围编码”（vibe coding）与孤立的外挂式单点方案这两大陷阱。他借鉴工业电气化的历史经验，详细阐述了涵盖数据、平台、人员和数字化流程的四支柱运营模型，指出企业必须借此进行结构性重构以适配AI。该演讲还概述了长效自主智能体的实用架构——包括治理控制平面、身份管理以及规范的Token经济学机制——进而探讨了流程数字孪生、具身机器人和边缘AI将如何从根本上重塑企业架构与新兴产业。 人工智能是一种强大的变革工具，而非目的本身；组织必须超越单纯的研究与实验阶段，创造出可衡量的业务成效。 (0:04) 企业领导层迫切要求快速部署AI，但员工队伍已分化为三大对立阵营：漠不关心的员工、充满热情的氛围编码者，以及心怀恐惧的怀疑者或阻挠者。 (2:37) 缺乏治理的氛围编码本质上是换了新马甲的影子IT，会引入严重的运营风险、身份管理缺失、脆弱的部署以及潜在的恶意流氓智能体威胁。 (4:13) 单点外挂式AI方案仅优化了狭窄的特定任务，却将运营瓶颈转移到了上游或下游，无法交付端到端的企业价值。 (4:13) 正如电动机直到工厂彻底重构物理布局后才真正提升工业生产力一样，企业AI同样需要从根本上重新设计运营模型。 (7:45)

Reading summary

Ashit Talukder of S&P Global argues that enterprise AI delivers durable business value only when organizations move past isolated point solutions and unmanaged prototypes to fundamentally rewire their operating models, workflows, and governance. Pitfalls of Shadow AI and Bolt-On Tools (4:13) Talukder observes that the 2026 enterprise landscape faces intense executive pressure for rapid deployment alongside a workforce split between disengaged staff, ad-hoc "vibe coders," and skeptical employees who fear job loss. In response, organizations frequently fall into two traps: ungoverned employee vibe coding and single-point bolt-on features. Ad-hoc prototypes created with generative tools function as shadow IT; they lack enterprise identity management, audit trails, and core stack compatibility. Scaled beyond pilot datasets, these fragile applications risk severe operational outages, data deletion, or cybersecurity breaches if autonomous agents act without oversight. Simultaneously, single-point additions—such as isolated chatbots or summarizers—fail to move enterprise metrics. Talukder compares bolt-on tools to widening a single section of a two-lane highway into five lanes: traffic accelerates briefly, only to stall immediately at upstream and downstream bottlenecks. To escape this trap, organizations must replace fragmented task automation with end-to-end workflow re-engineering and enforce a federated AI platform with centralized identity, role-based access, and real-time observability. Rewiring Operating Models Beyond Electric Motor Delays (7:45) To illustrate why point solutions fall short, Talukder cites the historical transition from steam to electricity. Early factories merely swapped central steam engines for electric motors while preserving mechanical drive shafts and floor layouts, resulting in flat industrial productivity for nearly 40 years. Output surged only in the 1920s when factories were completely redesigned to route power directly to individual workstations. Enterprise AI faces the same inflection point, requiring organizations to rewire four interconnected pillars: data, platforms, people, and processes. Talukder identifies business processes as the most difficult structural barrier. While data requires semantic discoverability and platforms require federated governance, an estimated 70% to 80% of enterprise processes exist as unwritten tribal knowledge rather than digitized, standardized workflows. Because AI systems cannot interact with invisible routines, companies must systematically map, standardize, and digitize these opaque workflows before realizing substantial productivity gains. Production Agent Governance and Cost Control (11:05) Transitioning from single-turn chatbots to multi-step autonomous agents can automate roughly 80% of repetitive manual tasks, freeing workers for higher-value analysis. However, deploying enterprise-grade agents requires a disciplined architectural stack. This framework includes execution harnesses using the Model Context Protocol (MCP) for secure database interactions, long-running context and memory management, and confidence-scoring thresholds that escalate uncertain decisions to human supervisors. Talukder emphasizes that agents must be provisioned with distinct digital identities, role-based permissions, and audit logs identical to the safeguards applied to junior human analysts. To counter corporate concerns over runaway inference and compute expenses, Talukder warns against letting unguided experiments proliferate. Leaders should focus resources on two or three initiatives tied to clear key performance indicators like cycle-time reduction and operational throughput. Organizations can maintain fiscal discipline through intelligent model routing—directing approximately 80% of routine workloads to compact, low-cost models while reserving expensive frontier models for complex reasoning—while leveraging operational telemetry to build a self-improving AI flywheel. Emerging Horizons: Digital Twins to Embodied Systems (18:10) Over a 6- to 18-month horizon, Talukder highlights enterprise process digital twins and self-learning loops as pivotal advances. Because system integration and change management represent the largest expenditures in enterprise AI, digital replicas of operational workflows allow organizations to automatically detect bottlenecks and test agent behaviors in simulated environments before live deployment, bypassing disruptive manual staff interviews. Looking 2 to 3 years ahead, Talukder forecasts that low-power edge inference chips, omni-modal world models, and embodied physical robotics will converge to create new commercial sectors, such as Robotics as a Service (RaaS) and digitally native materials discovery. Within corporations, this shift is expected to flatten traditional hierarchies into outcome-driven units, alongside the emergence of multi-billion-dollar single-operator businesses and ten-person enterprises where administrative overhead is managed almost entirely by autonomous agents. Talukder concludes that capturing enterprise value from AI requires treating it as an architectural operating-model transformation rather than a software add-on. Long-term success hinges on whether organizations can successfully digitize unwritten operational knowledge, enforce strict agent governance, and maintain economic discipline across model routing before transitioning to autonomous and physical AI frontiers.

精炼阅读摘要

标普全球评级业务AI负责人阿希特·塔鲁克达尔指出，企业AI转型若要超越单点实验并创造持久业务价值，必须摆脱无治理的自发编程与外挂陷阱，效仿工业电气化变革对数据、平台、人员与流程进行系统性重构。 规避组织撕裂与AI落地双重陷阱 (4:13) 塔鲁克达尔观察到，企业高层对AI落地的迫切诉求与基层的实际状态严重脱节，基层员工已分化为冷眼旁观者、利用生成式工具自制原型的“氛围程序员”（vibe coders），以及担忧岗位被取代的抵制者。在交付压力下，企业极易落入两大陷阱：一是员工自发构建的原型本质上是缺乏身份认证、审计追踪和技术栈兼容性的“影子IT”，在智能体环境中这类未受管制的原型甚至可能渗透内网并破坏核心数据；二是针对复杂工作流中的孤立环节加装聊天机器人等单点外挂工具，这如同在狭窄公路中仅拓宽极短路段，前后节点依然拥堵，无法推动整体业务回报。 针对这些结构性风险，塔鲁克达尔主张企业必须推行具备集中身份管理、基于角色的访问控制（RBAC）和端到端可观测性的联邦AI平台，同时将关注重点从孤立任务的微调转向重塑端到端的工作流拓扑结构。 汲取电气化经验重塑四大运营支柱 (7:45) 为阐明单点工具为何收效甚微，塔鲁克达尔引用工业史指出，早期工厂仅用电动机替换蒸汽机而保留原有的机械传动布局，致使工业生产力在近40年间停滞不前，直至20世纪20年代彻底重构车间布线与独立工位供电后才迎来生产力爆发。当前企业部署零散辅助工具同样面临这一瓶颈，必须全面重塑四大运营支柱：数据必须具备语义检索能力并拥有端到端血缘追踪；平台须具备标准化的安全协议与可观测性；人员层面需通过推广网络培育AI应用能力。 塔鲁克达尔强调，业务流程是重塑中最关键且最具挑战的一环。在典型企业中，仅有20%至30%的业务流程实现了标准化与数字化，其余70%至80%则作为未成文的“部落知识”（tribal knowledge）隐性存在。由于AI无法直接读取或自动化这些暗流工作，企业唯有主动梳理、标准化并数字化这些深层隐性流程，才能释放真正的结构性红利。 生产级智能体架构与工程经济学 (11:05) 塔鲁克达尔提出，企业必须从被动的单轮问答机器人，升级为能够自主规划、调用工具并全天候执行多步骤复杂任务的生产级智能体，以此接管约80%的重复性工作。在工程落地层面，严谨的智能体架构需要运行载体（harness）、模型上下文协议（MCP）工具链以及长周期上下文管理机制支持；同时必须设立独立的数字身份与RBAC权限，并配置置信度阈值机制，在遇到不确定场景时主动将决策权升级转交给人工介入。 面对财务管理者对计算与推理成本失控的担忧，企业必须坚持业务关键指标导向，避免陷入长达数月追求完美却无法落地的测试死循环。塔鲁克达尔建议运用工程经济学策略，通过智能模型路由将约80%的常规工作流分配给紧凑型低成本模型，仅将20%涉及深度推理的高难度任务交由前沿大模型处理；并依托生产环境遥测数据构建“AI飞轮”，实现系统的持续自学习与反馈迭代。 流程数字孪生与长远组织形态变革 (18:10) 展望技术演进路径，塔鲁克达尔预测在未来6至18个月内，构建企业“流程数字孪生”将成为破局关键。企业无需依赖繁琐的人工访谈来排查流程缺陷，而是利用全模态模型摄取多渠道业务信号，在数字环境中自动诊断瓶颈、模拟工况并测试AI应用，从而大幅削减企业AI转型中成本最高的系统集成与组织变革阻力。 在2至3年的长期维度上，低功耗边缘AI芯片、具身智能机器人及持续学习世界模型的成熟，将催生机器人即服务（RaaS）等新兴产业形态。塔鲁克达尔预测，随着智能体消化大量行政运营开销，商业领域或将诞生单人支撑的十亿美元级企业（solopreneur）或仅约十人规模的超高市值敏捷团队；传统企业内部等级森严的职能架构也将瓦解，全面转向更加扁平、以最终业务结果为导向的协同单元。 阿希特·塔鲁克达尔总结指出，企业AI落地的核心壁垒不在于单一模型能力，而在于底层运营模型的重塑与工程治理的协同。不过，该论述对于企业占比高达七至八成的隐性流程数字化的实际推行成本、跨部门长周期自主智能体的合规追责边界，仍留有待市场实践验证的现实挑战。

Detailed notes — zh

## Background 该演讲视频由科技大会频道 Ai4 于 2026 年 8 月 19 日发布在 YouTube 平台上。内容由标普全球（S&P Global）全球 AI 与自动化负责人 Ashit Talukder 主讲，围绕企业如何从前期的 AI 概念试验跨越至规模化落地展开，阐述通过重塑业务流程、构建联邦化生产平台与治理智能体系统，将 AI 战略切实转化为可衡量的业务价值与投资回报。 ## TL;DR 演讲指出企业 AI 转型当前面临管理层盲目求快与基层员工认知撕裂的矛盾，并警惕“氛围编码”催生的影子 AI 与单点拼贴方案所带来的局限。Ashit 借用工业电气化必须重构生产线的历史经验，强调企业必须系统性改造数据、平台、人才与流程四大支柱，尤其是把 70% 至 80% 隐藏在水面下的非标准隐性流程转化为 AI 可消费的服务。在实操与成本控制上，企业应告别“概念验证地狱”，借助智能模型路由、人在回路治理框架与 AI 飞轮实现可持续运行，并在未来 6 到 36 个月内积极布局流程数字孪生、具身智能与去层级化的敏捷商业组织形态。 ## Atmosphere & style 整场演讲沉稳务实且极富战略格局，Ashit 依托从 NASA 前沿科研到华尔街金融巨头的深厚履历，展现出兼具科学家严谨与企业高管敏锐的专业气度。叙事风格抽丝剥茧、见解辛辣，毫不留情地揭露了企业内盲目堆砌 Chatbot 与单点自嗨的浮躁乱象，随后又以工程经济学与历史长周期的维度冷静破局。整场分享带给听众一种剥除概念炒作迷雾、回归企业真实运营本质的清醒感，并在展望超级个体与数字孪生产业时传递出强烈的未来重塑紧迫感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=1htxOFU_bfI&t=4s) AI 转型从来不是单纯的技术替换，企业必须汲取工业革命从蒸汽动力转向电力的历史教训，进行全局性的端到端组织与流程重构才能释放真正的生产力。 - [02:37](https://www.youtube.com/watch?v=1htxOFU_bfI&t=157s) 企业在推进 AI 时普遍面临认知撕裂：管理层极度急切期望成果，而基层员工分化为漠不关心、自学成才的“氛围编码者”（Vibe Coders）与暗中抵触的三大阵营。 - [04:13](https://www.youtube.com/watch?v=1htxOFU_bfI&t=253s) 自建小工具催生的“影子 AI”（Shadow AI）缺乏身份管控与安全治理，且极易在生产环境崩溃，必须收拢至具备全局可观测性的联邦化平台中统筹纳管。 - [04:13](https://www.youtube.com/watch?v=1htxOFU_bfI&t=253s) “单点拼贴式”（Bolt-on）方案如同在拥堵的老旧双车道上修一段五车道高速，无法解决端到端瓶颈，无法真正为 CTO 和 CFO 交付商业价值。 - [07:46](https://www.youtube.com/watch?v=1htxOFU_bfI&t=466s) 流程是 AI 转型的最大短板：企业内部 70% 至 80% 的流程高度依赖员工个人经验（Tribal Knowledge），未被标准化与透明化，AI 根本无从介入。 - [11:06](https://www.youtube.com/watch?v=1htxOFU_bfI&t=666s) AI 必须从被动问答的反应式 Chatbot 演进为能够自主制定计划、调用 MCP 工具并长时间运行的智能体（Agent），才能承担 80% 的常规繁琐工作并实现 24/7 全天候扩展。 - [11:06](https://www.youtube.com/watch?v=1htxOFU_bfI&t=666s) 部署智能体需像对待新入职员工一样，建立完善的控制平面，明确赋予独立的身份、角色与访问权限，并配置置信度检查与人在回路（Human-in-the-loop）纠错机制。 - [14:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=884s) 避免陷入长达半年的“概念验证地狱”（Pilot Hell），应聚焦前三大核心项目尽早上线生产环境，借助真实数据反馈构建持续进化的“AI 飞轮”。 - [14:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=884s) 践行 Token 经济学与工程经济学：80% 的日常任务只需调用轻量级小模型进行智能路由，仅 20% 的高难度任务才需调用昂贵的前沿旗舰大模型。 - [17:04](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1024s) 企业流程“数字孪生（Digital Twins）”能在数字空间精准诊断隐性业务瓶颈并模拟集成效果，大幅降低企业在业务集成与变革管理（Change Management）上的巨额开销。 - [20:05](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1205s) 未来 2 至 3 年内，低功耗芯片、具身智能、世界模型与自主智能体将推动 AI 从数字软件环境深度渗透至实体物理世界与新材料等原生行业。 - [21:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1304s) AI 正在颠覆传统组织架构：催生单人 10 亿美元独角兽与 10 人百亿美元企业，同时促使大企业打破层级化汇报体系，转向高度扁平化、以结果为导向的敏捷团队。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=1htxOFU_bfI&t=4s) 演讲背景与三大主题：从研究先锋到企业商业价值落地 现任 S&P Global Ratings（标普全球评级）全球 AI 与自动化负责人的 Ashit Talukder 首先回顾了自己的职业履历。他自嘲是标普内部为数不多的“火箭科学家”之一，早年在 NASA 喷气推进实验室（JPL）从事机器学习与 AI 基础研究。在那个年代，AI 仍属于高度学术化的前沿课题，仅有少数资金雄厚的联邦政府机构有能力提供资助。随着深度学习技术的演进与成熟，他随后加入美国国家标准与技术研究院（NIST）主导启动数据科学项目，之后转战 Moody's Analytics（穆迪分析）领导 AI 战略与产品职能，最终加入标普。这段横跨国家级科研攻关与商业应用实践的复合经历，让他深刻理解技术突破与实际企业落地之间的鸿沟。 在本次演讲中，Ashit 明确指出，AI 本质上是一个功能极为强大的“工具”，但企业追求的核心始终是可衡量的商业价值。为了系统性剖析 AI 转型中的常见障碍与破局之道，他将演讲分为三大主线： 1. **汲取历史教训**：工业革命和过往历次技术浪潮的经验表明，组织转型从来不是单纯换上新技术，必须借鉴历史规律才能避开陷阱。 2. **立足当下执行**：深入剖析大型组织内部推动 AI 落地、打破阻碍的具体实操方案与运营模式。 3. **面向未来布局**：AI 演进极其迅速，解决方案绝不能只适配过往模式而缺乏持续演进的扩展能力。 --- ### [02:37](https://www.youtube.com/watch?v=1htxOFU_bfI&t=157s) 职场认知撕裂：管理层的迫切期待与员工的三大阵营 从历史视角回顾 AI 的认知演变，虽然早在 1950 年代斯坦福大学等机构就开启了 AI 研究，但在 2000 年至 2022 年长达二十余年的时间里，绝大多数企业组织对 AI 的感知几乎毫无波澜，它始终局限在一小部分机器学习研究员的小圈子里。直到 2022 年底 ChatGPT 的问世，生成式 AI（Generative AI）才真正引爆了企业高层的关注。 发展到当下，天平彻底摆向了另一端——管理层普遍表现出极度急迫的心态，甚至恨不得“昨天就已经上线了 AI 解决方案”。然而与高层的激进期待形成鲜明反差的是，企业基层员工在心态上分裂成了截然不同的三大阵营： - **漠不关心派**：在日常工作中依然对 AI 无动于衷，维持原有工作方式。 - **自学成才的“氛围编码者”（Vibe Coders）**：积极利用最新的自动化编程工具自主搭建各类小应用。 - **怀有戒心的抵触派**：深切担忧 AI 会夺走自己的工作，甚至在部分企业中已经出现了暗中抵制或故意破坏 AI 推广流程的真实案例。 ```mermaid flowchart TD accTitle: 组织内部对 AI 的期望与阵营分化 subgraph Leadership["管理层预期"] L["极度急迫<br/>期望'昨天就能上线'AI 落地成果"] end subgraph Employees["基层员工三大阵营"] E1["漠不关心<br/>（不参与/无感）"] E2["氛围编码者 (Vibe Coders)<br/>（自学工具/自行造轮子）"] E3["抵触与怀疑者<br/>（担忧失业/被动抵制）"] end L -.->|"期望与执行脱节"| Employees ``` --- ### [04:13](https://www.youtube.com/watch?v=1htxOFU_bfI&t=253s) 影子 AI 与拼贴式方案：企业落地面临的两大误区 在当前企业 AI 的具体推进中，普遍存在两大严重拖累战略成效的误区： #### 1. “氛围编码”（Vibe Coding）催生的影子 AI（Shadow AI） 许多技术爱好者利用代码生成工具迅速拼凑出炫酷的独立工具并展示给领导层，管理层往往大为惊艳并急切询问何时能全面上线。但这本质上是“换了新马甲的影子 IT（Shadow IT）”。这些自建工具没有经过技术栈兼容性审查，缺乏身份管理（Identity Management）与合规治理，且极其脆弱，在小样本试点时表现良好，一旦进入生产环境便频繁崩溃，最终维护包袱全甩给了正式的 AI 团队。尤其在智能体（Agentic AI）时代，如果失控的智能体攻破内网、引发网络安全漏洞或误删核心数据，可能直接导致价值数十亿美元的企业陷入停摆。 解决影子 AI 的唯一路径是将所有 AI 应用强制收拢至**联邦化平台（Federated Platform）**中，实行统一纳管、身份认证、安全治理与端到端可观测性，杜绝失控智能体的出现。 #### 2. “单点拼贴式”（Bolt-on）解决方案的瓶颈 许多团队为了尽快证明价值，倾向于在复杂的现有工作流中挑选一个所谓的“低垂果实”，直接生硬地拼贴一个单点 AI 工具。Ashit 用公路交通作了生动比喻：这就像一条原本只有双车道的拥堵老旧公路，团队仅仅将其中一小段扩建成了五车道的高速公路。局部通行速度虽然变快了，但前后瓶颈依然存在，整条道路依然拥堵。最终，AI 团队依然无法向 CTO 和 CFO 交代为什么巨大投入没有转化为端到端的切实商业价值。 --- ### [07:46](https://www.youtube.com/watch?v=1htxOFU_bfI&t=466s) 电气化重构的启示：摆脱单点工具，重构企业四大支柱 历史总在押着相同的韵脚。在工业史上从蒸汽动力向电力动力的切换过程中，早期的工厂仅仅是用电动机直接替换了原本的蒸汽机动力源，但整个厂房结构、流水线布局依然保持不变，导致生产率在数十年间几乎没有实质提升。直到 40 年后的 1920 年代，工业界彻底推倒重来，围绕电力特征重新设计了厂房结构与流水线（Rewiring），电力的全部生产力才真正得到释放。 当前的 AI 浪潮如出一辙：企业内部充斥着上百个彼此割裂的聊天机器人（Chatbots）、Copilot 工具与自制小脚本，但因为缺乏全局体系的重构，根本无法带来高层所期盼的根本性跃升。 ```mermaid flowchart LR accTitle: 蒸汽机向电力转型与 AI 重构的比喻 subgraph PastHistory["历史教训：电力革命"] direction TB P1["仅用电动机替换蒸汽源<br/>（厂房布局不变 → 生产率停滞）"] -->|"历经 40 年重构 (1920s)"| P2["围绕电力重新设计厂房与产线<br/>（生产力全面释放）"] end subgraph PresentAI["当前现实：AI 转型"] direction TB C1["单点拼贴/百个独立 Chatbot<br/>（双车道改五车道 → 仍遇瓶颈）"] -->|"必需的转型路径"| C2["重构端到端运营模型<br/>（释放全局商业价值）"] end ``` 要想完成真正的系统重构，企业必须同步打通四大核心支柱： - **数据（Data）**：数据是驱动 AI 引擎的燃料。数据必须具备 AI 就绪性（AI-ready），支持语义化检索（Semantically Searchable）、全局可发现（Discoverable），并具备完整的端到端血缘追踪（End-to-End Lineage），为模型提供准确的上下文。 - **平台（Platforms）**：构建联邦化平台，内嵌安全防护、身份管控与全局可观测性。 - **人才（People）**：在大企业中必须开展全员技能升级与 AI 素养（AI Fluency）培训，通过建立“AI 冠军网络”（AI Champions Network）进行跨部门布道，确保所有员工理解 AI 在未来工作中的定位。 - **流程（Processes）——转型的最大短板**： - 许多团队误以为只要流程存在就能被 AI 自动化，但现实中大多数企业只有 **20% 到 30%** 的流程是标准化的、数字化的、可测量和具备可观测性的； - 剩余 **70% 到 80%** 的业务流程高度依赖员工个人经验（Tribal Knowledge），处于隐蔽、不透明、非标准的状态，AI 根本无从介入。 Ashit 强调，实现 AI 商业价值转型的核心关键，就在于将这 70%~80% 隐藏在水面之下的流程全面透明化、数字化与标准化，唯有完成流程与组织的深度重构，AI 技术才能真正发挥效能。 ### [11:06](https://www.youtube.com/watch?v=1htxOFU_bfI&t=666s) 从反应式聊天机器人到能够自主行动的智能体（Agent） AI 已经从过去只能进行单轮问答、被动响应的传统 Chatbot，演进为能够主动采取行动、端到端完成复杂任务的智能体（Agent）。一个设计良好的 Agent 能够自主制定计划、决定检索哪些数据、通过测试框架与控制支架（Harness）管理上下文，并最终将完整任务交付完成。只有达到这种能力水平，企业才能真正实现 CFO 和高管层所期望的商业价值——包括效率提升、运营模式提速，以及打造具有高用户粘性的 AI 产品。在理想状态下，AI 可以承担人类过去所做的 80% 的繁琐手动工作，从而将员工解放出来，让他们专注于最具人本价值与创造力的事务，比如人际沟通、团队协作和客户深度交流。此外，优质的 Agent 解决方案能够实现 24/7 全天候无休扩展，打破了人类员工每天 8 小时的工作时间上限。 然而，构建成熟的 Agent 绝非仅靠大语言模型（LLM）本身那么简单。一个成功的 Agent 架构必须具备以下核心要素： - **执行支架与工具集成（Harnesses & MCP Tools）：** 必须具备清晰的指令集、上下文管理机制，以及通过模型上下文协议（MCP）工具对接企业内部数据集，让 Agent 能够真正对数据执行操作。 - **控制平面与治理（Control Plane & Governance）：** 部署 Agent 就像对待新入职的初级助理一样，必须明确赋予其独立的身份（Identity）、角色（Roles）与访问权限（Permissions），并配备完备的可观测性（Observability）和审计工具。 - **长时间运行与上下文管理（Long-Running Agents）：** 对于需要持续运行数小时甚至数天、包含多个子目标的大型任务，必须建立稳健的上下文管理层。同时，Agent 需要具备置信度判断机制——当自身信心不足时，能够主动向人类管理者请求协助（Human-in-the-loop），确认后方可继续执行。 ```mermaid flowchart TD accTitle: 生产级 Agent 核心架构与治理要素 subgraph Execution["执行与工具层 (Execution Layer)"] A["指令与上下文 (Instructions & Context)"] --> B["MCP 工具集 (MCP Tools)"] B --> C["内部企业数据交互 (Internal Data Action)"] end subgraph Control["控制平面与治理层 (Control Plane)"] D["身份与权限 (Identity & RBAC)"] --> E["可观测性与审计 (Observability & Audit)"] E --> F["置信度检查与人机协同 (Human-in-the-loop)"] end subgraph Value["业务价值转化 (Business Value)"] G["长时间自主运行 (Long-Running 24/7)"] --> H["端到端业务交付 (Efficiency & ROI)"] end Execution --> Control Control --> Value ``` --- ### [14:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=884s) 成本控制、工程经济学与 AI 飞轮效应 在 Agent 规模化运行后，企业管理层和 CFO 最关心的核心问题往往是：Token 消耗巨大且成本高昂，如何有效控制成本？应对这一挑战，首要原则是进行**无情的优先级排序（Ruthless Prioritization）**。企业绝不能陷入“百花齐放”却无一结果的散乱状态，而应将有限资源集中在能够通过具体业务指标（如生产力提升、运营速度加速、用户参与度提升）产生实质性改变的前三大核心项目上。团队必须尽快走出长达半年以上追求完美却无法落地的“概念验证地狱（Pilot Hell）”，推动方案尽早上线生产环境。 为了说服 CEO 和高级领导层批准上线尚不完美的初期方案，Ashit Talukder 提出了 **AI 飞轮（AI Flywheel）** 概念：AI 的核心特性在于“随数据积累而持续进化”。只要具备完善的治理规范和安全护栏（Guardrails），哪怕是不完全成熟的方案也可以先推向市场；随着真实用户行为数据和日志的回流，结合人在回路纠错机制，Agent 能够实现自主演进闭环，从错误中快速学习成长。 ```mermaid flowchart LR accTitle: 企业 AI 飞轮效应闭环 A["上线受控的初期产品"] --> B["收集用户交互与日志数据"] B --> C["人在回路反馈与纠错"] C --> D["模型与 Agent 自主迭代"] D --> A ``` 在技术层面的成本管理上，必须引入**工程经济学（Engineering Economics）与 Token 经济学（Tokenomics）**： - **智能模型路由（Intelligent Routing）：** 严禁对所有简单任务一律调用昂贵的前沿旗舰大模型（Frontier Models）。实践表明，**80% 的常规任务只需使用轻量级小模型即可胜任**，仅有 20% 的高度复杂任务才真正需要调用前沿大模型。 - **上下文窗口优化：** 并非每个任务都需要 100 万 Token（1M）的超长上下文窗口，精细化管理上下文长度可大幅降低推理开销。 - **避免工具泛滥（Tool Sprawl）：** 审慎评估供应商采购与自研（In-house）的边界，借助 AI 辅助编码能力合理自建内部关键能力，防止多供应商工具泛滥带来的管理失控。 --- ### [17:04](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1024s) 6 至 18 个月内的近期技术趋势与企业“数字孪生（Digital Twins）” 作为企业领导者，必须具备面向未来的前瞻性，构建能够随 AI 领域快速演进和转型的系统架构。在未来 6 到 18 个月内，行业将迎来一系列关键技术突破：长程规划与推理能力更强的智能体、自我演进与自我学习模型，以及能够直接操作和控制电脑界面的“计算机操作智能体（Computer-Use Agents）”。而在这些技术中，对于企业落地最为关键的概念是**企业流程的“数字孪生（Digital Twins）”**。 结合企业运营中大量未被记录的“冰山隐性流程”，数字孪生技术能够通过全模态模型（Omnimodal Models）捕获企业内部的多渠道信号，为组织的各项业务流程构建数字模型。这一机制带来的核心价值在于： - **瓶颈精准诊断：** 无需逐一走访业务分析师或基层员工，即可在数字空间中直接定位流程瓶颈。 - **低成本模拟与验证：** 企业集成 AI 方案的最大开销往往不是算法开发本身，而是后期的业务集成与变革管理（Change Management）。通过数字孪生，企业可以在系统正式上线前，在数字空间中充分模拟各种运营条件与 AI 交互效果，从而大幅削减变更管理成本并加速流程编排。 --- ### [20:05](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1205s) 未来 2 至 3 年的长期愿景：从数字世界走向实体智能与新产业 展望未来 2 到 3 年的发展路线，AI 正在从纯软件环境向实体物理世界和全行业底层逻辑深度渗透，并形成一系列具有连锁反应的大趋势： - **低功耗 AI 芯片（AI Chips）：** 端侧低功耗芯片的发展将摆脱对云端连接的绝对依赖，大幅压降 AI 的规模化部署成本。 - **具身智能（Embodied Intelligence）：** AI 深度融入实体机器人，能够直接与物理世界交互，并从现实世界的反馈中学习。 - **世界模型与全模态（World Models & Omnimodal Models）：** 具备更深层次理解物理世界运转规律并进行持续学习（Continual Learning）的模型将从学术论文走向实际落地。 - **物理与数字自主智能体（Autonomous Agents）：** 自主智能体不仅存在于云端软件系统，还将广泛应用在物理硬件中。 - **新兴原生行业与协同飞轮：** 催生“机器人即服务（Robotics as a Service）”以及数字原生的制药、制造与新材料企业。新材料的研发突破反过来将加速太空探索与量子技术的发展，形成相互促进的技术飞轮。 ### [21:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1304s) 组织架构与商业形态的重塑：从单人独角兽到扁平化结果导向团队 探讨人工智能对就业的影响，焦点不应仅仅停留在具体岗位的存废上，而更应该着眼于伴随技术演进所孕育出的崭新行业与蓬勃发展的数字经济。随着数字商务以极其迅猛的态势崛起，企业传统的组织架构与运作逻辑正迎来一场根本性的颠覆。我们正在加速步入一个由 AI 深度赋能的“超级个体创业者”（solopreneurs）时代——在这一全新格局下，完全可能出现仅凭单兵作战就创造出估值达到 10 亿美元的独角兽企业；而仅由 10 人左右规模构成的小型团队，也能依托强大的智能体（Agents）承担绝大部分繁琐的行政与日常运营支持工作，从而轻松支撑起数百亿美元的惊人估值。 这种由 AI 驱动的组织形态演变，同样在大型企业内部掀起深刻的结构重组浪潮。以往那种层级分明、职责固化的金字塔式汇报体系正逐渐瓦解，取而代之的是高度扁平化、以实际结果为导向（outcome-driven）的敏捷团队。在演讲的尾声，主讲人强调，尽管时间所限无法穷尽所有细节，但他希望在场的企业领袖能够带走这幅关于 AI 未来发展脉络的宏大图景，以极具前瞻性的战略思维，尽早为正在发生的行业重塑做好全面准备。 ```mermaid flowchart LR accTitle: AI驱动的企业组织形态变革 subgraph Traditional["传统企业架构模式"] direction TB T1["层级化汇报体系"] --> T2["固化部门分工"] T2 --> T3["大量人力消耗于行政与流程事务"] end subgraph Future["AI与智能体赋能的新模式"] direction TB F1["超级个体与极小团队<br/>(单人10亿 / 10人百亿估值)"] F2["AI Agents 全权承接行政支撑"] F3["扁平化、结果导向型敏捷团队"] F1 --- F2 F2 --- F3 end Traditional -->|"组织结构演进"| Future ``` - **超级个体崛起与价值重构**：单人创办 10 亿美元企业（"single person billion dollar companies"）与仅需 10 人团队实现数百亿美元估值的商业神话正成为现实。 - **智能体接管繁复事务**：AI 智能体（Agents）将全面接管企业日常中大量的行政性与支持性任务，释放核心人员的创造力与战略决策力。 - **大企业内部去层级化**：大企业正在打破僵化的层级制度（hierarchical fixed teams），向结构更加扁平、更加专注于终极业务成果的灵活协作团队（flatter organizations & outcome-driven teams）全面转型。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=1htxOFU_bfI&t=4s) AI 本质上是一个功能极为强大的“工具”，但企业追求的核心始终是可衡量的商业价值。 — Ashit Talukder > [02:37](https://www.youtube.com/watch?v=1htxOFU_bfI&t=157s) 管理层普遍表现出极度急迫的心态，甚至恨不得“昨天就已经上线了 AI 解决方案”。 — Ashit Talukder > [04:13](https://www.youtube.com/watch?v=1htxOFU_bfI&t=253s) 自建工具本质上是“换了新马甲的影子 IT（Shadow IT）”。这些自建工具没有经过技术栈兼容性审查，缺乏身份管理与合规治理，且极其脆弱。 — Ashit Talukder > [04:13](https://www.youtube.com/watch?v=1htxOFU_bfI&t=253s) 这就像一条原本只有双车道的拥堵老旧公路，团队仅仅将其中一小段扩建成了五车道的高速公路。局部通行速度虽然变快了，但前后瓶颈依然存在，整条道路依然拥堵。 — Ashit Talukder > [07:46](https://www.youtube.com/watch?v=1htxOFU_bfI&t=466s) 现实中大多数企业只有 20% 到 30% 的流程是标准化的、数字化的、可测量和具备可观测性的；剩余 70% 到 80% 的业务流程高度依赖员工个人经验，处于隐蔽、不透明、非标准的状态，AI 根本无从介入。 — Ashit Talukder > [11:06](https://www.youtube.com/watch?v=1htxOFU_bfI&t=666s) 在理想状态下，AI 可以承担人类过去所做的 80% 的繁琐手动工作，从而将员工解放出来，让他们专注于最具人本价值与创造力的事务。 — Ashit Talukder > [14:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=884s) 企业绝不能陷入“百花齐放”却无一结果的散乱状态，而应将有限资源集中在能够通过具体业务指标产生实质性改变的前三大核心项目上。 — Ashit Talukder > [14:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=884s) 80% 的常规任务只需使用轻量级小模型即可胜任，仅有 20% 的高度复杂任务才真正需要调用前沿大模型。 — Ashit Talukder > [17:04](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1024s) 企业集成 AI 方案的最大开销往往不是算法开发本身，而是后期的业务集成与变革管理（Change Management）。 — Ashit Talukder > [21:44](https://www.youtube.com/watch?v=1htxOFU_bfI&t=1304s) 在这一全新格局下，完全可能出现仅凭单兵作战就创造出估值达到 10 亿美元的独角兽企业；而仅由 10 人左右规模构成的小型团队，也能依托强大的智能体承担绝大部分繁琐的行政与日常运营支持工作，从而轻松支撑起数百亿美元的惊人估值。 — Ashit Talukder ## Mentioned Ashit Talukder, S&P Global (标普全球), S&P Global Ratings (标普全球评级), NASA JPL (喷气推进实验室), NIST (美国国家标准与技术研究院), Moody's Analytics (穆迪分析), Stanford University (斯坦福大学), ChatGPT, Model Context Protocol (MCP), Digital Twins (数字孪生), Computer-Use Agents, Robotics as a Service (RaaS) ## Video description Ashit Talukder - Global Head of AI and Automation, S&P Global AI is no longer just a technology initiative; it is becoming a strategic transformation agenda for the enterprise. This talk will explore how organizations can move from AI ambition and experimentation to scaled enterprise capabilities that deliver measurable business value. The talk will include a practical strategy for aligning AI vision, business priorities, data transformation, AI-native workflows, production platforms, governance, and adoption into a cohesive enterprise operating model. I will explain why many AI and agentic AI initiatives stall after pilots, and how enterprises can overcome these barriers by redesigning workflows, transforming enterprise assets into AI-consumable services, and managing context, memory, and agents at scale. The session will illustrate how leaders can connect AI investments directly to ROI, productivity, innovation, risk reduction, and competitive advantage. Attend this talk to learn how to turn enterprise AI strategy into scalable transformation and durable business value. [Source](https://www.youtube.com/watch?v=1htxOFU_bfI)

Source quotes

[0:04] Ashit Talukder: AI 本质上是一个功能极为强大的“工具”，但企业追求的核心始终是可衡量的商业价值。 [2:37] Ashit Talukder: 管理层普遍表现出极度急迫的心态，甚至恨不得“昨天就已经上线了 AI 解决方案”。 [4:13] Ashit Talukder: 自建工具本质上是“换了新马甲的影子 IT（Shadow IT）”。这些自建工具没有经过技术栈兼容性审查，缺乏身份管理与合规治理，且极其脆弱。 [4:13] Ashit Talukder: 这就像一条原本只有双车道的拥堵老旧公路，团队仅仅将其中一小段扩建成了五车道的高速公路。局部通行速度虽然变快了，但前后瓶颈依然存在，整条道路依然拥堵。 [7:46] Ashit Talukder: 现实中大多数企业只有 20% 到 30% 的流程是标准化的、数字化的、可测量和具备可观测性的；剩余 70% 到 80% 的业务流程高度依赖员工个人经验，处于隐蔽、不透明、非标准的状态，AI 根本无从介入。 [11:06] Ashit Talukder: 在理想状态下，AI 可以承担人类过去所做的 80% 的繁琐手动工作，从而将员工解放出来，让他们专注于最具人本价值与创造力的事务。 [14:44] Ashit Talukder: 企业绝不能陷入“百花齐放”却无一结果的散乱状态，而应将有限资源集中在能够通过具体业务指标产生实质性改变的前三大核心项目上。 [14:44] Ashit Talukder: 80% 的常规任务只需使用轻量级小模型即可胜任，仅有 20% 的高度复杂任务才真正需要调用前沿大模型。 [17:04] Ashit Talukder: 企业集成 AI 方案的最大开销往往不是算法开发本身，而是后期的业务集成与变革管理（Change Management）。 [21:44] Ashit Talukder: 在这一全新格局下，完全可能出现仅凭单兵作战就创造出估值达到 10 亿美元的独角兽企业；而仅由 10 人左右规模构成的小型团队，也能依托强大的智能体承担绝大部分繁琐的行政与日常运营支持工作，从而轻松支撑起数百亿美元的惊人估值。

Source quotes

Ashit Talukder: AI is no longer just a technology initiative; it is becoming a strategic transformation agenda for the enterprise. [0:04] Ashit Talukder: one of the few rocket scientists at S&P. [2:37] Ashit Talukder: AI solutions yesterday [4:13] Ashit Talukder: shadow IT with a new hat [4:13] Ashit Talukder: vibe coding [7:45] Ashit Talukder: AI fluency [14:44] Ashit Talukder: let a thousand flowers bloom [14:44] Ashit Talukder: pilot hell [18:10] Ashit Talukder: The greatest expense in enterprise AI is not the initial model development, but rather the integration into existing systems and the friction of change management.

Original transcript

[0:04] So I'm Ashit um um I lead the AI initiatives at  SMB global ratings. Uh just a quick background [0:11] about myself. Uh I started my career at NASA JPL.  Uh so I'm one of the few rocket scientists at S&P. [0:18] Um so I did uh ML research, AI research at NASA  JPL back when AI was still very much a research [0:25] topic. Um and the federal government was the only  one with deep pockets to fund uh machine learning [0:30] research right um and then over time as deep  learning matured I moved to NIST to start the [0:36] data science program initiatives there and then  moved to Muny's analytics to read the to lead the [0:42] AI uh strategy functions and product functions  at Moody's Analytics. So I have this breadth [0:47] of experience in um both AI research um and uh  applied AI in the commercial sector. So what I'll [0:55] do today in the next 15 minutes is I'll walk you  through my experiences in successfully managing [1:02] and leading AI initiatives uh across the board in  different organizations and what are the typical [1:08] hurdles that one face [snorts] uh that one faces  as a leader in actually effectively managing this [1:14] AI transformation right what are the obstacles we  face and how do you overcome these obstacles so [1:19] that's what I'll spend in the next 15 minutes or  so um and in order to do that I want to you know [1:26] what I did was I divided my discussions in into  three different paths. Uh so the first path is [1:32] more about learning from the past. So industrial  transformation is nothing new. Uh organizational [1:38] transformation is nothing new. This you know we've  been through waves of technology revolutions and [1:44] we need to learn from the results from the past to  successfully apply to AI transformation. So I'll [1:49] quickly walk you through that before we dive into  the nuts and bolts of what it takes to actually [1:53] successfully uh execute AI initiatives within  large organizations and create business value. [2:00] So AI is nothing more than a tool, right? And  ultimately what businesses want is they want to [2:04] create business value from the from this new tool  which is very powerful. It can do a lot of things. [2:09] So I'll walk you through uh some of my experiences  in uh at Moody's S&P and and at NIST and then end [2:16] with a future-looking perspective because as  an organization we have to be future ready and [2:21] AI is evolving so quickly that you can't build a  solution that only fits into the past solutions [2:25] and cannot evolve to future solutions. So I'll  end with that. So kind of start at the past uh [2:30] you know discuss the present and then uh walk into  the future and what to expect there. So if if you [2:37] take a look at the evolution of AI right and and  the expectations of AI and the perceptions of AI [2:43] throughout history I mean AI started in 1950s at  Stanford uh but even between 2000 and and 2022 the [2:50] perception of AI at an organizational level did  not change right except for a small sliver of ML [2:56] researchers now 2022 when Chad GV came out that's  when leaders started sort of noticing the value of [3:02] AI uh especially with gen generative AI right and  uh some employees you know started sort of getting [3:09] interested in this field and now fast forward into  2026 now it's almost a full swing in the pendulum [3:15] where leaders actually want AI solutions yesterday  right uh and and I'm sure most of you are seeing [3:21] that in your organizations but amongst the working  population we kind of have three camps one camp [3:27] still doesn't care too much about AI at work  then one camp actually uh is now self-taught AI [3:34] specialist They are the wipe coders of the world.  And the third camp are those who are kind of very [3:39] skeptical of AI. They fear that AI will take the  jobs away. And there's actually sort of this this [3:46] um this process where folks are actually I didn't  sorry I didn't click this where folks are actually [3:54] sabotaging AI uh in the workplace. Okay. And this  is real. I mean you can read articles about this. [4:01] So we have three camps of employees but leaders  want a different outcome as quickly as possible [4:05] whereas staff actually is is divided into three  different camps. So if you look at kind of the [4:13] current state of uh AI in the enterprise I kind of  see two different camps or types of initiatives. [4:19] So we have one population that is actually doing  a lot of active wipe coding that are developing uh [4:25] AI solutions through wipe coding tools through  the latest coding tools and uh that to me is [4:31] nothing but shadow IT with a new hat. So shadow  it in the past used to be where you know we had [4:37] teams that were that had their own databases the  private servers and I've been in organizations [4:42] where there were public facing servers and on on  desktops of of staff members so that is shadow [4:47] IT and shadow AI is very similar to shadow IT but  it it poses a lot of risks unforeseen risks which [4:57] I've kind of tried to highlight here. So one risk,  one significant risk is that you know wipe coding [5:03] is great. You come up with a tool that looks shiny  and you show it to leadership. Uh the leadership [5:08] gets wowed and they say when can we have it? Can  we have it tomorrow? But the danger with that [5:13] is that these vipod tools are not compliant with  your technology stacks. They don't have governance [5:17] control. They don't have identity management,  right? And they they're very brittle. So they've [5:22] been tested on some uh pilot data sets, but once  you roll them out into production, they actually [5:28] start failing. And the burden then falls on the  AI teams to actually maintain those white coded [5:33] apps. Right? So that's something that we have to  control especially now with agentic. Imaginary [5:39] agentic woded app that actually goes rogue and it  actually breaks into enterprise systems creates [5:44] some cyber security hacks or deletes data.  You're talking about a multi-billion dollar [5:50] business going down. Right now the other camp or  the other type of initiatives that we have seen [5:55] uh and I'll delve a bit deeper into this is is  this uh with the rush to prove value. We have a [6:03] lot of teams including AI teams that are creating  these bolt-on I call them bolt-on solutions [6:09] right but they are actually singlepoint solutions  where an AI team looks at a complex workflow they [6:16] identify kind of a lowhanging fruit and they plug  in a solution that they bolt on an AI solution [6:21] the problem with that approach is that even  as you kind of consider this this paradigm of [6:27] well it's creating value for an individual  task end to end it's not really creating [6:32] it's not moving the needle in terms of business  value. So the analogy is kind of a two-lane [6:37] highway which is a legacy process and then you  have this point solution that you're adding on. [6:43] So you now convert this this part of the two-lane  highway into five lane highway but the rest of the [6:48] so that that obviously you get fast faster traffic  there but the rest of the process is still slow. [6:53] So there are bottlenecks and then you have to go  back to the CTO to the CFO and answer why it's [6:59] not creating tangible business value right so  the answer to the first one to wipe coding is [7:06] you bring every AI application onto a federated  platform right you have AI platforms which are [7:12] managed governed uh identity management where you  have observability and so on and you literally [7:17] have to force everyone in the organization to  actually comply with this federated platform [7:22] to make sure that you don't have agents that  go wild. Okay, so that's how you that's my take [7:29] and my perspective on how you deal with wipe coded  apps and after talking to many of the leaders here [7:34] at the conference and even in other conferences  I kind of hear the same story right other [7:38] organizations are taking the same approach. Now if  you look at bolt on uh strategy that's happening [7:46] right now this is nothing new this has happened  before in history. So if you look at uh the the [7:51] shift from steam power to electric power, we have  had instances where you know factories came in [7:58] uh they replace the steam power source with  electric motors but they didn't change the [8:02] layout of the factory. So productivity was very  limited or productivity gains were very limited [8:07] until 40 years later in in uh in the 1920s when  the factory floors were redesigned and rewired [8:14] to accommodate the full power of electricity  and electric motors. The same solution, the [8:20] same analogy now exists with AI because we have  singleoint solutions across the board. We have 100 [8:25] different chat bots in different applications. We  have co-pilot solutions. We have wipe coded apps. [8:30] But there is these are single point solutions.  There is no cohesive rewiring of the system. And [8:35] until we do that rewiring, we're not going to get  the benefits that the leaders are looking for that [8:40] we are looking for as AI leaders. And you have  that you have to have that with governance with u [8:46] you know observability and all the other factors  that go in to create a successful AI solution. [8:51] So diving a little bit deeper into that right  so if you're looking at rewiring an operating [8:56] model there are many components to an operating  model. So one is you know so there are four main [9:02] components uh one is data the second one is um uh  people then it's about platforms and the third and [9:10] the fourth one is about processes so data is the  fuel for the AI engine you have to have AI ready [9:16] data you have to have data that's semantically  searchable discoverable where you have end to [9:20] lineage because then AI has a proper context  platforms is something I talked about you have [9:25] to have federated platforms with security identity  management and then people is a big part of this. [9:31] In larger organizations, you have to upskill the  workforce. [snorts] You have to add AI fluency to [9:36] the to the working uh staff. And uh we are going  through you know champions AI champions network [9:42] where we eval evangelizing AI to different parts  of the organization and making sure that everyone [9:47] is AI ready. They know what role AI is going  to play in the future of work. Um and that's [9:53] a huge part of it. Now process I would say is the  most important part. And why is that? So for the [10:00] most part whether it's uh you know large small or  medium enterprises are the processes that are that [10:07] are typically observable are the ones that AI can  touch and they can actually automate right but the [10:13] reality is that in in most organizations about 20  to 30% of these processes are observable they are [10:21] measurable they're standardized they're digitized  and the remaining 80 to 70% are actually hidden [10:26] somewhere they're opaque they're not visible  to the AI processes they're not standardized [10:30] they they exist through tribal knowledge and  therefore AI cannot impact that uh 70 to 80% [10:36] of the processes so one key part of re rewiring  your organization to create value from AI is [10:44] to be able to make these processes transparent  digitized standardized okay so as we and that's [10:52] one of the key takeaways I want you to remember  which is along with the AI technology rewiring [10:57] of the organization or the process processes is  a key part of the transformation process and and [11:01] creating their business value. The other thing is  I see organizations still kind of living in the [11:06] chatbot world but AI has evolved now to to go from  these reactive chat bots where uh you can just do [11:12] single point single turn Q&A to now agents that  can actually take action right they can complete [11:19] end to end tasks and so therefore you know you  give an a properly designed agent an activity a [11:26] task it'll actually plan it'll uh decide what  to retrieve it'll manage the context with the [11:32] harnesses and then it'll actually complete a  task and that's when you start to create the [11:37] get the business value that the CFO needs or the  leadership wants that you are looking for right [11:42] which is efficiency gains faster operating uh  models um uh getting you know user engagement with [11:50] uh products with AI solutions that are sticky  where the AI is doing 80% of the manual work [11:55] uh that the humans used to do and therefore that  frees up uh you know the users to do what they [12:00] enjoy it the most which is human creativity  talking to people, uh talking to others, [12:04] talking to clients and so on. Uh [snorts] and  the other advantage if you have a good agentic [12:09] solution is that it can actually you can scale  24/7 compared to humans who can actually only work [12:14] 8 hours a day or something, right? Um so creating  an agent is not that easy. It's not just the LLM. [12:22] So the first part of creating a good agent is  that you have to have the right harnesses. So [12:25] you have to have instructions, context, you have  to have MCP tools to actually access uh internal [12:32] data set so that the agents can actually action  on the data. Um and then you once you have an [12:37] agent that can actually take and complete a task,  you can't just deploy it. You have to have the [12:43] right control plane on it. So when you have an  agent that can that's about to complete a task [12:49] or that can complete a task, you have to make sure  that you give it an identity, you give it roles, [12:54] permissions just like you would a junior uh human  assistant, right? Um and once you do that and you [13:02] have to have uh uh the tools to observe the agent  to audit the agent and so on. So that governance [13:08] plane the control plane is absolutely a mustave  to have a successful deployed agent. Right? And [13:13] then once you have that harness and the control  plane, you then have to go to be able to have [13:20] these long running agents that can run for hours,  for days and can complete end to end tasks with [13:26] multiple subtasks and goals in it. Okay. So there  you have to have the right context management [13:32] layer. you you have to have the correct decision  points for the agent to actually say when it's not [13:38] uh you know when it's not confident enough  to to uh come out with an answer and and [13:43] uh basically sort of uh uh you know jog the human  uh uh owner to say hey I need an answer from you [13:49] before I can proceed right you have to have this  controls in place to have these longunning agents [13:54] and once you have all of these the harness the  control plane and the capability and the tools [13:59] to actually have these longrunning activities is  when you then start to create the business value. [14:06] You start to see the revenue growth. You start to  see productivity improvements and so on, right? [14:10] And we're getting there. I mean there are now so  many different solution providers I'm sure you're [14:14] walking around uh who provide those tools. But you  cannot have tools sprawl. So one thing you have [14:19] to be careful about is how do you manage uh you  know multiple vendor solutions uh and what do you [14:25] do inhouse right with coding as a tool available  can you build some of these solutions in house? [14:30] We cannot take I mean I cannot take that call  for you as a leader you have to take that call [14:34] and every organization is unique right but these  are the essential ingredients for you to create [14:38] these longunning agents and start to create to  start to actually measure value the other question [14:44] that I started to get from the CFO is well okay  now you have agents running how do you control [14:51] the cost because agents actually consume a lot of  tokens and tokens are expensive so there you have [14:57] to ruthlessly prioritize your initiatives. You  can't let a thousand flowers bloom. You have to [15:02] say here are the top three initiatives that my  team can with the limited resources can address [15:07] and you pick those initiatives that actually  move the needle with tangible business metrics. [15:12] So whether it's productivity, whether it's uh u  you know operating time and speed or whether it's [15:18] client engagement or user engagement. Define  your metrics, choose your initiatives wisely, [15:23] prioritize them ruthlessly, and then execute and  make sure that you you you have these solutions [15:29] in production as soon as possible. Get out of  the pilot hell where you're piloting things [15:33] for and trying to get things perfect for over  6 months or so. Right? The other thing about AI [15:38] that I I I actually is not on the slide, but  I use to convince uh you know, the CEO and [15:44] the senior leadership is something called the AI  flywheel, which is AI gets better with more data. [15:49] So convincing the leadership that hey even if  the solution is imperfect as long it has as it [15:54] has the right governance and guardrails you can  roll it out and then once you get user data and [15:59] user logs the AI solutions can get better and now  you have you know self-evolving agent loops where [16:06] the agent can learn from the mistakes as long  as the humans are are in the loop and they're [16:09] pointing out the mistakes so that has actually  caught on and you can use the term AFly wheel to [16:14] roll out imperfect solutions as long as you have  the right governance and controls in place Right. [16:19] So the other thing about managing cost is that you  have to use engineering economics there. Right? [16:26] You don't you don't have to use a frontier  model for every single simple task because [16:30] they're expensive. So you have intelligent routing  capabilities that route appropriate task to the [16:35] appropriate models. And what we found is that 80%  of the time you can make do with a smaller model [16:41] and it's only 20% of the time with complex tasks  when you have to really uh use a frontier model [16:46] and that's when the cost go up. But by by doing  this managing tokconomics by managing a context [16:51] window every task does not need a 1 million  context window. Right? So do that effectively and [16:58] and you'll come up with solutions that actually  um uh uh are controlled uh through cost. Right [17:04] now taking a little bit of sneak peek into the  future right so as leaders we have to be future [17:09] ready. you have to build solutions that can evolve  quickly that can pivot as uh AI as a field evolves [17:17] right so I'll spend the next two minutes talking  about what to expect in the future so um a couple [17:24] of things one is that AI capabilities are going  going to get better we'll have agents that can [17:29] do better actions that can do longerterm planning  and reasoning then we'll have AI models that can [17:36] understand processes much better so the concept  of digital twins is something that'll come up and [17:41] I'll try to cover that very quickly. Uh because to  me that's a central piece of having successful AI [17:46] outcomes. Uh and then now we have self-learning AI  uh models that can actually learn and get better [17:53] uh from its own mistakes and from human inputs.  And then we also have computer use agents where [17:59] uh agents can actually control computers. So with  the right governance can you design systems for [18:04] computer use as well? Right. So the this is just a  smattering of some of the near-term technologies. [18:10] This is not short, this is not long-term. This is  kind of things that are coming down the pike and I [18:15] would say uh six to 6 months to 18 months or so,  right? Um and I'll cover digital uh twins a bit [18:23] more. But again, self-evolving and self-learning  AI uh solutions is is going to be the next big [18:28] thing. So if you remember that uh that uh slide  that I showed you about the with the iceberg and [18:35] hidden processes that's where the digital twin  comes in. So if you have a solution where you can [18:42] create a digital model of most of the processes  in the organization right then you can actually [18:49] you can simulate different conditions. You can  find the the bottlenecks ma uh digitally rather [18:54] than have and go to talk to analysts go to talk  to staff members. Right? [snorts] So this digital [19:00] twin concept will actually give you a lot of  wins. It'll accelerate orchestration processes. [19:04] It'll identify bottlenecks. It'll allow you to  actually simulate different AI solutions digitally [19:09] before you roll them out operationally, which  is a big cost savings for you. Right? For us, [19:13] integrating AI solutions is a huge cost. It's  not developing the solutions, it's integrating [19:18] them in, right? So by by having digital twins, you  basically cut down your cost of change management [19:24] because most of your changes are done digitally.  the tools to create digital twins actually are are [19:31] getting more and more mature and I believe when  we as we have omnimodal models that can measure [19:37] multiple channels from the enterprise signals  we'll then be able to uh you know transform [19:43] them digitally create digital models or digital  components of each of the processes and then [19:48] ultimately end up creating a a pretty robust  digital twin that then can where you can do [19:53] the simulations you can accelerate the  bottlenecks overcome the bottlenecks [19:57] and then uh basically roll out solutions much more  much more easily now beyond the 18 months that I [20:05] just talked about what's coming down the pipe  about two to three years from now right and I [20:10] want to leave you with that vision so with the  rapid evolution in AI uh one here are sort of [20:18] my big picture ticket items so one is AI chips  where the cost of AI uh you have AI solutions [20:23] on chips with low power and it doesn't need  constant cloud connectivity. So to me, you know, [20:29] AI chips is a big thing that'll cut down the cost  of AI deployment. Embodied intelligence which is [20:34] AI embod uh you know embedded in physical robots  and interacting with the real world, learning from [20:39] the real world that is the next big thing. Um and  then um you know world models, omnimodel models [20:47] uh which interact with the world, which learn from  the world uh is another thing. And then continual [20:52] learning is a term you have heard in papers now  but that is going to become a reality in the next [20:58] two to three years. And then we'll have you know  autonomous agents not only in the in the digital [21:03] world but autonomous agents in the physical world  and that itself will actually create new we heard [21:09] a little bit of that in today's talk. I was  hoping to hear that in the keynote more today [21:12] but I think we'll have new industries come up.  We'll have robotics as a service. We'll have uh uh [21:18] digitally native companies like pharma companies,  manufacturing firms, materials firms. So new [21:24] materials will be discovered and when you have and  that itself will have a cascading effect, right? [21:28] It'll have a flywheel effect. Well, new materials  will lead to better space travel or better quantum [21:33] technologies and so on. So these are all not  independent technologies, but really it'll it'll [21:38] uh rise up and create new opportunities for  everyone. Right? So my my focus here is not [21:44] to talk about jobs per se but new industries  that'll be created and then digital economies will [21:49] uh you know digital commerce will come up in a big  way and [snorts] then the organizational structure [21:53] will also change. Uh we already will we are moving  into the era of solopreneurs where we have single [22:00] person billion dollar companies. We'll have  smaller teams uh uh you know 10 size teams that [22:05] are tens of billions of dollars in valuation  because you have agents doing a lot of the [22:09] administrative uh work right and team structures  within large organizations will change as well [22:14] uh we'll have flattered organizations and  outcomedriven teams rather than sort of [22:19] hierarchical fixed teams that we have now right  so my time is up but I want you to leave with a [22:25] big picture of where I see AI moving as a field  and how you as leaders need to plan for Thank you.