AI Agent Adoption Surges 233% as Multi-Agent Teams Cut Labor Costs

The Economic and Social Impact of AI With Economic Strategy Institute — Ai4

Ai4

Rapidly declining compute expenses and multi-agent systems are shifting corporate operations toward autonomous 24/7 workflows that deliver full productivity at a fraction of human labor costs.

AI-assisted summary of The Economic and Social Impact of AI With Economic Strategy Institute — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 29, 2025, this presentation features Robert Cohen, an AI Industry Analyst and Economist from the Economic Strategy Institute. The session examines how agentic AI architectures act as capital deepening, reshape enterprise competitiveness, and establish national economic advantages.

## TL;DR Economist Robert Cohen outlines the transition into the 'Turing economy,' driven by the rapid growth of cross-functional multi-agent AI systems across global enterprises. Deflationary compute costs and interoperability protocols are reorganizing workflows in finance, procurement, and pharmaceuticals, while necessitating robust supervisor-agent observability layers. Cohen addresses the evolving economics of labor, noting that future organizational demand will shift away from syntax programming toward business logic, agent orchestration, and economic ROI measurement.

## Atmosphere & style The presentation carries the pragmatic, analytically rigorous atmosphere of an enterprise strategy briefing. Cohen speaks with measured authority and accessible candor, grounding forward-looking macroeconomic concepts with practical industry benchmarks and occasional cultural humor. The delivery balances technological optimism with sober deliberation on labor displacement, productivity metrics, and institutional governance.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) The transition from static AI models to autonomous, reasoning agentic AI marks the emergence of the 'Turing economy,' where machine cognition becomes deeply integrated across business operations. - [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) Enterprise AI agent utilization surged by 233% in the first half of 2025, with 40% of adopting firms deploying agents across cross-functional operations. - [04:23](https://www.youtube.com/watch?v=KW-NPTqloqo&t=263s) Compute costs for running advanced models are experiencing an aggressive deflationary trend, dropping by an estimated factor of two annually as efficient GPU architectures emerge. - [04:23](https://www.youtube.com/watch?v=KW-NPTqloqo&t=263s) Enterprise governance increasingly relies on an 'AI observability' layer where dedicated supervisor agents monitor worker agents in real time to catch hallucinations and logic conflicts. - [06:52](https://www.youtube.com/watch?v=KW-NPTqloqo&t=412s) Core enterprise workflows—such as procurement, contract negotiations, and compliance monitoring—are being automated and coordinated via frameworks like Model Context Protocol (MCP). - [06:52](https://www.youtube.com/watch?v=KW-NPTqloqo&t=412s) Agentic AI adoption in life sciences and pharmaceuticals has drastically accelerated R&D, compressing initial drug discovery timelines from 4–5 years to under 12 months. - [11:46](https://www.youtube.com/watch?v=KW-NPTqloqo&t=706s) Operating a multi-agent team costs approximately 20% to 25% of a full-time human employee's compensation while providing continuous 24/7 productivity. - [11:46](https://www.youtube.com/watch?v=KW-NPTqloqo&t=706s) Enterprises like Citigroup are measuring AI agent ROI by the acceptance rate of generated outputs into production builds, setting an 80% to 90% merge rate as a high-value benchmark. - [15:35](https://www.youtube.com/watch?v=KW-NPTqloqo&t=935s) The transition to low-code and no-code environments shifts indispensable workforce skills away from low-level coding toward multi-agent orchestration, contextual oversight, and ROI modeling. - [15:35](https://www.youtube.com/watch?v=KW-NPTqloqo&t=935s) Agentic productivity gains may compress the standard workweek to 30–35 hours, while highlighting the need for safety nets for older workers who face friction adapting to digitized workflows.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) Entering the "Turing Economy" and the Rise of Cross-Functional Agents

Robert Cohen of the Economic Strategy Institute introduces his analysis of the economic and social ramifications of artificial intelligence, noting with a lighthearted remark that the conference agenda shifted his speaking slot forward by ten minutes within the last 24 hours. Cohen, an economist specializing in how industries adopt emerging technologies through detailed case studies of early-adopter enterprises, frames the current era as a fundamental shift from previous static AI models to agentic AI systems capable of gathering data, reasoning, and autonomously making operational decisions. He argues that this transition marks the dawn of what he terms the "Turing economy." This builds upon Alan Turing's 1950 prediction that by the year 2000, society would speak naturally of machines thinking in ways indistinguishable from human beings.

The adoption curve for agentic AI is accelerating at an unprecedented pace. In 2025, an estimated 40% of enterprises deploying agentic AI are utilizing it in a cross-functional capacity rather than confining it to isolated tasks like basic risk management. In practical terms, an agentic system in financial services can detect an anomaly—such as a sudden spike in the Russian ruble on a Friday—and immediately cross-correlate that event across the entire enterprise, alerting relevant desks, adjusting exposure, and cross-checking decisions against other institutional data sources to avoid errors. Across the broader economy, enterprise AI agent utilization expanded by 233% in the first six months of the year alone. Evoking the classic line from *The Wizard of Oz*, "Toto, I've got the feeling we're not in Kansas anymore," Cohen emphasizes that the structural reality of the corporate world has changed, even as conventional macroeconomic forecasters remain overly conservative in their projections for GDP and labor productivity growth.

- AI agent adoption across enterprise workflows grew by 233% in the first half of 2025. - 40% of firms utilizing agentic AI deploy it across cross-functional operations to integrate insights across risk, finance, and trading desks. - The concept of the "Turing economy" reflects Alan Turing’s 1950 thesis that machine intelligence would become functionally indistinguishable from human cognition.

---

[04:23](https://www.youtube.com/watch?v=KW-NPTqloqo&t=263s) Dynamics of Multi-Agent Systems, Observability, and Plummeting Compute Costs

In the Turing economy, substantial productivity leaps occur because agents do not operate in isolation; they function in coordinated agent teams. These teams identify required datasets, analyze operational processes in minute detail, and interact dynamically to resolve complex, multi-stage problems. Simultaneously, the underlying capital costs—encompassing software development, energy consumption, and infrastructure—are declining sharply. Compute expenses are following an aggressive deflationary curve reminiscent of the early semiconductor industry, experiencing steady step-function declines. Pointing to industry projections, Cohen notes that Nvidia CEO Jensen Huang has argued compute costs are dropping by a factor of two each year as high-efficiency GPU architectures come online.

A crucial, often underappreciated pillar of this agentic architecture is the necessity of robust governance frameworks and "AI observability." Aggressive enterprise adopters are deploying dedicated supervisor agents tasked specifically with monitoring worker agents and multi-agent teams. These observability layers scrutinize agentic decision paths in real time to detect hallucinations, flag logic conflicts, and catch flawed outputs before they execute.

``` [ Core Enterprise Data & Context Protocols (MCP) ] │ ┌────────────────────┴────────────────────┐ ▼ ▼ [ Agent Team Alpha ] [ Agent Team Beta ] (e.g., Risk & Compliance) (e.g., Logistics & Supply) │ │ └────────────────────┬────────────────────┘ ▼ [ AI Observability Layer ] (Supervisor Agents: Hallucination & Conflict Checks) │ ┌──────────────┴──────────────┐ [ Passed: Autonomous Action ] [ Flagged: Human-in-the-Loop ] ```

- Compute costs for running advanced models are dropping along an aggressive deflationary curve, estimated by Nvidia's Jensen Huang to decrease twofold annually with new GPU architectures. - Observability agents act as algorithmic guardrails, supervising peer agents to identify hallucinations, biased decisions, and operational conflicts. - Human-in-the-loop support teams are positioned downstream of the observability layer to adjudicate edge cases and correct flagged actions.

---

[06:52](https://www.youtube.com/watch?v=KW-NPTqloqo&t=412s) Autonomous Corporate Workflows and Industrial Case Studies

As agentic intelligence permeates the enterprise, corporate operating structures are undergoing fundamental reorganization. Cohen highlights how distinct functional roles are being delegated to autonomous agents across the entire value chain. In procurement, finance, and retail, agents now handle supplier onboarding, real-time performance monitoring, end-to-end supply chain management, autonomous contract negotiations, and contract lifecycle administration. These workflows are increasingly orchestrated using standardized interoperability frameworks like Model Context Protocol (MCP) and dedicated agent-to-agent communication channels.

Consultancies like Accenture are actively implementing tiered, multi-agent frameworks for enterprise clients, deploying specialized base-level agents to automate high-volume compliance checks, such as scanning cross-border financial transactions across Eastern Europe and Latin America for fraudulent behavior. In the pharmaceutical sector, the competitive imperative is even starker. Companies such as AstraZeneca are deploying agentic AI across drug discovery pipelines, compressing the initial target identification phase from the traditional four-to-five-year timeline down to a single year or less. For pharmaceutical giants relying on existing blockbuster revenues from therapeutics like Ozempic or Keytruda, failing to adopt these autonomous R&D workflows risks complete obsolescence within half a decade.

- Key functional domains undergoing agentic automation include supplier onboarding, autonomous price and terms negotiations, performance tracking, and contract lifecycle management. - Integration architectures are increasingly built on emerging standards such as Model Context Protocol (MCP) and direct agent-to-agent interfaces. - In life sciences, agentic AI has compressed the early-stage drug discovery timeline from 4–5 years down to under 12 months.

---

[11:46](https://www.youtube.com/watch?v=KW-NPTqloqo&t=706s) Social Transformations, Value Metrics, and the Economics of Labor

The widespread diffusion of autonomous systems introduces profound societal and institutional dilemmas. Cohen outlines the fundamental questions facing the near future: how aggregate labor demand will shift, what will happen to worker compensation, whether Universal Basic Income (UBI) programs will become necessary, and whether corporate structures will flatten or become dominated by elite technical hierarchies ("Will tech bros run the world?"). Beyond social anxieties about existential risk or rogue systems generating harmful outputs, organizations are currently grappling with how to quantify the financial surplus generated by agents relative to human labor.

Enterprise adopters are beginning to institute rigorous performance metrics to assess agent contributions. At Citigroup, where agents assist in software engineering, productivity is tracked by measuring the acceptance rate of agent-suggested code into final production builds. If an agent's code suggestions only achieve a 5% to 10% acceptance rate, its utility is negligible; if the acceptance rate reaches 80% to 90%, the value creation is massive. Furthermore, enterprises are modeling staffing structures where small agent teams—costing roughly one-fourth to one-fifth of a full-time human employee—work alongside human staff, delivering continuous 24/7 productivity. These economic advantages will expand as foundation models (such as GPT-5) place heavier emphasis on mitigating hallucinations and enhancing architectural reliability.

- Citigroup evaluates coding agents by the proportion of generated code merged into production builds, establishing a benchmark where 80–90% acceptance represents high enterprise value. - Operating costs for a multi-agent team are estimated at roughly 20% to 25% of the compensation of a single full-time employee, paired with continuous 24/7 operational capability. - Newer foundation model architectures are prioritizing hallucination minimization to make agent deployments economically defensible in high-stakes environments.

---

[15:35](https://www.youtube.com/watch?v=KW-NPTqloqo&t=935s) Q&A: Human Purpose, Evolving Workweeks, and Future Enterprise Skills

Opening the floor to audience questions, Cohen addresses a philosophical inquiry regarding human purpose and the future of work if automated agents handle the bulk of labor. He suggests that agentic tools will actually democratize job mobility rather than restrict it to elite computer science PhDs. As coding shifts toward low-code and no-code abstractions, the nature of employment will pivot toward business logic, economic decision-making, and contextual oversight. While standard workweeks may compress to 30 or 35 hours, Cohen acknowledges the necessity of robust social safety nets for older workers—such as a 60-year-old without a technical or mathematical background—who may struggle to transition into highly digitized workflows.

In response to a final question regarding the organizational skill sets that will be indispensable over the next one to five years, Cohen highlights the growing need for professionals who can measure the economic ROI of specific AI workflows and manage enterprise-wide agent networks. Networking distinct multi-agent systems across dispersed business units will yield major operational synergies. Cohen contrasts the gloomy unemployment forecasts of figures like Anthropic's CEO with the viewpoint of Salesforce's CEO Marc Benioff. Salesforce has noted that while traditional software engineering hiring has stabilized, the company is actively expanding headcount for strategic, analytical, and business-focused roles that direct agentic systems.

- Human-in-the-loop roles will increasingly center on AI observability, model output verification, and operational exception handling via low-code platforms. - Macroeconomic adjustments may drive standard full-time employment down to a 30- to 35-hour workweek. - High-demand future skills will prioritize workflow cost accounting, ROI modeling, and multi-agent network orchestration rather than pure syntax programming. - Perspectives on labor displacement remain sharply divided between AI lab executives predicting mass unemployment (e.g., Anthropic) and enterprise software leaders noting net hiring growth in strategic roles (e.g., Salesforce).

## Notable quotes > [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) Toto, I've got the feeling we're not in Kansas anymore — Robert Cohen

> [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) AI agent adoption across enterprise workflows grew by 233% in the first half of 2025. — Robert Cohen

> [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) 40% of firms utilizing agentic AI deploy it across cross-functional operations to integrate insights across risk, finance, and trading desks. — Robert Cohen

> [04:23](https://www.youtube.com/watch?v=KW-NPTqloqo&t=263s) Compute costs for running advanced models are dropping along an aggressive deflationary curve, estimated by Nvidia's Jensen Huang to decrease twofold annually with new GPU architectures. — Robert Cohen

> [04:23](https://www.youtube.com/watch?v=KW-NPTqloqo&t=263s) Observability agents act as algorithmic guardrails, supervising peer agents to identify hallucinations, biased decisions, and operational conflicts. — Robert Cohen

> [06:52](https://www.youtube.com/watch?v=KW-NPTqloqo&t=412s) In life sciences, agentic AI has compressed the early-stage drug discovery timeline from 4–5 years down to under 12 months. — Robert Cohen

> [11:46](https://www.youtube.com/watch?v=KW-NPTqloqo&t=706s) Will tech bros run the world? — Robert Cohen

> [11:46](https://www.youtube.com/watch?v=KW-NPTqloqo&t=706s) Operating costs for a multi-agent team are estimated at roughly 20% to 25% of the compensation of a single full-time employee, paired with continuous 24/7 operational capability. — Robert Cohen

> [11:46](https://www.youtube.com/watch?v=KW-NPTqloqo&t=706s) Citigroup evaluates coding agents by the proportion of generated code merged into production builds, establishing a benchmark where 80–90% acceptance represents high enterprise value. — Robert Cohen

> [15:35](https://www.youtube.com/watch?v=KW-NPTqloqo&t=935s) Macroeconomic adjustments may drive standard full-time employment down to a 30- to 35-hour workweek. — Robert Cohen

> [15:35](https://www.youtube.com/watch?v=KW-NPTqloqo&t=935s) High-demand future skills will prioritize workflow cost accounting, ROI modeling, and multi-agent network orchestration rather than pure syntax programming. — Robert Cohen

## Mentioned Robert Cohen, Economic Strategy Institute, Alan Turing, The Wizard of Oz, Jensen Huang, Nvidia, Model Context Protocol (MCP), Accenture, AstraZeneca, Ozempic, Keytruda, Citigroup, GPT-5, Marc Benioff, Salesforce, Anthropic

## Video description Robert Cohen, AI Industry Analyst and Economist, Economic Strategy Institute

Abstracts

AI is reshaping economic landscapes by enhancing productivity, accelerating value creation, and redefining industrial and national competitiveness. This talk assesses AI’s role in unlocking new forms of capital efficiency, increasing returns to scale, and strengthening national strategic advantage in the global economy as well as how these changes will change society. AI acts as a form of capital deepening: intelligent agents and machine learning models embody knowledge that can be replicated and scaled across markets. agentic AI architectures transform decision-making, linking operational data to strategic insights across manufacturing, logistics, and customer engagement. Cognizant Corporations with ubiquitous AI agents in their operations demonstrate how agentic systems can amplify competitiveness through continuous innovation and real-time decision-making. Nations investing in AI infrastructure and workforce adaptation will achieve sustained economic advantages, akin to the role of industrialization in past centuries.

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

中文简报

多智能体降低人力成本，AI智能体采用率激增233% 算力开支的迅速下降与多智能体系统正推动企业运营转向全天候自主工作流程，仅需人类劳动力成本的一小部分即可释放全部生产力。 经济学家罗伯特·科恩（Robert Cohen）概述了向“图灵经济”的转型过程，这一转变主要由全球企业跨职能多智能体AI系统的快速增长所推动。持续走低的算力成本与互操作性协议正在重组金融、采购和制药行业的工作流程，同时也催生了对稳健的主管智能体可观测层的必要需求。科恩探讨了不断演变的劳动力经济学，并指出未来组织的需求将从语法编程转向业务逻辑、智能体编排以及经济投资回报率（ROI）的衡量。 2025年上半年企业级 AI Agent 的采用率激增233%，且40%部署该技术的企业已将其应用于跨职能业务，标志着机器认知深度融入运营的“图灵经济”正式开启。 (0:04) 得益于高效 GPU 架构，运行前沿模型的算力成本正以每年约两倍的速度快速下降；同时企业高度依赖“AI 可观测性”层，通过主管 Agent 实时监控工作 Agent 的幻觉与逻辑冲突。 (4:23) 采购、合同谈判和合规等核心业务流正加速基于 Model Context Protocol (MCP) 等框架实现自动化，医药领域更借助 AI Agent 将初期药物发现周期从4至5年压缩到12个月以内。 (6:52) 多智能体团队的运营成本仅为一名全职人类员工的20%至25%并提供24/7连续产能，花旗集团等企业正将80%至90%的代码采纳率作为衡量 Agent 投资回报率的核心基准。 (11:46)

Reading summary

Robert Cohen of the Economic Strategy Institute argues that enterprise adoption of cross-functional, autonomous AI agents marks the start of a "Turing economy," outstripping conservative macroeconomic forecasts and reorganizing corporate workflows. The Rise of Cross-Functional Agentic Systems (0:04) Cohen frames the shift from static AI models to autonomous agentic systems as the emergence of the "Turing economy," fulfilling Alan Turing's 1950 prediction of machine intelligence matching human cognition. Rather than being confined to single, isolated tasks, enterprise adoption expanded by 233% in the first half of 2025. Approximately 40% of firms deploying agentic AI now use it in a cross-functional capacity, enabling real-time correlation across risk, finance, and operational desks to adjust exposures and resolve multi-stage problems. Declining Compute Costs and Governance Layers (4:23) The economic feasibility of agentic AI is accelerating due to deflationary hardware costs. Citing Nvidia CEO Jensen Huang, Cohen notes that compute expenses are dropping by roughly a factor of two each year with high-efficiency GPU architectures. Multi-agent teams operate collaboratively rather than in silos, supported by interoperability frameworks like the Model Context Protocol (MCP) to manage end-to-end tasks. Industrial Restructuring and Observability Guardrails (6:52) Enterprises are restructuring functional operations around autonomous agents across supply chain management, supplier onboarding, and contract negotiations. For instance, Accenture is deploying multi-agent frameworks for high-volume fraud and compliance checks across Eastern Europe and Latin America, while pharmaceutical companies like AstraZeneca have shortened early-stage drug discovery from 4–5 years to under 12 months. To prevent systemic errors, enterprises deploy dedicated supervisor agents in an observability layer to flag hallucinations and logic conflicts before passing unresolved exceptions to human adjudicators. Labor Economics, Productivity Metrics, and Skills (11:46) The economic impact on labor hinges on cost efficiency and performance verification. Cohen reports that operating a small multi-agent team costs approximately 20% to 25% of a full-time human employee's compensation while providing continuous 24/7 capacity. Enterprise adopters measure utility through concrete benchmarks, such as Citigroup tracking code acceptance rates, where an 80% to 90% merge rate into production builds represents high enterprise value. Uncertainty remains regarding the broader societal impact on employment, with contrasting views between AI lab executives predicting widespread displacement and software leaders like Salesforce observing net hiring growth in strategic roles. Cohen anticipates a potential shift toward a 30- to 35-hour workweek, emphasizing that future high-demand skills will center on ROI modeling, multi-agent orchestration, and governance rather than basic coding, while older non-technical workers may require targeted social safety nets.

精炼阅读摘要

美国经济战略研究所（Economic Strategy Institute）经济学家罗伯特·科恩（Robert Cohen）指出，企业人工智能正从静态工具演进为具备自主推理与决策能力的跨职能多智能体系统，标志着“图灵经济”（Turing economy）的开启。这一变革推动了算力成本的大幅通缩和跨行业研发周期的极限压缩，但同时也对传统劳动力结构、绩效评估及企业治理带来了深远挑战。 “图灵经济”到来与跨职能智能体扩张 (0:04) 科恩指出，当前的技术浪潮呼应了阿兰·图灵（Alan Turing）在1950年作出的预言，即机器智能在功能层面将达到与人类认知难以区分的状态，从而催生由自主决策系统主导的“图灵经济”。与过去局限于数据检索或单一任务的静态AI不同，具备自主行动能力的智能体（Agentic AI）能够主动收集数据、推演业务逻辑并独立做出运营决策。数据显示，2025年上半年，企业级AI智能体的使用规模激增了233%，且在已部署该技术的企业中，有40%将其应用于跨部门协同场景，而非局限于孤立的基础风控任务。 在金融服务等行业中，跨职能智能体展现出了全局关联分析的能力。例如，当系统在周五监测到俄罗斯卢布汇率异常波动时，能够迅速在企业全盘业务中进行交叉验证，主动向风险、财务与交易部门同步预警，动态调整风险敞口，并比对机构内部数据以规避执行错误。科恩强调，传统宏观经济预测机构在预估GDP和劳动生产率增速时普遍过于保守，未能充分捕捉到企业底层运营现实中已经发生的结构性剧变。 多智能体协同、算力通缩与可观测性治理 (4:23) 生产率的大幅飞跃源于多智能体团队的协同作战，而非孤立模型的单点运行。智能体团队能够自主识别必要的数据集，微观拆解作业流程，并通过多轮动态交互攻克多阶段的复杂业务。与此同时，支撑智能体运转的基础设施成本正在经历类似早期半导体行业的剧烈通缩。科恩援引英伟达（Nvidia）首席执行官黄仁勋（Jensen Huang）的行业预测指出，随着高能效GPU架构的迭代部署，运行先进模型所需的算力成本正以每年减半（成本降低两倍）的阶跃速度持续下降。 然而，自主智能体的广泛使用也使得“AI可观测性”（AI Observability）治理架构成为刚需。前沿企业正部署专门的监督智能体（Supervisor Agents），构建起算法层面的防护栏，实时审查作业智能体团队的决策链路。这套治理机制旨在执行前自动拦截逻辑冲突、偏见和模型幻觉；对于被标记的高风险或边缘异常案例，系统则交由下游的“人机协同”（Human-in-the-loop）支持团队进行最终人工裁决。 业务流程重组与研发周期的极限压缩 (6:52) 伴随智能体技术的落地，企业的运营结构正在整个价值链上发生重构。在采购、财务与零售等关键环节，自主智能体已接管供应商入驻核查、实时绩效监控、端到端供应链管理以及自主价格与条款谈判等事务。这些复杂的协作流程正日益依托模型上下文协议（Model Context Protocol, MCP）等新型标准化互操作协议以及专用的智能体间通信通道（Agent-to-agent interfaces）来平稳运转。 具体案例方面，埃森哲（Accenture）等咨询公司正为大型企业搭建分层的多智能体框架，利用专用底层智能体自动化扫描东欧和拉丁美洲等地区的跨境金融交易，高频排查潜在欺诈。在生命科学领域，阿斯利康（AstraZeneca）等制药巨头将智能体AI融入药物发现管线，将早期的药物靶点识别周期从传统的4至5年压缩至1年以内。科恩认为，对于依赖重磅药物（如Ozempic或Keytruda）销售额的制药企业而言，若未能在研发中及时融入自主工作流，恐将在未来五年内面临被市场彻底淘汰的风险。 劳动力经济学、价值评估与组织技能转型 (11:46) 自动化流程的大规模铺开引发了关于全社会劳动力总需求演变、薪酬水平变动以及是否推行全民基本收入（UBI）等深层社会讨论。为了精准量化智能体相对人类劳动力带来的财务增量，领先企业开始建立严格的量化指标。例如花旗集团（Citigroup）在软件工程实践中，通过衡量智能体生成的代码合入生产环境的采纳率来评估效能：若采纳率仅为5%至10%，其实际价值微乎其微；而若采纳率达到80%至90%，则代表巨大的生产力释放。此外，企业测算表明，一个多智能体协作团队的运营成本仅占一名全职人类员工薪酬的20%至25%左右，且具备全天候不间断运转的能力。随着下一代基础模型（如GPT-5）更加侧重抑制幻觉和提升可靠性，此类部署的经济性将更加确固。 在技能需求与就业前景上，科恩反驳了技术将仅由极少数技术精英垄断的观点，认为低代码与无代码抽象化正在降低技术门槛，未来岗位将向业务逻辑定义、经济ROI测算以及跨部门智能体网络协调等综合方向转移，标准全职工作周亦可能因此缩短至30至35小时。关于就业冲击，业内存在鲜明对立：Anthropic首席执行官等AI实验室高层预警可能出现大规模失业，而Salesforce首席执行官马克·贝尼奥夫（Marc Benioff）等企业软件领导者则指出，传统编程招聘趋稳的同时，负责指挥与分析智能体系统的战略性岗位正在积极扩招。科恩同时提醒，必须建立配套的社会安全兜底体系，以协助缺乏数字技能的高龄劳动者平稳度过转型期。 科恩总结认为，企业正不可逆转地跨入以多智能体协同和算力通缩为基础的“图灵经济”，这一演进正在重塑研发与业务效率的底层逻辑。然而，宏观经济统计与预测的迟滞、量化智能体ROI的指标体系构建，以及对传统雇佣体系的深远冲击，依然是企业决策者和全社会亟待应对的核心议题。

Detailed notes — zh

## Background 本视频由 Ai4 于 2025 年 8 月 29 日发布在 YouTube 平台上，主讲人为经济战略研究所（Economic Strategy Institute）的 AI 行业分析师兼经济学家 Robert Cohen。根据内容简介，该演讲主要评估了人工智能（尤其是智能体 AI）如何通过提升资本效率、重塑决策机制与跨职能运作，推动产业与国家竞争力的深刻变革，并对未来的社会结构带来深远影响。 ## TL;DR Robert Cohen 指出全球正迈入由智能体深度协作驱动的“图灵经济”，跨职能智能体在企业中的渗透率呈指数级跃升。随着算力成本快速通缩，企业正借助多智能体通信协议与 AI 可观测性护栏重构组织架构，将药物研发等复杂流程大幅压缩，并在金融等行业确立了“人机 1:4 混合协同”的全新成本产出比。面对就业与社会冲击的疑虑，他认为技术门槛的降低将释放更多业务创新与战略管理岗位，推动劳动时长缩短，但同时也必须为受冲击群体建立坚实的社会安全保障网。 ## Atmosphere & style 整场演讲洋溢着严谨、务实且兼具宏观视野的经济学分析氛围。Cohen 谈吐从容沉稳，既能引经据典并运用《绿野仙踪》式的生动隐喻点明技术范式转移的不可逆，又能迅速切入代码采纳率、人机成本比等冷峻的量化指标；在问答环节面对关于人类生存意义与社会阵痛的提问时，他的解答理性克制且充满建设性，展现出技术剧变前夜先锋学者的笃定与冷静。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) 经济正在从传统 AI 模型向具备自主思考与决策能力的智能体（AI Agents）演进，迈入全新的“图灵经济”（Turing Economy）时代。 - [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) AI 智能体迎来爆发式增长，2025 年上半年使用量激增 233%，已有 40% 的企业将其部署于跨职能业务协同中。 - [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) 主流宏观经济学家对 AI 提振 GDP 与生产率的预测偏向保守，但微观企业层面的智能体落地与重构已呈现出指数级推进态势。 - [04:23](https://www.youtube.com/watch?v=KW-NPTqloqo&t=263s) 底层算力分析成本以年均约 2.5% 的斜率稳步下降，新一代高能效 GPU 架构更推动算力边际成本向极低水平演进。 - [04:23](https://www.youtube.com/watch?v=KW-NPTqloqo&t=263s) “AI 可观测性”（AI Observability）成为企业多智能体落地的必备护栏，用于全天候监控逻辑冲突与模型幻觉，并配合人机回环机制兜底风险。 - [06:52](https://www.youtube.com/watch?v=KW-NPTqloqo&t=412s) 企业组织架构全面重塑，智能体深度接管供应商准入、绩效监测、自主谈判及合同全生命周期管理等中后台业务。 - [06:52](https://www.youtube.com/watch?v=KW-NPTqloqo&t=412s) 制药行业展现出显著的智能体提效范式，阿斯利康等先锋药企将早期药物发现与靶点筛选周期从 4~5 年压缩至 1 年甚至数月以内。 - [11:49](https://www.youtube.com/watch?v=KW-NPTqloqo&t=709s) 企业开始精准量化智能体边际价值，大型金融机构推行“1 名员工搭配 4~5 个智能体”的混合架构，以 1/5 的成本实现 7×24 小时不间断产能。 - [15:35](https://www.youtube.com/watch?v=KW-NPTqloqo&t=935s) 低代码与无代码工具大幅降低了技术准入门槛，提升了跨行业职业流动性，商业嗅觉与业务理解深度正取代纯技术背景成为核心竞争力。 - [15:35](https://www.youtube.com/watch?v=KW-NPTqloqo&t=935s) 全社会工作周未来有望缩减至 30~35 小时，但在转型阵痛期需为年长及非技术劳动力建立坚实的社会安全保障网络。 - [18:27](https://www.youtube.com/watch?v=KW-NPTqloqo&t=1107s) 未来 1 至 5 年企业极其渴求能够量化业务流改造成本与 ROI、精通跨职能网络化统筹的复合型商业管理人才。 - [18:27](https://www.youtube.com/watch?v=KW-NPTqloqo&t=1107s) AI 的发展并非单纯消灭工作岗位，而是在淘汰机械式研发与中后台流程的同时，将劳动力全面推向高维度的战略设计与业务运营领域。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) 迈向“图灵经济”：智能体驱动的跨职能演进 经济战略研究所（Economic Strategy Institute）的经济分析师 Robert Cohen 在开场时首先解释了议程时间的临时调整，随后切入主题：人工智能对经济与社会的深远影响。作为一名长期追踪行业技术落地的经济学家，Cohen 指出，当前我们正处于从传统 AI 模型向具备“思考与自主决策能力”的智能体（AI Agents）过渡的关键节点。这一范式转移正在重塑整个经济运作模式。他将当下正在形成的全新体系定义为“图灵经济”（Turing Economy），以此致敬艾伦·图灵（Alan Turing）在 1950 年提出的预言——即到 2000 年，人们谈论机器时将认为其行为与真正的人类思考者已毫无二致。 在图灵经济中，智能体的普及速度极其迅猛。根据 Cohen 引述的行业统计，在今年上半年，AI 智能体的使用量暴增了 233%；到 2025 年，已有 40% 采用智能体 AI 的企业将其部署在跨职能（Cross-functional）业务中。这种跨职能协同打破了以往孤立的业务壁垒。例如在银行风控场景中，当智能体监测到周五俄罗斯卢布汇率出现剧烈异动时，系统不再仅仅发出局部预警，而是能够跨部门联动并综合研判：如果某个智能体做出了激进的抛售决策，其他接入不同数据源的智能体能够迅速进行交叉校验与纠偏，最终形成最优决策。面对这种剧变，Cohen 引用《绿野仙踪》的经典台词感叹道：“托托，我觉得我们已经不在堪萨斯了（Toto, I've got the feeling we're not in Kansas anymore）”，以此强调传统经济分析模型对 GDP 和生产率预测的保守态度已难以跟上技术突变。 - **图灵经济核心驱动**：智能体以集群形式协同运作，自主识别所需数据、精细化拆解业务流程，并跨职能完成复合型任务。 - **智能体爆发增速**：2025 年上半年 AI 智能体应用规模增长 233%，40% 的企业已将其应用于跨业务协同。 - **保守预期的反差**：主流宏观经济学家对 AI 提振 GDP 与生产率的预测普遍偏向谨慎，但底层微观企业的落地速度已呈现指数级特征。 ```mermaid flowchart TD accTitle: 智能体跨职能协同与可观测性架构 subgraph MultiAgentSystem["智能体多维协作系统"] A["风控监测智能体"] -->|"汇率异动/欺诈预警"| B["跨职能决策智能体"] C["数据整合智能体"] -->|"多源市场数据"| B B -->|"执行初步决策"| D["业务执行系统"] end subgraph ObservabilityLayer["AI 可观测性与治理"] E["AI 可观测性智能体"] -->|"实时监控幻觉与冲突"| B E -->|"异常冲突升级"| F["人类专家介入 (Human-in-the-Loop)"] F -->|"纠偏指令与规则优化"| D end ``` --- ### [04:23](https://www.youtube.com/watch?v=KW-NPTqloqo&t=263s) 算力成本骤降与“AI 可观测性”护栏的兴起 探讨图灵经济的成本结构时，Cohen 指出企业的技术支出主要分布在软件开发、计算基础设施与电力能源上。伴随着底层计算成本的持续下行，AI 智能体的部署成本正在发生类似早期半导体行业的剧烈通缩。从长期趋势线来看，算力分析成本正以约 2.5% 的斜率稳步下降；而英伟达（NVIDIA）CEO 黄仁勋（Jensen Wang）甚至给出了更为激进的预测，认为在新一代高能效 GPU 架构的推动下，计算成本每年将实现翻倍降低（即性能功耗比翻倍）。 然而，成本骤降与智能体自主性的增强带来了一个此前被严重忽视的全新挑战：系统的安全护栏与监管。Cohen 强调，走在落地前沿的先锋企业正将大量精力投向“AI 可观测性”（AI Observability）。在复杂的智能体网络中，企业必须部署专门的审查智能体，全天候监视业务团队与基础智能体的决策链路，精准捕捉模型幻觉、逻辑漏洞或客观存在偏误的指令。一旦不同智能体之间产生决策冲突，系统需及时标记并建立一套清晰的降级与人工接入机制，由人类专家最终拍板，确保自动化系统不会给企业带来灾难性风险。 - **算力成本通缩**：计算成本呈现类似摩尔定律的下行曲线，新一代高算力 GPU 的能效提升加速了企业智能化边际成本的归零。 - **可观测性中枢**：AI Observability 成为企业部署多智能体系统的先决条件，重点解决模型幻觉、虚假决策与智能体间的逻辑冲突。 - **人机回环机制**：设立多级支撑团队，在自动化智能体输出异常警告时介入复核，实施动态纠偏。 --- ### [06:52](https://www.youtube.com/watch?v=KW-NPTqloqo&t=412s) 企业架构重构：职能下沉与制药业降本增效范例 随着智能体在企业内部的全面渗透，传统企业的组织架构正在发生深刻质变。Cohen 指出，企业内部正在涌现出一批全新的智能体职能角色：从采购端的供应商准入（Supplier Onboarding）、供应商绩效动态监测、供应链自主谈判，到合同全生命周期管理及自动化支付结算。智能体不仅接管了中后台的流程性事务，还通过模型上下文协议（Model Context Protocol, MCP）以及智能体对智能体（Agent-to-Agent, A2A）的通信协议，形成了横向跨职能融合、纵向深度集成的全新网状架构。 在咨询公司埃森哲（Accenture）协助客户搭建的多智能体案例中，系统设立了多层级的职能智能体：基础层智能体专注于特定微观任务（如定向排查东欧或拉美地区金融交易中的欺诈洗钱风险），向上则汇聚至协调层完成全局决策。这种架构革新在生物医药行业尤为显著。以阿斯利康（AstraZeneca）为代表的先锋药企已全面引入 Agentic AI 进行早期药物研发，将原本漫长的新药靶点与候选分子筛选周期从传统的 4 至 5 年断崖式压缩至 1 年甚至数月以内。Cohen 直言，哪怕药企目前正依靠 Ozempic（诺和泰）或 Keytruda（可瑞达）等重磅药物坐享巨额利润，如果不在当下加速推进智能体转型，在未来四到五年内就将在研发竞赛中彻底出局。 | 业务维度 | 传统企业运作模式 | 智能体赋能企业（Cognizant Corporation） | |:--- |:--- |:--- | | **供应链与采购** | 人工推进供应商准入、纸质/静态合同审批、人工核算谈判 | 智能体自主准入排查、动态绩效监控、自主跨协议谈判与支付 | | **药物发现周期** | 早期候选药物识别与验证耗时 4~5 年 | 基于生成式与智能体系统压缩至 1 年以内 | | **风控与反欺诈** | 抽样事后审计、跨国界跨币种数据滞后 | 分布式智能体 24/7 实时排查东欧/拉美等特定区域欺诈行为 | | **协同通信协议** | 人工邮件、传统 API 接口调用 | 依赖模型上下文协议（MCP）与 A2A 协议实现自主协作 | --- ### [11:49](https://www.youtube.com/watch?v=KW-NPTqloqo&t=709s) 社会冲击与劳动价值度量：从人机成本比到 GPT-5 防御机制 智能体经济的崛起不可避免地引发了对社会制度的深度拷问：未来的工作形态将如何演变？全民基本收入（UBI）是否成为必然？人类社会是会走向权力高度集中的“科技新贵”（Tech Bros）统治，还是演变为更加扁平化的社会结构？Cohen 指出，当智能体深入生产流程后，企业开始具备精确度量“智能体 vs 人类员工”边际产出的能力。 以花旗集团（Citigroup）的软件工程实践为例，管理层开始量化评估编程智能体生成的代码采纳率：如果一个智能体给出的代码建议仅有 5% 到 10% 被合并入最终代码库，说明该智能体价值有限；但若采纳率达到 80% 至 90%，则意味着其生产力极为惊人。目前在多家大型银行中，普遍推行“1 名人类员工带领 4 至 5 个智能体”的混合团队模式。从经济账来看，单个智能体的运行成本仅为全职人类员工的 1/4 到 1/5，却能提供 7×24 小时不间断的生产力溢出。与此同时，前沿基础模型也在安全可靠性上发力，例如新一代 GPT-5 将大量研发资源倾注于最大化抑制模型幻觉，使智能体在金融等严苛场景中的决策更具可防御性与合规性。 - **代码采纳率度量法**：通过建议代码入库比例（5%-10% 对比 80%-90%）精准评估智能体赋能开发的真实 ROI。 - **1:4 人机协同编制**：大型金融机构试行单个员工挂帅 4~5 个专属智能体，以 1/5 的人力成本实现全天候产能覆盖。 - **底层模型确定性强化**：前沿大模型将治理重点转向控制幻觉率，大幅提升企业级部署的可靠性边界。 --- ### [15:35](https://www.youtube.com/watch?v=KW-NPTqloqo&t=935s) 问答环节：人类价值重塑、工作时长缩短与兜底保障 ``` 问：您的分析主要集中在经济与就业层面。如果未来人类不再需要承担如此繁重的工作，人类的生活意义（Purpose）何在？我们该如何规划作为人类的未来？ ``` 针对现场观众关于人类生存意义与就业未来的提问，Cohen 给出了积极的展望。他认为智能体时代不仅不会封死职业道路，反而会极大促进“职业流动性”（Job Mobility）。过去高科技行业往往要求从业者拥有计算机科学或工程学博士学位，但随着低代码（Low-code）与无代码（No-code）工具的普及以及编程门槛的大幅降低，技术门槛正在瓦解。未来的核心竞争力将转向商业嗅觉、经济决策能力与业务理解深度，非技术背景的人才同样能够自如地调遣智能体解决复杂业务问题。 在宏观劳动形态上，Cohen 预测全社会的工作周可能会缩短至 30 到 35 小时。而在创新生态层面，初创公司正迎来巨大机遇，从边缘计算（Edge AI）、小型化定制模型到数据定义治理，各类细分赛道正在百花齐放。不过，他也坦承转型期必然存在阵痛，社会必须为缺乏数理背景的群体（例如 60 岁以上难以完成技能重塑的年长劳动者）建立坚实的社会安全兜底保障网络，确保技术红利能够公平普惠。 --- ### [18:27](https://www.youtube.com/watch?v=KW-NPTqloqo&t=1107s) 问答环节：未来 1 至 5 年的核心技能与就业辩论 ``` 问：在当前的企业组织内部，我们究竟应该培训员工掌握哪些技能？从 1 年、2 年到 5 年的跨度来看，哪些能力才是不可或缺的？ ``` 在回答全场最后一个问题时，Cohen 明确指出，尽管技术底座依旧需要算力基础设施和计算机工程人才，但企业越来越渴求能够“量化 AI 业务价值”的复合型人才。企业急需弄清每个具体业务流（Workflow）的改造成本是多少、能带来多少实际收益，因此具备指标构建能力、业务流程拆解能力与战略统筹思维的人才将极其抢手。此外，随着多智能体网络（Networks for AI）跨地域、跨业务单元的打通，具备跨部门协同与网络化统筹能力的复合管理者将创造巨大的协同降本空间。 在总结发言中，Cohen 针对当前科技巨头高管之间的就业争论表明了立场。针对 Anthropic CEO 对 AI 导致大规模失业的悲观预警，他更倾向于认同 Salesforce CEO 的观点——即技术变革在压减传统纯软件研发岗位的同时，正在催生大量全新的战略管理、业务流程设计与 AI 运营岗位。技术的演进并非单纯消灭就业，而是在淘汰单一机械劳动的同时，将劳动力全面推向更高维度的战略与商业决策领域。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) Toto, I've got the feeling we're not in Kansas anymore — Robert Cohen > [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) 图灵经济：智能体以集群形式协同运作，自主识别所需数据、精细化拆解业务流程，并跨职能完成复合型任务。 — Robert Cohen > [00:04](https://www.youtube.com/watch?v=KW-NPTqloqo&t=4s) 在今年上半年，AI 智能体的使用量暴增了 233%；到 2025 年，已有 40% 采用智能体 AI 的企业将其部署在跨职能业务中。 — Robert Cohen > [04:23](https://www.youtube.com/watch?v=KW-NPTqloqo&t=263s) 在新一代高能效 GPU 架构的推动下，计算成本每年将实现翻倍降低。 — 黄仁勋 > [04:23](https://www.youtube.com/watch?v=KW-NPTqloqo&t=263s) AI Observability 成为企业部署多智能体系统的先决条件，重点解决模型幻觉、虚假决策与智能体间的逻辑冲突。 — Robert Cohen > [06:52](https://www.youtube.com/watch?v=KW-NPTqloqo&t=412s) 哪怕药企目前正依靠 Ozempic 或 Keytruda 等重磅药物坐享巨额利润，如果不在当下加速推进智能体转型，在未来四到五年内就将在研发竞赛中彻底出局。 — Robert Cohen > [11:49](https://www.youtube.com/watch?v=KW-NPTqloqo&t=709s) 单个智能体的运行成本仅为全职人类员工的 1/4 到 1/5，却能提供 7×24 小时不间断的生产力溢出。 — Robert Cohen > [15:35](https://www.youtube.com/watch?v=KW-NPTqloqo&t=935s) 技术门槛正在瓦解。未来的核心竞争力将转向商业嗅觉、经济决策能力与业务理解深度，非技术背景的人才同样能够自如地调遣智能体解决复杂业务问题。 — Robert Cohen > [18:27](https://www.youtube.com/watch?v=KW-NPTqloqo&t=1107s) 技术的演进并非单纯消灭就业，而是在淘汰单一机械劳动的同时，将劳动力全面推向更高维度的战略与商业决策领域。 — Robert Cohen ## Mentioned Robert Cohen, 经济战略研究所 (Economic Strategy Institute), 艾伦·图灵 (Alan Turing), 黄仁勋 (Jensen Huang), 英伟达 (NVIDIA), 埃森哲 (Accenture), 阿斯利康 (AstraZeneca), 花旗集团 (Citigroup), 诺和泰 (Ozempic), 可瑞达 (Keytruda), Anthropic, Salesforce, GPT-5, 模型上下文协议 (Model Context Protocol / MCP), 智能体对智能体协议 (Agent-to-Agent / A2A), 绿野仙踪 (The Wizard of Oz) ## Video description Robert Cohen, AI Industry Analyst and Economist, Economic Strategy Institute Abstracts AI is reshaping economic landscapes by enhancing productivity, accelerating value creation, and redefining industrial and national competitiveness. This talk assesses AI’s role in unlocking new forms of capital efficiency, increasing returns to scale, and strengthening national strategic advantage in the global economy as well as how these changes will change society. AI acts as a form of capital deepening: intelligent agents and machine learning models embody knowledge that can be replicated and scaled across markets. agentic AI architectures transform decision-making, linking operational data to strategic insights across manufacturing, logistics, and customer engagement. Cognizant Corporations with ubiquitous AI agents in their operations demonstrate how agentic systems can amplify competitiveness through continuous innovation and real-time decision-making. Nations investing in AI infrastructure and workforce adaptation will achieve sustained economic advantages, akin to the role of industrialization in past centuries. [Source](https://www.youtube.com/watch?v=KW-NPTqloqo)

Source quotes

[0:04] Robert Cohen: Toto, I've got the feeling we're not in Kansas anymore [0:04] Robert Cohen: 图灵经济：智能体以集群形式协同运作，自主识别所需数据、精细化拆解业务流程，并跨职能完成复合型任务。 [0:04] Robert Cohen: 在今年上半年，AI 智能体的使用量暴增了 233%；到 2025 年，已有 40% 采用智能体 AI 的企业将其部署在跨职能业务中。 [4:23] 黄仁勋: 在新一代高能效 GPU 架构的推动下，计算成本每年将实现翻倍降低。 [4:23] Robert Cohen: AI Observability 成为企业部署多智能体系统的先决条件，重点解决模型幻觉、虚假决策与智能体间的逻辑冲突。 [6:52] Robert Cohen: 哪怕药企目前正依靠 Ozempic 或 Keytruda 等重磅药物坐享巨额利润，如果不在当下加速推进智能体转型，在未来四到五年内就将在研发竞赛中彻底出局。 [11:49] Robert Cohen: 单个智能体的运行成本仅为全职人类员工的 1/4 到 1/5，却能提供 7×24 小时不间断的生产力溢出。 [15:35] Robert Cohen: 技术门槛正在瓦解。未来的核心竞争力将转向商业嗅觉、经济决策能力与业务理解深度，非技术背景的人才同样能够自如地调遣智能体解决复杂业务问题。 [18:27] Robert Cohen: 技术的演进并非单纯消灭就业，而是在淘汰单一机械劳动的同时，将劳动力全面推向更高维度的战略与商业决策领域。

Source quotes

[0:04] Robert Cohen: Toto, I've got the feeling we're not in Kansas anymore [0:04] Robert Cohen: AI agent adoption across enterprise workflows grew by 233% in the first half of 2025. [0:04] Robert Cohen: 40% of firms utilizing agentic AI deploy it across cross-functional operations to integrate insights across risk, finance, and trading desks. [4:23] Robert Cohen: Compute costs for running advanced models are dropping along an aggressive deflationary curve, estimated by Nvidia's Jensen Huang to decrease twofold annually with new GPU architectures. [4:23] Robert Cohen: Observability agents act as algorithmic guardrails, supervising peer agents to identify hallucinations, biased decisions, and operational conflicts. [6:52] Robert Cohen: In life sciences, agentic AI has compressed the early-stage drug discovery timeline from 4–5 years down to under 12 months. [11:46] Robert Cohen: Will tech bros run the world? [11:46] Robert Cohen: Operating costs for a multi-agent team are estimated at roughly 20% to 25% of the compensation of a single full-time employee, paired with continuous 24/7 operational capability. [11:46] Robert Cohen: Citigroup evaluates coding agents by the proportion of generated code merged into production builds, establishing a benchmark where 80–90% acceptance represents high enterprise value. [15:35] Robert Cohen: Macroeconomic adjustments may drive standard full-time employment down to a 30- to 35-hour workweek. [15:35] Robert Cohen: High-demand future skills will prioritize workflow cost accounting, ROI modeling, and multi-agent network orchestration rather than pure syntax programming.

Original transcript

[0:04] I'm the the next speaker, Robert Cohen [0:06] from the Economic Strategy Institute and [0:09] I'm talking about the economic and [0:10] social impact of AI. Uh the agenda [0:14] changed the time slot I think just in [0:17] the last 24 hours. So that it's supposed [0:19] to start at at 11:30 and finish at [0:23] 11:50. Previously it started 11 at [0:26] 11:40. [0:28] So, I'll start at the at the new time [0:31] and uh hopefully uh people who come in [0:35] late can ask a question to catch up on [0:38] what I've talked about before. [0:43] [Music] [0:45] Okay. [0:47] So, I I'm going to be talking about the [0:49] economic and social impact of AI. I I'm [0:52] a economic [0:55] uh analyst who specializes in how [0:58] industries are adopting new technologies [1:02] and I I've done [1:05] a detailed study of a number of early [1:09] firms that are um [1:12] that are adopting uh these technologies. [1:16] So um what what I the theme of this is [1:20] that in a world where AI agents as we [1:24] make the shift from previous AI models [1:27] to ones that can actually not only [1:30] gather data uh but also think and make [1:34] decisions that they change the way our [1:37] economy and society is going to work in [1:40] the future. Um, and what I what I've [1:46] translated this into is saying that with [1:49] AI agents, we're now in what I call a [1:53] touring economy. Because Touring said in [1:57] 1950, that by 2000, we'd be thinking [2:02] about machines that are acting [2:04] indistinguishably [2:06] from the way a thinker acts. that we'd [2:08] be speaking about machines thinking [2:12] and in uh n 2025 this year 40% of all [2:16] firms using agentic AI in some crossf [2:20] functional were using AI agenic AI in a [2:23] crossf functional capacity meaning that [2:26] they were not just doing something like [2:28] looking at risk management but also [2:30] looking at um ways that they could uh [2:35] get this information to other people in [2:37] the bank that needed to act on changes [2:40] in risk that occurred. So for instance, [2:43] if the Russian ruble spikes up on Friday [2:47] um and the the model says, you know, [2:51] sell off all these uh countries that are [2:55] not uh benefiting from purchasing things [3:00] from Russia that there'd be a way to [3:04] crosscorrelate that within the firm and [3:07] it there wouldn't be if there was a [3:09] wrong decision made by one of these [3:11] these agents, there'd be an ability to [3:14] to integrate it with information coming [3:18] from other sources and make a a more [3:20] optimized decision. Uh, and then one of [3:24] the other statistics I've seen recently [3:26] is that in the first six months of this [3:29] year, AI agent use was up 233%. [3:33] So, we're seeing this kind of explosive [3:35] growth. And I put a quote down here, [3:37] Toto, I've got the feeling we're not in [3:40] Kansas anymore. from u [3:44] uh I can't quite quite read who it's [3:47] from but but it's but it's from uh [3:50] Wizard of Oz and I mean it's it's [3:53] essentially saying that the world has [3:55] completely changed and we need to [3:57] recognize this and the previous economic [3:59] panel that was here I I think tried to [4:02] be fairly cautious about what kind of [4:05] GDP and productivity changes we're [4:08] likely to see. Uh the estimates for this [4:11] have been kind of all over the board. Uh [4:14] and [4:16] I I think there's been a push more [4:19] recently to be more conservative. So let [4:23] let me just go through why I'm calling [4:26] this a touring economy. What's what's [4:29] changed? I think that productivity goes [4:31] up dramatically since agents work as [4:35] groups. Identify the data they they [4:38] need. lets them uh analyze processes in [4:41] detail and teams of agents interact as I [4:43] mentioned before and complete complex [4:46] tasks. Um in terms of costs, the costs [4:50] are spread over software development uh [4:54] computing infrastructure like energy [4:56] requirements and lower compute costs are [4:59] driving these declines. I've got a chart [5:01] of that coming up. And then uh the other [5:04] dimension of this that's not really [5:07] given a lot of play and has really only [5:09] been brought out recently by some of the [5:12] more aggressive adopters of agents is [5:15] that they need to pay a lot more [5:17] attention to essentially the guard rails [5:20] in in this new economy. uh that there [5:24] needs to be far more AI observability [5:26] with AI agents dedicated to looking at [5:30] the sort of decisions made by certain [5:33] teams or individual agents to see if [5:36] they're hallucinating or they're making [5:39] decisions that would be uh deemed wrong [5:42] by most objective people or to at least [5:45] flag where there are conflicts. So I [5:49] think that that's going to be a a much [5:51] greater part of of how AI develops going [5:55] forward. And this is almost similar [5:58] chart to what one could see if you [6:01] looked at the early days of the [6:04] semiconductor industry where there's a a [6:07] drop of price by two and a half% in [6:11] terms of what it costs to to analyze uh [6:15] to do the computing for for various [6:19] models that are being run. And the the [6:22] data is kind of really jumpy, but if you [6:25] put a trend line in, you can see that [6:27] it's dropping that long. Uh Jensen Wang [6:31] from from Nvidia has argued that there's [6:34] a drop of twofold every single year, [6:37] which I think is a little bit more [6:39] aggressive, but if the new uh GPU chips [6:44] are very efficient, that his uh [6:47] prediction could be true. [6:52] Okay. So, uh this is just to illustrate [6:55] I mean what one of the things that I've [6:57] seen in looking at some of these [6:59] individual corporations is that they're [7:02] also already identifying new [7:05] occupational roles or new functional [7:07] roles that agents are playing. So here [7:10] you have supplier onboarding which is [7:13] pretty common in finance and uh [7:17] retailing and other industries like [7:18] that. You've got supplier per [7:20] performance monitoring. The supply chain [7:23] management comes in here. Uh you've got [7:26] um a autonomous negotiations with [7:29] suppliers. You've got contract life [7:32] cycle uh evaluation and management and [7:36] payments and so forth. Uh I my feeling [7:41] is that we're not only going to see [7:42] these kind of middle level or higher [7:45] level functions come into play much more [7:49] significantly as we move to AI agents [7:52] but also that there'll be support [7:55] requirements needed for various teams. [7:57] So if you have an AI observatory team [8:01] and something comes up as you know this [8:04] this is does not look great there would [8:07] be a a number of uh teams backing up the [8:12] AI observability so that there'd be a [8:15] resolution that could be made by humans [8:18] intervening and looking at what the [8:20] decision was and sending in a correction [8:23] if there was a need to correct the the [8:26] uh the decision that was taken. So I [8:30] I've said that what's this is also going [8:32] to mean that corporations change [8:35] dramatically um that as we move into [8:38] embedding AI particularly AI agents [8:42] throughout uh corporate operations that [8:45] we're going to get much richer [8:46] intelligence about all of these [8:48] functions that went before. They're [8:50] going to leverage agents to gather data [8:53] and act autonomously. And they're going [8:55] to integrate new tools like MCP uh model [8:59] context protocols or agent to agent. [9:02] They're going to shift to agent systems. [9:05] Um that the I think I've got a a picture [9:10] that shows how this works in a bank. I [9:13] think it's after this one. But there [9:16] there's a neat what what happens here is [9:18] you get different levels of functions [9:21] within the agent teams. So if you look [9:24] at how Accenture has set things up with [9:26] some of their early uh agent team [9:29] clients, they have basic level sort of [9:33] agents who do specific functions like [9:37] check on all of the risk in uh financial [9:41] transactions from Eastern Europe or [9:44] Latin America to make sure there aren't [9:46] any uh real uh major [9:52] fraudulent things happening. happening. [9:54] Um they they're going to benefit greatly [9:58] from the decline in computing costs. I [10:00] gave some illustrations here, but I mean [10:03] it's very much like on the chart I [10:05] showed before, maybe even bigger. And [10:07] then uh they're going to enhance [10:10] competitiveness that companies like um [10:15] Astroenica that talked yesterday is [10:18] already implementing uh Agentic AI to do [10:23] drug discovery. A lot of the drug [10:25] companies are cutting the original [10:27] identification of new drugs which used [10:30] to take about four or five years down to [10:33] a year or less. So we're seeing the [10:36] early stage in just one specific [10:38] industry and that's the pressure. I mean [10:40] if you're a major drug discovery firm uh [10:46] you may be sitting on your profits from [10:48] Ozmpic or Katruda but if you don't start [10:51] to do this you're going to be losing the [10:52] race in four or five years. So a number [10:56] of the companies like Astroenica have [10:58] figured this out and they're moving [11:00] ahead very act actively and I think this [11:03] is going to transform this the corporate [11:06] structures that we see in the future. [11:08] This was a a single study that said, [11:11] okay, what's a bank going to look like [11:13] in the future if all of these agentic [11:16] capabilities are embedded throughout the [11:18] bank? that there's going to be uh not [11:22] only cross functional integration of uh [11:27] systems and models but there'll be [11:29] vertical which isn't shown very well in [11:32] here in terms of uh you know integrating [11:36] data much better and having operating [11:39] operating models in here there's not [11:41] much on observability in this in this [11:44] particular chart [11:46] uh okay so I'm I'm getting towards the [11:49] end of I'll leave some time for for [11:51] questions. Uh but u there are h a whole [11:56] lot of uh social implications of what I [11:59] just talked about. I mean what's going [12:01] to happen to work if we move into this [12:04] sort of an economy? Uh what happens to [12:06] incomes? How are people going to live? [12:09] Will everybody be getting a universal [12:11] basic income or are there going to be [12:13] enough jobs to pay people? Uh will [12:16] people fear that AI is going to take [12:18] over the world? that everyone that's [12:20] using it will make harmful decisions, [12:23] create, you know, bad drugs and and kind [12:26] of destroy society as we know it. Uh [12:29] will what's the society going to look [12:31] like? Will tech bros run the world? Uh [12:34] or will the the society be much flatter [12:39] and not so hierarchical as it's been in [12:42] the last few years? And then what kind [12:45] of skills will people need in the new [12:47] workforce in the new society and what [12:50] institutions are going to see the [12:51] biggest changes going forward. So I kind [12:55] of ran through these pretty quickly. I [12:57] mean ju just to review [13:00] what I said in the beginning was that [13:02] you know now that we can measure things [13:05] with agents we've got the ability to say [13:08] how much does an agent contribute versus [13:12] a regular employee doing the same thing. [13:15] So at at Cityroup they have agents [13:19] writing code and what they're doing is [13:21] they're saying okay here are all the [13:24] suggestions made by an agent. How many [13:26] of those suggestions actually get into [13:29] the final code? So it's five if it's 5 [13:32] to 10% that agent's not really helping [13:34] me very much. If it's 80 to 90% then [13:38] that agent is doing a lot to help my [13:40] final coding process. And so there's [13:44] there's an abil there's a beginning of [13:46] quantifying things in that way and then [13:48] looking at at it and saying okay if if I [13:52] pay a regular employee X how much do I [13:56] pay a [13:58] agent compared to X or pay a team of [14:01] agents. So in a number of the banks [14:03] they're running four or five agents [14:05] together with uh a regular employee and [14:09] so that then they can you know if if an [14:11] agent is going to cost roughly [14:15] a quarter or a fifth of what a a regular [14:18] employee costs are the agents really [14:22] providing more surplus as as they're [14:24] able to work 247 rather than just uh [14:29] eight time seven and what you know how [14:33] how are the costs going to change over [14:35] time? I mean if we're going to get much [14:37] better compute costs and if we're get [14:39] going to get much better observability [14:41] of how agents function then the the [14:45] overall cost of using agents will come [14:48] down dramatically and will block a lot [14:51] of the harmful operations. I mean, one [14:54] of the [14:56] characteristics that I saw in the new [14:57] GPT5 is that they spent a lot of time [15:02] figuring out how to minimize how [15:04] hallucinations much better than a number [15:07] of the other models. So, even with that [15:10] shift, we're seeing some early attention [15:13] to making these models much more [15:15] reliable and defensible. So, um, but [15:20] we've got about six minutes left. Why [15:23] don't I open up the floor for questions [15:25] and I'd be glad to to explore anything [15:29] more. Yeah. [15:35] >> Right. [15:38] >> So, your focus has been on the economy [15:40] and on employment. So, really the work [15:42] world. But I'm wondering if we don't [15:44] have to work so much, what about [15:46] purpose? How should people be thinking [15:48] about their futures as human beings? [15:51] >> Right. [15:54] I think this is going to open up a lot [15:56] more uh job mobility to people uh for [16:01] two reasons. I mean, I think that uh [16:05] beginning to become someone who works [16:07] with an agent is not going to require [16:09] that you just have a PhD in computer [16:11] science or engineering. Uh and I think a [16:14] lot of the coding will be simplified. [16:17] Um, so I I I think there's going to be a [16:21] lot more attention paid to, you know, do [16:23] you have really good kind of economic or [16:26] business decision-making skills so you [16:29] can fit into an organization because I [16:31] think that what we're seeing right now [16:32] is this is largely confined to big [16:36] aggressive companies and then a lot of [16:38] the the hyperscalers who are providing [16:41] the services. But I think if you look at [16:43] the startup world, you're seeing a lot [16:45] of companies come in at different [16:48] levels. How do I uh define the data? How [16:52] do I uh run models better? How do I [16:56] create uh limited size models? How do I [17:00] run this thing on the edge instead of in [17:02] the cloud? uh and I think the other [17:04] phenomenon that wasn't really given too [17:06] much play earlier is that there's this [17:09] big movement towards low code or no code [17:13] uh employment. So that there's going to [17:15] be opportunities for people to come in [17:17] as I mentioned before and check know AI [17:21] observability check the decisions that a [17:23] model has made and make a correction if [17:27] it's necessary. I I think that it may [17:30] mean that we go down to a 35 or 30 hour [17:33] work week if people are employed. It may [17:35] mean that some people who have [17:37] difficulty getting employment or never [17:40] really had much skills in the in the way [17:44] of what's coming into the world that we [17:47] we have to set up some sort of safety [17:49] net for them. and just recognize that a [17:52] 60 year old who's never had any math or [17:55] science background is is not going to [17:58] fit into this kind of level. Although I [18:01] think there'll be a lot of retail and [18:03] commercial and other changes, robotics [18:06] and so forth, which I can't, you know, [18:08] really get into too much detail here, [18:10] but I think they'll be far more like [18:13] agents than uh than they've been before. [18:17] Um, other questions? [18:19] >> Yes, we do. And due to time constraints, [18:21] Robert will be able to answer this last [18:23] question. Thank you. [18:27] >> So in the organization right now, what [18:29] are the skills that we need to train [18:31] people towards? What are the skills that [18:34] are going to be necessary a year, two [18:36] years from now, five years from now? [18:42] I think you'll still have a lot of need [18:44] for sort of compute and engineering [18:47] skills on the technical side, but I [18:49] think as we're getting closer to [18:52] measuring the impact of this both on the [18:55] economy and in the firm that there's [18:57] going to be more need to begin to figure [18:59] out what what value does this really [19:02] bring. So they're going to be there [19:05] there will be a need for more skills for [19:07] people who can measure and begin to [19:10] estimate you know what are the various [19:12] workflows costing us uh what benefits do [19:16] we get from specific workflows and and [19:19] then for different things like I I work [19:23] with a group that's beginning to develop [19:26] networks for AI and I think there'll be [19:30] um additional efficiencies gained if one [19:33] can network together a lot of the [19:35] different AI agent teams or multi- aent [19:38] systems that operate within a company [19:41] and maybe have them operate uh within [19:46] various locations uh together so that [19:50] that kind of synergy will will reduce [19:53] costs let you do things much faster so I [19:57] think that there'll be opportunities for [19:59] people with uh with good ability to do [20:02] metrics or business uh thinking about [20:05] this uh to get positions. [20:09] Um so I'm not so fearful of lots of [20:13] people being uh unemployed as the the [20:17] CEO of Anthropic has argued and the [20:20] opposite of that has been um the CEO of [20:24] Salesforce has argued that he's already [20:27] seeing that they're gaining many more [20:29] jobs. They're not hiring in some [20:31] categories that they've built up a lot [20:33] already like software developers or [20:35] software engineers, but they're hiring [20:37] people in these areas that I described [20:39] where they would do more kind of [20:41] business functions and strategic [20:43] functions for their company.