Brynjolfsson Says 5% of Users Drive 95% of Enterprise AI Value

Know and Grow the ROI of AI: From Enterprise Transformation to Business Value With Workhelix — Ai4

Ai4

Enterprise productivity gains stall on direct software purchases unless firms deploy nine times that capital into redesigning workflows around top-performing power users.

AI-assisted summary of Know and Grow the ROI of AI: From Enterprise Transformation to Business Value With Workhelix — original source. Check the original for full context.

Source published: 2026-08-15

Detailed notes

## Background Published on YouTube by Ai4 on August 15, 2026, this keynote features Erik Brynjolfsson—Director of the Stanford Digital Economy Lab and Co-Founder at Workhelix—addressing how enterprise leaders can move beyond model adoption to capture measurable business value. The presentation draws on findings from the Stanford Enterprise AI Playbook and deployment case studies to outline frameworks for task-level AI evaluation, causal KPI tracking, and navigating the productivity J-curve.

## TL;DR Erik Brynjolfsson explains that while AI capabilities have crossed human-level benchmarks, enterprise ROI lags due to the classic productivity J-curve and the requirement for nine dollars of complementary intangible investment—in process redesign and skills—for every dollar spent on tech. To bridge this gap, enterprises must adopt a granular task-based methodology that maps distinct workplace activities to AI applicability rather than treating occupations as monoliths. Data across Fortune 500 deployments, such as NASDAQ, reveals that AI usage follows a strict power law where just 5% of users generate roughly 95% of the total value. Organizations can achieve compounding productivity growth by identifying these hidden power users, turning their advanced workflows and prompts into company-wide playbooks, and pursuing top-line labor augmentation over simple headcount reduction.

## Atmosphere & style The session delivers the crisp, data-driven authority of a premier academic lecture blended with an actionable executive briefing. Brynjolfsson presents with composed clarity, demystifying enterprise anxiety through vivid physical analogies like boiling water thresholds and hidden investment icebergs. The mood is pragmatic yet encouraging, systematically shifting the room from frustration over stalled pilots to strategic optimism grounded in empirical measurement.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=4s) A significant 'capabilities overhang' exists as AI model performance rapidly crosses human baselines while broader macroeconomic and enterprise returns continue to lag. - [01:43](https://www.youtube.com/watch?v=57obkqdD2RM&t=103s) Every $1 spent on direct technology investments requires approximately $9 in complementary intangible assets, including workforce retraining, process redesign, and new product development. - [03:53](https://www.youtube.com/watch?v=57obkqdD2RM&t=233s) Deploying general-purpose technologies inevitably triggers a productivity J-curve, causing measured productivity to stagnate or dip during upfront restructuring before compounding into a harvest phase. - [03:53](https://www.youtube.com/watch?v=57obkqdD2RM&t=233s) Analysis of 51 enterprise implementations revealed that organizational obstacles—such as change management, process adaptation, and measurement—accounted for 77% of deployment challenges rather than technical limitations. - [06:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=376s) Evaluating AI's business impact requires a task-based approach, decomposing monolithic jobs into discrete atomic activities and weighting their AI exposure by economic value. - [08:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=484s) Most corporate activity sits in the 'opportunity zone,' where high-value enterprise tasks remain severely underutilized by AI models and require deliberate workflow migration. - [09:38](https://www.youtube.com/watch?v=57obkqdD2RM&t=578s) Enterprise AI adoption follows an extreme power-law distribution, exemplified at NASDAQ where just 5% of users generated approximately 95% of total business value. - [11:32](https://www.youtube.com/watch?v=57obkqdD2RM&t=692s) Systematically identifying top performers through telemetry data and codifying their prompts into enterprise-wide templates can effectively double overall workforce productivity. - [13:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=796s) The greatest enterprise ROI stems from labor augmentation and end-to-end business process reconfiguration focused on top-line revenue growth rather than narrow headcount reduction. - [15:44](https://www.youtube.com/watch?v=57obkqdD2RM&t=944s) Crude executive metrics like 'token maxing' and generic mandates fail to drive adoption compared to providing workers with proven, department-specific peer prompts mapped to granular tasks. - [18:29](https://www.youtube.com/watch?v=57obkqdD2RM&t=1109s) The power-law distribution of AI value creation is consistent across all industry sectors, meaning the primary variance and upside reside inside individual organizations rather than between industries.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=4s) The Capabilities Overhang and the Boiling Water Threshold

Erik Brynjolfsson introduces himself by clarifying the pronunciation of his Icelandic surname and reflecting on how rapidly executive mindsets have shifted. Just two years prior, boardrooms were consistently asking him whether artificial intelligence actually worked. Today, that skepticism has vanished completely. The question dominating leadership discussions has transitioned to: "Where's the ROI? How do we get value from it?" Brynjolfsson notes that despite the technology's undeniable promise, a massive "capabilities overhang" persists. Models are advancing at a breakneck pace, yet the broader macroeconomy has barely registered the impact, leaving many individual enterprises feeling understandably disappointed.

To demonstrate the speed of technical progress, Brynjolfsson points to the Stanford AI Index Report—an annual, freely available 520-page document packed with benchmark data. Across virtually every core category, model performance curves are rocketing up and to the right. To make the charts more intuitive, he asked his team to include a dotted reference line representing human performance on the exact same benchmark examinations. Benchmark after benchmark shows models crossing above human-level capability. Brynjolfsson compares this moment to water reaching the boiling point: just as water suddenly shifts from a liquid to a gas at a specific temperature threshold, cognitive tasks across industries are undergoing a phase change from human execution to machine capability. The problem, however, is that this technical inflection has failed to translate into proportional economic returns.

- The Stanford AI Index Report tracks global AI capabilities across 520 pages of empirical benchmarks. - Model performance on standardized cognitive and task evaluations is crossing the human baseline across an expanding set of domains. - The core challenge facing executive leadership is converting technical capabilities into measurable bottom-line value.

---

[01:43](https://www.youtube.com/watch?v=57obkqdD2RM&t=103s) The ROI Gap and the 9-to-1 Intangible Asset Iceberg

The disconnect between surging model capability and lagging business value forms what Brynjolfsson terms the "ROI gap" or "chasm." Enterprise leadership watches technical benchmarks climb, licenses get deployed, and tool usage numbers expand, but the aggregate activity often amounts to low-value busywork. Drawing on a conversation with a writer from *The New Yorker*, Brynjolfsson illustrates how generative AI frequently ends up weaponized for low-value tasks like generating deluge volumes of email spam. Alongside the unchecked proliferation of "shadow AI," these patterns can actively degrade organizational productivity rather than unlock it. Crossing this chasm requires rigorous measurement, anchored in the classic management maxim that "you can't manage what you don't measure."

```mermaid flowchart LR accTitle: The Tangible vs Intangible Enterprise Investment Iceberg subgraph Visible["Visible Assets (10%)"] direction TB T["$1 Tech Investment Hardware & Software"] end subgraph Intangible["Intangible Assets (90%)"] direction TB HC["Human Capital & Skills"] BPR["Business Process Redesign"] NPD["New Products & Services"] end T -->|"Requires 9x Complementary Capital"| Intangible ```

Bridging this gap demands looking far beyond the technical stack to the rest of the organization. Referencing foundational research conducted with Lorin Hitt and subsequent colleagues, Brynjolfsson emphasizes that the classic "iceberg" analogy is not mere poetry—it is an empirical ratio. For every single dollar spent on tangible information technology investments, an organization must invest roughly nine dollars in complementary intangible assets: retraining human capital, radically redesigning business processes, and developing new products and services. While CEOs and CTOs easily spot technology line items because they pay hard dollars for software and hardware they can physically "knock on with your knuckles," the intangible investments are almost an order of magnitude larger. Trillions of dollars in intangible assets drive the United States economy yet remain completely absent from national economic accounts and corporate balance sheets.

- Empirical research demonstrates an approximate 9:1 ratio of intangible organizational investment needed for every $1 of direct technology expenditure. - Trillions of dollars in organizational capital, human skills, and redesigned workflows remain unrecorded on traditional corporate financial statements. - Without proper measurement and workflow alignment, enterprise AI deployments easily degenerate into shadow AI and spam generation.

---

[03:53](https://www.youtube.com/watch?v=57obkqdD2RM&t=233s) Escaping the Productivity J-Curve and the Enterprise AI Playbook

Because AI is a foundational general-purpose technology—what economists call a "GPT," rather than merely a Generative Pre-trained Transformer—its economic rollout adheres to a classic productivity J-curve. When transformative technologies first enter a firm, the organization must devote immense effort to restructuring workflows, hiring consultants, and consuming valuable worker and management hours on internal reorganizations. During this phase, expenses rise while immediate output remains flat, causing measured productivity to stagnate or even decline. Only later, as the complementary intangible assets mature and compound, does the curve hook upward into a harvest phase. Brynjolfsson argues that most modern enterprises are currently sitting right at the bottom inflection point of this J-curve, having incurred upfront integration costs without yet realizing the harvest.

```mermaid flowchart TD accTitle: The Enterprise AI Productivity J-Curve subgraph Dip["Initial Investment Dip"] direction TB I["Deploy AI & GPTs"] --> C["Incur Costs: Consultants, Management & Worker Time"] C --> P["Measured Productivity Flattens or Declines"] end subgraph Rise["The Harvest Phase"] direction TB M["Systematic Task Measurement & Process Redesign"] --> H["Compounding Value Creation"] H --> S["Steep Productivity Growth"] end P -->|"Workhelix Mission: Shorten & Shallow the Dip"| M ```

The mission of Workhelix, the company Brynjolfsson co-founded, is to shorten the duration of the J-curve, shallow the investment trough, and accelerate the upward slope. To identify how leading companies achieve this, Brynjolfsson and his Stanford colleagues developed the *Enterprise AI Playbook*, analyzing 51 successful enterprise implementations. While public discourse often obsesses over failed deployments, this research isolated what successful projects shared in common. The analysis revealed that in 77% of cases, technical hurdles were not the limiting factor; the make-or-break determinants were organizational barriers surrounding people, process adaptation, and, above all, empirical measurement.

- General-purpose technologies reliably induce a productivity J-curve characterized by short-term performance dips followed by delayed, compounding growth. - Stanford's *Enterprise AI Playbook* evaluated 51 real-world enterprise deployments to isolate the core drivers of implementation success. - Organizational dynamics—spanning change management, workflow redesign, and causal performance tracking—accounted for 77% of deployment challenges over technical issues.

---

[06:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=376s) The Task-Based Approach: Deconstructing Occupations

To manage organizational transformation systematically, Brynjolfsson advocates for the "task-based approach," an analytical framework he originally pioneered alongside Tom Mitchell in the journal *Science* and later expanded with his former student and Workhelix co-founder Daniel Rock in the landmark paper *GPTs are GPTs* (co-authored with researchers from OpenAI). Attempting to evaluate AI's impact across an entire enterprise or monolithic job title invariably stalls out because broad job descriptions obscure the underlying mechanics of day-to-day operations. Instead, every profession must be recognized as an aggregated bundle of distinct tasks.

When an occupation is disassembled into its granular components, analytical traction becomes straightforward. A radiologist's job breaks down into 26 distinct tasks; an economist performs over 20 distinct activities; similar decompositions apply to software engineers and journalists. Once mapped, each task can be empirically evaluated for its exposure to artificial intelligence on a low, medium, or high scale. While large language models excel at drafting analytical memos, they offer virtually zero utility for physical duties like lifting boxes. By scoring each task's exposure, weighting it by the economic value generated by that specific task, and aggregating those scores back up through occupations and business units, leadership gains unprecedented visibility into where true automation and augmentation opportunities reside.

- Jobs are operationalized as bundles of discrete tasks (e.g., 26 tasks for a radiologist, 20+ for an economist). - AI exposure scoring evaluates the susceptibility of individual tasks to model assistance (e.g., high for memo drafting, zero for lifting heavy boxes). - Weighting task-level AI exposure by economic value allows organizations to roll granular data back up to the enterprise level.

---

[08:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=484s) The Conversation vs. Value Matrix and Opportunity Zones

Translating the task-based framework into practical enterprise software reveals clear patterns in how employees interact with AI tools. Workhelix maps enterprise activity along a two-dimensional grid: the vertical axis measures adoption volume—specifically, the number of user conversations with AI systems—while the horizontal axis reflects the objective task value derived from the task-based methodology. Plotting enterprise usage across these axes exposes three distinct zones that dictate where intervention is needed.

Brynjolfsson explains that the primary objective for an enterprise is to migrate workflows out of the lower-right opportunity zone and push them vertically up toward the 45-degree diagonal. In nearly every corporate environment Workhelix inspects, vast reservoirs of high-value tasks show minimal model interaction because workers have not yet incorporated AI into their core workflows. The purpose of Workhelix's diagnostic tooling is to pinpoint these stranded opportunities and provide an analytical roadmap for driving adoption precisely where leverage is greatest.

---

[09:38](https://www.youtube.com/watch?v=57obkqdD2RM&t=578s) The Power Law of AI Usage and the NASDAQ Case Study

When analyzing user behavior within enterprises, a striking empirical regularity appears without exception: enterprise AI adoption follows a classic, heavily skewed power law. In every single company examined, a tiny vanguard of power users on the extreme tail extracts extraordinary returns, while the vast majority of employees sit in a long tail, deriving negligible benefit. In typical Fortune 500 organizations, the median worker captures only 1% to 3% of the potential value available through AI systems.

A dramatic illustration of this dynamic emerged from Workhelix's engagement with NASDAQ. Evaluating user interactions with AI agents revealed that roughly 5% of the user base was responsible for generating approximately 95% of total business value. While many executives view such extreme skew as an implementation failure, Brynjolfsson frames it as a massive, low-hanging economic opportunity. Increasing bottom-line value does not require discovering uninvented capabilities; it simply requires moving workers along the distribution curve so that long-tail users begin emulating the behaviors of top-quintile power users. While casual users limit their prompts to drafting slightly more eloquent emails, advanced users deploy models to architect comprehensive corporate marketing strategies, generate robust production code, and orchestrate automated workflows.

- Enterprise AI usage mirrors a strict power-law distribution across all organizations studied. - At typical Fortune 500 companies, the average employee captures only 1% to 3% of the technology's addressable value. - At NASDAQ, an internal audit demonstrated that just 5% of users generated roughly 95% of the total value produced by enterprise AI agents. - Shifting users from low-leverage tasks (polishing email prose) to high-leverage workflows (strategic planning, code generation) represents the fastest path to expanding AI return on investment.

[11:32](https://www.youtube.com/watch?v=57obkqdD2RM&t=692s) Creating a Learning Organization: Identifying Power Users and Scaling Prompts

Lifting the broader workforce up to match the performance of top-tier adopters—represented visually by the green uplift area on the chart—has a profound impact: that seemingly modest sliver of improvement can effectively double a company's total productivity. Crucially, this is not a one-time exercise. Enterprise AI transformation requires a dynamic, continuous cycle where organizations systematically assess what their highest-performing users are doing, distill those practices into structured templates, and distribute them across the rest of the company. Making this work hinges on measurement and visibility.

As the underlying models evolve, use cases and top performers continuously shift. The power users of two years ago are often not the power users of today or next year. Furthermore, high performers are frequently introverted employees who may not vocalize their breakthroughs, but organizational data objectively highlights who they are. Once these individuals are pinpointed, companies can examine their workflows to isolate highly effective prompts across coding, marketing strategy, or customer service. Turning these techniques into standardized templates transforms the enterprise into a continuously improving learning organization.

```mermaid flowchart TD accTitle: Power User Playbook Scaling Loop accDescription: Workflow showing how top-performer practices are identified through data and distributed across an enterprise.

subgraph Identify ["1. Observe & Identify"] A["Telemetry & Observability Data"] --> B["Identify Hidden Power Users"] end

subgraph Extract ["2. Codify & Template"] B --> C["Extract High-Value Prompts & Workflows"] C --> D["Build Department Playbooks & Templates"] end

subgraph Distribute ["3. Scale & Learn"] D --> E["Distribute to Broader Workforce"] E --> F["Achieve Productivity Uplift"] F -.->|"Continuous Model Evolution"| A end ```

[13:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=796s) The Playbook for Action: Task-Level Analysis, Process Redesign, and Top-Line Augmentation

To implement these gains immediately, enterprise leaders must connect AI capabilities directly to concrete business value starting the next morning. This begins with a comprehensive scan of all occupational roles across the organization, deconstructing each job into its constituent, atomic tasks—a methodology detailed in Erik Brynjolfsson's published research in *Science* and operationalized through Workhelix across generative AI, agentic AI, and predictive AI. Once tasks are mapped, progress must be monitored using causal inference, observable KPIs, and rigorous measurement rather than guesswork.

However, task-level optimization is merely the initial phase. True transformation demands a complete reconfiguration of overarching business processes. Because AI automates certain legacy components while opening up novel capabilities, job descriptions must fundamentally change. Rather than treating AI primarily as a headcount-cutting mechanism, forward-looking organizations focus on labor augmentation to expand top-line growth. Brynjolfsson recounts advising a senior executive who was fixated purely on reducing headcount, demonstrating that the far greater ROI lay in carefully measuring and orchestrating how AI can complement human workers to create new lines of value.

- **Actionable Steps for Enterprise Value:** - Map every organizational occupation down to its granular, atomic tasks. - Evaluate task applicability across generative, agentic, and predictive AI. - Track real business impact through causal inference and clear KPIs. - Redesign end-to-end processes and job scopes instead of merely automating isolated legacy steps. - Prioritize labor augmentation and top-line revenue expansion over pure headcount reduction.

[15:44](https://www.youtube.com/watch?v=57obkqdD2RM&t=944s) Q&A: Moving Beyond 'Token Maxing' to Concrete, Task-Level Adoption

During the Q&A session, an attendee from Amazon's internal evangelism team asked how to drive deeper adoption across an organization when evangelists encounter diminishing returns within their limited span of influence. Brynjolfsson praised Amazon’s progress while addressing a widespread executive mistake: issuing generic, top-down memos commanding employees to "use AI more" or instituting crude metrics like "token maxing" (counting raw token consumption). While well-intentioned, broad executive exhortations fail to move the dial because they lack tangible direction.

Real adoption requires getting deep into the operational weeds of specific departments. Workhelix evaluates roughly 25,000 distinct enterprise tasks to pinpoint exactly which activities should incorporate AI and which should not. When a developer, HR manager, or marketer is handed specific, proven prompts from high-performing peers delivering 2x, 3x, or 10x performance gains, adoption naturally accelerates because workers are equipped with an actionable guidebook. Because AI capabilities change so rapidly—looking substantially different in the summer of 2026 compared to late 2025—companies must employ tools that continuously update task suitability and monitor ongoing ROI.

[18:29](https://www.youtube.com/watch?v=57obkqdD2RM&t=1109s) Q&A: Overcoming Middle-Management Inertia and Cross-Industry Power Laws

Responding to an audience question regarding organizational pushback, Brynjolfsson validated a common corporate dynamic: enthusiastic boards and CEOs eagerly greenlight AI initiatives, while middle management remains skeptical and hesitant, hoping the trend will "blow over" like past tech fads. This resistance stems from the natural inertia of successful enterprises, whose established processes were intentionally designed for stability and replication. Overcoming this friction requires CEOs to provide concrete, instrumented metrics that prove feasibility and remove the ambiguity surrounding AI integration.

A subsequent question explored whether the heavy skew in AI adoption—such as the power-law distribution observed at Nasdaq, where 5% of power users drive 99% of generated value—varies across sectors like manufacturing versus finance. Brynjolfsson explained that this qualitative power-law pattern is surprisingly consistent across all industries; the meaningful variance is observed *inside* companies rather than *between* them. While curve steepness may differ, the universal presence of this skewed distribution represents enormous low-hanging fruit, offering a repeatable formula for enterprises to systematically capture trapped ROI over the next five or more years.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=4s) Where's the ROI? How do we get value from it? — Erik Brynjolfsson

> [00:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=4s) capabilities overhang — Erik Brynjolfsson

> [01:43](https://www.youtube.com/watch?v=57obkqdD2RM&t=103s) you can't manage what you don't measure. — Erik Brynjolfsson

> [01:43](https://www.youtube.com/watch?v=57obkqdD2RM&t=103s) knock on with your knuckles — Erik Brynjolfsson

> [03:53](https://www.youtube.com/watch?v=57obkqdD2RM&t=233s) Enterprise AI Playbook — Erik Brynjolfsson

> [06:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=376s) task-based approach — Erik Brynjolfsson

> [06:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=376s) GPTs are GPTs — Erik Brynjolfsson

> [08:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=484s) Opportunity Zone — Erik Brynjolfsson

> [15:44](https://www.youtube.com/watch?v=57obkqdD2RM&t=944s) use AI more — Erik Brynjolfsson

> [15:44](https://www.youtube.com/watch?v=57obkqdD2RM&t=944s) token maxing — Erik Brynjolfsson

> [18:29](https://www.youtube.com/watch?v=57obkqdD2RM&t=1109s) blow over — Erik Brynjolfsson

## Mentioned Erik Brynjolfsson, Stanford Digital Economy Lab, Stanford Institute for Human-Centered AI, Workhelix, Ai4, Stanford AI Index Report, The New Yorker, Lorin Hitt, Tom Mitchell, Daniel Rock, OpenAI, Science, GPTs are GPTs, Enterprise AI Playbook, NASDAQ, Amazon

## Video description Erik Brynjolfsson - Director of the Stanford Digital Economy Lab, a Professor at the Stanford Institute for Human-Centered AI, Co-Founder at Workhelix

AI is advancing rapidly, but the central question for enterprise leaders is no longer whether it works—it is how to convert capability into measurable business value. This talk will show why AI transformation requires more than model adoption, drawing on evidence from Canaries in the Coal Mine on how AI is already reshaping work and from the Stanford Enterprise AI Playbook’s lessons from 51 enterprise deployments.

The talk will explain how to identify high-value opportunities at the task and process level, measure ROI with credible KPIs and causal inference, and move from scattered pilots to scaled productivity and revenue growth. It will also highlight why complementary investments—in skills, workflows, and organizational redesign—are essential to escaping the productivity J-curve.

Attendees will leave with a practical framework for knowing where AI is creating value today and growing that value across the enterprise.

[Source](https://www.youtube.com/watch?v=57obkqdD2RM)

中文简报

Brynjolfsson称5%的用户创造了95%的企业AI价值企业若仅进行直接的软件采购，生产力增长便会停滞，除非投入九倍于此的资金围绕表现优异的超级用户重塑工作流程。 Erik Brynjolfsson解释称，尽管AI能力已突破人类基准，但受经典的生产力J曲线效应影响，加之技术本身每投入1美元就需要在流程重构与技能培训等互补性无形资产上投入9美元，企业的投资回报率仍显滞后。为缩小这一差距，企业必须采用颗粒度更细的基于任务的方法，将各项具体工作与AI适用性相对应，而非将各类职业视作铁板一块。纳斯达克等《财富》500强企业的实践数据表明，AI的使用遵循严格的幂律分布，仅5%的用户就创造了大约95%的总价值。企业可以通过识别这些隐藏的超级用户，将其先进的工作流程与提示词沉淀为全公司通用的操作手册，并追求以提升营收为导向的劳动力赋能增效而非单纯裁员，从而实现生产力的复合增长。尽管 AI 模型的基准测试性能迅速超越人类基线，但企业回报普遍滞后；研究表明，在直接技术投资上每支出 1 美元，往往需要配套投入约 9 美元的组织无形资产（如员工培训与业务流程重构）。 (1:43) 通用目的技术的引入会引发生产力 J 曲线效应，导致重组初期的生产力停滞或下降；针对 51 个企业级部署案例的分析显示，77% 的实施障碍源自人员管理与流程变革等组织因素，而非纯技术限制。 (3:53) 衡量 AI 的商业价值需要采取基于任务的拆解方法，将单一职业细分为具体的原子任务并按经济价值加权，从而帮助企业系统性锁定并迁移大量仍未充分利用 AI 的高价值“机会区”工作流。 (6:16) 企业内的 AI 应用遵循极端的幂律分布（如纳斯达克仅 5% 的超级用户创造了约 95% 的业务价值）；通过数据识别顶尖用户并将其提示词固化为企业模版向全员推广，有望将整体生产力提高一倍。 (9:38)

Reading summary

Erik Brynjolfsson argues that enterprise artificial intelligence has created an economic ROI chasm because technical breakthroughs outpace organizational adaptation. Realizing measurable business value requires managing AI as a general-purpose technology through granular task-level mapping, intangible process redesign, and systematically scaling the practices of internal power users. The Intangible Asset Iceberg and Productivity J-Curve (1:43) Although the Stanford AI Index Report documents machine capabilities crossing human baselines across diverse benchmarks, macroeconomic returns remain muted. Brynjolfsson frames this capabilities overhang through foundational research conducted with Lorin Hitt: for every single dollar spent on tangible hardware and software, organizations must invest roughly nine dollars in complementary intangible assets, including human capital retraining, business process redesign, and new product creation. These intangible investments total trillions of dollars in the broader economy but remain omitted from corporate balance sheets. Because artificial intelligence operates as a general-purpose technology, its implementation produces a classic productivity J-curve. Enterprises face an initial trough where restructuring costs, consultant fees, and managerial hours lower measured output before organizational investments compound into a harvest phase. Brynjolfsson notes that Stanford’s analysis of 51 enterprise deployments identified that in 77% of cases, organizational and change-management barriers—rather than technical limitations—determined project outcomes. Deconstructing Jobs Through Task-Level Economic Exposure (6:16) To navigate the J-curve, Brynjolfsson, alongside collaborators like Daniel Rock and Tom Mitchell, advocates abandoning monolithic occupational titles in favor of a task-based framework. Every profession decomposes into atomic duties—such as 26 discrete tasks for a radiologist or over 20 for an economist. Tasks are scored on their exposure to generative, predictive, or agentic AI and weighted by the economic value they produce, allowing leadership to roll task-level diagnostics back up to the enterprise level. Mapping organizational activity across conversation volume and task value exposes critical misalignments. Workhelix's diagnostic matrices show that enterprise usage frequently pools into an "Opportunity Zone" characterized by low AI conversation frequency on high-value business tasks. Left unmanaged, deployments risk degenerating into shadow AI and automated email spam that degrade productivity, instead of channeling machine capabilities toward high-leverage workflows. Exploiting Power-Law Adoption to Multiply Productivity (9:38) Across every enterprise analyzed, employee AI adoption strictly mirrors an extreme power law. In typical Fortune 500 environments, the median worker captures only 1% to 3% of addressable AI value. A Workhelix internal audit at NASDAQ illustrated this skew: approximately 5% of users generated roughly 95% of total business value from enterprise AI agents, while the remaining long tail used systems for basic drafting. Brynjolfsson emphasizes that shifting this distribution represents immediate economic leverage rather than an implementation failure. By using telemetry data to identify high performers—who are frequently quiet contributors rather than vocal evangelists—enterprises can extract their prompt architectures and workflows across strategic planning, customer service, and production code. Standardizing these workflows into departmental templates can effectively double aggregate corporate productivity. Replacing Top-Down Token Maxing With Process Redesign (13:16) Meaningful organizational adoption cannot be achieved through executive exhortations or crude usage metrics like "token maxing," which measure raw token consumption without evaluating business impact. Instead, management must apply causal inference and clear KPIs to pinpoint adoption across roughly 25,000 discrete operational tasks, equipping staff with peer-proven templates that have demonstrated twofold to tenfold performance gains. Furthermore, long-term returns depend on complete process reconfiguration rather than isolated task automation. While middle management often exhibits institutional inertia—hoping AI trends will blow over—leadership must actively realign job descriptions. Brynjolfsson cautions against treating AI solely as a headcount-reduction tool, asserting that superior returns come from orchestrating labor augmentation to capture new top-line revenue. While the power-law distribution of adoption offers a clear blueprint for multiplying enterprise value, sustaining these gains requires continuous measurement. Because underlying model capabilities shift rapidly from year to year, static playbooks quickly become obsolete, requiring organizations to perpetually re-evaluate task exposure and overcome persistent middle-management friction.

精炼阅读摘要

Erik Brynjolfsson指出，企业AI应用面临严重的投资回报率（ROI）鸿沟；要跨越生产力J曲线并释放技术价值，组织必须摆脱空泛的号召，通过任务级拆解、9比1的无形资产重组，以及规模化推广头部超级用户的经验来实现。能力溢出、ROI鸿沟与9比1无形资产冰山 (0:04) Brynjolfsson表示，董事会对AI的态度已从怀疑技术是否有效，全面转向追问投资回报率究竟在何处。尽管《Stanford AI Index Report》以520页的基准测试数据证实AI模型在多项认知任务上正在超越人类基准线并引发能力相变，但宏观经济与微观企业层面的收益却严重滞后。在缺乏严密衡量的情况下，企业往往将生成式AI用于处理低价值任务（例如生成大量垃圾邮件）或滋生影子AI，导致技术活动沦为空忙，甚至实质性损害组织生产力。 Brynjolfsson引用与Lorin Hitt等学者的实证研究指出，企业在IT硬件与软件等有形资产上每投入1美元，就必须向配套的无形资产补充投入约9美元。这些无形投资涵盖人力资本再培训、业务流程再造（BPR）以及新产品与服务的开发。尽管美国经济运行依赖数万亿美元的组织资本，但这些资产在传统企业财务报表与国民经济核算中完全缺失。只关注看得见的IT采购支出而忽视庞大的无形配套重组，是导致企业陷入ROI鸿沟的根本原因。生产力J曲线与基于任务的微观拆解框架 (3:53) 作为一种通用目的技术（GPT），AI在企业内部的落地必然引发典型的生产力J曲线效应。在初始集成阶段，企业需要投入大量顾问成本、管理者精力与员工工时来重构流程，此时账面成本激增而短期产出停滞，甚至导致统计上的生产力出现下滑。Brynjolfsson创立Workhelix的目的即在于缩短这一投资低谷并加速拐点到来。Stanford对51家企业成功部署案例的《Enterprise AI Playbook》研究显示，77%的实施阻碍源于组织层面的流程再造、人员变革与因果绩效追踪，技术本身并非限制因素。为了克服这一挑战，Brynjolfsson主张采用其在《Science》及《GPTs are GPTs》论文中阐述的“任务级方法”（task-based approach）。该方法将宽泛的职位描述解构为具体的工作任务（例如放射科医师拆解为26项任务，经济学家拆解为20余项活动），并对各项任务的AI暴露度进行低、中、高评级。大型语言模型虽然擅长起草分析备忘录，但对搬运箱子等物理劳动几乎毫无助益。通过将任务暴露度与经济价值加权聚合，企业能够穿透岗位迷雾，清晰定位自动化与劳动增强的真实落地点。对话与价值矩阵及AI应用的极端幂律分布 (8:04) 在企业实际运行中，Workhelix将AI应用绘制在二维坐标系上：纵轴代表AI对话量，横轴代表任务的客观业务价值。健康企业的理想状态是沿45度对角线分布，但绝大多数企业均深陷右下角的“机会区”（Opportunity Zone）——该区域聚集着极高业务价值的任务，但实际的模型交互频率却极低。与此同时，企业内部的AI采纳呈现出严格的幂律分布：在典型的《财富》500强企业中，中位数员工仅捕获了AI潜在价值的1%到3%；而在Workhelix对NASDAQ的评估中，仅占5%的头部超级用户便创造了AI智能体总业务价值的约95%。 Brynjolfsson强调，这种极端的价值倾斜并非落地的失败，而是巨大的低垂果实。普通员工往往仅使用AI润色邮件文案，而顶尖用户则利用模型架构企业营销战略、编写生产环境代码并编排自动化流程。企业提升ROI最直接的路径，不是等待尚未发明的外部技术，而是通过组织内部的持续学习机制，推动长尾用户模仿前20%的高阶操作，这种行为迁移足以将企业的整体生产力有效翻倍。行动指南：走出代币迷思并化解中层阻力 (13:16) 在具体的实施路径上，Brynjolfsson反对管理层下达“多用AI”的空泛指令，也反对单纯统计Prompt数量或Token消耗量的粗放考核。Workhelix通过评估约25,000项具体企业任务，为不同职能部门提供经过验证的实战模板。例如，直接为软件工程师、人力资源或市场营销人员提供带来2倍至10倍效能提升的同行Prompt，能够提供清晰的实操路径。此外，企业必须借助客观的可观测数据发现那些不善言辞的隐性超级用户，并将其工作流固化为组织资产。针对中层管理人员因习惯既有稳定流程而持观望态度的阻力，Brynjolfsson指出，CEO必须提供经过因果推断与明确KPI验证的实证数据来消除不确定性。他强调，跨行业的幂律分布模式高度一致，行业内不同员工间的差距远大于行业间的差距。具有前瞻性的企业应当将重心放在劳动增强与开拓新营收线条上，而非单纯追求削减人员编制；通过在生成式、代理式及预测式AI中持续重构端到端业务流程，企业才能在长达五年的周期内系统性获取AI红利。 Brynjolfsson总结认为，企业AI价值实现的核心矛盾在于技术性能的超前突破与组织重塑的相对滞后；企业唯有摆脱对技术本身的盲目崇拜，转向任务级因果度量与配套无形资产的系统重组，才能真正跨越中层阻力与生产力J曲线。

Detailed notes — zh

## Background 该视频由 YouTube 频道 Ai4 于 2026 年 8 月 15 日发布，主讲人为斯坦福数字经济实验室主任兼 Workhelix 联合创始人 Erik Brynjolfsson。视频页面简介指出，演讲聚焦于企业如何将强大的 AI 能力转化为切实的商业 ROI，基于斯坦福《企业 AI 实战手册》的 51 个真实落地案例，系统讲解如何通过任务级拆解、因果推断度量与配套组织变革来逃离“生产力 J 曲线”。 ## TL;DR Erik Brynjolfsson 深入剖析了当前技术能力飞跃与企业实际生产力平滞之间的“ROI 鸿沟”，指出每投入 1 美元的技术成本往往需要 9 美元的组织重构与技能等无形资产投资来配套。他提出“基于任务的方法论”，主张将职业拆解为细分任务束以精准识别高价值落地场景，并警惕“空喊口号”与盲目追求“Token 最大化”的管理误区。通过揭示企业内部普遍存在的“5% 超级用户创造 95% 价值”的幂律分布，他强调企业应通过识别、提炼超级用户的最佳实践与高效 Prompts，构建动态学习型组织，从而以人机协同拓展业务并实现整体生产力成倍跃升。 ## Atmosphere & style 整场演讲兼具顶尖经济学家的严谨实证底色与企业落地咨询的实战锋芒，语调从容睿智、一针见血。Erik 善于用幽默生动的冰山隐喻和清晰的坐标图表化解高管层的焦虑，将复杂的宏观生产力理论转化为立即可落地的行动指南。问答环节务实而富有穿透力，既犀利指出了中层管理者的组织惯性与避险心理，又以普适的跨行业数据为企业发掘内部增量指明了极具说服力的破局路径。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=4s) 企业关注焦点已从“AI 是否管用”转向“如何实现商业价值”，技术基准的爆发与企业实际生产力迟滞之间形成了显著的“ROI 鸿沟”。 - [02:53](https://www.youtube.com/watch?v=57obkqdD2RM&t=173s) 技术投资遵循“冰山法则”：每 1 美元的技术硬性支出，通常需要配套投入约 9 美元在人力资本、流程再造与组织无形资产上。 - [02:53](https://www.youtube.com/watch?v=57obkqdD2RM&t=173s) AI 作为通用目的技术必然经历“生产力 J 曲线”，落地初期因组织调整成本会导致账面生产力下滑，77% 的落地阻碍源于流程、人员与度量等组织性挑战。 - [06:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=376s) 采用“基于任务的方法（Task-Based Approach）”，将职位拆解为原子任务束，能够精准量化每项任务对 GenAI、Agentic AI 及预测性 AI 的暴露度与商业价值。 - [06:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=376s) 企业价值诊断显示，绝大多数企业的大量极高价值任务严重滞留在低频使用的“机会区”，转型的关键在于推动其向高频交互的“甜区”对角线迁移。 - [09:38](https://www.youtube.com/watch?v=57obkqdD2RM&t=578s) 企业内部 AI 采纳普遍呈现极度陡峭的“幂律分布”：约 5% 的超级用户创造了高达 95% 的业务价值，而绝大多数长尾普通用户仅捕获了极少潜在价值。 - [11:32](https://www.youtube.com/watch?v=57obkqdD2RM&t=692s) 提升企业 ROI 最直接的抓手是构建动态学习型组织：通过数据观测提炼超级用户的 Prompts 与工作流模板，规模化赋能长尾员工以实现全员生产力翻倍。 - [13:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=796s) AI 转型应从单点任务替代跃迁至端到端业务流程重构，跳出单纯裁员削减成本的狭隘视角，转向人机协同增强能力与拓展全新营收空间。 - [15:44](https://www.youtube.com/watch?v=57obkqdD2RM&t=944s) 告别空泛的全员号召与盲目追求“Token 最大化”的考核误区，企业必须深入业务具体细节，为不同岗位的细分任务提供明确、经过验证的实操操作手册。 - [18:29](https://www.youtube.com/watch?v=57obkqdD2RM&t=1109s) 跨行业的超级用户幂律特征惊人一致，用切实的量化指标突破中层管理者的组织惯性与观望心态，是持续释放未来 5 至 10 年生产力红利的核心策略。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=4s) AI 的能力过剩与“ROI 鸿沟” 演讲者 Erik Brynjolfsson（斯坦福数字经济实验室主任、Workhelix 联合创始人）以自己名字的冰岛语发音开场，并分享了一个有趣的现象：两年前他在董事会，大家还在询问“AI 这东西到底管不管用？”；而如今再也没人问这个问题了，所有管理者最核心的关切变成了——“ROI（投资回报率）究竟在哪里？我们到底怎么从中获取业务价值？” 当前技术与商业之间存在巨大的“能力过剩（Capabilities Overhang）”：AI 的技术潜力极其巨大，在斯坦福每年发布的 520 页《AI Index 报告》中，各项核心模型指标图表都在以更快的速度“向右上方狂飙”，甚至在越来越多的基准测试上大幅超越了人类水准（图表上标注的人类基准虚线）。这就像水温达到沸点瞬间从液体汽化一样，许多领域正发生从“人类能做”到“机器能做”的相变跃迁。然而在宏观经济和微观企业层面，宏观生产力数据并没有发生明显变化，许多企业感到极度失望。管理层虽然看到技术基准亮眼、工具持续部署、使用量节节攀升，但实际上积累的往往是大量“低价值使用”与影子 AI（Shadow AI）泛滥——比如《纽约客》员工抱怨收到更多 AI 生成的垃圾邮件。AI 不仅没有释放出预期的生产力，反而在一些场景下拉低了效率。这种技术能力与商业价值之间的脱节形成了一条巨大的“ROI 鸿沟（ROI Gap）”。 - 克服 ROI 鸿沟的核心钥匙是“度量（Measurement）”，正如管理学名言：“你无法管理你不能度量的事物（You can't manage what you don't measure）”。 ```mermaid flowchart LR accTitle: 技术能力与商业回报之间的 ROI 鸿沟 subgraph S1["技术端供给"] A["核心技术模型能力 (AI 指标指数级上升)"] --> B["企业部署工具 & 使用频次上升"] end subgraph S2["ROI 鸿沟 (ROI Gap)"] B -.->|"低价值泛滥 / 缺乏度量"| C["现实瓶颈 (生产力平平 / 失望)"] end subgraph S3["商业端转化"] C ==>|"科学度量与组织变革"| D["可衡量的商业 ROI & 生产力释放"] end ``` --- ### [02:53](https://www.youtube.com/watch?v=57obkqdD2RM&t=173s) 冰山法则、无形资产与生产力 J 曲线要跨越这条鸿沟，必须认识到技术投资与配套投资之间的比例关系。Erik 引用自己与 Lorin Hitt 等学者的长期实证研究指出：技术投资就像一座冰山，硬件或软件的硬性支出仅仅是浮出水面的一角。 > “在人力资本、业务流程重构、新产品与服务开发上，你实际需要投入大约 9 美元，来匹配每 1 美元的技术投资。” 美国经济中存在数万亿美元未在国家账户或企业资产负债表上体现的无形资产（Intangible Assets），但它们往往比硬资产更为值钱。AI 作为一种通用目的技术（GPT，经济学上的 General Purpose Technology，而非仅仅指 Generative Pre-trained Transformer），在引入初期需要企业投入高昂成本去重塑业务流程、引入顾问、培训员工技能，这必然占用大量管理与员工时间。在支出增加而短期产出尚未显现的阶段，企业生产力在账面上会呈现停滞甚至下滑，随后当无形资产投资开始集中变现时，生产力才会迎来爆发式上升——这就是著名的“生产力 J 曲线（Productivity J-Curve）”。根据斯坦福团队对 51 个成功落地案例编写的《企业 AI 实战手册》（Enterprise AI Playbook）研究，高达 77% 的阻碍并非源于技术本身，而是来自人员、流程以及度量等组织性挑战。企业的核心任务是缩短 J 曲线的摸索期、减浅下滑的低谷，并让上升轨迹更加陡峭迅速。 ```mermaid flowchart TD accTitle: 生产力 J 曲线的形成与跨越 subgraph SG1["1. 初期技术与无形投资 (J 曲线下行)"] A["引入 AI 通用技术 (1 美元)"] --> B["组织配套重构 (9 美元): 流程再造 + 人力技能 + 管理时间"] B --> C["短期成本上升，产出未见增长 (账面生产力下滑/平滞)"] end subgraph SG2["2. 价值收割与跃升 (J 曲线陡峭上扬)"] C --> D["无形资产积累完成"] D --> E["业务价值集中变现与 生产力大幅释放"] end ``` --- ### [06:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=376s) 基于任务的方法论与企业价值诊断模型为了精准度量并识别 AI 赋能的机会，Erik 与 Tom Mitchell（在《Science》发表）以及 Daniel Rock 等学者（与 OpenAI 团队合作撰写《GPTs are GPTs》）共同开创了“基于任务的方法（Task-Based Approach）”。该方法主张不要将一个职业或整个组织视为不可分割的黑盒，而是将每个岗位拆解为一整套“任务束（Bundle of Tasks）”： - 放射科医生包含 26 项截然不同的任务； - 经济学家包含 20 多项任务； - 记者、程序员等职业同样由细分任务组成。将岗位拆解到任务粒度后，便可评估单项任务对 AI 的暴露程度（低、中、高）。例如，撰写备忘录等文本任务对大语言模型（LLM）高度暴露，而搬运箱子等体力任务则几乎不受 LLM 影响。将各项任务的 AI 潜力得分乘以其业务价值贡献，再自下而上汇总到岗位、部门直至整个组织，即可清晰洞察 AI 机会的具体分布。在基于该逻辑构建的企业价值散点图（纵轴为“AI 交互使用频次”，横轴为“任务业务价值”）中，企业通常呈现三大区域： 1. **左上方区域（低价值/低频）**：基本为空白，属于无需过多关注的低价值区间。 2. **中间甜区（Sweet Spot）**：分布在 45 度对角线上，使用频次与任务价值成正比，处于理想状态。 3. **右下方机会区（Opportunity Zone）**：聚集了大量极高价值但使用频次严重偏低的任务。绝大多数企业的大量任务都滞留在此区域，Workhelix 的目标就是通过工具驱动企业将这些高价值任务从机会区向对角线“甜区”迁移提升。 --- ### [09:38](https://www.youtube.com/watch?v=57obkqdD2RM&t=578s) 企业内部的幂律分布与超级用户杠杆通过对大量企业的实际数据分析，Erik 指出所有企业内部都存在一种极其经典的“幂律分布（Power Law）”：极少数超级高价值用户（Power Users）构成了价值贡献的核心，而占据绝大多数的长尾普通用户（Typical Users）仅获取了微薄的价值。在 Erik 走访的一家财富 500 强企业中，测算显示普通用户捕获的潜在价值仅为 1% 左右，其他企业通常也仅在 2% 至 3% 之间徘徊；在纳斯达克（NASDAQ，Workhelix 的客户之一），数据显示大约 5% 的用户通过 AI Agent 创造了高达 95% 的业务价值。 | 用户群体 | 典型占比 | 行为特征 | 产出价值量级 | |:--- |:--- |:--- |:--- | | **超级用户 (Power Users)** | 约 5% | 制定全公司复杂营销方案、高阶代码生成、构建工作流工具 | 创造约 95% 的业务价值（2x-3x 生产力释放） | | **长尾普通用户 (Long Tail)** | 约 95% | 仅用来润色邮件、编写简单文本等低杠杆场景 | 仅获取潜在价值的 1% ~ 3% | 许多企业管理者会因普遍的使用低效而沮丧，但从经济学与管理角度来看，这种悬殊的分布恰恰蕴藏着巨大的增长空间——企业提升 ROI 最直接、最高效的路径，并非盲目采购新工具，而是识别并提取那 5% 超级用户的行为范式，系统化地带动 95% 的长尾用户向超级用户进阶。 ### [11:32](https://www.youtube.com/watch?v=57obkqdD2RM&t=692s) 从超级用户到全员赋能：构建动态学习型组织在展示图表时，Erik 指着屏幕上的绿色区域强调，图表上那一小块看似不起眼的绿色区域，实际上代表着仅通过这一层面的提升，就能让整个公司的生产力直接翻倍。而且这绝非终点，企业仍有巨大的空间去持续拉升这一曲线。他指出，AI 的赋能绝非一次性的静态项目，而是一个“动态机遇”（dynamic opportunity）——企业需要建立持续评估机制，观察最优秀的用户在做什么，提炼出可复制的模板，并将其开放给整个组织的其余成员。提升生产力的打法非常清晰直观，但一切的前提都始于“测量与可见性”（measurement and visibility）。如果企业能不断拓展这一闭环，就能逐步打造出一个自我进化的学习型组织。随着底层技术的持续演进，全新的用例与可能性会不断涌现，企业内部的“超级用户”（power users）群体也会随之发生动态变化。Erik 幽默地提到，有些超级用户性格非常内向，平时并不多话，但在系统数据的微观分析下，他们的卓越表现会一览无余；两年前的超级用户与今天的超级用户大不相同，而明年的超级用户又将是另一批人。一旦识别出这些人，企业就能制定出针对性的操作手册（playbook），挖掘那些真正有效的提示词（effective prompts）：“无论是用它写出更高质量的代码、制定更优越的营销方案，还是显著改善客户服务，我们都可以把这些做法提炼出来，转化为标准化模板分享给部门同事，从而带动全员跃迁。” ```mermaid flowchart TD accTitle: 动态学习型组织演进闭环 subgraph S1["1. 发现与提炼"] A["数据全量观测与分析"] --> B["识别超级用户与高效 Prompts"] end subgraph S2["2. 赋能与推广"] B --> C["构建部门标准化 Playbook / 模板"] C --> D["推广至普通员工拉动生产力翻倍"] end subgraph S3["3. 迭代与进化"] D --> E["业务流程重构与新能力涌现"] E --> A end ``` ### [13:16](https://www.youtube.com/watch?v=57obkqdD2RM&t=796s) 行动指南：从任务分解到端到端业务重构 Erik 梳理了企业管理者在“明天一早回到公司即可着手推进”的行动清单。首先，必须将底层技术能力与切实的商业价值挂钩。企业需要全面扫描内部所有的职业岗位，将其彻底拆解至“任务”（tasks）这一原子级颗粒度。无论针对生成式 AI（Generative AI）、智能体 AI（Agentic AI）还是预测性 AI（Predictive AI），都可以采用这一套分析框架，随后建立清晰的 KPI、因果推断（causal inference）和系统可观测性来追踪实际进展。然而，任务级别的拆解仅仅是转型的起点。Erik 强调，企业最终必须重塑整个端到端业务流程，绝不能仅仅停留在替换孤立的零部件上，而是要全面重新配置流程，这意味着工作岗位的本质将被深刻重塑。在这一过程中，某些旧任务被自动化取代，而全新的能力与增量空间被创造出来。他特别提醒，很多企业眼光过于狭隘，往往只盯着裁员削减成本（headcount reduction），但真正巨大的机会在于利用 AI 赋能并补充（augment and complement）员工，拓展以前无法涉足的全新业务，从而带动顶线营收（top line）增长。他分享道，自己曾与一位高级管理层交流，对方最初只想利用 AI 来削减团队人数，但经过梳理后才发现，真正能带来可观 ROI 的突破口，其实在于更精准地衡量并释放 AI 对业务增量的推动价值。 - **三大技术维度覆盖**：在任务颗粒度上全面扫描评估 Gen AI、Agentic AI 及 Predictive AI 的适用边界。 - **量化衡量核心体系**：依托因果推断、高清晰度 KPI 与全链路可观测性构建科学 ROI 评估模型。 - **避免单一裁员思维**：推动从“劳动力替代”转向“人机协同增强”，着眼于开拓新业务能力与扩大营收规模。 ### [15:44](https://www.youtube.com/watch?v=57obkqdD2RM&t=944s) 提问环节一：告别泛泛号召与“Token 最大化”，精准落地具体任务在现场问答环节，一位来自亚马逊（Amazon）AI 布道团队的观众率先提问。该观众表示自己所在团队正全力推动公司内部的 AI 落地，但目前自身影响力有限，且整体采纳正面临边际效益递减（diminishing returns）的瓶颈，迫切想知道如何进一步扩大突破口。Erik 盛赞这是一个价值“数万亿美元的核心问题”，并肯定了亚马逊在落地方面已经走在绝大多数企业前列。随后，Erik 犀利剖析了过去一两年中屡见不鲜的企业管理盲区。他直言，很多 CEO 常常群发全员备忘录，泛泛地喊口号说“大家要多用 AI”，但这种空洞的号召根本无济于事；更荒谬的是，有些企业甚至走向极端去考核所谓的“Token 最大化”（token maxing），按员工消耗的 Token 数量论英雄。他指出：“这些高管的初衷固然是好的，他们深知 AI 至关重要，但空泛的说教根本无法带来实际改变。你必须深入泥泞的细节（get down into the weeds）。”企业必须精确回答各职能部门的具体落地路径——HR 经理究竟能用 AI 做什么？程序员能用 AI 做什么？销售与市场人员又该怎么用？通过将典型企业的业务解构成多达 25,000 项具体任务，明确划定哪些任务绝对不该用 AI、哪些任务必须立即使用，并直接提供同行已验证能带来 2 倍、3 倍甚至 10 倍效能飞跃的优质提示词。当员工拿到这样具体、可落地的实操指南时，才会真正发自内心地拥抱技术。此外，由于技术演进极为迅猛，当前的能力版图与半年前相比已大相径庭，企业必须借助常态化的动态测量与更新工具，才能持续维系 ROI 的爆发式增长。 ### [18:29](https://www.youtube.com/watch?v=57obkqdD2RM&t=1109s) 提问环节二：克服组织惯性与各行业普遍存在的“幂律分布” 第二位观众随后提问：在实际推行中是否会遭遇来自中层或基层员工的阻力？例如 CEO 可能非常狂热，但要真正把具体任务梳理清楚并让员工采纳落地往往充满困难。Erik 深表赞同，直言这是他在各类企业中最常目睹的典型现象。董事会和 CEO 往往充满激情，觉得 AI 是一场不可错过的技术巨浪；但往下走一级，中层管理者却往往充满疑虑与戒备，甚至心存侥幸：“这玩意儿会不会像之前的比特币一样？也许我们只要把头埋在沙子里按兵不动，这阵风刮过去就算了。”然而 Erik 明确指出：“AI 绝不可能随风而逝。”颇具讽刺意味的是，一家极其成功的成熟企业往往内置了强大的组织惯性，企业越成功，就越倾向于固守并延续以往行之有效的传统流程。为了打破这种阻力，高管必须拿出切实可见的成效，用具体量化的事实替代抽象概念来重塑沟通对话。最后一位观众就此前演讲中提到的数据提问：演讲中展示了纳斯达克（NASDAQ）内部“前 5% 的超级用户驱动了 99% 的 AI 业务价值”，在不同行业之间是否会存在不同的偏度趋势？Erik 给出了一个出人意料的洞察：这一价值倾斜在跨行业间表现得“惊人地一致”（surprisingly similar）。他坦言：“我也很想告诉你制造业与金融业截然不同，但实际数据分布形态几乎如出一辙。企业内部不同人员之间的鸿沟，远远大于不同行业企业之间的差异。”在他们调研过的每一家企业中，无一例外都遵循着这种极度陡峭的“幂律分布”（power law）。Erik 总结道，这种普适的幂律规律其实是一个天大的利好，因为它意味着企业内部蕴藏着极其庞大且极易采摘的“低垂果实”（low-hanging fruit）。只要有一套程式化、标准化的科学打法将少数超级用户的能力复刻至全员，企业就能在未来至少五到十年内，持续源源不断地释放巨大的生产力红利。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=4s) 两年前他在董事会，大家还在询问“AI 这东西到底管不管用？”；而如今再也没人问这个问题了，所有管理者最核心的关切变成了——“ROI（投资回报率）究竟在哪里？我们到底怎么从中获取业务价值？” — Erik Brynjolfsson > [00:04](https://www.youtube.com/watch?v=57obkqdD2RM&t=4s) 你无法管理你不能度量的事物（You can't manage what you don't measure） — Erik Brynjolfsson > [02:53](https://www.youtube.com/watch?v=57obkqdD2RM&t=173s) 在人力资本、业务流程重构、新产品与服务开发上，你实际需要投入大约 9 美元，来匹配每 1 美元的技术投资。 — Erik Brynjolfsson > [11:32](https://www.youtube.com/watch?v=57obkqdD2RM&t=692s) 无论是用它写出更高质量的代码、制定更优越的营销方案，还是显著改善客户服务，我们都可以把这些做法提炼出来，转化为标准化模板分享给部门同事，从而带动全员跃迁。 — Erik Brynjolfsson > [15:44](https://www.youtube.com/watch?v=57obkqdD2RM&t=944s) 这些高管的初衷固然是好的，他们深知 AI 至关重要，但空泛的说教根本无法带来实际改变。你必须深入泥泞的细节（get down into the weeds）。 — Erik Brynjolfsson > [18:29](https://www.youtube.com/watch?v=57obkqdD2RM&t=1109s) 这玩意儿会不会像之前的比特币一样？也许我们只要把头埋在沙子里按兵不动，这阵风刮过去就算了。 — Erik Brynjolfsson > [18:29](https://www.youtube.com/watch?v=57obkqdD2RM&t=1109s) AI 绝不可能随风而逝。 — Erik Brynjolfsson > [18:29](https://www.youtube.com/watch?v=57obkqdD2RM&t=1109s) 我也很想告诉你制造业与金融业截然不同，但实际数据分布形态几乎如出一辙。企业内部不同人员之间的鸿沟，远远大于不同行业企业之间的差异。 — Erik Brynjolfsson ## Mentioned Erik Brynjolfsson, Stanford Digital Economy Lab, Stanford Institute for Human-Centered AI, Workhelix, Lorin Hitt, Tom Mitchell, Daniel Rock, OpenAI, NASDAQ, Amazon, AI Index Report, Enterprise AI Playbook, Canaries in the Coal Mine, GPTs are GPTs, Science, The New Yorker ## Video description Erik Brynjolfsson - Director of the Stanford Digital Economy Lab, a Professor at the Stanford Institute for Human-Centered AI, Co-Founder at Workhelix AI is advancing rapidly, but the central question for enterprise leaders is no longer whether it works—it is how to convert capability into measurable business value. This talk will show why AI transformation requires more than model adoption, drawing on evidence from Canaries in the Coal Mine on how AI is already reshaping work and from the Stanford Enterprise AI Playbook’s lessons from 51 enterprise deployments. The talk will explain how to identify high-value opportunities at the task and process level, measure ROI with credible KPIs and causal inference, and move from scattered pilots to scaled productivity and revenue growth. It will also highlight why complementary investments—in skills, workflows, and organizational redesign—are essential to escaping the productivity J-curve. Attendees will leave with a practical framework for knowing where AI is creating value today and growing that value across the enterprise. [Source](https://www.youtube.com/watch?v=57obkqdD2RM)

Source quotes

[0:04] Erik Brynjolfsson: 两年前他在董事会，大家还在询问“AI 这东西到底管不管用？”；而如今再也没人问这个问题了，所有管理者最核心的关切变成了——“ROI（投资回报率）究竟在哪里？我们到底怎么从中获取业务价值？” [0:04] Erik Brynjolfsson: 你无法管理你不能度量的事物（You can't manage what you don't measure） [2:53] Erik Brynjolfsson: 在人力资本、业务流程重构、新产品与服务开发上，你实际需要投入大约 9 美元，来匹配每 1 美元的技术投资。 [11:32] Erik Brynjolfsson: 无论是用它写出更高质量的代码、制定更优越的营销方案，还是显著改善客户服务，我们都可以把这些做法提炼出来，转化为标准化模板分享给部门同事，从而带动全员跃迁。 [15:44] Erik Brynjolfsson: 这些高管的初衷固然是好的，他们深知 AI 至关重要，但空泛的说教根本无法带来实际改变。你必须深入泥泞的细节（get down into the weeds）。 [18:29] Erik Brynjolfsson: 这玩意儿会不会像之前的比特币一样？也许我们只要把头埋在沙子里按兵不动，这阵风刮过去就算了。 [18:29] Erik Brynjolfsson: AI 绝不可能随风而逝。 [18:29] Erik Brynjolfsson: 我也很想告诉你制造业与金融业截然不同，但实际数据分布形态几乎如出一辙。企业内部不同人员之间的鸿沟，远远大于不同行业企业之间的差异。

Source quotes

[0:04] Erik Brynjolfsson: Where's the ROI? How do we get value from it? [0:04] Erik Brynjolfsson: capabilities overhang [1:43] Erik Brynjolfsson: you can't manage what you don't measure. [1:43] Erik Brynjolfsson: knock on with your knuckles [3:53] Erik Brynjolfsson: Enterprise AI Playbook [6:16] Erik Brynjolfsson: task-based approach [6:16] Erik Brynjolfsson: GPTs are GPTs [8:04] Erik Brynjolfsson: Opportunity Zone [15:44] Erik Brynjolfsson: use AI more [15:44] Erik Brynjolfsson: token maxing [18:29] Erik Brynjolfsson: blow over

Original transcript

[0:04] Hi, I'm Eric Brinolson. That's how you pronounce it. Uh, it's an Icelandic name. [0:09] I'm delighted to be here. Two years ago, I was in a boardroom and the folks there were asking me, [0:17] "Hey, does this AI stuff really work?" Nobody's asking me that anymore. Now, the question is, [0:25] where's the ROI? How do we get value from it? And that's what I'm going to tell you about today [0:30] because there's a huge capabilities overhang. The technology has enormous potential but companies [0:37] are not converting that into real value. Uh the overall macroeconomy is not changing a [0:42] whole lot and individual companies many of them are understandably very disappointed. [0:48] We do see huge improvements in capabilities at the AI index. We had this report uh you can get [0:54] a free copy of it. We publish it every year. It's 520 pages filled with charts like this [1:00] one. Um, and you can see what's happening in the core capabilities. Those lines are all going up [1:05] and to the right faster and faster. I had them put in that dotted line just for fun. That's [1:12] how well humans do on the same um exams as these models are doing. And as you can see in more [1:18] and more categories, machines are outperforming humans. And it's almost, I think, like crossing [1:24] a threshold when water, you know, crosses the boiling point. It converts from a liquid to a gas. [1:30] What we're seeing is in many places there's a switch from having a machine, a human be able [1:36] to do it to a machine be able to do it. But the problem is that companies are not converting that [1:43] into real value. We call it the ROI gap. It's what work helix is focused on doing. leadership sees [1:49] these impressive temp tech technical benchmarks, the tools being deployed, usage usage numbers [1:54] going up. They see all of that, but at the same time, what's actually accumulating is a lot of low [2:00] value usage. I was just talking to a uh someone from the New Yorker who was describing how they [2:05] were getting spammed with a lot of email that wasn't really valuable to them. And there are so [2:10] many other lowv value uses out there. Shadow val shadow AI is proliferating. And so instead of this [2:16] unlocking productivity, sometimes it's doing just the opposite. So there's a gap between [2:21] them. I'd really call it a chasm. And we need to cross that chasm. How do you cross the chasm? [2:28] Measurement. Measurement is the key. You can't manage what you don't measure. So my research [2:35] over a couple of decades has focused on understanding the value of it and now AI. [2:41] How do we measure that? How do we translate capabilities into performance? And it's why [2:46] we started work helix to create a set of tools to make it easier for companies to do that. [2:53] And it's not just the technical parts, it's the whole rest of the organization. In work I did [2:58] with Lauren hit going back and and successive work with other colleagues, what we found [3:03] is that this iceberg is not just a metaphor. It's literally about $9 of investment you need to make [3:10] in human capital, business process redesign, developing new products and services for every $1 [3:17] of technology investment. People see the technology investment, the CEO, the CTO see that [3:24] you pay hard dollars for it. You can often touch it and you know knock on it with your knuckles. [3:29] But the intangible assets are much bigger. The US economy has trillions of dollars of intangible [3:37] assets, business processes, and human capital that are not measured anywhere in the national [3:41] accounts. They're not measured in most companies balance sheets, but they're actually more valuable [3:46] than the hard assets. Part of my research agenda has been to quantify those more explicitly. [3:53] They also have some practical implications because these intangible assets are so important [3:58] especially when you introduce a new powerful generalpurpose technology like AI. AI is a [4:05] GPT. Economists use GPT to mean general purpose technology not just generative pre-trained [4:11] transformer. Um because of those intangible assets are costly but they don't immediately pay off. [4:19] What you tend to see is kind of a J curve in performance. So we have a series of papers on this [4:24] as well. And the idea of the J curve is that in your organization you get these technologies but [4:30] then you need to rethink your business processes. You need to rethink what kinds of skills the [4:35] workforce has. Um that's costly. Sometimes you bring in consultants to do it. Always [4:40] you involve management time and worker time. While you're doing that you have more expense, [4:46] no increase in output. So by definition your productivity looks worse or flat at best. [4:53] Later you start harvesting those investments, those intangible investments and that's where you [4:59] get the upward part of the J curve. I think most companies are just at the turning point of that J [5:04] curve. They haven't fully taken advantage of the potential the harvest that they could be doing. [5:09] And my mission in life and with the company is to shorten that J curve to make the dip shallower [5:16] to make the uh upward part upward trajectory faster and steeper and the way you do that [5:22] is by measuring and identifying the opportunities more explicitly. So for instance [5:28] uh with a couple of colleagues at Stanford we put together what we called the uh enterprise [5:34] AI playbook and we looked at 51 successful deployments. There have been other people who [5:39] talked about all the failures and there's lots of failures out there. We decided to take a different [5:43] perspective. What did the successful deployments have in common and what we found was that in most [5:49] cases 77% of the cases it was not the technical part. It were these organizational challenges [5:55] that really needed to be overcome people, processes and especially measurement. And we'll [6:01] if you guys hang on at the end of the talk I'll give you a QR code so you can download [6:05] the uh enterprise playbook and see for yourself. You can dive into all those case studies. Uh there [6:10] will probably be several companies in your industry that you can learn from. [6:16] The key methodology that we found has been really helpful is what we call the taskbased approach. [6:21] This is something I pioneered with Tom Mitchell. We wrote about it in uh in the journal Science. uh [6:27] one of my students Daniel Rock a co-founder of uh work helix uh fleshed it out further with a [6:33] team from open AAI in a paper called GPTs or GPTs. Let me explain briefly the idea of the taskbased [6:40] approach. You can read about it in those papers if you'd like. The idea is that if you try to look at [6:45] a whole organization or a whole occupation, it's really hard to make headway of how [6:50] AI is going to help you. But if you break it down into the task, you can do much more. Every job, [6:57] every occupation is a bundle of tasks. You know, a radiologist does 26 distinct tasks. An economist [7:05] does over 20 tasks. A journalist, a coder. You break it down into those individual tasks. [7:12] When you get at that level, it's so much easier to make progress because individual tasks, you [7:17] can evaluate them, as we show here, as to whether or not they're exposed to AI from low to medium [7:23] to high. There are some tasks like writing a memo. AI can do a lot for that. LLMs are great at that. [7:29] Other tasks like I don't know, lifting a box. LM's aren't going to help you a whole lot with that. [7:34] As you evaluate each of the tasks, you get a score. And then you multiply those by how much [7:40] value you get from that task. And then you roll it back up to the occupation. You roll it back up [7:46] to the whole economy or the whole organization. And you start getting incredible visibility into [7:51] where your opportunities are. It's as simple as that. Simple in conception. Execution is trickier. [7:58] That's why we developed software to automate all of this. Let me give you a peek at how it works. [8:04] So look at the graph there. On the left you have um um you have the vertical axis is how many [8:12] conversations people are using the technology for. On the horizontal axis is how much value you get. [8:19] Basically that taskbased approach. And what you see is there sort of like a more to the upper [8:25] right. That's good because what that means is that in this organization there are more conversations [8:31] in tasks that are more useful. But I could divide it into three zones. The upper left zone is [8:38] mostly empty. Those are places where there's low value and there aren't a lot of conversations. [8:45] uh the middle zone there, call it the sweet spot. Those are places on that sort of [8:51] 45 degree line where the number of conversations is roughly proportional to the value. [8:58] And then the lower right, that's your opportunity zone and that's where most conversations are in [9:04] most companies we visit. Those are conversations that are way below what they could be. So [9:09] people are not using the technology to its full potential. What you want to do and what we want [9:15] to help you do is move them up to that sweet spot there. You don't want to go too far, but what you [9:21] want to do is when there's opportunities, you increase the use of the technology. [9:26] And most companies aren't doing that nearly nearly fast enough. So, work helix basically helps you [9:32] move from the opportunity zone up to the diagonal. And we have a set of tools to help you do that. [9:38] How do we do it? Well, one of the first things is we look at what people are actually doing there. I [9:43] see a lot of you taking pictures, which is great. I'd be happy to send these slides to you as well. [9:48] Again, if you if you fill out the form with the QR code at the end, I will send you all these slides. [9:53] So, um, here's a pattern that we see. I was going to say in most companies, actually, we see this [9:57] in every single company I've looked at, and I've looked at a lot. And this is a classic power law. [10:03] A power law has a few super high value users on the upper left there. Um, and then there's [10:09] this long tail of the typical user, and that's basically how much value they're getting from [10:15] it. We can score that. So, the typical user is getting maybe 1% of the value. I was just visiting [10:21] a big Fortune 500 company um and uh we did the calculation for them and it was it was almost [10:27] exactly 1% of the value that they were getting. And that is not atypical. Some companies it's 2 or [10:32] 3%. Um, but there's also this big distribution um, where some people are getting a lot of value. At [10:39] NASDAQ, one of our work helix customers, what we found was that about 5% of the users were creating [10:45] about 95% of the value with AI agents. It was super skewed. That's the classic power law. So, [10:52] I see this as opportunity. A lot of companies are disappointed. Why aren't more people getting value [10:56] from it? But that that low usage is actually an opportunity because it's a very straightforward [11:02] way of increasing value. And that's to get more of those people who are at the long tail. [11:09] Get them to behave more like the power users, those super users that are getting 2x 3x value. [11:15] The longtail users maybe were only using it to write more eloquent emails. [11:19] the power users were creating a marketing plan for the whole company that was much more sophisticated [11:25] uh cycling through or better code or other tools. And what you want to do is get that blue area [11:32] to raise it up and be like the green area. That little sliver of green area may not look like [11:37] much, but that's doubling the productivity of that company just from doing that uplift there. And [11:42] that's not the end of it. You've got room to keep increasing it more and more. And this is a dynamic [11:48] opportunity. You don't simply do this one off. You're constantly assessing what the best users [11:55] are doing, identifying those templates, making it accessible to the rest of the organization. [12:00] It's a really straightforward game plan for improving the productivity of your organizations. [12:05] But it all starts with measurement and visibility. And that's one of the things we want to do. [12:09] And I'd love to take your questions and comments on this, by the way. So [12:15] if you continually expand this, you create a learning organization that just keeps getting [12:18] better and better. As the technology evolves, what you see is that there's new use cases. There's new [12:24] possibilities. Also, the set of users that are power users change over time. Some of them are [12:30] introverted. They don't talk a lot, but the data can reveal who they are. And the ones that were [12:35] power users two years ago may be different than the ones that are power users today and the ones [12:39] that are going to be power users next year. Once you've identified them, you can create a playbook [12:46] and show what's doing. So you can go through and identify the effective prompts. What are those [12:52] effective prompts? And you can go through and see, okay, here's somebody who's using it to create [12:59] better code, to create a better marketing plan, to improve customer service. let's highlight what [13:04] they're doing and share that with other people in their department and get them to have a template [13:10] so they can improve it. And this is something that we uh can help organizations with. So [13:16] I want to take some time to get some questions and comments from all of you. But let me just wrap up [13:20] by saying what you can do tomorrow morning is first off connect the capabilities to business [13:26] value. And what you want to do is scan through the organization all of the occupations you have. [13:32] break them down through the tasks. You can look at my writing in science and other articles or [13:37] you can use the tools that we have available. You can do it for Gen AI, Aentic AI, predictive AI. [13:45] Then start tracking progress, causal inference, clear KPIs, and observability. [13:53] The taskbased approach is a great way to do this. You break things down into their atomic unit [13:59] and measure them more carefully. But it's just the start. Actually, you also eventually have to [14:05] reinvent your entire processes. You can't just take the individual components. You also need [14:11] to think about how to reconfigure them. And what that means is changing the nature of jobs. [14:17] Parts of jobs no longer are needed. AI can do them. New capabilities are there. And you want [14:23] to expand those. What that means is that instead of simply reducing employment for people, in [14:29] many cases, there's an opportunity to uh increase employment and have them do new things that they [14:36] weren't doing before. And most companies aren't sufficiently focused on that top line. And um in [14:43] particular, I like the idea of using AI to augment what people are doing, complement what people are [14:48] doing rather than substitute. I'll I'll wrap up with a little story about one uh senior executive [14:54] I was talking to and and who wanted to use AI just to reduce the headcount of their organization [15:02] and instead there was an opportunity there to measure more carefully the way that AI was adding [15:08] value to the top line and that was something that they needed help with to get that ROI. [15:14] So, if you'd like to learn more, here are some URLs where you can download my research papers. [15:18] You can get more about Work Helix. As promised, if you scan that QR code there, I'll wait give [15:24] you a chance to do that. Um, we'll send you the enterprise playbook. Um, we can send you more [15:29] information. I'll be happy to send you a copy of these slides as well. So, I've left a few minutes [15:34] here for questions. Happy to take any questions or uh uh comments you may have. Do we have a [15:39] microphone? Yeah. Let's go ahead and and see what questions you have. There's one right over here. [15:44] So, what do you see as the most effective way to drive adoption? On that chart, you showed [15:49] two-level increase. I work at Amazon. I'm part of an evangelist team to drive adoption. Yes. And [15:57] my span of influence is limited and I'm reaching diminishing returns. Well, how can I expand? [16:02] That's a great question. It's a trillion dollar question and and um Amazon's been doing it better [16:06] than most. I have to give you guys credit uh and maybe you specifically. Uh that said, [16:11] I'll tell you one of the big frustrations is that a year or two ago, I kept seeing all these memos [16:16] go out by CEOs saying, "Use AI more and then that wasn't really doing much." And they said, "Let's [16:22] token max. You know, we're going to count how many tokens you do." That was not the right way either. [16:26] I I think that the their heart was in the right place. They understood that AI was important. [16:31] They wanted people to use it more. But these generic ex, you know, exertations to use AI more [16:37] don't really move the dial. What do you have to do? You have to get concrete and tangible [16:41] about specific things and that's what it gets you. So you got to get down into the weeds. [16:46] What exactly can this person in this department use AI for? What can the HR uh manager use AI for? [16:54] What can the coder use AI for? What can the sales and marketing person use AI for? [16:58] And that's what we try to do at work is we get down to the specific tasks. Everybody gets [17:03] a different set of opportunities. We can assess 25,000 different tasks that company, the typical [17:10] company does and we can say here are the ones you should not be using AI for and here's the ones you [17:15] should be using AI for. And go one step further. I showed you a couple of slides back. Here are [17:20] some uh prompts that some of your colleagues are using and we can see they're getting 2x, 3x, 10x [17:27] more performance by doing that. You give people actionable work like uh opportunities like that [17:34] and wow they they are so happy because now they they want to be productive and you're giving them [17:40] a a a a guide book to doing that. So that's the night and day difference between these generic [17:46] you know use AI more versus here's something you specifically can do. And I want to stress [17:51] this is not a oneanddone kind of thing. You know the you know the technology is evolving so rapidly [17:57] so you've got to constantly be updating. We are constantly updating our set of tasks [18:01] of what AI can do well and not so well. It's it's substantially different now in the summer of 2026 [18:09] than it was last fall. And so you just got to be keeping up. And so you want to have a tool [18:14] that constantly measures this and constantly updates it. You do that and we're seeing, [18:18] you know, uh, huge increases in the ROI of companies that adopt that approach. [18:22] Does that answer your question? Yes. Great. Okay. Other questions? I think right here over there. [18:29] Could you say something on whether there's potential resistance? Maybe the CEO loves [18:34] your idea, but to really get the tasks out the individuals may not be easy all the time. Yeah, [18:41] I have to say uh what you just described is such a common pattern that I see. [18:46] It's it's very frequent that I I talk to boards and CEOs and they tend to be pretty enthusiastic [18:51] about AI. They've heard about, hey, AI is going to be big. Let's get let's get involved. At the [18:56] same time, the next level down middle managers are a little more apprehensive. They're like, [19:00] "Oh, is this going to be like Bitcoin? Maybe if we just keep our heads down, this will blow over." [19:04] You know, AI is not going to blow over. It's not going to go away. But the irony is that a [19:09] successful organization often is designed to have a lot of inertia. The more successful a company [19:15] is, you want to preserve those business processes that have been effective and perpetuate them. But [19:20] that means that the CEO is going to have to work extra hard to overcome that inertia. And you need [19:25] to have you need to show people that this is this is real and tangible. And it gets again down to [19:31] concrete measures. Most companies just don't have the measures. But it it's it's a solvable problem. [19:37] And if you get in there and you instrument and identify concretely what can be done, it's it's a [19:42] whole different conversation. Do we have time for one more question? We have one more question. And [19:47] based on your experience working with different companies and industries, are you seeing a [19:52] different trend in terms of the skewed value? It was really interesting to see that NASDAQ 5% of [19:58] users are driving 99. That's right. Are you seeing any trends different industries? Um, you know, [20:04] it's surprisingly similar. I'd love to be able to tell you that manufacturing is totally different [20:08] than finance. Um, the pattern looks surprisingly similar. The difference is between is inside of [20:14] companies more than between companies. It is this power law in every company we see. Now to be fair, [20:21] the exact shape and the steepness of the curve is different, but qualitatively it's it's what we [20:28] see everywhere. Um you know maybe maybe someday I'll find a company that's not like that. Um, [20:33] but so far it's it's a common pattern and I actually think it's good news because it it to me [20:39] it says that there's a ton of opportunity there. A lot of frankly very lowhanging fruit that you [20:45] you know it it's it's it's formulaic almost to implement this approach and get the benefits from [20:51] that. So it's it's a it's a common pattern and I think it's going to continue at least for for for [20:57] you know five or more years that we'll keep seeing that pattern. So thank you very much. [21:01] Please uh you know do check out the QR code and we'll send you more information.