Cursor, Moloco Replace SaaS Tools With In-House AI Agents
Ai4
Enterprise software is transitioning from passive tools into autonomous service layers that execute operational workflows directly, enabling internal teams to replace specialized third-party software with custom-built agents.
## Background
Published on YouTube by Ai4 on August 29, 2025, this fireside chat brings together AI leaders Adam Nahari of Pinegrove Venture Partners, George Hou of Cursor, and Tom Rikert of Moloco. According to the video description, the panel explores how top teams leverage AI and automation to elevate productivity, drive meaningful ROI, and navigate the shift toward agentic systems across organizational structures and talent models.
## TL;DR
Leaders from Moloco, Cursor, and Pinegrove share concrete operational playbooks for driving AI adoption across go-to-market, HR, and investment underwriting functions. Moloco outlines how sanctioned model access and grassroots experimentation led to internally built agents—such as a Glean-powered peer review assistant—without purchasing costly point solutions, while Cursor details how hands-on rep shadowing yielded real-time Slack sales concierges and automated follow-up drafts. Pinegrove highlights custom API tooling that digests debt diligence packages into first-draft investment memos. Looking ahead, the panel forecasts that enterprise software will evolve from passive SaaS tools into autonomous service layers, redefining talent around domain taste, data curation, and multidisciplinary creativity.
## Atmosphere & style
The session moves at a brisk, pragmatic pace, driven by informal audience polling and an easy camaraderie between the speakers. The tone deliberately sidesteps speculative AI hype, leaning instead into practical trade-offs, grassroots enablement, and the tactical realities of prompt calibration and IT governance. It feels like an authentic, behind-the-scenes executive debrief where seasoned operators candidly compare notes on what genuinely drives internal enterprise adoption.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=ATisFik8yo0&t=4s) Enterprise AI adoption is largely driven by non-technical teams, with the majority of stakeholders operating in go-to-market, business, and operational functions rather than pure engineering.
- [04:02](https://www.youtube.com/watch?v=ATisFik8yo0&t=242s) Moloco mitigates unmonitored 'shadow AI' by providing enterprise-sanctioned access to top foundation models (ChatGPT, Claude, Gemini) and fostering a bottom-up culture of organic experimentation and prompt sharing.
- [05:50](https://www.youtube.com/watch?v=ATisFik8yo0&t=350s) Gaining user trust for AI automations requires solving mundane, high-frequency daily pain points rather than pursuing complex, underutilized moonshots.
- [05:50](https://www.youtube.com/watch?v=ATisFik8yo0&t=350s) Cursor achieved 100% weekly GTM team adoption by deploying an in-call knowledge agent in Slack and cutting post-call recap drafting time from 15 minutes down to 2 minutes.
- [10:12](https://www.youtube.com/watch?v=ATisFik8yo0&t=612s) Enterprise search and agent builders can eliminate the need for costly point solutions, exemplified by Moloco using Glean to synthesize peer review data across Slack, email, and docs directly into Workday.
- [11:24](https://www.youtube.com/watch?v=ATisFik8yo0&t=684s) Fostering an AI-curious culture requires IT organizations to allow safe, grassroots experimentation first before stepping in to standardize and secure winning use cases.
- [12:32](https://www.youtube.com/watch?v=ATisFik8yo0&t=752s) Effective prompt engineering relies on shadowing domain experts, structuring role-specific operational context, and calibrating voice and tone rather than feeding raw data dumps.
- [14:37](https://www.youtube.com/watch?v=ATisFik8yo0&t=877s) Pinegrove Venture Partners automated venture debt workflows by engineering an internal tool powered by GPT and Claude APIs to parse diligence collateral and draft first-version investment memos.
- [15:53](https://www.youtube.com/watch?v=ATisFik8yo0&t=953s) The agentic AI shift is moving enterprise software from passive SaaS tools to autonomous service layers capable of executing end-to-end operational workflows.
- [18:24](https://www.youtube.com/watch?v=ATisFik8yo0&t=1104s) The agentic era redefines organizational talent around three archetypes: domain experts with nuanced taste, data curators and context architects, and creative out-of-the-box thinkers.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=ATisFik8yo0&t=4s) Setting the Stage and Audience Demographics
Adam Nahari opens the session, framing the discussion around translating AI capabilities from theoretical vision into operational business value. Rather than focusing purely on abstract futures, the conversation centers on the practical realities of how Tom Rikert and George Hou have implemented AI inside their respective organizations to drive tangible innovation and operational efficiency. The agenda also looks ahead at how the maturation of agentic systems will reshape internal organizational structures, talent acquisition, and long-term corporate strategy.
To establish the room's baseline, Adam conducts an informal poll of the audience. Only about 20% of attendees indicate that they write code on a daily basis, with the vast majority representing business, operational, and go-to-market functions. When polling organizational scale, roughly 70% of the audience raises their hands for companies with more than 500 employees, framing the discussion around scaling AI adoption inside mid-to-large enterprise environments.
[01:21](https://www.youtube.com/watch?v=ATisFik8yo0&t=81s) Panel Introductions: Moloco, Cursor, and Pinegrove
Tom Rikert opens the introductions with a self-deprecating quip about the crowded room, joking that attendees are either seeing zero return on investment and looking for answers or experiencing massive success and eager to compare notes. Tom leads product management at Moloco, a decade-old machine learning advertising platform with approximately 800 employees. The company manages over $2 billion in annual advertising spend across mobile advertising, retail media networks on e-commerce platforms, and streaming media channels. While Moloco has operated as an ML-native platform for recommendations and ad-matching for ten years, it is now aggressively incorporating generative and agentic workflows. Tom’s career spans early work as a computer vision engineer, product leadership at Autodesk and Google/YouTube, early-stage venture investing at Andreessen Horowitz, and founding a startup acquired by Moloco.
George Hou introduces his role leading enterprise growth at Cursor, the AI-powered code editor currently utilized by more than half of the Fortune 1000, including engineering teams at Nvidia, Adobe, and Uber. George focuses on scaling enterprise distribution and building internal AI systems to capture inbound demand. His background includes private equity investing at Silver Lake, product leadership at Archive, managing creator Jake Paul's venture capital fund, and studying statistics at Harvard, where he originally met Adam. Adam Nahari introduces his dual mandate at Pinegrove Venture Partners: leveraging AI internally to streamline deal sourcing and investment underwriting, while advising portfolio companies on integrating AI directly into their operational stacks.
- **Moloco Scale**: 800 employees, 10-year track record, powering $2B+ in annual ad spend across retail media, mobile, and streaming.
- **Cursor Penetration**: Adopted across >50% of Fortune 1000 engineering organizations (e.g., Nvidia, Adobe, Uber).
- **Pinegrove Mandate**: Integrating AI into venture underwriting/sourcing and driving portfolio-level AI adoption.
[04:02](https://www.youtube.com/watch?v=ATisFik8yo0&t=242s) Bottom-Up Experimentation and the "Search Engine" Culture at Moloco
When asked how Moloco approaches internal AI adoption, Tom highlights their deliberately bottom-up strategy. Instead of imposing rigid top-down mandates, leadership aimed to create "fertile ground" across all business units—including finance, HR, legal, product, and engineering. Moloco provided enterprise-sanctioned access to major foundation models like ChatGPT, Claude, and Gemini to prevent unmonitored "shadow AI" while giving employees the freedom to solve daily friction points on their own terms.
```mermaid
flowchart TD
accTitle: Moloco Internal AI Experimentation Loop
subgraph Enterprise["Sanctioned Tooling Layer"]
A["Enterprise-Approved Models<br/>(Claude, ChatGPT, Gemini)"]
end
subgraph Community["Internal Sharing & Validation"]
D["Slack Channels &<br/>Informal Meetups"]
E["'Pride of Craftsmanship'<br/>& Peer Model Benchmarking"]
end
A --> B
B --> C
C --> D
D --> E
```
Rather than orchestrating mandatory learning-and-development training, Moloco relies on organic internal advocacy. Employees share prompts, database integrations, and custom automations across dedicated Slack channels and informal meetups. Tom likens the current environment to the early days of the web before Google's dominance, when users enthusiastically compared search results between Infoseek, Lycos, and Ask Jeeves: employees now compare nuances between Gemini, Claude, and ChatGPT across corporate data, cultivating genuine "pride of craftsmanship" and a sense of pioneering their own tools.
[05:50](https://www.youtube.com/watch?v=ATisFik8yo0&t=350s) Cursor's Playbook: Building Trust with Everyday GTM AI Workflows
George outlines Cursor's hybridized approach, pairing grassroots experimentation with an opinionated focus on high-volume go-to-market workflows. He cautions against chasing flashy "moonshot" ideas that end up underutilized. To earn the trust of seasoned professionals with 10 to 15 years of deep domain experience, teams must first solve mundane, highly repetitive daily tasks. Earning trust on foundational workflows creates the credibility required to deploy deeper agentic automations later.
```mermaid
flowchart LR
accTitle: Cursor GTM Real-Time Knowledge Agent
subgraph Sources["Enterprise Knowledge Hubs"]
S1["Salesforce"]
S2["Gong Transcripts"]
S3["Notion & Slack"]
end
subgraph Agent["Cursor AI Agent"]
AG["Real-Time Slack Agent<br/>Context Retrieval & Inference"]
end
subgraph Impact["Live GTM Execution"]
R1["Instant In-Call Answers<br/>(Hours/Days down to seconds)"]
R2["Recap Email Drafting<br/>(15 mins down to 2 mins)"]
end
S1 & S2 & S3 --> AG
AG --> R1
AG --> R2
```
George shares two live operational implementations within Cursor’s sales organization:
1. **Real-Time Knowledge Concierge in Slack**: Account Executives (AEs) onboarding or handling customer calls frequently encounter evolving technical questions. Previously, AEs had to ping product managers or engineers, creating communication bottlenecks that delayed deal cycles by hours or days. Cursor built an internal agent indexing Salesforce, Gong call recordings, Notion documentation, and Slack history. AEs now query the agent via Slack during live customer calls and receive instantaneous answers, eliminating deal friction.
2. **Automated Post-Call Recap Drafting**: Writing personalized follow-ups previously required 15 minutes of manual review per call. With strong inbound volume creating AE bandwidth constraints, Cursor automated the generation of recap drafts matched to each rep's tone and voice. This reduced post-call administrative overhead from 15 minutes down to 2 minutes.
Public Slack channels reveal continuous employee engagement, with queries firing every few minutes and reaching 100% weekly active adoption across the go-to-market organization.
- **AE Knowledge Query Latency**: Reduced from hours/days of cross-functional pinging to real-time Slack responses on live calls.
- **Post-Call Follow-Up Time**: Reduced from 15 minutes per deal to 2 minutes of human review.
- **Adoption Rate**: 100% weekly active usage across the go-to-market team, with the majority querying the system daily.
[09:30](https://www.youtube.com/watch?v=ATisFik8yo0&t=570s) The Build vs. Buy Dilemma and Foundation Model Evolution
Adam transitions the discussion to the strategic trade-off between out-of-the-box foundation models and proprietary internal tool development. He questions how organizations should evaluate whether to build custom software when underlying frontier models may commoditize and resolve those same problems out of the box within a few months. Tom responds by noting that Moloco’s initial expectation was that they would need to purchase an extensive suite of specialized third-party point solutions, but rapid baseline model improvements quickly challenged that assumption.
[10:12](https://www.youtube.com/watch?v=ATisFik8yo0&t=612s) Internal Agents and Grassroots Automation: The Glean Performance Review Case
Organizations often assume that unlocking advanced workflows requires purchasing specialized software for every HR or legal niche. Tom Rikert shares how internal teams at Moloco bypassed this assumption using Glean, their enterprise search provider that includes an agent-building tool. Performance review season typically drains enormous amounts of time as employees comb through past notes, documents, and messages to evaluate peers. In response, Moloco’s HR team built an agent capable of scanning an employee’s Slack conversations, emails, calendar invites, and shared documents to summarize direct peer collaboration.
This agent was integrated directly into their performance review tool connected to Workday. While the agent "didn't write it for you," it aggregated all necessary collaboration data securely and privately inside their existing compliance boundary. Building this capability internally saved significant time and eliminated the need to procure an expensive, standalone HR system.
subgraph AgentLayer["Glean Agent Engine"]
E["Peer Collaboration Summary Agent"]
end
subgraph Workflow["HR Execution Layer"]
F["Performance Workflow Tool"]
G["Workday Integration"]
end
A --> E
B --> E
C --> E
D --> E
E --> F
F --> G
```
[11:24](https://www.youtube.com/watch?v=ATisFik8yo0&t=684s) Fostering an "AI-Curious" Culture and IT Flexibility
Adam Nahari asks how leaders can cultivate a culture where team members feel empowered to build custom agents rather than just consuming foundation models. At Pinegrove Venture Partners, employees use enterprise ChatGPT daily for deal evaluation and investment research, yet custom tool creation had not spontaneously emerged across the team.
Tom explains that fostering this mindset requires encouraging employees to be "AI curious" and giving them room to experiment without immediate bureaucratic constraints. He credits Moloco's IT team for refusing to "squash" early, unorthodox grassroots projects:
- IT allowed employees to experiment with unconventional workflows in the short term.
- Once an experiment proved successful, IT stepped in to standardize the implementation into an official, secure best practice.
- Creating a safe environment for bottom-up exploration surfaced high-ROI internal use cases before establishing top-down corporate standards.
[12:32](https://www.youtube.com/watch?v=ATisFik8yo0&t=752s) Human-in-the-Loop Shadowing and Prompt Engineering at Cursor
When evaluating when to rely on out-of-the-box LLMs versus purpose-built agentic workflows, George Hou outlines Cursor's hands-on discovery process. Rather than building in a vacuum, team members shadow account executives and sales leaders to map repetitive, predictable tasks step by step.
For predictable workflows, foundation models excel when given structured prompts, explicit instructions, and high-quality reference examples "as if you were teaching a kid." George emphasizes that effective AI adoption hinges on prompt design that provides deep operational context and role-specific interpretation rather than raw data dumps:
> "It's not just giving it the data that it needs, but telling it how to interpret it as if they were, for example, a sales leader or an account executive, and really helping them think through and remove as much cognitive load as possible."
In high-touch tasks such as drafting post-meeting follow-up emails, style and craft are paramount. George sat directly next to account executives to calibrate the LLM's voice and tone until the generated drafts sounded indistinguishable from the individual reps.
```mermaid
flowchart LR
accTitle: Workflow Mapping and Prompt Calibration
subgraph Discovery["Workflow Discovery"]
A["Shadow AE / Sales Leader"] --> B["Map Monotonous Steps"]
end
subgraph Calibration["Prompting & Voice Design"]
B --> C["Structure Explicit Rules & Examples"]
C --> D["Pair with Rep to Tune Voice & Craft"]
end
subgraph Deployment["Human-in-the-Loop Execution"]
D --> E["Automated Draft Generation"]
E --> F["Human Review & Send"]
end
```
[14:37](https://www.youtube.com/watch?v=ATisFik8yo0&t=877s) Automating Venture Debt Investment Memos at Pinegrove
Adam Nahari shares a parallel initiative at Pinegrove Venture Partners. Beyond company-wide enterprise ChatGPT access, Pinegrove engineered a proprietary internal application utilizing GPT and Claude APIs to streamline their venture debt strategy:
- Incoming deal packages contain extensive, disparate collateral and diligence materials.
- The custom tool parses the raw deal documents, extracts relevant operational and financial metrics, and drafts a comprehensive first version of the investment memo.
- Building the tool required deep collaboration with the investment team to define exact structural requirements, ensuring the AI output directly matched the team's underwriting workflow.
[15:53](https://www.youtube.com/watch?v=ATisFik8yo0&t=953s) The Agentic Shift: From Software Tools to Autonomous Services
Looking forward at the broader trajectory of AI, Tom Rikert argues that the current agentic movement is unprecedented across previous technology cycles. In enterprise software, companies traditionally deliver tools to customers, leaving the difficult change management and operational hiring to the customer.
Using Moloco’s retail media enterprise platform as an example, retailers must transition into functioning media companies by building advertising operations and sales teams from scratch:
- Software alone cannot execute change management or build operational sales muscle.
- Agentic systems can evolve beyond pure software tools to act as autonomous ad sales and ad operations teams.
- Enterprise software models are shifting from passive SaaS interfaces toward high-touch, full-solution service layers where agents execute the underlying business operations.
When asked by Adam how this impacts monetization, Tom notes that software vendors are actively experimenting with evolving their business models to capture the higher value delivered by end-to-end autonomous business execution.
[18:24](https://www.youtube.com/watch?v=ATisFik8yo0&t=1104s) Redefining Talent: Three Essential Archetypes for the Agentic Era
George Hou explains that while AI adoption will not replace the necessity of hiring top-tier talent, it completely redefines what "great" looks like in an organization. He outlines three foundational talent archetypes required to scale organizations in the AI era:
1. **Domain Experts with "Taste" and Deep Intuition**: Foundation models generate plausible outputs instantly, but evaluating good versus exceptional work requires years of domain reps and seasoned judgment.
2. **Data Curators and Context Architects**: Because model performance strictly obeys the principle of "good data in, good response out," teams need rigorous thinkers who understand what metrics matter, how to measure them, and how to structure context so the AI can digest it cleanly.
3. **Out-of-the-Box Creative Connectors**: Highly creative individuals who understand domains outside standard LLM training data can bridge unexpected concepts (such as pop culture and multidisciplinary analogies) that an AI would not surface organically.
Adam closes the session by thanking Tom and George, remarking on how quickly the 20-minute fireside chat concluded.
## Notable quotes
> [04:02](https://www.youtube.com/watch?v=ATisFik8yo0&t=242s) fertile ground — Tom Rikert
> [04:02](https://www.youtube.com/watch?v=ATisFik8yo0&t=242s) pride of craftsmanship — Tom Rikert
> [05:50](https://www.youtube.com/watch?v=ATisFik8yo0&t=350s) To earn the trust of seasoned professionals with 10 to 15 years of deep domain experience, teams must first solve mundane, highly repetitive daily tasks. — George Hou
> [10:12](https://www.youtube.com/watch?v=ATisFik8yo0&t=612s) didn't write it for you — Tom Rikert
> [11:24](https://www.youtube.com/watch?v=ATisFik8yo0&t=684s) AI curious — Tom Rikert
> [12:32](https://www.youtube.com/watch?v=ATisFik8yo0&t=752s) as if you were teaching a kid — George Hou
> [12:32](https://www.youtube.com/watch?v=ATisFik8yo0&t=752s) It's not just giving it the data that it needs, but telling it how to interpret it as if they were, for example, a sales leader or an account executive, and really helping them think through and remove as much cognitive load as possible. — George Hou
> [15:53](https://www.youtube.com/watch?v=ATisFik8yo0&t=953s) Software alone cannot execute change management or build operational sales muscle. — Tom Rikert
> [18:24](https://www.youtube.com/watch?v=ATisFik8yo0&t=1104s) good data in, good response out — George Hou
## Video description
Adam Nahari, Partner, Chief Data and AI Officer, Pinegrove Venture Partners
George Hou, Growth Lead, Cursor
Tom Rikert, Senior Director, Product Management, Moloco
Abstract
Fireside chat featuring AI leaders from Cursor, Moloco, and Pinegrove Venture Partners. The discussion will explore how top teams leverage AI and automation to elevate team productivity and drive meaningful ROI. Whether you're leading a sales team, managing engineering projects, or scaling your operations, learn how AI is revolutionizing core business functions and serving as a force multiplier. The conversation will also look ahead to the agentic AI future, examining how AI adoption will influence organizational structure, talent models, user trust, and the evolving definition of product excellence.
Deploying generative and agentic AI inside the enterprise delivers the highest return when organizations focus on high-frequency, mundane workflows rather than theoretical moonshots, according to leaders from Moloco, Cursor, and Pinegrove Venture Partners.
Bottom-Up Experimentation and Permissive IT (4:02)
Rather than relying on top-down directives or rigid learning programs, Moloco provided sanctioned enterprise access to foundational models such as ChatGPT, Claude, and Gemini to prevent shadow IT while encouraging organic problem-solving. Employees across HR, legal, finance, and engineering independently test workflows and share prompts across Slack channels. Tom Rikert noted that the environment mirrors the early web search era, with employees benchmarking different models against proprietary corporate data.
Targeting Repetitive Tasks to Earn Trust (5:50)
Cursor takes a targeted approach by addressing high-volume, mundane tasks to win over skeptical domain experts before attempting ambitious automations. George Hou detailed how Cursor mapped sales workflows by shadowing account executives and calibrating prompts to match individual rep styles. This hands-on process produced an internal Slack-based agent that indexes Salesforce, Gong, Notion, and Slack to answer technical questions in real time during live prospect calls.
Internal Agents Over Third-Party Point Solutions (10:12)
Rapidly improving baseline foundation models have challenged the initial enterprise assumption that companies must procure specialized third-party point solutions. At Moloco, the HR team avoided purchasing dedicated review software by using their enterprise search provider, Glean, to build an agent that summarizes peer collaboration across Slack, emails, calendar invites, and documents, feeding context into Workday within existing compliance boundaries. Similarly, Adam Nahari described how Pinegrove Venture Partners built a proprietary internal tool using GPT and Claude APIs to parse diligence collateral and generate first drafts of venture debt investment memos.
Autonomous Services and Evolving Talent Profiles (15:53)
Rikert argued that agentic systems will transform software from passive SaaS tools into autonomous service layers that run complex business operations directly, such as staffing advertising sales and operations for retail media platforms. This shift is also redefining enterprise hiring. Hou identified three core talent profiles necessary for operating alongside AI: domain experts with refined taste to evaluate model outputs, context architects skilled at structuring clean data pipelines, and creative thinkers who bridge non-standard multidisciplinary concepts.
While enterprises are already seeing administrative hours cut to minutes through custom agents, software vendors and enterprise buyers are still actively testing new business and monetization models to capture the value of end-to-end autonomous business execution.
精炼阅读摘要
Adam Nahari、Tom Rikert 与 George Hou 认为，企业实现从人工智能技术愿景向实际业务价值的转化，关键在于摒弃激进的宏大构想，转而通过合规的自下而上实验解决日常高频痛点，并依托深度上下文工程推动软件由被动工具逐步演进为自主化服务。
基层全员实验与安全合规演进 (4:02)
针对企业内部采纳人工智能的路径，Moloco 产品副总裁 Tom Rikert 强调了一种自下而上的推进模式。为避免无序的“影子 AI”，管理层向财务、HR、法务及工程等全功能部门提供经过企业安全核准的 ChatGPT、Claude 与 Gemini 访问权限，支持员工自主探索日常工作中的摩擦点。Moloco 员工自发在 Slack 与非正式聚会中交流提示词技巧和模型表现，形成了类似早期互联网搜索引擎百家争鸣的比对文化，从而培育了员工自主构建工具的认同感。
Tom Rikert 指出，IT 部门在其中扮演了关键的“包容与赋能”角色，对早期的基层探索未加盲目压制，而是在验证可行性后再介入推行标准化治理。例如，Moloco 人力资源团队利用企业搜索服务商 Glean 内置的代理构建能力，搭建出可抓取 Slack 对话、邮件及文档记录的协作总结代理，并无缝接入 Workday 绩效考核流程。该内部方案不仅大幅节约了员工整理考核材料的时间，也在合规边界内避免了向外采购昂贵专用人事系统的开销。
聚焦高频流程与上下文精准校准 (5:50)
Cursor 企业增长负责人 George Hou 指出，在推行技术落地时，团队应警惕使用率低下的宏大设想，优先攻克枯燥、高频的基础任务，以此在资深业务人员中建立信任。Cursor 团队采用跟岗模式深入销售一线，拆解重复性环节，并通过如同“教授儿童”般包含详尽规则与高质量范例的提示词设计引导模型。George Hou 强调，提示词工程的关键不仅在于提供数据，更在于向模型注入特定业务角色的解读逻辑，例如与销售代表并排校准语气，使自动生成的跟进邮件在风格上与人工草拟毫无二致。
这一策略带来了显著的效率突破。Cursor 内部搭建了对接 Salesforce、Gong 通话录音与 Notion 文档的实时问答代理，使销售代表在客户通话期间的知识检索延迟从数小时或数天缩减至实时回复；会后纪要草拟时间由每单 15 分钟降至 2 分钟的人工核对，全业务团队周活跃度达到 100%。Pinegrove 同样践行了这一逻辑，通过自研结合 GPT 与 Claude API 的专用工具，自动化解析分散的底层材料并生成风险债权备忘录初稿，有效消除了投资团队的认知负荷。
从被动软件工具迈向自主化运营 (15:53)
在展望技术演进时，Tom Rikert 提出，代理系统的成熟正对传统软件模式形成根本性颠覆。传统 SaaS 软件仅向客户交付功能工具，而将繁琐的组织变革管理与岗位招聘甩给客户自行承担。以 Moloco 赋能的零售媒体网络为例，零售商在搭建广告业务时，往往面临缺乏专业广告销售与运营体系的结构性挑战，单纯的软件交付无法解决组织维度的能力缺失。
Tom Rikert 预测，未来的代理系统将超越被动辅助的工具界面，直接作为端到端的自主化业务运营实体运作。这种转变推动企业软件由单一功能层演化为高附加值的全托管服务层，促使软件供应商积极探索新型商业化与计费模式，以便直接捕获自动化业务执行所带来的更高溢价。
代理系统驱动的组织核心人才重塑 (18:24)
George Hou 认为，人工智能的普及并不会减少组织对优秀人才的依赖，但会彻底改变组织对“卓越”的评判维度。他提出了适应代理时代的三类核心人才画像：一是具备深厚业务直觉与鉴赏“品味”的领域专家，负责对基础模型瞬间生成的貌似合理的输出进行优劣把控；二是数据策划师与上下文架构师，能够严谨定义关键业务指标，并将现实问题精准结构化为模型可消化的上下文输入。
第三类则是具备跳出固有框架能力的跨界创意连接者。George Hou 阐述道，这类人才拥有大语言模型标准训练集之外的跨学科视野，能够将看似无关的多领域概念与业务逻辑相串联，从而提供人工智能依靠既有数据无法自发构想的创新方案。
针对企业在底层基础模型不断升级背景下是否应自研工具的权衡，对话参与者指出了通用模型迭代带来的商品化挑战。尽管各方在基层实验与流程自动化上取得了切实收益，但面对外置模型能力的快速演进，企业如何在维持自建代理技术壁垒的同时确立自主服务的定价标准，仍是留待后续实践探索的核心命题。
Detailed notes — zh
## Background
本视频由行业峰会频道 Ai4 于 2025 年 8 月 29 日发布在 YouTube 平台。内容为 Pinegrove Venture Partners 首席数据与 AI 官 Adam Nahari 主持的炉边对话，邀请了 Moloco 高级产品总监 Tom Rikert 与 Cursor 增长负责人 George Hou，共同探讨顶尖团队如何利用 AI 与自动化提升核心业务生产力、获取实质性 ROI，并展望了智能体时代对企业组织架构、人才模型与产品交付形态的深远重构。
## TL;DR
对话深入分享了 Moloco 与 Cursor 在企业内部规模化落地 AI 的核心路径与实战方法论。Moloco 采取自下而上的创新策略，由 IT 部门统一授权主流模型以杜绝“影子 AI”，并利用 Glean 等工具低成本打通跨部门绩效评估等日常流程；Cursor 则主张从销售跟进邮件、会议实时答疑等极度枯燥且重复的高频业务切入，通过“影子跟岗”与深度上下文注入快速建立用户信任。此外，Pinegrove 结合自身投资备忘录生成的自动化实践，共同探讨了开箱即用模型与定制构建的权衡。展望未来，嘉宾们指出软件行业正从“售卖功能工具”转向“交付由 Agent 构成的端到端业务服务”，组织未来将更加依仗具备业务品味、严谨数据度量意识以及跨界创意思维的三类卓越人才。
## Atmosphere & style
整场炉边对谈节奏紧凑、务实且充满极客洞见，开场针对现场大企业管理层比例的摸底迅速确立了接地气的交流基调。Tom Rikert 以幽默生动的比喻阐述了内部创新的工匠自豪感，George Hou 则以一线战术拆解展现了极度严谨的业务颗粒度，主持人穿针引线的提问让两家领先企业的碰撞极具张力。现场既毫无空洞的概念浮夸，又在探讨 Agent 终局交付形态与人才变革时展现出宏大的前瞻视野与深刻思辨。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=ATisFik8yo0&t=4s) 大型企业内部推动 AI 落地需面向非技术人员：现场摸底显示仅约 20% 为硬核技术人员，70% 来自 500 人以上大型组织，AI 赋能商业与管理职能是规模化落地的核心。
- [04:02](https://www.youtube.com/watch?v=ATisFik8yo0&t=242s) Moloco 采取自下而上的探索模式，由 IT 部门统一批准并官方背书主流大模型以杜绝影子 AI，通过内部 Slack 频道与非正式 Meetup 激发员工的工匠自豪感与自驱创新。
- [05:50](https://www.youtube.com/watch?v=ATisFik8yo0&t=350s) AI 落地应避免脱离实际的“登月计划”，Cursor 强调必须从极度重复枯燥的日常工作流切入，逐步建立资深业务人员对 AI 工具的信任。
- [05:50](https://www.youtube.com/watch?v=ATisFik8yo0&t=350s) Cursor 打通 Salesforce、Gong、Notion 等企业数据源构建 GTM 销售答疑智能体，将原本耗时数小时至数天的跨部门求助转变为开会通话中的秒级即时答复。
- [05:50](https://www.youtube.com/watch?v=ATisFik8yo0&t=350s) 通过定制邮件智能体自动模仿销售个人语气生成跟进邮件，Cursor 将单次跟进用时从 15 分钟降至 2 分钟以内，实现 GTM 团队 100% 的周渗透率。
- [10:12](https://www.youtube.com/watch?v=ATisFik8yo0&t=612s) 企业无需盲目采购专用 SaaS，利用 Glean 等已有搜索平台的 Agent 功能即可自动检索整合 Slack、邮件与日程记录，安全高效地生成跨部门绩效回顾报告。
- [10:12](https://www.youtube.com/watch?v=ATisFik8yo0&t=612s) 包容实验的 IT 文化是 AI 创新的基石，IT 团队对员工自发的探索尝试采取“先让子弹飞一会儿”的态度，待摸索出高价值实践后再进行官方规范与推广。
- [12:32](https://www.youtube.com/watch?v=ATisFik8yo0&t=752s) Cursor 采用“影子跟岗”深入一线拆解业务细节，结合深度的 Prompt 上下文注入与“人类在回路”持续校准，确保模型输出精准契合专业调性。
- [12:32](https://www.youtube.com/watch?v=ATisFik8yo0&t=752s) Pinegrove 基于大模型 API 定制自动化投资备忘录工具，自动解析项目包并生成初版投资备忘录，显著降低分析师初始信息整理的认知负荷。
- [15:53](https://www.youtube.com/watch?v=ATisFik8yo0&t=953s) Agent 时代的软件交付范式正从“售卖功能与工具”向“直接代运营业务结果”转变，软件厂商未来有望直接向客户交付由 AI 驱动的端到端虚拟业务团队。
- [15:53](https://www.youtube.com/watch?v=ATisFik8yo0&t=953s) AI 时代对顶尖人才的定义发生重塑，组织将格外倚重三类核心人才：具有行业品味与鉴别力的业务专家、把关输入质量的数据架构师，以及能打破常规的跨界创意连接者。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=ATisFik8yo0&t=4s) 开场与现场互动调查
在讲座正式开始前，主持对话的 Adam Nahari 首先点明了这场对谈的主旨——虽然议题标题定为“AI 的落地实践：从愿景到价值”（AI in Action: From Vision to Value），但今天的核心将聚焦于更具实操意义的内幕分享。他将与来自 Moloco 的 Tom Rikert 和来自 Cursor 的 George Hou 一同探讨，这两家处于 AI 前沿的公司究竟如何在内部全面落地 AI 以驱动创新和效率提升。此外，对话还将具有前瞻性地展望 AI 能力的快速演化将如何倒逼企业的人才组织架构与底层战略发生深刻变革。
为了精准掌握台下听众的背景构成，Adam 在开讲前进行了一轮即兴的现场举手摸底。调查结果显示，每天实际编写代码的硬核技术人员仅占现场听众的 20% 左右，绝大多数参会者来自商务、产品及管理等商业职能岗位。随后关于企业规模的摸底则进一步揭示，现场约有 70% 的听众就职于员工人数超过 500 人的大型组织，仅有 30% 左右来自 500 人以下的中小团队。这一受众结构直接为后续探讨“大型企业内部如何推动 AI 规模化落地”奠定了现实基调。
### [01:28](https://www.youtube.com/watch?v=ATisFik8yo0&t=88s) 嘉宾背景：广告平台、代码新贵与风险投资
轮到嘉宾自我介绍环节时，Tom Rikert 以一句幽默的调侃开场，引得全场会心一笑：“很高兴看到现场座无虚席。我想原因大概只有两个：要么是大家在 AI 投资上还没看到实际的 ROI，急着来这里取经；要么就是你们已经拿到了漂亮的 ROI，迫不及待想来互相印证。”随后他介绍了自身所在的 Moloco。这家拥有约 800 名员工、成立逾 10 年的企业是一家典型的“AI 原生”广告技术平台，每年驱动超过 20 亿美元的广告交易流水，业务横跨移动广告、零售媒体（Retail Media）以及流媒体网络。Tom 表示，Moloco 虽长期依赖机器学习算法完成广告匹配与精准推荐，但近期正在极度激进地拥抱生成式 AI 与智能体（Agentic AI）。Tom 本人最早是计算机视觉与机器学习领域的工程师，后来在 Autodesk、Google 和 YouTube 担任产品管理职位，之后曾加入顶尖风投机构 Andreessen Horowitz（a16z）投资早期 AI 项目，再后来开启创业之路，并在公司被 Moloco 收购后加入其中。
接棒介绍的 George Hou 是新晋 AI 代码编辑器独角兽 Cursor 的增长负责人（Growth Lead）。Cursor 发展极为迅猛，目前客户已涵盖半数以上的《财富》1000 强企业，包括 Nvidia、Adobe 和 Uber 等科技巨头。George 在 Cursor 的核心职责是推动企业级客户的深度采纳，并通过自研 AI 系统在规模化层面精准捕获企业端需求。George 的职业履历同样横跨投资与产品：他最早在私募股权投资巨头银湖资本（Silver Lake）从事 PE 投资，随后在 Archive 负责产品，期间还曾协助运营知名网红 Jake Paul 的风投基金；在学术背景上，他毕业于哈佛大学统计学专业，并在求学期间结识了主持人 Adam。
主持人 Adam Nahari 则介绍了自己的背景。他目前担任 Pinegrove Venture Partners 的合伙人兼首席数据与 AI 官（Chief Data and AI Officer）。Pinegrove 是一家总部位于旧金山的风险与成长期投资平台。Adam 的工作分为双轨驱动：一方面深入思考如何借助前沿 AI 技术更高效地发掘与评估投资标的，另一方面则紧密协助旗下的被投企业将 AI 深度嵌入其日常商业运营之中。
- **Tom Rikert**：Moloco 高级产品总监；公司年支撑广告支出逾 20 亿美元，团队规模约 800 人。
- **George Hou**：Cursor 增长负责人；产品服务超过 50% 的财富 1000 强企业（如英伟达、Adobe、Uber）。
- **Adam Nahari**：Pinegrove Venture Partners 首席数据与 AI 官；负责投研 AI 化及被投企业的赋能集成。
### [04:02](https://www.youtube.com/watch?v=ATisFik8yo0&t=242s) Moloco 的自下而上探索与工匠自豪感
在被问及 Moloco 内部落地 AI 的经验、踩坑与核心认知时，Tom Rikert 直言公司的策略是高度“自下而上”（bottoms-up）的。管理层的第一要务并非自上而下指手画脚，而是为广泛的自主探索提供“肥沃的土壤”。在具体落地层面，Moloco 为全员开通并统一了企业官方批准的顶级主流模型通道，包括 ChatGPT、Claude 和 Gemini。Tom 强调，公司必须通过 IT 部门给予正式背书，坚决避免各团队在私底下使用未经审核的“影子 AI”（Shadow AI）。企业要让员工确信，自己在日常工作中所做的任何探索都拥有管理层 100% 的合法授权与安全保障。
在这种宽松而受保护的环境中，最生动、最接地气的创新往往直接迸发自财务、人力资源、法务、产品和工程等一线人员的具体业务需求中。为了沉淀并放大这些零星的火花，Moloco 搭建了专门的 Slack 讨论频道，并不定期举办非正式的内部 Meetup。Tom 观察到，这种分享机制极大地激发了员工内心的“工匠自豪感”（pride of craftsmanship）——大家热衷于向同事炫耀自己利用 AI 搭建的新流程或小工具，人人都能从中体验到充当业务“拓荒先锋”（pioneers）的成就感。
Tom 还形象地将当下的内部生态比作早期搜索引擎大战的时代：“如果大家还记得 Google 诞生之前的旧时光，那时有各式各样的搜索引擎，比如 Infoseek、Lycos、Ask Jeeves。当年大家最喜欢的乐趣就是互相比较用这个搜出的结果和那个有什么不同。”如今 Moloco 内部也在上演完全相同的一幕：员工们兴致盎然地对比同一个任务在 ChatGPT 和 Gemini 上的输出优劣，或者比拼谁能更好地将大模型接入内部企业数据库。值得注意的是，Moloco 从未组织过任何强制性的培训与技能发展课程（L&D），全体员工完全凭借自驱力自发掌握了最新工具的深度应用。
### [05:50](https://www.youtube.com/watch?v=ATisFik8yo0&t=350s) Cursor 的落地哲学：从乏味日常中建立用户信任
与 Moloco 纯粹的自下而上相比，George Hou 透露 Cursor 采取的是一种“混合模式”：在保持基层自由探索的同时，公司在特定业务领域——特别是客户拓展与市场开拓（Go-to-market, GTM）端——带有非常强烈的“主观意图”（opinionated），会主动划定最具业务冲击力的高潜力场景。George 指出，他们在落地过程中总结出的最核心教训是：必须从“极度重复的枯燥日常工作流”切入。行业里的大多数人往往沉迷于宏大性感的“登月计划”（moonshot ideas），言必称某些天马行空的颠覆性概念，但 Cursor 深刻认识到，AI 落地最艰难也最关键的一环在于赢得终端用户的“信任”。特别是在许多专业岗位上，老员工往往拥有 10 到 15 年打磨特定技艺的资深经验，要打破他们的心理防御、建立对其工作习惯的信任，唯有从最基础、最底层的琐碎事务做起。
George 现场分享了 Cursor 内部两个极具代表性的高 ROI 实战案例：
第一个案例是销售客户经理（Account Executive, AE）的入职培训与日常业务答疑。过去，新 AE 或销售在与客户交涉时，遇到复杂的产品功能或棘手的问题，往往只能在 Slack 上向资深 AE、销售主管甚至核心研发工程师求助。这种传统沟通方式往往会导致长达数小时乃至一整天的信息等待，不仅严重拖慢了交易推进周期（deal cycles），甚至可能让即将敲定的单子陷入停滞。为此，Cursor 构建了一个专有的 AI 智能体，直接挂载于企业核心知识库与底层数据源之上，全面打通 Salesforce、Gong、Notion 以及 Slack 历史记录。如今，销售人员在与客户开会的通话过程中，只要实时向 Slack 中的智能体发送指令，便能瞬间获取精准答案。这一改动直接将原本容易拖延数天的谈判停滞，化解为通话现场的“当场关闭问题”。
第二个案例则是通话后的纪要邮件（Recap Email）。按照标准销售规范，AE 在结束每场客户会议后，都需要复盘会议记录、整理细节并撰写跟进邮件，单次通常耗时 15 分钟左右。由于 Cursor 面临着极其庞大的 inbound 销售线索涌入，销售人员被大量存量商机事务缠身，无法迅速切入下一个潜在客户，跟进环节直接沦为业务瓶颈。Cursor 顺势推出了定制的邮件智能体，能够根据通话内容自动模仿该销售本人的专属语气和说话风格（tone and voice），一键生成完整的邮件草稿。这一自动化将销售撰写邮件的用时从 15 分钟骤降至 2 分钟以内，彻底释放了 GTM 团队的产能瓶颈。目前，Cursor 内部这套基于 Slack 的智能体达到了每周 100% 的团队渗透率，绝大多数员工每日重度依赖；在公开的 Slack 频道中，几乎每隔数分钟就会跳出一条员工的提问与调用请求。
```mermaid
flowchart TD
accTitle: Cursor GTM 智能体处理日常业务流程
subgraph DataSources["企业数据源与核心知识库"]
A["Salesforce 销售数据"]
B["Gong 通话录音与纪要"]
C["Notion 文档中心"]
D["Slack 历史沟通记录"]
end
subgraph AIAgent["内部 GTM AI 智能体"]
E["中央检索与语义理解引擎"]
end
subgraph Workflows["高频日常工作流赋能"]
F["会议中即时答疑<br/>(秒级反馈，杜绝交易推迟)"]
G["个性化跟进邮件生成<br/>(耗时从 15 分钟缩减至 2 分钟)"]
end
A --> E
B --> E
C --> E
D --> E
E --> F
E --> G
```
### [09:30](https://www.youtube.com/watch?v=ATisFik8yo0&t=570s) 开箱即用 vs 定制构建的抉择悬念
基于两家公司的实操演进，Adam Nahari 抛出了一个深具战略分量且困扰着众多企业决策者的核心命题：企业究竟该如何划分“开箱即用”（out-of-the-box）的大模型使用边界，与需要企业投入真金白银“定制构建”（bespoke custom tools）的自研工具边界？
Adam 进一步细化了这一纠结的心态：一方面，现有的现成模型（如商用版 ChatGPT）在特定场景下或许还不完全成熟，但可能在一年之内就会迎来质的飞跃；另一方面，企业如果投入重兵自行研发定制系统，往往面临巨大的战略风险——很可能辛辛苦苦开发数月之后，底座基础大模型（Foundation Models）发布一次版本迭代，就直接让几个月前耗费心血自研的定制工具变得毫无必要甚至沦为废品。在面对飞速演进的底层能力时，企业该如何理性评估投资的时机与深度？
面对这个直击行业痛点的问题，Tom Rikert 接过话筒，透露了 Moloco 在这一探索过程中遇到的巨大意外：“在这一路走来中，最让我们感到惊讶的发现之一就是，我们原本以为必须去市面上大规模采购很多……”正当 Tom 准备揭晓 Moloco 在采购标准 SaaS 工具与自研定制工具之间的抉择反思时，对话在此留下了关键悬念。
### [10:12](https://www.youtube.com/watch?v=ATisFik8yo0&t=612s) 企业级内部定制 Agent：从 Glean 自动绩效总结到鼓励创新的 IT 文化
Moloco 的产品管理高级总监 Tom Rikert 分享了团队在日常工作中挖掘 AI 潜力的真实案例。很多团队最初试图寻找专门针对 HR 或法务场景的现成 SaaS 工具，但随后发现利用手头的工具就能构建出强大的内部 Agent。例如，Moloco 使用企业搜索平台 Glean，而 Glean 本身自带构建智能体（Agent）的功能。撰写员工年度/季度绩效考评（Performance Review）通常是一件极耗时间的苦差事，员工需要翻遍以往的笔记、会议纪要来拼凑工作细节。Moloco 的 HR 人员便利用 Glean 构建了一个专属 Agent，它能够自动抓取并整合员工在 Slack 聊天、工作邮件、日历邀请和文档中的协作记录，生成一份详尽的跨部门合作总结。该功能直接内嵌到了与 Workday 对接的绩效考评工作流中。它并没有直接替人代笔写评价，而是完成了所有繁琐的信息检索与整合，并且整个过程严格保证了企业内部数据的安全与隐私，从而免去了企业额外采购一套全新系统的开销。
针对 Pinegrove Venture Partners 首席数据与 AI 官 Adam Nahari 关于“如何营造鼓励员工自行创建定制 GPT 文化”的提问，Tom 表示关键在于培养“对 AI 保持好奇（AI-curious）”的企业氛围。他特别赞赏了 Moloco 的 IT 团队：面对员工一些略显非主流或实验性的 AI 尝试，IT 人员并不会立刻叫停或压制，而是“先让子弹飞一会儿（let it ride for a little bit）”，给予员工足够的自由度去探索。等摸索出真正有价值的最佳实践后，再由官方介入将其规范化并推广到全公司。
```mermaid
flowchart LR
accTitle: HR绩效回顾智能体工作流
subgraph DataSources["多源企业数据"]
A["Slack 对话"]
B["工作邮件"]
C["日历日程"]
D["共享协作文档"]
end
subgraph AgentEngine["Glean 企业 Agent 引擎"]
E["检索跨部门协作痕迹"]
F["提炼整合 & 生成工作摘要"]
end
subgraph Output["集成与输出"]
G["绩效评估工作流系统"]
H["Workday 核心人力平台"]
end
DataSources --> E
E --> F
F --> G
G --> H
```
---
### [12:32](https://www.youtube.com/watch?v=ATisFik8yo0&t=752s) 定制工具 vs 开箱即用模型：影子跟岗、上下文注入与投资备忘录生成
当谈到何时使用开箱即用的基础模型（如 ChatGPT、Claude），何时需要自建定制工具时，Cursor 的增长负责人 George Hou 介绍了他们独特的“影子跟岗（Shadowing）”方法论。George 或团队成员会亲自坐在业务人员（如客户主管 AE 或销售主管）身边，全程观察并完整绘制出对方的日常工作流。对于那些重复、单调且高度可预测的任务，直接调用现成的 LLM 即可高效解决；但成功的关键在于极其详尽的 Prompt 工程与上下文注入——如同手把手教小孩子一样，不仅要喂给模型数据，更要指导它像资深销售主管一样去理解和解读这些数据。George 强调，这绝非“设好就撒手不管（set it and forget it）”的自动化，而是必须有“人类在回路（Human in the loop）”的紧密协作。例如在为销售生成通话跟进邮件时，由于风格和调性（Tone and voice）直接决定了成败，团队会反复校准，确保生成的内容不仅事实准确，而且读起来就像该业务员本人的亲笔表达。
Adam 对此深表赞同，并分享了 Pinegrove 的自身实践。除了全员日常使用 ChatGPT Enterprise 外，他们利用 GPT 和 Claude API 专门为旗下的风险债权（Venture Debt）投资团队定制了一套工具。每当有新项目进来时，该工具能自动解析堆积如山的项目资料，精准提取核心数据并自动生成第一版投资备忘录（Investment Memo）草稿，从而大幅缩减分析师的初始整理时间。这一工具之所以能顺利落地，正是因为技术团队与投资业务团队紧密配合，吃透了备忘录中每一个关键模块的构建逻辑。
- **工作流拆解与跟岗**：针对销售等强风格依赖的场景，技术人员通过实地结对（Pairing）与影子跟岗（Shadowing），精准捕捉业务风格与工艺细节（Craft）。
- **降低认知负荷**：Prompt 不仅提供业务数据，还需赋予 AI 专业角色视角与解读框架，以最大程度减少员工的认知负荷。
- **投资备忘录自动化**：Pinegrove 基于 GPT/Claude API 解析项目包，自动化提炼并输出初版投资备忘录，实现了特定垂直业务流程的高效重塑。
---
### [15:53](https://www.youtube.com/watch?v=ATisFik8yo0&t=953s) Agent 时代的组织重构：从交付软件到交付“端到端服务”
展望 Agent 技术的未来演进，Tom 认为当前的 Agentic 浪潮是前所未有的技术周期。作为面向零售商的企业级软件提供商，Moloco 正在帮助传统大型零售商转型兼具媒体属性的广告平台。在以往的传统企业软件模式下，即便向客户交付了软件，客户仍需艰难地推进组织变革（Change Management），自己培养组建广告销售与广告运营团队。而在 Agent 时代，软件提供商不仅能交付工具，更有可能直接向客户交付一套由“AI Agent 构成的广告销售团队与运营团队”，实现端到端解决客户业务问题的全流程高阶服务（High-touch service）。虽然针对这种全新交付形态的商业模式与收费机制仍在探索中，但软件行业从“售卖功能”向“直接代运营业务结果”的转变已成必然趋势。
在团队人才结构方面，George 指出 AI 并不会降低企业招募顶尖人才的标准，但彻底重塑了对“卓越人才（Great talent）”的定义。未来组织中将格外依赖三类核心人才画像：
| 人才画像分类 | 核心特质与能力要求 | 核心价值体现 |
|:--- |:--- |:--- |
| **业务深耕与品味专家** | 对底层业务逻辑有深刻洞察，具备极高的行业“品味（Taste）”与实战经验积累（Reps）。 | AI 能够快速生成内容，但只有经验丰富的专家才能精准鉴别什么是“合格（Good）”，什么是真正的“卓越（Great）”。 |
| **数据架构与度量把关人** | 具备严谨的数据思维，深谙“高质量数据输入才能带来高质量输出（Good data in, good response out）”。 | 准确识别并量化业务核心指标，确保将最有效、结构最合理的数据以最易消化的方式输入给 AI。 |
| **跨界创意连接者** | 具备极强的创意思维与广博的领域知识，善于将流行文化等多元概念跨界融合。 | 能够打破常规、连接 LLM 训练数据中未曾直接建立联系的新颖概念，提供 AI 无法自行推演的创新解法。 |
## Notable quotes
> [01:28](https://www.youtube.com/watch?v=ATisFik8yo0&t=88s) 很高兴看到现场座无虚席。我想原因大概只有两个：要么是大家在 AI 投资上还没看到实际的 ROI，急着来这里取经；要么就是你们已经拿到了漂亮的 ROI，迫不及待想来互相印证。 — Tom Rikert
> [04:02](https://www.youtube.com/watch?v=ATisFik8yo0&t=242s) 如果大家还记得 Google 诞生之前的旧时光，那时有各式各样的搜索引擎，比如 Infoseek、Lycos、Ask Jeeves。当年大家最喜欢的乐趣就是互相比较用这个搜出的结果和那个有什么不同。 — Tom Rikert
> [09:30](https://www.youtube.com/watch?v=ATisFik8yo0&t=570s) 在这一路走来中，最让我们感到惊讶的发现之一就是，我们原本以为必须去市面上大规模采购很多…… — Tom Rikert
> [10:12](https://www.youtube.com/watch?v=ATisFik8yo0&t=612s) 先让子弹飞一会儿（let it ride for a little bit） — Tom Rikert
> [12:32](https://www.youtube.com/watch?v=ATisFik8yo0&t=752s) 设好就撒手不管（set it and forget it） — George Hou
> [15:53](https://www.youtube.com/watch?v=ATisFik8yo0&t=953s) 高质量数据输入才能带来高质量输出（Good data in, good response out） — George Hou
## Mentioned
Adam Nahari, George Hou, Tom Rikert, Cursor, Moloco, Pinegrove Venture Partners, Google, YouTube, Autodesk, Andreessen Horowitz (a16z), Silver Lake, Archive, Jake Paul, Nvidia, Adobe, Uber, ChatGPT, Claude, Gemini, Slack, Salesforce, Gong, Notion, Glean, Workday, Infoseek, Lycos, Ask Jeeves
## Video description
Adam Nahari, Partner, Chief Data and AI Officer, Pinegrove Venture Partners
George Hou, Growth Lead, Cursor
Tom Rikert, Senior Director, Product Management, Moloco
Abstract
Fireside chat featuring AI leaders from Cursor, Moloco, and Pinegrove Venture Partners. The discussion will explore how top teams leverage AI and automation to elevate team productivity and drive meaningful ROI. Whether you're leading a sales team, managing engineering projects, or scaling your operations, learn how AI is revolutionizing core business functions and serving as a force multiplier. The conversation will also look ahead to the agentic AI future, examining how AI adoption will influence organizational structure, talent models, user trust, and the evolving definition of product excellence.
[Source](https://www.youtube.com/watch?v=ATisFik8yo0)
Source quotes
[1:28] Tom Rikert: 很高兴看到现场座无虚席。我想原因大概只有两个：要么是大家在 AI 投资上还没看到实际的 ROI，急着来这里取经；要么就是你们已经拿到了漂亮的 ROI，迫不及待想来互相印证。
[4:02] Tom Rikert: 如果大家还记得 Google 诞生之前的旧时光，那时有各式各样的搜索引擎，比如 Infoseek、Lycos、Ask Jeeves。当年大家最喜欢的乐趣就是互相比较用这个搜出的结果和那个有什么不同。
[9:30] Tom Rikert: 在这一路走来中，最让我们感到惊讶的发现之一就是，我们原本以为必须去市面上大规模采购很多……
[10:12] Tom Rikert: 先让子弹飞一会儿（let it ride for a little bit）
[12:32] George Hou: 设好就撒手不管（set it and forget it）
[15:53] George Hou: 高质量数据输入才能带来高质量输出（Good data in, good response out）
Source quotes
[4:02] Tom Rikert: fertile ground
[4:02] Tom Rikert: pride of craftsmanship
[5:50] George Hou: To earn the trust of seasoned professionals with 10 to 15 years of deep domain experience, teams must first solve mundane, highly repetitive daily tasks.
[10:12] Tom Rikert: didn't write it for you
[11:24] Tom Rikert: AI curious
[12:32] George Hou: as if you were teaching a kid
[12:32] George Hou: It's not just giving it the data that it needs, but telling it how to interpret it as if they were, for example, a sales leader or an account executive, and really helping them think through and remove as much cognitive load as possible.
[15:53] Tom Rikert: Software alone cannot execute change management or build operational sales muscle.
[18:24] George Hou: good data in, good response out
Original transcript
[0:04] Hi everybody. So, uh the title of our
[0:07] talk today is AI and action from vision
[0:09] to value. But really where we're going
[0:11] to focus is on um learning more from uh
[0:15] Tom and George about how they've adopted
[0:18] AI internally to drive innovation and
[0:21] efficiency. And then maybe being a bit
[0:24] more forwardlooking as we think about
[0:27] these AI capabilities evolving over
[0:29] time, how we think that that might
[0:31] affect internal talent structure,
[0:34] strategy. So that's where we're going to
[0:37] focus today. Uh before we get started, I
[0:40] thought that we would do a quick by show
[0:42] of hands um just to better understand
[0:45] our audience. Who here is technical
[0:49] codes on a daily basis? Maybe raise your
[0:52] hands.
[0:53] Okay. Maybe what is that like 20% maybe?
[0:56] Uh more business folks. Raise your
[0:59] hands.
[1:01] Yeah, more people. Okay. Um what about
[1:04] this? Who in here works at an
[1:06] organization with under 500 people?
[1:12] Over 500.
[1:15] Okay. Maybe like 70% over 500. Something
[1:18] like that. Okay, great. Maybe I thought
[1:21] we'd start with a quick round of
[1:23] introductions. If Tom and George, you
[1:25] could tell us a bit about your
[1:27] background.
[1:28] >> Sure. Um, great to see such a packed
[1:30] house. So, I think either, you know, no
[1:32] one's getting ROI from AI and and you
[1:34] all want to get it or you know, you've
[1:36] got it and you're excited to tell about
[1:37] it. Um, but great to see everybody. Uh,
[1:40] my name is Tom Riker. I'm uh product
[1:43] management at Malokco. If you're not
[1:45] familiar with Malco, I think it's going
[1:46] to be pretty relevant for the people in
[1:48] this room because we're about 800
[1:49] people. We've been around for about 10
[1:51] years and we're an AI native advertising
[1:54] platform. U so we power over $2 billion
[1:57] of annual ad spend through mobile
[1:59] advertising uh retail media which is
[2:02] advertising on uh retailer websites and
[2:05] also on uh streaming media networks. So
[2:08] um we've been a machine learning company
[2:09] for a long time because we've always
[2:11] used ML for recommendations and matching
[2:13] people with ads but we've been adopting
[2:16] uh generative AI and agentic AI pretty
[2:19] aggressively. So we'll talk about that
[2:21] uh quickly my background um started life
[2:23] started life as an engineer in computer
[2:26] vision and machine learning a long time
[2:27] ago moved to product management at
[2:29] companies like Autodesk and also Google
[2:30] and YouTube. um have been a VC. I was at
[2:33] Andre and Horowitz and invested in some
[2:35] of the early AI companies and then
[2:37] became an entrepreneur uh and then ended
[2:39] up at Maloco through an acquisition. So
[2:41] that's my background. Over to George.
[2:44] Hey everyone, my name is George. Uh lead
[2:46] growth at Curser. Uh we're an AI code
[2:48] editor. Uh work with over half of the
[2:50] Fortune 1000 including companies like
[2:52] Nvidia, Adobe, and Uber. Um and at
[2:55] Curser I focus on accelerating
[2:57] enterprise adoption and figuring out how
[2:59] we can capture um enterprise uh demand
[3:02] at scale uh by building AI systems. Um
[3:05] and before Curser um started off as a
[3:08] private equity investor at Silverlake,
[3:10] led product at Archive and uh as a fun
[3:13] fact actually helped run uh Jake Paul's
[3:15] venture capital firm. Um and before that
[3:19] uh studied statistics at Harvard where I
[3:21] met the incredible Adam.
[3:24] uh Adam Nihari, chief data and AI
[3:26] officer at Pinerrow Venture Partners,
[3:29] which is a San Francisco based venture
[3:31] and growth investment platform, working
[3:34] both on thinking through how we can
[3:37] leverage AI to source and underwrite
[3:40] investment opportunities more
[3:43] effectively as well as working with our
[3:45] portfolio companies on integrating AI
[3:49] into their processes. So Tom, maybe we
[3:51] can get started by learning more about
[3:54] how you've been thinking about uh
[3:55] adopting AI internally in Malokco.
[3:58] What's worked? What hasn't worked? What
[4:00] have you learned?
[4:02] >> Sure. Yes. So uh our approach in Maloka
[4:04] has been very much bottoms up. So we
[4:06] wanted to set the stage and kind of
[4:09] create fertile ground for a lot of uh
[4:12] experimentation. And so what that's
[4:14] looked like is we've enabled people to
[4:17] uh from their own teams be creative and
[4:19] use a few kind of enterprise approved um
[4:23] AI solutions. So we do have Jad CPT and
[4:26] Claude and Gemini and we wanted to make
[4:28] sure those were blessed by it and we
[4:32] didn't want any like shadow AI. remember
[4:34] the term shadow IT. It's kind of
[4:36] unmanaged in the background, but we
[4:37] wanted um people to feel whatever
[4:39] they're doing, they have full approval.
[4:41] And really, we thought the innovation
[4:43] would come from people in their daily
[4:45] jobs, whether they be in finance or HR
[4:48] or legal or product or engineering
[4:50] coming up with like, hey, this is
[4:52] something I want to try. So really we've
[4:55] we've sought to just enable people to
[4:57] experiment and then share the results
[5:00] through Slack channels through kind of
[5:02] informal meetups to to be able to bring
[5:05] like kind of pride of craftsmanship
[5:07] craftsmanship because people are excited
[5:09] to talk about what they built and so
[5:11] they start to feel like empowered and
[5:13] they're like pioneers uh and people
[5:15] funny enough if you remember the old
[5:17] days of like prior to Google there were
[5:20] different search engines you know
[5:21] there's infosek and lyos as jeves
[5:24] And people used to have fun comparing
[5:26] like their results of using one versus
[5:27] the other. So now we're seeing that with
[5:30] people saying, "Hey, I got this result
[5:31] with chat GPT or this one with Gemini
[5:34] and this one, you know, integrating with
[5:36] this uh corporate database." So it's
[5:38] really been uh exciting to see people
[5:41] voluntarily jump in and do this. Uh
[5:43] we're not doing L & D sessions, you
[5:45] know, people are learning how to use the
[5:46] tools on the on their own.
[5:49] George.
[5:50] >> Yeah, I think for us we have a bit of of
[5:53] both. Uh, you know, there's a bottoms up
[5:55] approach, people doing people using and
[5:57] experimenting with it themselves. Uh,
[5:59] but separately, we will also be pretty
[6:01] opinionated on which areas we think um
[6:04] have kind of the most potential for
[6:06] making an impact, especially on the go
[6:08] to market side of things. Um, so for us,
[6:10] I think one of the core learnings we've
[6:11] had is really start with the really
[6:15] really repetitive everyday workflows. I
[6:18] think everybody's very excited about the
[6:19] next moonshot idea, like wouldn't it be
[6:21] super cool if we did this? And everybody
[6:24] loves ideidating, but I think what we
[6:25] found is trust, developing trust with
[6:28] the end user um is incredibly important,
[6:31] especially um when the person that's in
[6:33] seat often has more than 10 15 years of
[6:36] experience in crafting um perfecting
[6:39] their craft and winning that trust uh
[6:42] starts um at the very very uh most basic
[6:46] fundamental task. Um so for us uh we
[6:49] think about if you know there's anything
[6:51] that we want to spend time in building
[6:53] AI workflows around um this has to be
[6:56] something that is uh an everyday
[6:58] workflow um because that's a very that's
[7:00] that's a pretty good reflection of how
[7:02] important that is. Um so to give a few
[7:04] examples uh one is uh every time an
[7:08] account executive gets on boarded um
[7:10] they have plenty of questions that they
[7:11] want to ask around the product around
[7:13] what are the best ways to handle a
[7:14] specific question and in the past um
[7:17] they would ping either a fellow account
[7:19] executive a manager or if they had a
[7:21] question about the product they would
[7:22] ping an engineer. Um but for us what we
[7:25] ended up finding was one um that would
[7:26] often delay deal cycles. Um, so from
[7:29] going to have a question on a call to
[7:31] getting a response in Slack often times
[7:33] would take a few hours or potentially up
[7:34] to a day. Uh, so instead what we built
[7:36] was an AI agent that sits on top of our
[7:39] core knowledge bases and data sources.
[7:40] So Salesforce, Gong, notion, Slack, etc.
[7:44] Um, and effectively it in the background
[7:46] is constantly able to answer any
[7:48] questions. So um, you know, while you're
[7:50] on a call, you might get a question on
[7:52] product and the product is constantly
[7:53] evolving as well. Um, so before even
[7:55] hopping off the call, um, you're pinging
[7:57] this agent in Slack and you're getting
[7:58] an answer immediately. So rather than
[8:00] potentially delaying that deal cycle and
[8:02] pushing that back, you have an instant
[8:04] answer right then and there on the call.
[8:06] Another example is after every call,
[8:09] account executives need to write a recap
[8:11] email uh, following up and typically
[8:13] that means um, kind of reviewing your
[8:15] notes uh, then drafting kind of the
[8:17] email, putting everything together. That
[8:19] can take, you know, 15 minutes. Um, but
[8:21] for us, I think we've been pretty lucky
[8:22] to have pretty strong uh inbound, you
[8:25] know, demand and pipeline. And a lot of
[8:26] it was, you know, the bottleneck for AES
[8:29] to be able to move on to the next deal
[8:30] was working their existing pipeline. Uh,
[8:32] so being able to have, uh, an AI agent
[8:35] automatically draft that in your tone,
[8:36] in your voice, um, and have that ready
[8:38] for you to send. So instead of taking 15
[8:40] minutes to send, um, being able to cut
[8:42] that down to 2 minutes, um, has been
[8:44] kind of a very, very big lift in kind of
[8:46] driving, um, kind of like adoption at
[8:49] the everyday level. So the AI agent that
[8:50] I was talking about earlier um in Slack
[8:53] um we have basically 100% 100% of the
[8:56] team is using it on a weekly basis um
[8:58] and most of the team is using it on a
[9:00] daily basis and we have these public
[9:02] Slack channels uh where we can kind of
[9:04] see everybody asking their questions and
[9:05] it's like every few minutes someone is
[9:07] asking a question um so I think from
[9:10] from that perspective we've found to
[9:12] really tackle the everyday monotonous
[9:14] kind of boring workflows um start there
[9:17] build trust and then over time um you
[9:19] you see that your team is very receptive
[9:21] and adopting it and then gradually build
[9:23] on top on top um but it all comes down
[9:26] to really kind of mapping out what that
[9:27] initial workflow looks like.
[9:30] How do you guys think about um use cases
[9:34] that maybe can be handled just by out of
[9:38] the box chat GPT today or maybe down the
[9:42] line? Uh even if the technology isn't
[9:46] there today, it might be there in a year
[9:48] versus use cases that require more
[9:51] bespoke custom tool building. Um and and
[9:56] how do you think about well maybe we
[9:57] don't want to make that investment
[9:59] because the foundation models are going
[10:01] to be good enough to tackle that use
[10:03] case in just a few months. Any thoughts
[10:06] on that? Yeah, we've been one of the
[10:08] biggest surprises for us is that we
[10:10] thought we need to buy a number of
[10:12] specialized tools for like HR use cases
[10:14] or for legal use cases and it turns out
[10:18] like the team's been finding ways of
[10:20] building their own uh kind of custom GPT
[10:23] or we use glean for kind of enterprise
[10:25] search and glean came bundled with a way
[10:27] to make an agent. So a great example
[10:29] like if you have spent a lot of time
[10:31] writing performance reviews for your
[10:33] colleagues. Has anyone spent a lot of
[10:36] time doing that? Just going through your
[10:37] notes and finding everything you need to
[10:39] bring together to write that review.
[10:41] Some of our HR folks said, "Won't we
[10:43] make an agent which can look through all
[10:45] of your Slacks, all of your emails, all
[10:47] of your calendar invites, all of your
[10:49] docs and see how you collaborated with
[10:51] one of your peers and bring that all
[10:53] together and summarize it." And they
[10:55] built this into our performance workflow
[10:57] tool which is connected to workday. And
[11:00] then that saved everybody so much time
[11:02] of having to do the performance review
[11:04] step. Didn't write it for you but it
[11:06] collected all the information and it was
[11:08] already integrated into workflow was
[11:11] already secure and private and that was
[11:13] someone someone did that using a feature
[11:15] of glean and we thought we would have
[11:17] had to buy a whole new system. Um, so
[11:19] there's been a number of examples like
[11:20] that where you can do it today if you
[11:22] just get creative. H
[11:24] >> how do you create a culture though, Tom,
[11:26] where people feel encouraged and
[11:28] empowered to create these custom GPTs?
[11:32] Because I can tell you at Pinerove,
[11:33] everybody's using our enterprise version
[11:36] of chat GPT daily as they're evaluating
[11:40] deals, as they're doing investment
[11:41] research, but we haven't had anybody
[11:43] create custom GPTs yet. So um maybe tell
[11:47] us a bit more about how do you encourage
[11:49] that culture maybe training or how have
[11:51] you thought about that?
[11:53] >> Yeah, I think it's um it's partly like
[11:55] we've encouraged people to be AI curious
[11:58] and just like try stuff and it's not
[12:00] like uh and I think I really respect our
[12:03] IT guys who and gals who are like hey we
[12:05] don't want to uh squatchch this. So even
[12:08] if you know people are doing things a
[12:10] little unorthodox like they let it ride
[12:12] for a little bit and then they would
[12:13] make sure they're doing it kind of the
[12:15] official way. So I think being flexible
[12:17] would probably be the main takeaway and
[12:19] then sometimes it's just like hey like
[12:21] we will uh allow for some freedom and
[12:24] then ultimately we'll find the best
[12:26] practice and then we'll we'll know
[12:28] that's best and we'll want everyone to
[12:29] use it. So that that will come
[12:30] eventually.
[12:32] George, any thoughts on um use cases
[12:35] that require custom tool building versus
[12:38] out of the box foundation model? Chad,
[12:41] GPT or Claude?
[12:43] Yeah, I think
[12:46] I think for us the way we approach it
[12:48] and maybe this also touches the second
[12:49] part of the question is we'll often
[12:52] basically me or someone else on the team
[12:54] will shadow uh an account executive or a
[12:57] sales leader and see just see what their
[12:58] workflow is um and map that out
[13:00] entirely. Um and I think to answer your
[13:03] question more directly some part of
[13:04] those um kind of are a little bit more
[13:06] repetitive, monotonous um and pretty
[13:09] predictable. And for that, I think
[13:11] that's a great way to kind of use the
[13:12] off-the-shelf LLMs. Maybe there's like
[13:14] some custom agent building. Uh, but it's
[13:16] pretty much like figuring out what what
[13:18] is the workflow, how can we communicate
[13:21] those instructions and steps very
[13:23] explicitly with very very good examples
[13:24] as if you were teaching a kid. Um, and
[13:27] mapping that out entirely. I feel like a
[13:29] lot of uh kind of having uh people build
[13:32] successful uh and you know being able to
[13:34] leverage AI effectively really comes
[13:36] down to prompting um giving it all the
[13:39] context that you need. Um it's not just
[13:41] giving it the data that it needs but
[13:43] telling it how to interpret it as if
[13:44] they were for example a sales leader or
[13:47] an account executive and really helping
[13:49] them think through and remove as much
[13:50] cognitive load as possible. Um so I
[13:53] think for us it's uh it's often a human
[13:55] in the loop process. It's not just a set
[13:58] it and forget it and just hope that it
[13:59] works. So, for example, when we were um
[14:02] drafting emails uh for account
[14:04] executives after every call. Um it was
[14:06] me basically sitting next to this person
[14:08] um and making sure that the tone and the
[14:10] voice was accurately reflecting um
[14:12] because I think there are certain kind
[14:13] of fields of work where style um and
[14:17] craft matter a lot more and often times
[14:19] it's not just the content and whether
[14:21] it's factually correct but it's whether
[14:23] or not it sounds like the individual. So
[14:25] I think the way I think it's been pretty
[14:27] effective for us is uh basically pairing
[14:29] shadowing um and very very much ensuring
[14:32] that the output is very much uh expected
[14:35] uh for whoever you're building for
[14:37] >> and that that resonates with me as well
[14:40] at Pinerove in addition to deploying
[14:43] chatpt across the organization and
[14:45] letting everybody just use the out of
[14:47] the box tool we've also built a custom
[14:50] tool uh that leverages the GPT and cloud
[14:56] APIs and for every new deal that our one
[15:00] of our strategies is a venture debt
[15:02] strategy. So they get a bunch of when a
[15:05] new deal comes in, they get a bunch of
[15:08] materials related to that deal. And
[15:11] we've built a tool that it goes through
[15:14] all of those materials, extracts the
[15:17] relevant information, and creates a
[15:20] first version of an investment memo,
[15:24] which hopefully uh
[15:27] cuts the time that the team needs to
[15:29] create that memo and extracts all of
[15:32] that information. Um
[15:36] and we've found that working very
[15:38] closely with the team to understand
[15:40] exactly what elements need to be in that
[15:42] memo. How do we
[15:44] build that so that it's part of their
[15:48] process? Uh so working very closely with
[15:50] them has been very important. Okay,
[15:53] great. We have five minutes. Let's talk
[15:55] about how we think that as AI
[15:58] capabilities develop, how will that
[16:01] affect your future respective
[16:04] organizations? Any thoughts on that,
[16:06] Tom?
[16:07] >> Yeah. So, um, and I think that, you
[16:10] know, the agentic movement, it really
[16:12] really is, uh, unprecedented. So, I
[16:15] think, you know, I've seen a few
[16:17] different technology cycles, but this
[16:18] really does seem like be completely new
[16:21] and exciting in a a grand way. Okay, so
[16:23] you probably heard the hyperbole, but I
[16:24] I really think it's it's becoming true
[16:26] quickly. So, one of the things we've
[16:28] done is we we built an enterprise
[16:30] product um for retailers. And so, we're
[16:33] a we're an enterprise software solution,
[16:36] but as we look at the Agentic
[16:38] opportunity instead of just providing
[16:40] software to a big retailer. Uh and the
[16:44] reason we're doing this is because
[16:45] they're using it to run their own
[16:46] advertising business. So, we're helping
[16:48] them transition from being a retailer to
[16:50] also being a media company. So a big
[16:53] challenge is we can give them the
[16:55] software but they still have to do the
[16:56] change management. They have to build
[16:57] the muscle to know how to sell
[16:59] advertising and run a media business and
[17:01] we can't really do that for them. But
[17:04] with with agent tech software we we can
[17:07] right like it's possible that we could
[17:09] give them agents who could be the ad
[17:12] sales team and be the ad operations team
[17:14] and deliver a whole solution right like
[17:16] solve the business problem because we
[17:18] can start to do that what used to be
[17:20] people and process with agent-based
[17:22] software so this is really exciting and
[17:25] I think this is just an example of one
[17:26] kind of enterprise software uh but for
[17:29] whatever software you're building if
[17:30] you're in the software business I think
[17:32] thinking of building not
[17:33] or providing a a software solution but a
[17:36] you know like a hightouch service you
[17:39] know like where you have a whole
[17:41] organization doing you know helping you
[17:42] run your business that's kind of the
[17:44] model we can we can move to I think with
[17:46] agents
[17:47] >> that that's really exciting are you so
[17:50] that that sounds like a great way to add
[17:53] value to your customers how are you
[17:56] thinking about that commercially is that
[17:58] an additional product that your
[18:00] customers are going to pay for or is
[18:03] that going to be part of your existing
[18:05] offering?
[18:07] >> Great questions and you know things that
[18:08] we're we're working through. Um but I
[18:11] think you know that this is this has
[18:12] become possible is kind of my point and
[18:14] so how we how you build a business model
[18:16] around it or adapt your business model
[18:18] is is a is kind of the next question.
[18:22] George.
[18:24] >> Yeah, I think um I think AI does not
[18:28] change the fact that you still need to
[18:30] hire the great and the best people, but
[18:32] I think it does redefine uh and shift
[18:35] like what does great actually mean. Um
[18:37] so I think going back to what you were
[18:39] disc you know you were saying earlier
[18:40] Adam um I think there are a few kind of
[18:44] archetypes of people that will be really
[18:45] really important. And I think the first
[18:46] archetype of the of person is one who
[18:50] intimately understands the underlying
[18:51] business problem and understands what is
[18:54] good versus great because ultimately AI
[18:56] will spit something out but to be able
[18:58] to actually understand what is good
[18:59] versus great often times comes down to
[19:01] experience and reps um taken for for a
[19:04] given domain. Um and I think the second
[19:06] maybe taste
[19:08] >> 100% 100%. Um and then I think the
[19:11] second bucket is um ultimately the data
[19:14] that you feed in to the AI uh matters so
[19:17] much. Um good data in good response out.
[19:21] Um so having someone who's very very
[19:23] thoughtful in understanding uh what what
[19:25] are the most important things to measure
[19:28] measuring that and ensuring that it is
[19:31] fed to the AI uh in the most digestible
[19:34] way. Um, and I think the third bucket is
[19:37] um, kind of in left field, which is
[19:39] people that are just incredibly
[19:40] incredibly creative and probably
[19:41] understand the domain very well. Uh,
[19:43] things that the LMS aren't trained on
[19:45] um, but are just able to connect
[19:47] different things like pop culture, etc.
[19:49] together that the AI probably wouldn't
[19:51] kind of immediately suggest. So, I think
[19:53] um, those kind of three uh, general
[19:56] buckets of folks I think is kind of how
[19:58] I would probably think about uh,
[19:59] constructing an org in today's world.
[20:02] >> Thank you, Tom and George. Thank you,
[20:03] everybody. Those 20 minutes passed a lot
[20:05] faster than I thought they would. Thank
[20:08] you,
