Momentum Finds 8% of Firms Run Real AI as Bad CRM Data Stalls Workflows
Ai4
Enterprise AI initiatives fail because manual CRM logging captures less than 1% of customer conversation data, starving downstream agents of the structured context required to automate revenue operations.
## Background
Published by Ai4 on YouTube on August 26, 2025, this session features Jonathan Kvarfordt, Head of GTM Growth at Momentum. The presentation covers how organizations can move past AI hype by using AI orchestration to convert first-party conversational interactions into structured, prompt-ready data for autonomous agents across sales, customer success, and RevOps.
## TL;DR
Addressing the gap between AI hype and internal reality, Jonathan Kvarfordt reveals that only 7.6% of organizations have genuinely operationalized AI deployments, primarily due to unstructured and fragmented data foundations. Because manual CRM entries capture only a fraction of rich customer conversations, downstream go-to-market workflows frequently suffer from hallucinations and inaccurate forecasting. Kvarfordt presents an orchestration architecture where multi-LLM agents act as a 'System of Action' to automate deal scoring, action items, customer handoffs, and real-time coaching. He introduces the GRACE prompting framework, emphasizing that concrete good and bad examples are essential to eliminate hallucinations, alongside background agent tools like Smart Clips and Account Briefs.
## Atmosphere & style
The presentation is energetic, pragmatic, and candid, with 'Coach K' using self-deprecating humor and empirical call data to puncture executive AI FOMO and validate the audience's shared frustrations. Moving at a brisk, focused pace, the session transitions from diagnosing messy operational realities to delivering structured, prescriptive playbooks. It feels like a practical, no-nonsense coaching clinic that grounds modern AI capabilities in tangible enterprise workflows.
## Key takeaways
- [01:41](https://www.youtube.com/watch?v=S_89SWcV-a0&t=101s) Empirical analysis reveals an adoption gap: while high-profile announcements fuel executive FOMO, only roughly 7.6% of companies have truly operationalized AI, with 60% severely lagging behind.
- [04:20](https://www.youtube.com/watch?v=S_89SWcV-a0&t=260s) AI initiatives stall primarily due to poor, uncleaned, and siloed data rather than model limitations, with data preparation consuming 70% to 80% of total project time.
- [05:30](https://www.youtube.com/watch?v=S_89SWcV-a0&t=330s) Traditional CRM data is notoriously unreliable because human reps summarize 8,000-word customer calls into just 20 to 25 words, leaving leadership disconnected from real deal health and pipeline reality.
- [08:18](https://www.youtube.com/watch?v=S_89SWcV-a0&t=498s) Production-ready enterprise agents require a three-part architecture: a multi-model ensemble (e.g., OpenAI, Claude, Cohere), tool integrations (CRM, Slack, email), and an orchestration layer for continuous validation.
- [08:18](https://www.youtube.com/watch?v=S_89SWcV-a0&t=498s) High-impact GTM agent use cases include predictive churn risk scoring, automated next-step execution, multi-month CS handoff dossiers, and real-time coaching tied to revenue outcomes.
- [11:39](https://www.youtube.com/watch?v=S_89SWcV-a0&t=699s) Modern orchestration transforms passive 'Systems of Record' into active 'Systems of Action' by extracting structured conversation data and routing customized context to relevant stakeholders.
- [11:39](https://www.youtube.com/watch?v=S_89SWcV-a0&t=699s) The GRACE framework (Goal, Role, Advanced Context, Constraints, Examples & Evaluation) provides a reliable prompting structure to eliminate single-prompt failures in enterprise agents.
- [14:36](https://www.youtube.com/watch?v=S_89SWcV-a0&t=876s) Providing explicit few-shot examples of both 'good' and 'bad' outputs acts as a rigid generation anchor, effectively preventing LLM hallucinations.
- [15:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=904s) 'Invisible agents' operate autonomously in the background without user adoption friction, powering workflows like automated video Smart Clips, deal Account Briefs, and coaching dashboards.
- [16:08](https://www.youtube.com/watch?v=S_89SWcV-a0&t=968s) Curated GTM prompt libraries, especially when paired with deep research tools for competitive battle cards, enable revenue teams to immediately deploy structured AI workflows at scale.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=4s) Introduction: The Reality Behind AI Frustrations
Jonathan Kvarfordt opens the session by addressing the widespread frustration with artificial intelligence, asking the audience how many have felt burned by AI hallucinations, reliability bugs, and broken promises. Acknowledging that he has personally experienced these issues numerous times, he jokes that anyone who has never had a single problem with AI should talk to him afterward so he can learn their secret. With only twenty minutes on the clock, he aims to unpack why these failures happen and how to move past the hype into real, working automation.
Kvarfordt introduces his background, explaining why he goes by the moniker "Coach K." With roughly ten people named John in his former office and a difficult-to-pronounce Swedish surname, the nickname stuck—especially fitting given his background in coaching sales and customer success teams on applying AI practically. Today, he serves as Head of GTM Growth at Momentum and leads the GTM AI Academy, an educational initiative that has trained tens of thousands of professionals across marketing, sales, and customer operations. Having started as an enthusiastic customer of Momentum two years prior, Kvarfordt literally begged the CEO to let him join the company to lead growth and marketing.
He outlines the agenda for the talk across four key pillars:
- Uncovering the core operational bottleneck facing modern AI initiatives.
- Identifying the "golden data" organizations already possess but consistently fail to leverage.
- Examining real-world production examples of AI agents executing go-to-market workflows.
- Deconstructing a structured, prompt-driven architecture that attendees can implement in their own stacks.
---
[01:41](https://www.youtube.com/watch?v=S_89SWcV-a0&t=101s) The Hype vs. Reality: Why Most Companies Are Far Behind
The tech industry has been swept up in intense anxiety and fear of missing out (FOMO) driven by splashy executive announcements. Kvarfordt highlights major headlines—such as Klarna's widely publicized customer service automation announcements, Marc Benioff claiming that Salesforce has 30% to 50% of its internal workloads handled by AI, and Duolingo's public push to AI-first operations. These high-profile statements have sent executive teams into panic mode, terrified that competitors are miles ahead while they lag behind.
To determine whether companies are genuinely running sophisticated AI workloads or merely marketing them, Kvarfordt analyzed Momentum's proprietary dataset of over 1,000 discovery calls conducted over the preceding six months across roughly 150 industries, spanning small-to-medium businesses up to large enterprises:
While broad market surveys from management consultancies like McKinsey suggest that upwards of 78% of companies have adopted at least one AI use case, Momentum's empirical data reveals that true operationalized deployment sits at roughly 7.6%. The vast majority (60%) are severely lagging, while another large segment is merely in the exploratory "emerging" stage. Kvarfordt reassures the audience: if you feel behind, you are, "but so is everybody else." The outward corporate positioning rarely matches internal operational reality.
---
[04:20](https://www.youtube.com/watch?v=S_89SWcV-a0&t=260s) The Foundation Problem: Why AI Projects Die on Unstructured Data
The primary reason AI initiatives fail or stall is rarely the intelligence of the underlying model—it is poor, uncleaned, unstructured, and fragmented data. Nearly half of all organizations surveyed remain bogged down in manual tasks that should theoretically be automated, but their internal infrastructure cannot support it. Studies indicate that 70% to 80% of total project time in AI deployments is consumed strictly by cleaning and preparing data.
When organizations lack a solid data foundation, downstream AI systems generate inaccurate or shallow outputs. This triggers process breakdowns across sales and customer success pipelines, leading teams to conclude that the multi-thousand-dollar AI software they purchased is failing to deliver value. In reality, the AI is starved of reliable, contextual inputs.
```mermaid
flowchart LR
accTitle: AI Data Failure Cycle
accDescription: Flowchart illustrating how poor data foundations lead to AI initiative failure.
subgraph S1["Data Ingestion"]
A["Unstructured / Siloed Data"] --> B["Poor Foundation"]
end
subgraph S2["Downstream Impact"]
B --> C["Hallucinations & Flawed Insights"]
C --> D["Broken GTM Workflows"]
D --> E["Failed ROI & Tech Abandonment"]
end
```
---
[05:30](https://www.youtube.com/watch?v=S_89SWcV-a0&t=330s) The Golden Asset: Unlocking Value from Customer Conversations
Kvarfordt argues that an enterprise's most valuable asset is the first-party conversational signal generated during daily customer interactions. However, traditional workflows rely on human beings to extract, summarize, and enter that information into systems of record. This creates an enormous data bottleneck:
- An average sales or customer call generates an 8,000-word transcript.
- The average account executive enters only 20 to 25 words into Salesforce or HubSpot.
- When polled, zero sales professionals in the room enjoy manually filling out CRM fields.
Because the CRM relies on exhausted humans to summarize complex conversations, "your CRM is probably lying." This informational breakdown cascades across the entire organization:
- **C-Suite Executives:** Blind to deal health, forced to manually call account executives to find out what is actually happening on flagship quarterly deals.
- **Sales Managers:** Left guessing revenue forecasts, hoping tomorrow's 1-on-1 will clarify if a pipeline commit is real.
- **RevOps & Enablement:** Unable to verify if sales methodologies (e.g., MEDDPICC) are practiced or if training initiatives move the needle.
- **Product & Marketing:** Completely disconnected from actual voice-of-customer feedback and ideal customer profile (ICP) pain points.
Rather than leaving data extraction to human memory, modern orchestration platforms must continuously capture conversational signals, validate the contents, and route them to CRMs, data warehouses (like Snowflake), and team collaboration tools (Slack and Microsoft Teams).
---
[08:18](https://www.youtube.com/watch?v=S_89SWcV-a0&t=498s) AI Agent Architecture and Real-World GTM Use Cases
Referencing Google's research framework on autonomous agents, Kvarfordt explains that a functioning enterprise agent requires three core layers: an underlying Foundation Model (Momentum utilizes an ensemble including Claude, OpenAI, and Cohere), integrated Tools (CRM, Google Calendar, email, Slack), and an Orchestration Layer to coordinate workflows and manage state.
```mermaid
flowchart TD
accTitle: Google Agent Framework in GTM
accDescription: Architectural model showing LLM ensemble, tools, and orchestration powering GTM workflows.
subgraph LLMs["Multi-LLM Ensemble"]
L1["OpenAI / Claude / Cohere"]
end
subgraph Orch["Momentum Orchestration Layer"]
O1["Identification & Validation Engine"]
end
subgraph Tools["Tools & Systems of Record"]
T1["Salesforce / HubSpot"]
T2["Slack / Teams"]
T3["Snowflake / BI Dashboards"]
end
L1 <--> O1
O1 --> T1
O1 --> T2
O1 --> T3
```
Kvarfordt details four production use cases executing across customer pipelines:
1. **Churn Risk & Deal Health Scoring:** The agent scans incoming meeting transcripts for subtle conversational cues (e.g., non-committal or cagey buyer behavior). If disengagement is detected, it scores the opportunity down and alerts deal desks before time is wasted on phantom pipeline.
2. **Automated Next-Step Execution:** Instead of reps manually setting tasks, the agent isolates agreed-upon action items from the call, formats them, and prompts the rep with a one-click confirmation to update Salesforce, Google Calendar, and Slack.
3. **CS Enablement Packets:** Long sales cycles spanning 5 to 20 calls over 3 to 9 months contain crucial technical details that get lost during post-sale handoffs. The agent synthesizes all historical conversations into a customized handoff dossier covering customer tech specs, metrics, and persona goals for the incoming Customer Success Manager.
4. **Real-Time Performance Coaching:** Enablement managers cannot review every recorded call. The agent evaluates rep execution against baseline top-performer behaviors, links coaching feedback directly with closed-won revenue data inside Salesforce or Snowflake, and provides contextual suggestions directly to reps.
---
[11:39](https://www.youtube.com/watch?v=S_89SWcV-a0&t=699s) The System of Action and the GRACE Prompting Framework
Kvarfordt differentiates between a traditional "System of Record" and a "System of Action." While Salesforce and HubSpot store static fields, Momentum operates as an intelligent execution engine that validates data before writing to the database:
- **Structured 1st-Party Data Extraction:** Cleans, extracts, and validates specific transcript entities before mapping them to custom CRM objects.
- **Context-Specific Stakeholder Routing:** Customizes notifications based on persona—giving a CEO immediate alerts on stalled $100k+ enterprise deals while delivering tactical operational summaries to front-line account executives.
- **Business Action Orchestration:** Automates multi-step cross-platform tasks directly triggered by conversational events.
To build reliable agents across platforms like n8n, Make, or custom LLM stacks, Kvarfordt introduces the **GRACE** prompting framework:
| Component | Definition & Best Practice |
|:--- |:--- |
| **G — Goal** | The precise objective the agent must achieve, enriched with as much situational context as possible. |
| **R — Role** | Who the AI represents (e.g., combining complementary personas like a "McKinsey strategist" and a "SaaS finance director" to elevate output quality). |
| **A — Advanced Context** | Deep domain, company, and situational knowledge—the single area where most DIY enterprise prompts fail. |
| **C — Constraints / Instructions** | Explicit operational guardrails, edge-case rules, and output constraints governing what the agent must not do. |
| **E — Examples & Evaluation** | Few-shot inputs and structured rubrics ensuring deterministic, repeatable outputs. |
Kvarfordt stresses that naive "single-prompt-to-output" workflows inevitably fail in production. Enterprise-grade AI agents require continuous validation checks at every step of the orchestration pipeline to ensure junk data never pollutes the underlying CRM.
[14:36](https://www.youtube.com/watch?v=S_89SWcV-a0&t=876s) Prompting with Examples: Defining Good vs. Bad to Prevent Hallucinations
When designing prompts for AI workflows, the single most critical area to invest time in is providing concrete examples. Jonathan Kvarfordt emphasizes that teams must explicitly show the model both what "good" and "bad" outputs look like. By providing an exact response template and instructing the model to "respond exactly like this," teams provide a rigid anchor for generation. This few-shot approach dramatically reduces hallucinations because the model is no longer guessing the expected format or depth—it simply aligns with the provided pattern.
- **Example-Driven Prompting:** Clear templates eliminate ambiguity and prevent the model from deviating or inventing unsupported claims.
- **Defining Quality Boundaries:** Explicitly contrasting acceptable outputs against subpar examples establishes clear quality guardrails.
```mermaid
flowchart LR
accTitle: Few-Shot Example Grounding Flow
subgraph Inputs["Prompt Context"]
A["Task Definition"]
B["Negative Examples<br/>(What Bad Looks Like)"]
C["Positive Templates<br/>(What Good Looks Like)"]
end
subgraph LLM["Agent Execution"]
D["Model Processing"]
end
subgraph Output["Output"]
E["Structured, Hallucination-Free Response"]
end
A --> D
B --> D
C --> D
D --> E
```
[15:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=904s) Invisible Agents and New Feature Releases: Smart Clips, Account Briefs, and Coaching
The discussion shifts toward the core philosophy of "invisible agents"—systems that run autonomously in the background across go-to-market workflows without requiring continuous human babysitting or active dictation. To bring this concept to life, Momentum unveiled a series of product features designed to extract and distribute intelligence automatically across standard business channels.
Leading the announcements is **Smart Clips**, described as a hybrid of "Opus Clip and TikTok combined into one." The tool continuously monitors recorded customer conversations, automatically identifies key moments defined by the user, and trims them into bite-sized 15-second clips ready for internal distribution or enablement. Alongside video clipping, Momentum introduced automated **Account Briefs**, which operate similarly to closed-won or closed-lost post-mortems. As a deal progresses through opportunity stages, the agent bundles all recent communications—aggregating emails, calls, and meeting transcripts—and delivers real-time updates directly to stakeholders. Finally, an in-depth **Coaching Dashboard** aggregates these signals to give leadership granular visibility into rep execution and performance trends.
- **Smart Clips:** Automatically surfaces key conversation moments and trims them down into short, shareable clips (e.g., 15 seconds) for enablement and collateral.
- **Account Briefs:** Packages ongoing deal activity from calls and emails into contextual summaries that automatically update as deals move through CRM pipeline stages.
- **Coaching Dashboards:** Provides deep visibility into rep performance and interaction patterns without manual reporting.
[16:08](https://www.youtube.com/watch?v=S_89SWcV-a0&t=968s) The GTM Prompt Library and Closing Takeaways
After briefly scanning the room for representative Catalina from customer ElevenLabs, Jonathan concludes the presentation by sharing an extensive, personally authored prompt library containing over 200 tailored prompts. Designed specifically for go-to-market functions—including sales, customer success, and marketing—the repository provides ready-to-deploy frameworks for diverse revenue operations.
He highlights the "competitive battle card" prompt as a standout favorite, particularly when paired with modern deep research tools to synthesize competitor intelligence. With the session wrapping up right on schedule, the audience is invited to leverage the prompt library to immediately implement structured agentic workflows within their own teams.
- **200+ GTM Prompt Library:** Curated prompts targeting sales discovery, customer success handoffs, marketing workflows, and competitive analysis.
- **Competitive Battle Cards:** Recommended for pairing with deep research AI models to generate comprehensive market intelligence.
## Notable quotes
> [01:41](https://www.youtube.com/watch?v=S_89SWcV-a0&t=101s) if you feel behind, you are, but so is everybody else. — Jonathan Kvarfordt
> [05:30](https://www.youtube.com/watch?v=S_89SWcV-a0&t=330s) your CRM is probably lying. — Jonathan Kvarfordt
> [15:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=904s) Opus Clip and TikTok combined into one — Jonathan Kvarfordt
> [14:36](https://www.youtube.com/watch?v=S_89SWcV-a0&t=876s) respond exactly like this — Jonathan Kvarfordt
## Mentioned
Jonathan Kvarfordt, Momentum, GTM AI Academy, Klarna, Marc Benioff, Salesforce, Duolingo, McKinsey, HubSpot, Snowflake, Slack, Microsoft Teams, Google, Claude, OpenAI, Cohere, Google Calendar, n8n, Make, GRACE, Smart Clips, Opus Clip, TikTok, Account Briefs, ElevenLabs
## Video description
Jonathan Kvarfordt, Head of GTM Growth, Momentum
Abstracts
AI agents promise automation and decision-making superpowers—but most fall short due to one overlooked problem: the data. In this panel, you’ll learn how leading companies use AI Orchestration are turning first-party interactions into structured, prompt-ready fuel for agents that actually do the work. We’ll cover real use cases where agents orchestrate workflows, accelerate revenue, and drive performance across sales, CS, and RevOps—with no code, no dashboards, and no user adoption required.
From prompt frameworks to proven playbooks, this session will show you how to move past AI hype and into measurable execution—at scale
Jonathan Kvarfordt of Momentum argues that corporate claims regarding artificial intelligence adoption obscure an operational reality where few organizations run functioning autonomous workflows. To capture value, enterprises must transition from manual systems of record to automated execution engines that extract and validate conversational customer data using structured prompt architecture.
Adoption Hype Versus Enterprise Operational Reality (1:41)
Public announcements from major corporations—such as Klarna's customer service automation and Salesforce's claim that AI handles 30% to 50% of internal workloads—have generated widespread corporate anxiety. However, Kvarfordt contends that outward positioning rarely reflects operational reality. An analysis of Momentum's dataset comprising over 1,000 discovery calls across roughly 150 industries revealed that approximately 60% of organizations lag behind with no operational AI adoption, 32.4% occupy an emerging testing phase, and only 7.6% maintain fully operational deployments.
This empirical finding contrasts sharply with broader market surveys, such as McKinsey reports suggesting that up to 78% of companies have adopted at least one AI use case. Kvarfordt notes that while organizational leaders frequently fear their competitors have pulled ahead, the vast majority of enterprises remain stalled in exploratory stages rather than running mature, production-grade workloads.
The Conversational Data Bottleneck in CRMs (4:20)
AI initiatives stall primarily due to poor, unstructured data foundations rather than limitations in model intelligence. Studies cited in the presentation show that 70% to 80% of project time in AI deployments is consumed by cleaning and preparing data. Without clean, contextual inputs, downstream models produce shallow or inaccurate outputs, leading teams to abandon expensive software investments.
Kvarfordt identifies first-party customer conversations as an organization's most critical data asset, yet manual entry severely bottlenecks its capture. An average sales call generates an 8,000-word transcript, but account executives manually log only 20 to 25 words into customer relationship management (CRM) platforms like Salesforce or HubSpot. This information deficit blinds C-suite executives to deal health, forces sales managers to guess revenue forecasts, prevents revenue operations from verifying sales methodologies like MEDDPICC, and isolates product teams from authentic customer feedback.
Multi-Layered Architectures for Revenue Workflows (8:18)
Drawing on Google's autonomous agent framework, Kvarfordt describes a functional enterprise system comprising three components: a multi-model ensemble (such as Claude, OpenAI, and Cohere), integrated enterprise tools (CRMs, email, Slack, calendars), and an orchestration layer that coordinates state and validates data. Rather than acting as static databases, these systems operate as execution engines that validate conversational signals before routing them to data warehouses like Snowflake.
In production, this architecture supports four primary go-to-market workflows: scanning transcripts for buyer disengagement to score churn risk and pipeline health; isolating agreed action items to trigger one-click next-step updates; compiling technical specifications across multi-month sales cycles into customer success handoff dossiers; and scoring rep behavior against top performers to deliver real-time coaching linked to closed-won revenue data.
Structured Prompt Engineering and Background Automation (11:39)
To build deterministic workflows, Kvarfordt advocates the GRACE prompting framework: Goal (clear objectives and context), Role (complementary personas, such as a strategist paired with a finance director), Advanced Context (deep domain and company knowledge, cited as the most common DIY failure point), Constraints (explicit negative guardrails), and Examples and Evaluation (few-shot inputs). Explicitly contrasting 'good' and 'bad' response templates anchors generation and sharply curbs model hallucinations.
These prompt structures underpin background automation—termed 'invisible agents'—that operate continuously without manual prompting. Recent product applications include automated tools that clip notable 15-second conversational moments for training, account briefs that aggregate calls and emails into deal updates across CRM stages, and coaching dashboards that monitor rep execution trends across standard business channels.
Deploying reliable AI agents in go-to-market operations requires moving away from single-prompt direct generation and manual CRM data entry toward continuous, intermediate validation layers that clean and route conversational inputs automatically.
## Background
该视频由 Ai4 于 2025 年 8 月 26 日发布在 YouTube 平台上，内容为 Momentum 的 GTM 增长负责人 Jonathan Kvarfordt 的主题演讲。根据页面简介，演讲深入剖析了绝大多数 AI Agent 因数据缺陷而难以落地的核心痛点，并展示了领先企业如何通过 AI 编排将第一方真实交互转化为结构化数据，从而在无需代码和繁琐界面的前提下，跨销售、客户成功和 RevOps 驱动真正产生业务价值的“隐形智能体”。
## TL;DR
Jonathan 通过调研 1000 场企业真实咨询通话的数据揭示，仅有 7.6% 的企业真正实现了 AI 的常态化业务运营，而阻碍落地的最大症结在于人工录入导致的 CRM 脏数据与信息断层。为此，他提出应将客户对话信号作为第一方核心数据源，借助由大语言模型、工具生态和编排层构成的 AI Agent 架构，在流失预警、行动归档、交付赋能和销售辅导等场景中实现自动化闭环。最后，演讲系统拆解了专为智能体流程设计的 GRACE 提示词框架与少样本校准技巧，并发布了 Smart Clips 智能高光剪辑、动态账户简报以及覆盖 200+ 业务场景的开源 GTM 提示词库。
## Atmosphere & style
整场演讲务实生动、节奏明快，Jonathan 展现出极具亲和力的实战教练风格，开场便以直击痛点的行业焦虑调查牢牢抓住全场注意力。随着从刺破市场宣传泡沫的硬核数据，层层递进到 CRM 底层逻辑解构与 GRACE 提示词实操，现场氛围从对行业通病的深刻共鸣转向获得方法论后的豁然开朗。整体观感干货密集且极具实战指导意义，让听众在摆脱“AI 落后焦虑”的同时，清晰掌握了构建隐形智能体工作流的具体路径。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=4s) 行业中普遍弥漫着落后焦虑，但真实数据显示高达 60% 的企业在 AI 落地方面严重滞后，真正实现常态化业务运营落地的仅占 7.6%。
- [04:20](https://www.youtube.com/watch?v=S_89SWcV-a0&t=260s) 绝大多数 AI 项目夭折的根源在于数据基础设施崩塌与 CRM 脏数据，70% 至 80% 的项目时间被迫耗费在原始数据清洗与准备上。
- [04:20](https://www.youtube.com/watch?v=S_89SWcV-a0&t=260s) 客户真实对话是企业最有价值的黄金数据资产，但传统人工录入导致信息严重断层与流失，令高管决策陷入盲区。
- [06:54](https://www.youtube.com/watch?v=S_89SWcV-a0&t=414s) 高效的 AI Agent 体系依赖三大核心支柱：底层多模型混合底座（LLMs）、跨系统业务工具集（Tools）以及核心编排层（Orchestration Layer）。
- [08:52](https://www.youtube.com/watch?v=S_89SWcV-a0&t=532s) 通过对非结构化对话信号进行多阶段清洗与校验，AI Agent 可实现谨慎买家识别、商机流失预警及低意向线索过滤。
- [08:52](https://www.youtube.com/watch?v=S_89SWcV-a0&t=532s) AI Agent 能自动提取跟进待办并一键同步多平台，同时在长达数月的交付周期中自动聚合生成跨部门的客户赋能包与销售辅导分析。
- [11:40](https://www.youtube.com/watch?v=S_89SWcV-a0&t=700s) 企业应将 CRM 升级为推动业务前进的“行动系统”，并借助专为 Agent 设计的 GRACE 框架（目标、角色、高阶上下文、约束、范例）构建高可靠提示词。
- [14:36](https://www.youtube.com/watch?v=S_89SWcV-a0&t=876s) 遏制大模型幻觉的关键在于少样本（Few-Shot）正反示例设计，通过向模型明确展示优质与糟糕标准的对比来锚定输出格式。
- [15:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=904s) “隐形智能体”（Invisible Agents）理念主张让 AI 在业务流程后台无感自动化运行，无需业务人员步步下达指令或手动干预。
- [15:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=904s) Momentum 推出 Smart Clips 智能片段提取与 Account Briefs 账户全周期动态聚合简报，并开源了 200 余个覆盖完整 GTM 场景的提示词库。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=4s) 破除 AI 焦虑：行业炒作与企业落地的真实差距
演讲开场，Momentum 的 GTM 增长负责人 Jonathan Kvarfordt（因姓氏难念且长期培训销售与客户成功团队而被昵称为“Coach K”）向现场观众抛出一个直击痛点的问题：有多少人在使用 AI 时遭遇过幻觉、翻车或技术故障？台下几乎全员共鸣。作为 GTM AI Academy 的主理人，Jonathan 培训了数以万计的销售与营销人员使用 AI，他指出当前整个行业弥漫着严重的“落后焦虑”——媒体头条不断宣扬重大突破，Salesforce 首席执行官 Marc Benioff 宣称其 30% 到 50% 的工作量已由 AI 代劳，Duolingo 等公司也大谈 AI 优先。企业管理层和普通员工都在焦虑地认为“别人都跑在前面，只有我们落后了”。
为了探寻真相，Momentum 团队对过去六个月中来自跨越 150 多个行业、涵盖从中小企业（SMB）到大型企业的 1000 场关于 AI 自动化与 Agent 咨询的真实通话记录进行了深度挖掘。数据呈现出的真实情况彻底颠覆了外界宣传：
- **落后群体（Lagging）占 60%**：绝大多数企业在 AI 落地方面严重滞后，无论是大公司还是初创团队，均处于起步困难状态。
- **觉醒群体（Emerging）在增长**：部分企业意识到趋势并渴望采取行动，但仍缺乏落地路径。
- **先进群体（Advanced）凤毛麟角**：即使是极少数自认为先进的企业，其实际用例深度也远低于行业设想。
- **运营化落地仅为 7.6%**：尽管咨询机构（如麦肯锡）报告称 78% 的企业已尝试过某种形式的 AI 用例，但真正将其转化为常态化业务运营的仅有 7.6%。
- **半数企业受困于手动流程**：近 50% 的团队仍深陷本可通过 AI Agent 自动化的繁琐手工操作中。
```mermaid
flowchart TD
accTitle: AI Adoption Reality vs Perceived Gap
subgraph MarketPerception["市场认知与宣传"]
M1["媒体头条 & 巨头言论<br/>'30%-50% 工作由 AI 完成'"]
M2["麦肯锡报告<br/>78% 企业尝试 AI 用例"]
end
subgraph ActualData["Momentum 1000场真实通话分析"]
A1["60% 企业严重落后 (Lagging)"]
A2["30%+ 企业处于探索意识期 (Emerging)"]
A3["仅 7.6% 真正实现常态运营化 (Operational)"]
end
MarketPerception -.->|"产生认知偏差与焦虑"| ActualData
```
---
### [04:20](https://www.youtube.com/watch?v=S_89SWcV-a0&t=260s) 核心症结所在：脏数据与 CRM 的“谎言”
绝大多数 AI 项目夭折的根源并非模型能力不足，而是数据基础设施的崩塌。约 70% 至 80% 的 AI 项目时间被耗费在原始数据清洗与准备上，许多企业根本没有准备好迎接高阶 AI。缺乏坚实的数据底座，AI 只能吐出不准确、价值受限的伪洞察，进而导致销售与客户全生命周期业务流程断裂。最终，企业斥巨资购买的昂贵 AI 工具被弃之不用。
Jonathan 强调，企业最有价值的“黄金资产”是每天与客户进行的真实对话。然而在传统模式下，数据沉淀完全依赖人工录入，从而产生了巨大的信息断层：
- **信息严重流失**：一场包含 8,000 词的客户交流录音，销售代表在通话后可能仅在 Salesforce 中草草填上 25 个词。
- **高管层决策“盲飞”**：C-Suite 完全无法掌握核心交易动态，不得不频繁打电话催问销售代表季度大单的进展。
- **销售主管全凭猜测**：团队管理者在向上汇报时只能靠直觉预估“这个月大概能关单 100 万美元”，真实进度必须等第二天私下沟通。
- **RevOps 与赋能团队无从验证**：运营团队无法评估销售方法论是否被实际执行，亦无法量化培训效果。
- **客户心声与产品反馈湮灭**：最真实、最宝贵的理想客户画像（ICP）洞察与产品改进建议在口头沟通后随风消散。
---
### [06:54](https://www.youtube.com/watch?v=S_89SWcV-a0&t=414s) AI Agent 的核心三要素与对话信号挖掘
各个业务职能部门——财务、AE（客户经理）、BDR（业务拓展代表）以及技术高管（CIO/CTO）——都需要跨平台的实时业务数据，但这些系统往往彼此割裂、无法互通。解决该问题的关键，在于将对话信号（Conversational Signals）作为第一方核心数据源，并通过 Agent 引擎完成清洗、校验与自动分发。
根据 Google 发布的 AI Agent 白皮书，构建高效的 Agent 体系必须具备三大支柱：
1. **底层大语言模型（LLM）**：支持多模型混合调用（如 Momentum 同时接入 Cohere、Claude、OpenAI 等 5 至 6 种主流模型）。
2. **工具集成生态（Tools）**：连接 Salesforce、HubSpot、日历、邮箱、Slack、Teams 等日常生产力工具。
3. **编排层（Orchestration Layer）**：负责跨系统连接、数据流转与流程编排的核心中枢。
```mermaid
flowchart LR
accTitle: AI Agent Three-Pillar Architecture
subgraph Engine["Google 白皮书：AI Agent 三要素"]
direction TB
LLM["多模型底座 (LLMs)<br/>OpenAI / Claude / Cohere"]
Tools["业务工具集 (Tools)<br/>Salesforce / Slack / 日历 / 邮箱"]
Orch["Momentum 编排层 (Orchestration)<br/>清洗、校验、路由"]
LLM --- Orch
Tools --- Orch
end
Input["客户原始对话<br/>(8,000 词非结构化语料)"] --> Orch
Orch --> Out1["CRM 结构化对象<br/>(Account / Opportunity / Lead)"]
Orch --> Out2["自动化业务动作<br/>(Slack 预警 / 赋能包 / 会议日程)"]
```
---
### [08:52](https://www.youtube.com/watch?v=S_89SWcV-a0&t=532s) 落地实战：Momentum 的四大高价值业务场景
与传统简单调用单一 Prompt 处理文本并直接输出不同，Momentum 采用深度的多阶段校验与清洗机制，确保流向下游系统的数据真实可靠，避免向 CRM 注入垃圾信息。Jonathan 列举了四个已经在企业中产生巨大商业价值的场景：
- **客户流失预警与谨慎买家（Cagey Buyer）识别**：Agent 自动深入分析全量通话记录，捕捉客户参与度下降或模棱两可的潜在风险信号，触发评分模型并实时提醒销售团队，避免在虚假商机上浪费精力。
- **下一步行动（Next Steps）无缝归档**：Agent 从对话中自动提取双方约定的待办事项与后续跟进计划，销售只需一键确认即可同步写入 Salesforce、Slack 和日程表，省去手工整理时间。
- **跨周期交付与客户赋能包（Enablement Packets）**：对于周期长达 3 至 9 个月、历经 5 到 20 次沟通的复杂项目，Agent 自动汇总所有历史会话，按客户成功（CS）团队定制的行业、画像及核心指标模板生成交接知识包，打破部门间的信息孤岛。
- **实时销售辅导与成单表现分析（Coaching & Performance）**：管理者无需耗费数小时逐一听取录音，Agent 会自动对标团队销冠的沟通策略，精准指出当前代表的优劣势并提供针对性建议；同时将对话表现与 Salesforce 的最终营收结果进行归因关联，数据可直接在 Momentum 面板展示或同步至 Snowflake / Salesforce。
---
### [11:40](https://www.youtube.com/watch?v=S_89SWcV-a0&t=700s) 从记录系统到行动系统：GRACE Prompt 框架
Jonathan 指出，现代企业应当将 CRM 视为推动业务前进的“行动机制”（Business Action Mechanism），而不是单纯的静态记录库。Momentum 的核心定位即是“行动与数据系统”（System of Action and Data）。构建这类高可靠 Agent 工作流时，提示词工程扮演着关键角色。基于对前沿大模型 Prompt 指南的研究，Jonathan 提出了专为 Agent 流程设计的 **GRACE** 结构化框架：
- **G - Goal（目标）**：明确 AI 需要达成的具体业务产出，上下文信息越详实，效果越精准。
- **R - Role（角色）**：设定 AI 扮演的专家身份（例如复合角色：“你是一位麦肯锡战略咨询专家兼资深财务分析师”）。
- **A - Advanced Context（高阶上下文）**：注入业务全貌、系统状态及背景信息，这是多数失败 Prompt 最常缺失的核心环节。
- **C - Constraints / Instructions（约束与指令）**：定义严格的业务边界、格式规范与执行逻辑。
- **E - Examples / Evaluation（范例与评估标准）**：提供优质少样本范例（Few-shot）和预期评价指标以校准输出。
### [14:36](https://www.youtube.com/watch?v=S_89SWcV-a0&t=876s) 提示词设计关键：用正反示例（Few-Shot）遏制模型幻觉
在构建高质量的 AI 工作流时，Jonathan Kvarfordt 强调大家应该将大量精力投入到“提供示例”上。要让大模型稳定输出预期结果，核心在于明确告诉它“什么是优秀的回答，什么是糟糕的回答”（"Does the AI know what good looks like? Does it know what bad looks like?"）。如果能直接提供一套标准模板，并明确要求 AI “严格按照这种格式和标准来回复”，模型在执行任务时就会有据可依，从而大幅降低幻觉（Hallucination）的发生概率。
```mermaid
flowchart LR
accTitle: 提示词工程示例驱动机制
subgraph Input["输入标准与上下文"]
Good["优质示例 (Good)"]
Bad["反面示例 (Bad)"]
Tpl["目标响应模板"]
end
subgraph Engine["AI 执行与校准"]
LLM["大语言模型"]
end
subgraph Output["业务产出"]
Result["无幻觉 / 高结构化输出"]
end
Good --> LLM
Bad --> LLM
Tpl --> LLM
LLM --> Result
```
---
### [15:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=904s) “隐形智能体”与新功能发布：Smart Clips、账户摘要与 200+ GTM 提示词库
Jonathan 随后阐述了 Momentum 的核心理念——“隐形智能体”（Invisible Agents）。所谓隐形智能体，就是让 AI 能够在整个业务流转过程中默默自动化处理各项任务，无需业务人员步步下达指令或手动干预。紧接着，他现场公布了 Momentum 的几项最新功能与产品特性：
首先是全新发布的 **Smart Clips** 功能。Jonathan 将其形容为“Opus Clip 与 TikTok 的结合体”，系统会自动分析销售或客户沟通视频，根据用户设定的关键节点，智能提取并自动剪辑出 15 秒左右的高光片段，方便团队随时调取与分发。其次是 **Account Briefs（账户简报）**，它沿用了赢单/输单分析（Closed Won / Closed Lost）的底层逻辑，当交易在销售漏斗各阶段流转推进时，智能体能够自动聚合最新的往来邮件、通话记录及各类沟通信息，为团队实时更新全局交易全貌。此外，系统还提供深入的辅导看板（Coaching Dashboard），帮助管理者洞察团队业绩表现。
演讲尾声，Jonathan 特别提到了其标杆客户 ElevenLabs（并在现场寻找未能到场的嘉宾 Catalina），随后向全场与会者分享了一份由他个人精心编写、包含 200 多个提示词的官方 **Prompt Library**。该库覆盖了销售（Sales）、客户成功（CS）、市场营销（Marketing）等全部 GTM 业务场景，其中他最推荐的是“竞争对手对战卡（Competitive Battle Card）”提示词，尤其建议搭配 Deep Research 功能使用以发挥最强效果。
- **隐形智能体理念**：业务执行全程无需人类反复提示或干预，实现后台无感驱动。
- **Smart Clips**：自动将长视频通话提炼并裁剪为 15 秒左右的重点片段。
- **Account Briefs**：随着商机阶段推进，跨邮件、通话等多渠道自动聚合生成最新的账户进展报告。
- **辅导与绩效看板**：提供团队维度的深度洞察与绩效分析。
- **200+ GTM 提示词库**：全面开源供现场观众使用，重点推荐结合深度研究工具（Deep Research）运行竞品对战卡模板。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=4s) 有多少人在使用 AI 时遭遇过幻觉、翻车或技术故障？ — Jonathan Kvarfordt
> [00:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=4s) 别人都跑在前面，只有我们落后了 — Jonathan Kvarfordt
> [00:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=4s) 落后群体（Lagging）占 60%：绝大多数企业在 AI 落地方面严重滞后，无论是大公司还是初创团队，均处于起步困难状态。 — Jonathan Kvarfordt
> [00:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=4s) 运营化落地仅为 7.6%：尽管咨询机构（如麦肯锡）报告称 78% 的企业已尝试过某种形式的 AI 用例，但真正将其转化为常态化业务运营的仅有 7.6%。 — Jonathan Kvarfordt
> [04:20](https://www.youtube.com/watch?v=S_89SWcV-a0&t=260s) 绝大多数 AI 项目夭折的根源并非模型能力不足，而是数据基础设施的崩塌。 — Jonathan Kvarfordt
> [04:20](https://www.youtube.com/watch?v=S_89SWcV-a0&t=260s) 一场包含 8,000 词的客户交流录音，销售代表在通话后可能仅在 Salesforce 中草草填上 25 个词。 — Jonathan Kvarfordt
> [04:20](https://www.youtube.com/watch?v=S_89SWcV-a0&t=260s) 这个月大概能关单 100 万美元 — Jonathan Kvarfordt
> [11:40](https://www.youtube.com/watch?v=S_89SWcV-a0&t=700s) 现代企业应当将 CRM 视为推动业务前进的“行动机制”（Business Action Mechanism），而不是单纯的静态记录库。 — Jonathan Kvarfordt
> [14:36](https://www.youtube.com/watch?v=S_89SWcV-a0&t=876s) Does the AI know what good looks like? Does it know what bad looks like? — Jonathan Kvarfordt
> [14:36](https://www.youtube.com/watch?v=S_89SWcV-a0&t=876s) 严格按照这种格式和标准来回复 — Jonathan Kvarfordt
> [15:04](https://www.youtube.com/watch?v=S_89SWcV-a0&t=904s) Opus Clip 与 TikTok 的结合体 — Jonathan Kvarfordt
## Mentioned
Momentum, Jonathan Kvarfordt (Coach K), Ai4, Salesforce, Marc Benioff, Duolingo, McKinsey & Company (麦肯锡), Google, Cohere, Claude, OpenAI, HubSpot, Slack, Microsoft Teams, Snowflake, ElevenLabs, Catalina, Opus Clip, TikTok
## Video description
Jonathan Kvarfordt, Head of GTM Growth, Momentum
Abstracts
AI agents promise automation and decision-making superpowers—but most fall short due to one overlooked problem: the data. In this panel, you’ll learn how leading companies use AI Orchestration are turning first-party interactions into structured, prompt-ready fuel for agents that actually do the work. We’ll cover real use cases where agents orchestrate workflows, accelerate revenue, and drive performance across sales, CS, and RevOps—with no code, no dashboards, and no user adoption required.
From prompt frameworks to proven playbooks, this session will show you how to move past AI hype and into measurable execution—at scale
[Source](https://www.youtube.com/watch?v=S_89SWcV-a0)
Source quotes
[0:04] Jonathan Kvarfordt: 有多少人在使用 AI 时遭遇过幻觉、翻车或技术故障？
[0:04] Jonathan Kvarfordt: 别人都跑在前面，只有我们落后了
[0:04] Jonathan Kvarfordt: 落后群体（Lagging）占 60%：绝大多数企业在 AI 落地方面严重滞后，无论是大公司还是初创团队，均处于起步困难状态。
[0:04] Jonathan Kvarfordt: 运营化落地仅为 7.6%：尽管咨询机构（如麦肯锡）报告称 78% 的企业已尝试过某种形式的 AI 用例，但真正将其转化为常态化业务运营的仅有 7.6%。
[4:20] Jonathan Kvarfordt: 绝大多数 AI 项目夭折的根源并非模型能力不足，而是数据基础设施的崩塌。
[4:20] Jonathan Kvarfordt: 一场包含 8,000 词的客户交流录音，销售代表在通话后可能仅在 Salesforce 中草草填上 25 个词。
[4:20] Jonathan Kvarfordt: 这个月大概能关单 100 万美元
[11:40] Jonathan Kvarfordt: 现代企业应当将 CRM 视为推动业务前进的“行动机制”（Business Action Mechanism），而不是单纯的静态记录库。
[14:36] Jonathan Kvarfordt: Does the AI know what good looks like? Does it know what bad looks like?
[14:36] Jonathan Kvarfordt: 严格按照这种格式和标准来回复
[15:04] Jonathan Kvarfordt: Opus Clip 与 TikTok 的结合体
Source quotes
[1:41] Jonathan Kvarfordt: if you feel behind, you are, but so is everybody else.
[5:30] Jonathan Kvarfordt: your CRM is probably lying.
[15:04] Jonathan Kvarfordt: Opus Clip and TikTok combined into one
[14:36] Jonathan Kvarfordt: respond exactly like this
Original transcript
[0:04] So, my name is Jonathan. I'll get more
[0:05] into myself in a second. We only have 20
[0:07] minutes. So, I'm going to try and keep
[0:08] going and hopefully this is valuable for
[0:10] you. So, anyone here been burned, had
[0:12] hallucinations, problems, issues with
[0:15] AI? Anyone? I have many times. Um,
[0:19] anyone not who has not ever had a
[0:20] problem with AI? I want to talk to you
[0:22] afterwards. Okay, cool. I just want to
[0:24] make sure. So, um, this happens often.
[0:27] You want to feel like this. You want
[0:29] just blow blow the thing off, right? Um
[0:31] I can't tell how many times I have felt
[0:33] this way many many times. So I just want
[0:34] to take it out. So my name is Jonathan
[0:37] Carford. My last name is obviously hard
[0:39] to spell. So I got a nickname called
[0:41] Coach K because I coach sales teams and
[0:42] CS teams on how to use AI and and sales.
[0:45] Um and at long story short, there was
[0:47] about 10 of us in the same office named
[0:49] John. My name got nicknamed Coach
[0:51] because like the actual Coach K, my last
[0:54] name's hard to say or or pronounce. So I
[0:55] got my nickname is Coach K. So, I am the
[0:57] head of growth at Momentum. I own the
[0:59] what's called the GTM AI Academy. I've
[1:01] had several hundreds, not hundreds, tens
[1:03] of thousands of people go through the
[1:04] academy to learn how to use AI for
[1:06] marketing, sales, CS, etc. And then I do
[1:08] growth with momentum. I was a customer
[1:10] before two years ago and I loved it so
[1:12] much that I literally begged the CEO to
[1:13] let me come on and thankfully let me
[1:15] join and I'm now leading marketing
[1:16] growth for momentum. So, let's get into
[1:19] it. Um, I want to go into four main
[1:21] topics. I want to talk about what I view
[1:23] as the core problem or challenge, the
[1:26] golden data that you probably already
[1:28] have access to. Um, real life examples
[1:31] of what we've seen AI agent be used, and
[1:33] then lastly, the power of a prompt and
[1:35] and hopefully give you a structure you
[1:36] can leverage in whatever technology you
[1:38] might be using. So, everyone see this
[1:41] headline, April?
[1:44] I'm sure everyone's seen this, right?
[1:45] This one headline caused like this
[1:49] ripple effect throughout the industry
[1:50] that I've seen more and more people try
[1:51] to say, "Okay, if they're going to do
[1:53] this, then we should be doing this." And
[1:54] so people have been this freakout mode
[1:56] of trying to figure out how do we do
[1:58] this because we don't want to fall
[1:59] behind. So then you have like Mark Beni
[2:01] off come out and say, "Hey, we have 30
[2:02] to 50% of our workloads being done by
[2:04] AI." You guys seen this?
[2:06] Um, I have a theory behind what's
[2:09] happening here, but still the point is
[2:10] this comes out and more pressure adds
[2:11] on. Then you have this from Dualingo.
[2:14] They talk about, hey, we got to do more
[2:15] with AI, spec specifically this
[2:16] paragraph. If you have not read their
[2:18] updates, you probably should. So, has
[2:21] anyone felt this pressure of like, I
[2:22] gotta get AI in now? Anyone felt this
[2:25] now? I have much very much so.
[2:29] So, there's this feeling as I talk to
[2:30] people, both individuals and companies,
[2:33] that we're falling behind and and that
[2:35] everyone else is ahead of me. Um, anyone
[2:37] feel this way? Okay, good. Now, I did
[2:40] some data because I was curious like, is
[2:41] this the truth? Is it really that
[2:43] everyone else is ahead and you're behind
[2:45] or what what's the deal? So, we took a
[2:48] thousand calls we had this last six
[2:50] months and it ranges from SMB to
[2:52] enterprise. There was um a good mix
[2:54] about 150 different industries and so
[2:56] these are people who came in and talked
[2:57] to us about AI and AI adoption and
[3:00] agents and automation. So, I have this
[3:01] rich data set inside of Momentum that I
[3:03] was like I want to do some research on
[3:04] it. So, we tore it apart and we figured
[3:06] out a couple things. 60% of the people
[3:08] we talk to are way behind like lagging
[3:12] across the board. Didn't matter if it's
[3:14] SMB or enterprise way behind in the
[3:17] times. Next would be like the kind of
[3:19] emerging like they're actually aware of
[3:20] this happening and they want to do
[3:22] something about it and very very few is
[3:24] what I'd call advanced. Even those use
[3:26] cases are still not as advanced as it
[3:29] could be. So what we found was McKenzie
[3:33] and some others talk about that 78% of
[3:36] companies have adopted one use case with
[3:38] AI of some kind. What we found in our
[3:41] data, it's more like 7.6 is operational.
[3:45] So if you feel like you're behind, you
[3:47] are, but so is everybody else. So I want
[3:50] to help hopefully take some pressure off
[3:52] of you that like the reality what's
[3:54] being positioned or marketed is not
[3:56] actually the truth. What's happening
[3:57] internally? And I know that because we
[3:59] sell to people all the time who are not
[4:00] very advanced.
[4:02] Same thing here of like um a lot of
[4:04] people are facing challenges. They
[4:06] describe themselves as like we're
[4:08] struggling with this concept of AI. How
[4:10] do we do this? Um and almost half are
[4:13] still dealing with a lot of manual
[4:15] processes that could be done with some
[4:17] sort of AI agent platform of some kind.
[4:20] So here's another data set. A lot of
[4:22] reasons why projects die is is based on
[4:25] bad data or unclean data or unstructured
[4:28] or not integrated something around data
[4:30] is what's holding back a lot of AI
[4:32] capabilities.
[4:33] So project time is another one about 70
[4:35] to 80% of time is spent on just
[4:38] preparing the data for an AI project and
[4:41] people are organizations are not even
[4:42] ready for the AI in general. So when
[4:45] they come in and talk to us at Momentum
[4:47] I'm asking questions I'm like you're not
[4:49] ready for advanced AI stuff. We got to
[4:51] get the basics in involved in
[4:53] fundamentals.
[4:54] So here is if you have no solid data
[4:57] foundation that means you have very
[4:58] inaccurate insights or very um limited
[5:01] insights which leads to a lot of process
[5:04] breakdowns I've seen throughout sales
[5:05] and customer journeys and then AI is
[5:08] falling short of what it could be valued
[5:10] not because the AI is at fault but
[5:12] because the data is not there to power
[5:13] the AI. That make sense so far? Okay. So
[5:17] without this foundation, you produce
[5:19] inaccurate insights, failing to improve
[5:21] efficiency, and then you buy this really
[5:23] expensive AI tech that you're thinking,
[5:25] well, this is not doing what was sold to
[5:27] me. What's going on?
[5:30] We believe in momentum, the most
[5:32] valuable data you have or information
[5:33] you have at your disposal is the
[5:35] conversations you have every single day
[5:37] with your customers. The challenge is it
[5:40] relies currently at least in the past on
[5:42] human to take that data and dispose it,
[5:45] communicate it, enter into CRM, whatever
[5:46] the case might be. So your CRM is
[5:50] probably lying. Does anyone have a
[5:51] Salesforce or HubSpot or something that
[5:53] they feel like is like really rich of
[5:55] good data?
[5:58] Is there any sales people in here who
[5:59] absolutely love to fill in Salesforce
[6:01] data?
[6:03] Is the data inside of Salesforce
[6:04] important to have? So we have this gap,
[6:07] right?
[6:08] So this is the challenge. If humans are
[6:11] the data source, which we've been
[6:12] expecting to be, we have all sorts of
[6:14] problems. You have a 8,000word
[6:16] transcript that maybe 25 words gets put
[6:18] into Salesforce. When you have that,
[6:20] that means the seuite has no idea what
[6:22] in the world's going on. They have to
[6:23] call their top sales up and be like,
[6:24] "Hey, what's happening with the biggest
[6:25] deal this quarter? I have no idea what's
[6:27] happening." Which obviously slows down
[6:29] the process. Then you have the sales
[6:30] leaders themselves are trying to guess
[6:32] and saying I think we're going to close
[6:33] a million dollars this month but we'll
[6:35] find out when I talk to my sales rep
[6:37] tomorrow and then revops and enablement
[6:40] they can't prove impact they don't know
[6:41] if the sales process is being followed
[6:43] we don't know if the methodology is
[6:44] being adopted we don't know what's
[6:46] happening and then lastly customer
[6:47] insights and the most best information
[6:50] from your ICP and product feedback is
[6:52] gone
[6:54] so this every single team and I'm going
[6:56] to zip through these real quick every
[6:58] team needs this data
[6:59] finance, AEES, BDRs, whether it's
[7:02] themselves or cross communication, they
[7:04] need this information to be able to
[7:06] communicate to each other what's going
[7:07] on. So, you can have this engine be
[7:09] running without you having to have a
[7:10] human get in the way of things. [snorts]
[7:13] for for CIOS and CTO's um it's all this
[7:16] rich data in these different platforms
[7:18] that don't necessarily integrate or talk
[7:20] to each other very well that you can
[7:22] have integrated and communicating so
[7:23] that a tool like momentum can feed that
[7:25] into different places whether that's a
[7:27] human or to a system of record of some
[7:29] kind
[7:30] so I believe the gold mine is your
[7:32] conversational signal there is so much
[7:34] there so whether or not you work with
[7:36] momentum or not I highly suggest you do
[7:38] something with some sort of n make or
[7:40] whatever platform you have to capture
[7:42] more insight from conversations because
[7:44] there is a lot there that you're missing
[7:45] that you could be capitalizing on. So
[7:47] for us, what Momentum does, it's an
[7:49] engine that identifies, validates,
[7:52] cleans, and then moves data so that it's
[7:54] not just this weird random prompt result
[7:56] you're getting from a conversation, but
[7:58] it's actually validated data going into
[8:00] your CRM. She can rely on it. That way
[8:02] you can have like opportunity objects
[8:04] into your Salesforce or opportunities or
[8:06] accounts or leads, whatever the case is.
[8:09] You have the account objects going into
[8:11] Salesforce itself and then you have
[8:13] email or Slack or Teams, whatever your
[8:15] communication platform is. You can
[8:16] automate that process.
[8:18] Um, here I'm going to switch gears just
[8:20] a hair because I think it's important to
[8:21] cover this. Google released a white
[8:23] paper talked about AI agents. They
[8:25] talked about three things you need to
[8:26] have an AI agent. One is an LLM of some
[8:29] kind. U Momentum uses five or six
[8:31] different LLMs from Coher to Claude to
[8:33] OpenAI to whatever the case is. Um, then
[8:36] you have some sort of tool. This an
[8:38] example is Salesforce, but it could be
[8:40] your calendar, could be your email,
[8:41] could be Slack. There's a lot of
[8:42] different tools. And then lastly, you
[8:44] need some sort of orchestration layer.
[8:46] So that is what momentum is at its core.
[8:49] It connects things and moves data and
[8:50] processes around. So here's an example
[8:53] of one thing you could do when you have
[8:55] this data hooked up. You have a call. I
[8:57] know it says call summary, but imagine a
[8:58] call transcript. Momentum will go in and
[9:00] identify what parts of that call is
[9:03] important to know with this subject,
[9:04] which in this example is a churn risk or
[9:06] a KG buyer. Like maybe someone's not
[9:08] really as invested as you thought they
[9:10] would be. Then it can identify are they
[9:13] engaged or not? And if they're not,
[9:14] let's send them into a scoring model of
[9:16] some kind and saying, okay, they're
[9:18] saying some things that's telling us
[9:19] they're not necessarily as involved as
[9:20] we thought they were. And then we need
[9:22] to make sure that the team is aware of
[9:24] this so that we're not spending time on
[9:25] a deal that may not be as real as we
[9:27] thought. Does that make sense so far?
[9:29] Okay, cool. So, another one is like next
[9:32] steps. So, for sales reps, this is
[9:34] really nice so that you don't have to
[9:35] manually think about, okay, what's my
[9:36] next step? Momentum can go in capture
[9:39] the next step from the conversation
[9:40] already put it in there and make sure
[9:41] that it's ready to go. So, the
[9:43] salesperson just says, "Yep, that's
[9:44] correct." And it clicks over to your
[9:46] Slack or Salesforce or calendar,
[9:48] whatever the case might be, which is
[9:50] this here.
[9:52] Um, use case number three is like here's
[9:54] enablement packets. So think about all
[9:56] the information you have in um
[9:59] opportunity of five calls, 20 calls over
[10:02] three months or nine months or whatever
[10:03] the case might be. A lot of information
[10:05] that the sales rep is like required to
[10:07] then communicate to the team. That's a
[10:09] lot of information to do over three
[10:11] months even of time. So what it does
[10:13] instead is we have momentum take all the
[10:15] calls from the last however many times
[10:17] make like a packet that has all the
[10:19] things that the CS team has asked for.
[10:22] It could be very specific to the company
[10:23] whether it's industry specific or
[10:25] persona specific or metrics and then
[10:27] give them that packet so then they can
[10:29] have everything they need so the sales
[10:31] rep is not in the bottleneck anymore and
[10:32] they can have all this information
[10:34] automated either through Slack or
[10:35] against Salesforce or Teams or whatever
[10:37] the communication mechanism is. Um I'll
[10:41] give I think one more. Well, here's one
[10:43] from coaching performance. This is um I
[10:45] came from the enablement coaching world
[10:46] so I love this one. So there's a lot of
[10:48] skills being shown inside of sales calls
[10:51] or customer success calls. It's really
[10:54] hard for an enablement person or a
[10:55] manager to watch every single call and
[10:57] identify this is a really good golden
[10:58] nugget. I need to share this with the
[10:59] team. It's really hard to do that. So
[11:01] what momentum does is do it for them. So
[11:04] we can take that same thing, analyze the
[11:06] same call, say here's where you're
[11:07] strong, here's where you're weak. Um
[11:09] here's some suggestions based on the
[11:10] other five reps in your team of what
[11:12] they're doing well or not. and then
[11:14] gives that coaching to them real time.
[11:15] So you can say like that we know this is
[11:16] working because of the results because
[11:19] we have that correlation with Salesforce
[11:20] data. So we know if like someone has a
[11:22] higher revenue or not and then we can
[11:25] either do one of two things. We can
[11:26] either create the dashboard for you so
[11:28] you can use momentum to see the visual
[11:30] or we feed the data into Snowflake or
[11:32] Salesforce and you can do whatever you
[11:33] want with it. You can categorize it or
[11:36] correlate however you like.
[11:39] So this comes down to what I call the
[11:41] three capabilities. How am I doing on
[11:42] time?
[11:44] We okay? I want to keep my time check.
[11:46] We're good. Okay. Let me know if we want
[11:47] to get to 510ish. So, three
[11:49] capabilities. One is a structured
[11:51] firstparty data provider. So, the
[11:54] information you have needs to be
[11:56] cleaned, validated, and then structured
[11:58] in a way so you go to the right data to
[12:00] the right field, the right information
[12:01] inside your CRM. And that needs to
[12:03] happen across your ent your entire
[12:05] customer set. Then you have the ability
[12:07] to communicate. So, what does a CEO need
[12:10] to know about your deals versus their
[12:12] sales rep? They're way different
[12:13] communication. And so, you need to make
[12:15] sure we can automate and say, "Okay,
[12:16] CEO, you want to know when the deal was
[12:18] $100,000 that may be at risk versus
[12:21] something else like when do you want to
[12:22] know this information? We can automate
[12:24] that or just go down to the team members
[12:26] so we can help them automate their own
[12:27] workflows." Then lastly is the
[12:29] orchestration piece. So, we like to call
[12:31] it business action. Your CRM in my
[12:34] opinion should be used as a business
[12:35] action mechanism. So anything that
[12:38] happens inside of Salesforce should
[12:39] moving the business forward. So momentum
[12:42] literally helps move the business
[12:43] forward faster through business action.
[12:46] So this is what we call a system of
[12:49] action and data. So CRM is a system of
[12:52] record. We are the system of action and
[12:53] data.
[12:55] So you're probably familiar, I'm sure
[12:57] everyone's prompted something here, I'm
[12:59] sure, right? Everyone's prompting
[13:00] something or other. Um, a lot of people
[13:02] think this is what's happening where
[13:03] it's just a transcript, one prompt, you
[13:05] have an output. Ours is a little bit
[13:07] different. We take it very seriously to
[13:08] go through a lot of validation points to
[13:10] make sure there is um again that
[13:12] cleansing validating process is
[13:14] happening so that the data you get
[13:16] inside of your CRM is not junk because
[13:18] the last thing you want is a lot of junk
[13:19] data in your CRM instead of no data at
[13:21] all. So we take it very seriously to
[13:23] make sure we we go through each data
[13:25] point to make sure it's as clean and as
[13:27] valid as possible.
[13:29] So I'm going to switch gears again. Um
[13:30] if you're looking at agents, I do
[13:32] prompting for six or seven different
[13:34] companies. I teach prompting at GTM
[13:36] Academy at Brian University and a couple
[13:38] other places. Um, I love prompting. It's
[13:40] like one of my things I geek out on all
[13:42] the time. And ever since GBT5 came out,
[13:44] I've been studying the GBT5 prompting
[13:46] guide and learning things from them. And
[13:48] so I found like a framework that works
[13:50] pretty well with agents. So again,
[13:51] whether you use our platform or NAN or
[13:53] anything else, this is a good one to
[13:55] use. So it's just called Grace. It's
[13:57] just an acronym for a prompting
[13:59] framework. Um, the goal. I'm sure you
[14:01] guys are probably know um, know this
[14:02] already, but hopefully it'll give you
[14:04] some nuggets. What are you wanting it to
[14:06] achieve? Obviously, the more context the
[14:08] better. Ro or sorry, rules. It says
[14:11] rules, but should say role. My
[14:12] apologies. Who is the AI? What do you
[14:14] want it to conduct? I also find that
[14:16] using several different um, roles. You
[14:19] say like you're a Mckenzie expert and
[14:20] you're a finance expert and combining
[14:22] roles together makes a really good
[14:23] output. You have advanced context. This
[14:26] is where I find most people lack a lot.
[14:29] Thank you. Um, is they lack a lot of
[14:31] context on what in the world is
[14:32] happening inside the system. Then you
[14:34] have instructions. And then lastly is
[14:36] examples. And this is the one I think
[14:37] you should spend a lot of time on. Does
[14:39] the AI know what good looks like? Does
[14:41] it know what bad looks like? And do you
[14:43] have an example template you can give it
[14:45] to say, "Okay, I want you to respond
[14:46] exactly like this." If you do that, it
[14:49] does much better job at not
[14:50] hallucinating because it has something
[14:52] to follow. Is this helpful? Hopefully.
[14:54] Okay, good. I have a few minutes left,
[14:56] so I'm going to go through a couple
[14:57] things pretty quickly. So, I'm going to
[14:58] skip through this because I want to get
[15:00] to some good stuff. So, I'm g skip
[15:02] through that. Thank you. Um, this is
[15:04] what we call invisible agents. Making
[15:06] sure we can have things happen
[15:07] throughout the process without the human
[15:09] having to dictate what's happening as it
[15:11] goes.
[15:12] So, as part of this, my one of my
[15:14] announcements today is we're just
[15:15] released what's called smart clips.
[15:16] Think about Opus Clip and Tik Tok
[15:18] combined into one where we take your
[15:20] videos and auto clip it down. So you
[15:22] decide what moments you want to have and
[15:23] we auto clip it down to a 15 seconds
[15:25] whatever it is to then send to you and
[15:27] then you can use it for whatever you
[15:28] want to use it for which is nice. We
[15:30] have also like account briefs which
[15:32] means that it's kind of like the same
[15:33] theory as the close loss or close one
[15:35] where we just packet something together
[15:36] and so as a deal moves through the
[15:38] opportunities it updates you on all the
[15:41] newest and latest things that's happened
[15:42] through email or through calls or
[15:44] whatever the communication is. [snorts]
[15:46] Um, and then obviously the coaching
[15:47] where we give more of a in-depth
[15:48] dashboard of of what they're what you
[15:50] can see inside of performance. Um,
[15:53] Caroline, are you here?
[15:56] Where is she? Not Carolina. Catalina.
[15:59] Okay. I didn't think she was able to
[16:00] come. So, 11 Labs is one of our
[16:02] customers, and I I was going to have one
[16:04] of them join me today, but I don't see
[16:05] them. So, I think I'm going to pause.
[16:08] So, with that being said, this is a
[16:11] prompt library that I personally wrote
[16:13] 200 pro plus prompts. um because you're
[16:16] all here, please go there. This is all
[16:18] like go to market use cases, customer
[16:20] success, marketing, sales, etc. So, if
[16:22] you want to use that, please enjoy.
[16:24] Hopefully, it helps you in some way with
[16:26] prompts, but there's a ton of good meaty
[16:27] stuff in there. One of my favorites is
[16:29] the competitive battle card. It works
[16:30] awesome on deep research, by the way.
[16:32] So, if you use that, please use deep
[16:33] research. It does a really good job. And
[16:35] I think that's it unless some questions.
[16:38] I got through in time. Be good.
