Zig AI Replaces Fragmented SaaS Stacks With Autonomous Sales Agents

Empowering the Sales Teams of Tomorrow with Org Approved AI Agents With Zig AI — Ai4

Ai4

Sales reps spend 70% of their workweek managing software instead of selling, driving quota attainment down to 16% unless companies replace reactive copilots with governed autonomous agents.

AI-assisted summary of Empowering the Sales Teams of Tomorrow with Org Approved AI Agents With Zig AI — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 17, 2026, this presentation features Steve Ancheta, Founder and CEO of Zig AI. The talk explores how the traditional sales software stack is being superseded by org-approved, AI-native autonomous agents that manage research, outreach, and pipeline workflows within strict enterprise governance boundaries.

## TL;DR Addressing the steep collapse of B2B sales quota attainment caused by SaaS bloat and administrative overhead, Steve Ancheta explains why 95% of generative AI pilots fail and reactive copilots pull reps away from revenue-generating work. He presents the case for 'Org-Approved' autonomous multi-agent systems anchored to a compounding context graph, governed by least-privilege access and human-in-the-loop approval gates. Drawing on Zig AI's forward-deployed deployment framework, the presentation demonstrates how grounding agents in proprietary data transforms reps into managers of virtual employees, validated through an on-stage live voice orchestration demo.

## Atmosphere & style The presentation carries a sharp, authoritative keynote rhythm, grounded in the practical realities and scars of enterprise sales leadership. Ancheta blends sobering statistics on quota failures and data leakage with engaging audience polls, creating an informative yet urgent narrative arc. The session concludes on an energetic, high-tech note with a live voice-agent demonstration that makes the proposed future of autonomous sales operations feel immediate and tangible.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=4s) B2B sales quota attainment has collapsed from 53% in 2012 to just 16%, largely because reps spend most of their time on administrative friction rather than live customer conversations. - [01:58](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=118s) The explosion of fragmented SaaS point solutions forces sales reps to act as manual human middleware, juggling an average of 10 tools and spending 70% of their time managing software. - [04:10](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=250s) Enterprise AI adoption spans four stages from curiosity to maturity, but 95% of GenAI pilots currently deliver zero P&L return due to poor data readiness. - [06:43](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=403s) Shadow AI poses massive data security risks, with 71% of employees admitting to feeding confidential company data into public consumer models, driving average annual insider risk costs to $17.4 million. - [08:30](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=510s) Unlike reactive copilots that require constant prompt engineering, org-approved autonomous agents own end-to-end workflows while operating under strict least-privilege access and mandatory human approval gates. - [10:23](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=623s) High-performing autonomous AI requires co-building a single grounded knowledge layer and compounding context graph rather than deploying generic self-serve chatbots. - [12:10](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=730s) A disciplined four-step implementation framework—structuring data, establishing governance, redesigning workflows, and scaling with champions—enables specialized multi-agent architectures to succeed. - [14:22](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=862s) Sales representatives equipped with mature AI systems are 3.7 times more likely to achieve quota, with data maturity generating up to an $8 return per dollar invested. - [14:22](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=862s) Building internal custom agents often backfires by forcing sales reps into ongoing tool maintenance rather than enabling them to manage virtual employees and focus on closing deals. - [16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) Live multi-agent voice orchestration demonstrates how field reps can effortlessly leverage historical institutional context to draft, approve, and execute follow-ups hands-free.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=4s) The Quota Crisis and the Breakdown of the Modern Sales Rep

Drawing on 16 years of experience across the entire B2B sales spectrum—serving as an account executive, sales engineer, sales manager, head of sales, and founder of two sales-driven companies that were built and sold over the past decade—Steve Ancheta introduces the critical problem facing contemporary revenue teams. Looking at an average sales representative's weekly schedule, only a tiny sliver of time (represented as green) is actually spent having sales conversations and speaking with prospective buyers. Everything else is consumed by administrative overhead and operational friction.

Despite heavy market noise claiming that artificial intelligence is poised to replace sales professionals, initial corporate attempts to lay off sales staff in favor of automation have repeatedly backfired, forcing companies to rehire reps. AI cannot replicate human connection or the nuanced dynamics of interpersonal relationship-building. Instead of addressing the root issue, the tech sector's historical reaction has been to throw more tools or more bodies at sales organizations. This approach has led to a catastrophic decline in sales productivity over the last decade: while 53% of sales reps were hitting quota in 2012, that number plummeted to just 16% nationwide.

- In 2012, 53% of B2B sales reps hit quota; by the previous year, quota attainment collapsed to 16%. - Most of a salesperson's workweek is swallowed by non-selling tasks rather than customer conversations. - Artificial intelligence excels at automating the operational perimeter around a transaction, but cannot replace human-to-human relationship building.

---

[01:58](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=118s) From Spreadsheets to Agents: The Evolution of the Sales Stack

Tracing the history of sales tooling illustrates how incremental technological solutions frequently introduced new operational complications. The industry began in the "spreadsheet era," an early effort at organization that remains an astonishingly durable habit among many reps today because of its sheer comfort and tactical simplicity. That transitioned into the CRM era, moving from customized on-premises systems like Siebel to cloud dominance led by Salesforce.

```mermaid flowchart LR accTitle: The Evolution of the Sales Technology Stack subgraph S1["1990s–2000s"] A["Spreadsheets"] --> B["On-Prem & Cloud CRM<br/>(Siebel, Salesforce)"] end subgraph S2["2010s–2023"] B --> C["The Stack Boom<br/>(Fragmented Point Tools)"] C --> D["Copilots (2023)<br/>(Conversational Prompts)"] end subgraph S3["Present"] D --> E["Agentic Era<br/>(Org-Approved Agents)"] end ```

The cloud CRM gave rise to the "stack boom," characterized by dozens of venture-backed startups addressing hyper-specific, narrow slices of the sales cycle in isolation. By 2023, the industry introduced generative copilots, culminating today in what Ancheta views as one of the most transformative tech shifts in our lifetime: the agentic era. However, the modern enterprise is currently drowning in SaaS sprawl. With roughly 30,000 SaaS products on the market and about 1,000 running in the average enterprise, the only integration layer tying these disparate point solutions together is the human rep.

- Account executives, RevOps specialists, and GTM engineers act as manual human middleware to stitch siloed applications together. - Only about one-third of software tools deployed inside an enterprise are actually integrated into daily operational workflows. - Sales reps juggle an average of 10 distinct software tools to close a single deal, spending 70% of their workweek simply managing and updating their software stack rather than selling.

---

[04:10](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=250s) The AI Adoption Curve and Why 95% of GenAI Pilots Fail

Enterprise AI adoption follows a four-stage progression. It starts at Stage 0, where organizations are merely curious and exploring without financial investment, resulting in a net-zero ROI impact. The danger begins in Stage 1—experimentation—where companies bleed money and destroy ROI trying to parse through vendor noise. As Ancheta tells his team at Zig AI, their greatest competitor is not another startup or vendor, but the sheer market noise confusing enterprise buyers. The objective of true AI-native architectures is guiding enterprises out of Stage 1 experimentation into Stage 2 deployment and Stage 3 scaled maturity.

The urgency of this transition is underscored by the widespread failure of early generative AI deployments: 95% of generative AI pilots produce zero P&L impact, and 60% of all gen AI initiatives are abandoned or slated for abandonment before reaching full operational maturity. The root cause is data readiness. While an estimated 80% to 90% of companies actively dabble with AI, only 7% of organizations possess data that is genuinely "AI-ready." The 5% of implementations that do succeed share clear characteristics: they prioritize deep workflow integration, maintain ruthless data quality, ensure end-user adoption is owned directly by the front lines, and deploy systems that continuously learn from specific human interaction rather than generic web corpora.

- **Stage 0 (Curious):** Zero financial commitment, zero expected ROI, net-zero balance sheet impact. - **Stage 1 (Experimenting):** Unfocused testing, high vendor noise, and negative ROI where most mid-market and enterprise companies are stuck. - **Stage 2 & 3 (Integrated & Mature):** Deep production integration, front-line line ownership, and tangible P&L returns. - **The Readiness Gap:** Only 7% of enterprise data is structured and ready for AI models, explaining why 95% of gen AI pilots deliver zero financial return.

---

[06:43](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=403s) Shadow AI, Data Leakage, and the RFP Anecdote

During an interactive audience poll, nearly the entire room raises their hands when asked if they use AI in their day-to-day work, but virtually every hand drops when asked if that AI delivers the correct output every single time. This disconnect highlights the hazards of "Shadow AI." Approximately 81% of enterprise employees admit to using unauthorized AI tools in their daily jobs, a figure that jumps to 93% among senior executives. Alarmingly, 71% of employees acknowledge pasting confidential, sensitive company data directly into open, consumer-grade models—driving enterprise insider risk costs to an average of $17.4 million annually.

To illustrate the concrete risks of unapproved AI usage, Ancheta shares an industry anecdote regarding a high-stakes government Request for Proposal (RFP). A vendor suspected their fierce, perennial competitor might have fed proprietary bid documentation into a public commercial LLM to generate responses. By strategically prompting the open model, the sales rep successfully reverse-engineered key pricing and technical details of the competitor's submission, allowing them to outmaneuver the rival and win the contract. Whether apocryphal or fully confirmed, the story reflects the severe exposure created when unmanaged data flows into external platforms.

- 81% of employees and 93% of executives admit to leveraging unapproved AI tools in their enterprise workflows. - 71% of personnel have fed sensitive corporate information into public generative models. - Enterprise insider data leakage and associated risk costs average $17.4 million per company per year.

---

[08:30](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=510s) Copilots vs. Org-Approved Autonomous Agents

The fundamental operational distinction between copilots and true agents lies in their interaction model. Copilots are purely reactive—they only "answer when asked." While conversational prompts suit engineers who understand inputs, constraints, and algorithmic arithmetic—where "input plus context equals output"—they run counter to the instinctual rhythm of top-performing sales professionals. Forcing a salesperson to stop talking to clients in order to meticulously engineer prompts for a chat interface pulls them away from pipeline generation and directly saps revenue.

```mermaid flowchart TD accTitle: Copilots vs Org-Approved Autonomous Agents subgraph Reactive["Reactive Copilots"] R1["Rep stops selling to prompt chat"] --> R2["Model answers query"] R2 --> R3["Rep manually executes & pastes"] end subgraph Autonomous["Org-Approved Autonomous Agents"] A1["Agent owns end-to-end goal"] --> A2["Autonomous research, drafting & CRM updates"] A2 --> A3["Least-Privilege Approval Gate"] A3 --> A4["Rep Reviews & Approves Outcome"] A4 --> A5["Continuous System Learning"] end ```

By contrast, an effective AI agent takes full ownership of both the overarching goal and the end-to-end workflow: researching target accounts, drafting tailored communication, updating CRM pipeline stages, and managing follow-ups—a critical area where even seasoned reps routinely drop the ball. Crucially, autonomy does not mean abandoning oversight; human beings must always remain in the loop. Under the "Org-Approved" standard, agents operate with least-privileged system access, every external action requires explicit human sign-off, and all steps are comprehensively logged and fully reversible.

- Copilots require the rep to initiate work via conversational prompts, pulling them away from high-value prospect interactions. - Autonomous agents execute the complete operational loop: account research, email drafting, pipeline updates, and structured follow-up cadences. - "Org-Approved" architecture relies on three non-negotiable governance pillars: least-privilege system permissions, mandatory human approval gates for finished outputs, and fully auditable, reversible action histories.

[10:23](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=623s) Procuring Org-Approved AI and Laying the Grounded Knowledge Layer

When procuring AI-native software, organizations must demand clear operational boundaries before deploying anything into their live environments. Critical questions must be answered: What is the least-privileged access policy? What human approval gates exist? Are actions fully logged and reversible? And how is data grounded, governed, and tied to measurable outcomes? The biggest failure mode among AI companies today is delivering provable, quantifiable results. At Zig AI, contractual commitments are tied directly to hard statistical benchmarks, because clear outcomes are the only way an AI system can reliably understand how to operate.

```mermaid flowchart TD accTitle: Zig AI Forward-Deployed Knowledge Architecture subgraph DataFoundation["Data Foundation & Governance"] A["Fragmented Enterprise Data"] --> B["Forward-Deployed Engineer"] B --> C["Compounding Context Graph<br/>(Governed Knowledge Layer)"] end subgraph ExecutionLayer["Operating System & Agents"] C --> D["Zig AI Operating System"] D --> E["Research Agent"] D --> F["Outreach Agent"] D --> G["Pipeline Agent"] end ```

Agents are only as effective as the ground they walk on. The solution is to co-build a single governed knowledge layer between the enterprise and the AI provider, verifying security postures and testing extensively. Zig AI executes this via a forward-deployed engineering model: an engineer embeds directly into the customer's organization, maps out fragmented data silos, and synthesizes them into a compounding context graph that acts as the company's "AI brain." On top of this foundation, specialized agents—for research, outreach, and pipeline orchestration—are deployed. While this requires a deeper upfront investment than plugging into a simple self-serve chatbot, the compounding returns are massive: within three to six months, the system understands the inner workings of the organization better than most human employees.

---

[12:10](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=730s) The Four-Step Implementation Framework and Multi-Agent Design

Successfully rolling out autonomous systems requires a disciplined four-step implementation process:

1. **Structure the Data:** Map and aggregate all internal information assets. 2. **Set the Governance Bar:** Determine explicit boundaries regarding what data is exposed and what remains restricted. 3. **Redesign Workflows:** Realign operational paths around how data natively flows through the system. 4. **Prove and Scale:** Validate the model with a dedicated champion group, iterating constantly through feedback loops and on-the-ground human interviews with revenue-generating reps who know the end-to-end customer journey firsthand.

Once the underlying plumbing is established, agents transition from novelty features into engines solving core organizational problems. A compounding context graph extends far beyond sales, unlocking automated workflows across RevOps, HR, and payroll. In terms of agent architecture, a single agent assigned to tackle multiple disconnected tasks performs poorly. High-performing systems rely on a multi-agent framework where each distinct workflow is assigned a specialized agent operating under one-tap approval gates, complete audit logging, and reversible action histories.

---

[14:22](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=862s) The Changing Sales Landscape and the Pitfalls of Building In-House

Sales organizations equipped with grounded AI show dramatic performance advantages, with reps operating alongside AI becoming 3.7 times more likely to hit their quotas and high data maturity yielding an estimated $8 return for every $1 invested. As repetitive administrative burdens shift from humans to agents, the rep's work week flips: CRM records update autonomously, sales cycles compress, and institutional intelligence compounds over time.

| Metric / Trend | Projected Impact | |:--- |:--- | | **Quota Attainment Multiplier** | Reps selling with AI are **3.7x** more likely to hit quota | | **Data Maturity ROI** | Yields **$8 return** per $1 spent | | **Agent-to-Rep Ratio (by 2028)** | AI agents will outnumber human sales reps **10 to 1** | | **AI-Mediated B2B Purchases** | **$15 Trillion** in B2B transactions mediated by AI | | **Current Brand Adoption** | **60%** of brands actively using agentic AI for 1:1 interactions |

In the enterprise debate between buying dedicated AI platforms versus building internal agent tooling, custom in-house solutions often backfire. Companies that instruct sales teams to build and script their own agents inevitably force reps right back into the manual overhead of tool maintenance. When the foundational plumbing is managed correctly by a dedicated platform, reps are freed to act as managers of virtual employees, prioritizing relationship building, live conversations, and closing deals.

---

[16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) Live Voice Demonstration: Managing a Field Agent at AI4

Demonstrating real-time voice orchestration on-stage at the AI4 conference, Steve Ancheta executes an end-to-end conference follow-up entirely hands-free:

- Steve activates the agent via voice: *"Hey, I just ran into Carlos over at Sinch here at AI4. Wanted to send him an email so we can connect next week."* - The agent immediately references historical institutional intelligence: *"Tony reached out to Sinch three months back and their Gong renewal is at year-end. That's right around the corner, so timing's perfect. I'll pull our Gong intel into the follow-up."* - When Steve asks to verify that his meeting link is included, the agent confirms it is already attached and presents the completed draft for authorization. - Upon Steve's verbal confirmation (*"Yeah, go ahead and send it"*), the agent dispatches the email and commits to monitoring the thread: *"Sent. Want me to call you the moment Carlos replies?"* - Steve confirms, and the agent schedules the real-time phone call trigger before closing: *"Have a good rest of your day at AI4, Steve."*

## Notable quotes > [08:30](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=510s) input plus context equals output — Steve Ancheta

> [08:30](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=510s) answer when asked — Steve Ancheta

> [10:23](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=623s) Agents are only as effective as the ground they walk on. — Steve Ancheta

> [16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) Hey, I just ran into Carlos over at Sinch here at AI4. Wanted to send him an email so we can connect next week. — Steve Ancheta

> [16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) Tony reached out to Sinch three months back and their Gong renewal is at year-end. That's right around the corner, so timing's perfect. I'll pull our Gong intel into the follow-up. — Zig AI Agent

> [16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) Yeah, go ahead and send it — Steve Ancheta

> [16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) Sent. Want me to call you the moment Carlos replies? — Zig AI Agent

> [16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) Have a good rest of your day at AI4, Steve. — Zig AI Agent

## Mentioned Steve Ancheta, Zig AI, AI4, Salesforce, Siebel, Sinch, Gong, Carlos, Tony

## Video description Steve Ancheta - Founder and CEO, Zig AI

The Sales Software stack as we know it is being replaced. The next generation of sales teams won't toggle between a dozen SaaS tools, they'll operate through a single AI-native system where autonomous agents handle research, outreach, follow-up, and pipeline management while compounding institutional knowledge with every interaction. I'll explore why org-approved agents — ones that act independently but operate within defined boundaries, are the architecture separating high-performing teams from everyone else, and share what we've proven at zig.ai building in this space. If you're a sales leader wondering why your current stack feels like it's working against you, this talk will show you what's already replacing it.

[Source](https://www.youtube.com/watch?v=7MXJHMxwI3w)

中文简报

Zig AI以自主销售智能体取代碎片化SaaS技术栈 销售代表每周将70%的时间花在管理软件而非销售上，导致业绩达标率跌至16%，除非企业用受治理的自主智能体取代被动式Copilot。 针对SaaS膨胀与行政开销导致的B2B销售业绩达标率骤降问题，Steve Ancheta解释了为何95%的生成式AI试点项目失败，以及被动式Copilot为何反而使销售代表偏离创收工作。他论证了引入“企业认可”的自主多智能体系统的必要性——该系统立足于持续累积的上下文图谱，并受最小权限访问和人在回路审批关卡的严格治理。结合Zig AI的前沿部署框架，本次演讲展示了将智能体植根于专有数据如何使销售代表转型为虚拟员工的管理者，并通过现场实时语音编排演示进行了验证。 B2B 销售的业绩达标率已从 2012 年的 53% 暴跌至仅 16%，销售代表平均需要频繁切换 10 款 SaaS 工具，将 70% 的工作时间耗费在软件管理等行政琐事而非客户沟通上。 (0:04) 目前 95% 的生成式 AI 试点项目未能带来实际 P&L 收益，且 60% 的项目在落地前遭搁置，核心症结在于仅 7% 的企业数据真正达到了 AI 就绪标准。 (4:10) 影子 AI（Shadow AI）带来巨大数据泄露风险，71% 的员工承认曾将公司机密数据输入公共消费级模型，导致企业内部人风险平均每年造成 1740 万美元的损失。 (6:43) 相较于打断销售节奏且需要频繁编写提示词的被动 Copilot，获得企业批准的自主 Agent 能够在最小权限与强制人工审核的管控下，全权接管从客户调研到 CRM 更新的端到端闭环。 (8:30) 配备成熟 AI 系统的销售代表实现业绩达标的概率可提高 3.7 倍，而高水平的数据成熟度预计能实现每投入 1 美元换取 8 美元回报的商业效益。 (14:22)

Reading summary

Steve Ancheta of Zig AI argues that the collapse in B2B sales quota attainment stems from software sprawl and administrative friction, which can only be reversed by replacing reactive copilots with governed, autonomous multi-agent systems grounded in enterprise data. The Quota Collapse and Enterprise Software Sprawl (0:04) Ancheta reports that B2B sales quota attainment fell from 53% in 2012 to 16% nationwide, driven not by a decline in sales talent but by an accumulation of administrative overhead. Looking at average weekly schedules, sales representatives spend approximately 70% of their time updating and managing software stacks rather than speaking with prospective buyers. This operational burden reflects severe software sprawl: the average enterprise operates around 1,000 SaaS products, and reps routinely juggle roughly 10 distinct point solutions to advance a single deal. Because only about one-third of deployed enterprise software is integrated into daily workflows, sales representatives, RevOps teams, and go-to-market engineers effectively serve as manual middleware to connect disconnected tools. Data Readiness Deficits and Shadow AI Risks (4:10) Enterprise artificial intelligence deployments face acute delivery challenges: 95% of generative AI pilots produce zero P&L impact, and 60% are abandoned before reaching operational maturity. Ancheta attributes this failure rate to data hygiene; while an estimated 80% to 90% of organizations experiment with AI, only 7% possess data that is properly structured and ready for model ingestion. In the absence of functional enterprise solutions, employees turn to unauthorized tools. Approximately 81% of staff and 93% of executives admit to using unapproved AI in their daily workflows, with 71% acknowledging they paste confidential corporate data into public consumer models. This data exposure contributes to enterprise insider risk costs averaging $17.4 million annually, illustrated by an industry incident where proprietary bid documentation entered into a public model enabled a competitor to reverse-engineer and win a government RFP. Reactive Copilots Versus Org-Approved Autonomous Agents (8:30) Ancheta distinguishes between reactive copilots and autonomous agents based on workflow ownership. Copilots require users to initiate prompts, forcing sales reps to stop communicating with prospects to engineer queries. Conversely, effective agents take end-to-end ownership of specific operational goals, including account research, draft generation, CRM updates, and structured follow-up cadences. To manage enterprise risk, Ancheta outlines an 'Org-Approved' governance framework founded on three pillars: least-privileged system permissions, mandatory human approval gates before executing external actions, and fully auditable, reversible action histories. High-performing deployments structure these workflows across specialized multi-agent teams rather than relying on a single generalized model. Context Graph Architecture and Implementation Discipline (10:23) Deploying dependable agents requires establishing a grounded enterprise knowledge layer. Zig AI implements this through forward-deployed engineers who map internal data silos into a compounding context graph that captures institutional intelligence over three to six months. Rollouts follow a four-step framework: structuring raw data, defining governance boundaries, redesigning workflows around native data flows, and scaling alongside front-line champion reps. Ancheta cautions against directing internal sales teams to build custom agent tooling in-house, noting that it forces representatives back into maintenance overhead. According to presented metrics, sales reps operating with grounded AI are 3.7 times more likely to hit quota, while mature data foundations yield an estimated $8 return for every dollar spent. Ancheta anticipates that AI agents will outnumber human sales reps ten-to-one by 2028 as AI mediates a projected $15 trillion in B2B transactions, though realizing this shift depends on whether enterprises can move past superficial experimentation to resolve data hygiene and governance constraints.

精炼阅读摘要

Zig AI 创始人 Steve Ancheta 指出，现代 B2B 销售正遭遇配额达成率暴跌与软件工具泛滥的双重危机，绝大多数生成式 AI 试点也因缺乏就绪数据而陷入负回报困境。企业唯有摒弃被动交互的 Copilot，通过驻场工程建立严格治理的专用知识底座，转向经组织批准（Org-Approved）且受人类监督的自主多智能体，才能真正释放一线销售生产力。 销售配额危机与软件堆栈碎片化困局 (0:04) 依据 Ancheta 跨越 16 年的 B2B 销售及两度创业退出的行业经验，销售代表每周的时间日程中，仅有极小一部分用于直接与潜在买家沟通，其余时间均被管理冗余与操作摩擦吞噬。过去十年间，科技行业习惯于通过堆砌工具或增加人手来应对增长压力，但这直接导致生产力急剧崩溃：2012 年全美尚有 53% 的 B2B 销售代表能够完成业绩配额，而至去年这一比例断崖式跌至 16%。早期部分企业试图以自动化裁撤销售人员的做法屡屡受挫，最终不得不重新雇佣，事实证明 AI 无法复制人与人之间的情感连接与复杂关系网络。 从早期电子表格、本地化及云端 CRM，发展到针对销售细分环节的单点工具大爆发，再到 2023 年的生成式 Copilot，销售技术栈在演进中造成了严重的 SaaS 蔓延。市场上约有 3 万款 SaaS 产品，普通大型企业内部运行着约 1000 款工具，但其中仅约三分之一真正融入日常业务流。销售代表为了促成单笔交易平均需要切换 10 款独立软件，将每周多达 70% 的工作时间耗费在更新与管理系统上，不得不充当维系孤岛系统的“人工中间件”。 生成式试点失败根源与影子 AI 安全隐患 (4:10) 企业 AI 采纳通常经历四个阶段：从零资金投入的阶段 0（好奇探索），进入大量企业被供应商营销噪音围困的阶段 1（盲目实验），最终走向深入生产的阶段 2（部署集成）与阶段 3（规模化成熟）。Ancheta 披露，当前有 95% 的生成式 AI 试点项目未能对企业的损益表（P&L）产生任何实质影响，且多达 60% 的 AI 项目在成熟前即被废弃或计划废弃。其根本症结在于企业数据的准备程度：尽管有 80% 到 90% 的企业在涉足 AI，但全行业仅有 7% 的企业数据达到了供模型有效运转的“AI 就绪”状态。 在成效停滞的同时，“影子 AI”（Shadow AI）正引发严重的数据泄露风险。现场调查与统计数据显示，81% 的普通员工和高达 93% 的高管承认在日常工作中使用未经合规授权的 AI 工具，71% 的人员曾直接将企业保密数据复制粘贴至公共消费级模型中，使企业平均每年承受约 1740 万美元的内部风险成本。Ancheta 引述了一起行业案例：某供应商推测其竞争对手可能将高额政府招标提案（RFP）草案输入了公共商业大模型，销售代表通过针对性提示词反向推导出了竞对的关键报价与技术细节，最终抢下订单，揭示出未受控数据外流的致命隐患。 被动副驾驶与组织批准型自主智能体之辨 (8:30) 当前普遍应用的 Copilot（副驾驶）本质上属于被动响应式工具，遵循“有问才答”的逻辑。对话式提示词机制契合习惯“输入加上下文等于输出”的软件工程师，却严重违背顶尖销售人员的直觉节律；强制要求销售人员停止客户交流去精心编写 Prompt，本质上是在消耗核心营收机会。相反，真正的自主智能体能够全流程接管目标导向的复杂工作，自主完成目标客户调研、个性化沟通草拟、CRM 阶段同步以及日常销售极易遗漏的后续跟进。 驻场知识图谱搭建与多智能体分工实施框架 (10:23) 为确保系统安全性，Ancheta 提出了“组织批准”（Org-Approved）的三大治理支柱：系统执行最小权限访问策略、所有外部交互与成稿必须设立人类强制审批门控（Human-in-the-loop），以及全链路记录具备完全可逆性的审计日志。智能体的有效性完全取决于其底座知识库的质量。Zig AI 采用驻场部署工程师模式，直接进驻企业打破数据孤岛，构建具备复利效应的“上下文图谱”（Compounding Context Graph），并在其上并行部署调研、外联与管线管理等多款专用智能体，而非依赖难以胜任多任务的单一模型。 针对自主系统的落地，Ancheta 总结了四步实施法：首先结构化映射企业内部所有数据资产，其次设立明确的数据暴露与权限隔离红线，随后围绕原生数据流重构业务路径，最终通过由一线绩优销售组成的核心试点团队进行验证与迭代推广。他特别警告企业切勿要求内部销售团队自行编写智能体脚本，这只会让员工重新沦为工具维护者；只有将底层技术外包给专业平台，团队才能作为“虚拟员工的管理者”专注于建立客户信任。 销售人机配比预期与语音驱动自动化现场验证 (14:22) 依据 Ancheta 引用的行业数据，借助成熟 AI 辅助的销售代表达成业绩配额的可能性高出 3.7 倍，高数据成熟度的企业在 AI 领域每投入 1 美元可产生 8 美元回报。其行业预测认为，至 2028 年 AI 智能体与人类销售代表的比例将达到 10 比 1，并有 15 万亿美元的 B2B 交易将由 AI 介导参与。目前已有 60% 的品牌在利用智能体开展一对一交互。 在 AI4 大会现场，Ancheta 通过纯语音交互展示了移动端代理调度流程：销售通过口述告知智能体刚在会场接触了 Sinch 公司的客户 Carlos 并计划下周联络；智能体瞬间调取组织历史情报，提示同事 Tony 三个月前曾接触该公司且其核心软件 Gong 的续约期就在年底，因此契机极佳；在核对日历邀请链接已附入邮件后，销售口头确认发送，智能体随即执行派发并主动建立监测机制，承诺在客户回复的瞬间向销售发起即时电话提醒，完成了全流程闭环。 Steve Ancheta 展示了以治理规范与企业私有知识底座驱动自主智能体解决销售内耗的路径，但在企业广泛存在数据极度分散与系统割裂的现状下，依赖驻场工程师深度定制的交付模式其规模化推广成本仍有待检验；此外，关于数年内 AI 将介导 15 万亿美元交易及智能体人机比达 10:1 的长期预测，在复杂且高度依赖人际博弈的重大采购中能否如期全面实现，仍面临市场与技术的持续考验。

Detailed notes — zh

## Background 本视频由 Ai4 频道于 2026 年 8 月 17 日发布在 YouTube 平台上。内容为 Zig AI 创始人兼 CEO Steve Ancheta 的主题演讲，主要探讨传统销售软件栈正在被组织级批准的原生 AI 智能体系统所取代，以及企业如何通过自主智能体实现背景调研、自动化触达和商机管线管理。 ## TL;DR Steve Ancheta 结合自身 16 年 B2B 销售与创业经验，深度剖析了销售团队因工具碎片化导致业绩达标率暴跌的现状，以及 95% 的企业 GenAI 试点因数据未就绪而失败的根源。他对比了被动式 Copilot 与主动闭环 Agent 的本质差异，提出了以最小权限、人工审批关卡和全量可逆审计为核心的“组织级批准智能体”标准。此外，他详细阐述了通过前向部署构建复合上下文图谱的落地四步法与多智能体分工原则，并通过现场语音交互演示了智能体自主关联历史情报、起草邮件与电话跟进的全流程闭环。 ## Atmosphere & style 整场演讲务实干练、直击痛点，主讲人以资深销售老兵的视角打破行业炒作，兼具犀利的行业洞察与扎实的工程治理思维。现场互动热烈，既有对“影子 AI”与盲目自建陷阱的警示，也有对数据底座建设的严谨剖析。最后的实时语音 Demo 克服了网络挑战，流畅展现了人机协同的未来形态，给观众带来了强烈且真实的科技落地冲击感与信服力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=4s) AI 的核心价值在于处理围绕交易的辅助繁杂事务，无法也不可能取代销售人员与客户之间真实的人际信任与商业情感连接。 - [00:04](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=4s) 销售技术栈的过度堆叠导致效率下滑，全美销售代表达成业绩指标（Quota）的比例从 2012 年的 53% 骤降至去年的 16%。 - [03:15](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=195s) 企业软件极度碎片化，大型企业平均运行上千款 SaaS 却仅有 1/3 真正集成，导致销售 70% 的工作时间浪费在系统维护与跨工具操作上。 - [03:15](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=195s) 企业 AI 采纳经历四个阶段（好奇探索、盲目实验、定向落地、全面成熟自治），当前绝大多数企业停留在阶段 1 且持续流失 ROI。 - [05:24](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=324s) 高达 95% 的 GenAI 试点未能改善企业损益，核心瓶颈在于全美仅 7% 的企业内部数据达到了“AI 准备就绪”的标准。 - [05:24](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=324s) 工具体验不佳催生了庞大的“影子 AI”安全隐患，超过 80% 的员工和 90% 的高管曾使用未经批准的 AI 工具并输入敏感数据。 - [08:30](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=510s) Copilot 属于被动应答且依赖销售编写 Prompt，而 Agent 则由目标驱动并在后台自主完成背景调研、撰写草案与跟进回访。 - [08:30](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=510s) 组织级批准的 AI Agent（Org Approved Agent）必须严守三大红线：最小权限原则、人类把关审批机制、全量操作可审计与可逆。 - [10:23](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=623s) 智能体的表现取决于底层数据地基，Zig 采用前向部署工程师深入企业构建复合上下文图谱（Compounding Context Graph）。 - [12:10](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=730s) 落地 AI 智能体应遵循四步法：数据结构化、确立治理红线、重塑业务工作流，以及先锋小组验证后规模化推广。 - [12:10](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=730s) 在架构设计上应避免单一全能智能体，推崇“单工作流单智能体”的多智能体分布式协作分工体系。 - [14:22](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=862s) AI 赋能的销售代表业绩达标率是普通代表的 3.7 倍，AE 的角色正从兼职数据录入员转变为管理虚拟员工的交易促成者。 - [14:22](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=862s) 大企业盲目自建 AI 智能体容易陷入维护陷阱，导致销售人员从“点状工具使用者”沦为“智能体代码维护工”。 - [16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) 现场演示展示了语音驱动的多智能体协同：系统秒级调取客户历史交互与竞品情报，并在销售人员语音确认后自主发送邮件并设置回信电话提醒。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=4s) 销售困境与销售软件的技术演进史 演讲者 Steve Ancheta（Zig AI 创始人兼 CEO）开门见山地介绍了自己的背景。在过去的 16 年里，他深耕 B2B 销售领域，历任客户专员（Account Executive）、销售工程师（Sales Engineer）、销售经理以及销售负责人（Head of Sales），并在过去 10 年里创办并成功出售了两家重度依赖销售驱动的公司。他指出，目前行业中充斥着“AI 将彻底取代销售人员”的论调，甚至有企业因 AI 裁减销售团队后又不得不重新招人的闹剧。AI 的核心优势在于处理“围绕交易（Surround the sale）”的繁杂辅助事务，但它根本无法、也永远无法取代人与人之间真实的商业情感连接和信任建立。 当前销售团队面临着严峻的效能危机。在普通销售代表的工作周内，仅有一小部分时间被真正用于同客户沟通与签单，绝大部分精力都被工具管理与杂务吞噬。技术供应商多年来试图通过不断增加新工具来解决问题，反而导致问题层层加剧：2012 年全美有 53% 的销售代表能够完成业绩指标（Quota），而到去年，这一比例已暴跌至 16%。回顾销售技术的演变历程，行业经历了从最原始的电子表格时代（Spreadsheet Era，至今仍有许多销售人员因习惯而固守此道）、到以 Siebel 为代表的本地部署 CRM，再到以 Salesforce 为代表的云端 CRM 时代。随后，市场迎来了垂直单点工具层出不穷的“销售技术栈大爆发（Stack boom）”；进入 2023 年后，行业迅速转向 Copilot（副驾驶助手），如今则正式迈入基于 Agent（智能体）的全新时代。 ```mermaid flowchart LR accTitle: Sales Technology Evolution subgraph S1["工具与流程演进"] A["电子表格时代<br/>(本地表格记账)"] --> B["CRM 兴起<br/>(Siebel / 本地定制)"] B --> C["云端 CRM 时代<br/>(Salesforce 崛起)"] C --> D["技术栈大爆发<br/>(垂直单点 SaaS)"] D --> E["Copilot 时代<br/>(2023 对话交互)"] E --> F["Agent 时代<br/>(自主智能体协作)"] end ``` - **16 年 B2B 销售老兵**：历任 AE、SE、销售经理及销售负责人，创办并退出两家销售驱动型企业。 - **业绩达标率滑坡**：全美销售代表达成 Quota 的比例从 2012 年的 53% 骤降至去年的 16%。 - **时间分配失衡**：销售人员每周仅有极少量时间专注于向客户卖货，其余时间都在应付各类周边工具与杂务。 --- ### [03:15](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=195s) 技术栈碎片化与企业 AI 落地的四阶段 当前企业软件市场极度碎片化，全市场约有 30,000 款企业级 SaaS 产品，平均每家大型企业内部运行着 1,000 款工具。然而，企业中唯一的“集成层”竟然是活生生的人类员工——收入运营人员（RevOps）、AE、GTM 工程师不得不充当“胶水”，手动将各系统拼凑在一起。在企业内部，仅有约三分之一的软件完成了真正的日常流程集成；销售人员为了推动单笔交易的关闭，平均需要跨越使用 10 种工具，导致整周高达 70% 的时间被浪费在系统维护与跨工具操作上。 在构建 Zig 的过程中，团队归纳出了企业 AI 采纳的四个演进阶段： 1. **阶段 0（好奇探索，Curiosity）**：企业对 AI 保持好奇，没有实际预算投入，属于净零影响（Net Zero）。 2. **阶段 1（盲目实验，Experimenting）**：绝大多数中型市场与企业级客户目前所处的阶段。企业不断尝试新工具，在此阶段由于缺乏明确方向而持续流失 ROI。市面上最大的竞争对手不是竞品，而是铺天盖地的“市场噪音（Noise）”。 3. **阶段 2（深度落地与赋能）**：通过真正具有差异化的原生 AI 方案切入业务，帮助企业扭亏为盈。 4. **阶段 3（完全成熟自治）**：组织级 Approved AI Agent 体系完全成型。 ```mermaid flowchart TD accTitle: Stages of AI Adoption subgraph Stages["AI 采纳四阶段模型"] S0["阶段 0：好奇探索<br/>零投入 / 零 ROI / 净零影响"] --> S1["阶段 1：盲目实验<br/>大量中大企业所处阶段<br/>工具噪音多 / ROI 持续流失"] S1 --> S2["阶段 2：定向落地<br/>原生 AI 系统切入业务核心"] S2 --> S3["阶段 3：全面成熟<br/>组织级合规智能体协同运转"] end ``` - **企业 SaaS 泛滥**：市场上约 30,000 款 SaaS，平均每家大型企业内部运行 1,000 款，但仅 1/3 实现日常整合。 - **工具切换成本高昂**：销售代表关闭一笔交易平均需使用 10 款工具，70% 的工作时间被用于管理系统而非面对客户。 --- ### [05:24](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=324s) 95% 生成式 AI 试点失败的根源与“影子 AI”的安全黑洞 行业数据显示，高达 95% 的生成式 AI（GenAI）企业试点项目未能为公司带来任何实质性的损益表（P&L）改善，60% 的项目在走向成熟前就已被放弃或搁置。这一现象的核心根源在于：全美仅有 7% 的企业其内部数据真正达到了“AI 准备就绪（AI Ready）”的标准。绝大部分公司虽然嘴上谈论或部署了 AI，其数据层却是一片混乱。要跻身那真正取得成效的 5%，必须满足四个先决条件：深度集成、数据质量第一、业务一线主导落地，以及建立能通过人类日常互动行为进行持续反馈学习的闭环系统。通用大模型是基于公开网页和文献训练的，只懂得抽象的“大众标准”，完全不理解具体某家企业内部独特的“优质业务标准”。 在现场互动中，Steve 询问台下观众有多少人在日常工作中使用 AI，几乎全场举手；但当问及“有多少人觉得 AI 每次给出的产出都是完全准确的”时，举手者寥寥无几。这种工具的不顺手催生了庞大的“影子 AI”风险： - **81% 的员工**承认在工作中使用未经公司批准授权（Unapproved）的 AI 工具。 - **93% 的高管**同样存在类似的不合规使用倾向。 - **71% 的员工**曾将公司敏感数据直接输入公共大模型。 - **内部风险成本**：单家大型企业每年因内部数据泄露和不规范行为面临的平均潜在风险成本高达 1,740 万美元。 Steve 分享了一个流传于业内的惊人轶事：两家企业竞标一份政府 RFP（招标需求书），其中一方的销售代表敏锐地推测老对手可能会将敏感的招标文件塞给未受监管的公共大模型辅助撰写，于是该销售通过对开源模型进行逆向信息提取，成功套取了对手关于该 RFP 的核心策略细节并一举赢下单子。虽然故事真伪尚存争议，但这鲜活地反映了企业敏感数据外泄所带来的现实商业威胁。 --- ### [08:30](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=510s) Copilot 与 Agent 的本质分水岭及“组织级批准”规范 深入剖析交互机制可以发现，Copilot 与 Agent 之间存在不可逾越的鸿沟： - **Copilot 模式（被动应答）**：类似于 ChatGPT、Claude 或 Gemini，必须由人类主动向其提问或输入提示词才会反应。AI 的底层逻辑非常工程师化，遵循极简算术公式：`输入 (Input) + 上下文 (Context) = 输出 (Output)`。然而销售人员并非软件工程师，让销售人员停下手中与客户的真实交谈，去思考如何编写精确的 Prompt，这本质上是把能用来创造营收的宝贵时间白白浪费掉。 - **Agent 模式（目标驱动与闭环执行）**：理想的 Agent 能够端到端接管销售流程中的业务目标，在后台自主完成背景调研、撰写草稿、更新系统数据以及跟进回访（Follow-up）。销售人员往往忙于建立人际关系与追逐新商机，常常疏于机械性的后续跟进，而 Agent 恰好能够补齐这块短板。 ```mermaid flowchart TD accTitle: Copilot vs Agent Workflow subgraph CopilotFlow["被动式 Copilot 逻辑"] C1["人类销售中断业务沟通"] --> C2["手动编写输入与上下文"] C2 --> C3["Copilot 生成单次回答"] end subgraph AgentFlow["自主闭环 Agent 逻辑"] A1["设定销售目标 / 触发事件"] --> A2["Agent 自主推进调研与草拟"] A2 --> A3["Agent 自动化跟进与系统更新"] A3 --> A4["人类在环 (HITL) 审查与一键批准"] end ``` Steve 特别强调，真正的企业级智能体必须坚守“人类在环（Human-in-the-loop）”底线，AI 负责产出成型的交付物，由销售人员进行最终的审查与批准（Approve）。所谓“组织级批准的 AI Agent（Org Approved AI Agent）”，必须严格遵循三大工程红线： 1. **最小权限原则（Least Privilege）**：严密限制智能体的数据访问与执行边界。 2. **人类审批关卡（Human Approval Gates）**：关键决策与对外沟通必须由人类把关确认。 3. **全量审计与可逆性（Logged and Reversible）**：智能体采取的每一个步骤都必须完整记录在日志中，并且所有操作必须可撤销、可回滚。 ### [10:23](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=623s) 采购企业级 AI 智能体的核心考量与 Zig 的前向部署模式 企业在采购 AI 原生软件时，必须在将系统接入内部环境之前厘清一系列关键问题。企业不能只看演示效果，而要严查软件的安全与治理标准：系统是否遵循最小权限访问原则？提供怎样的由人类把关的审批机制？哪些操作被完整记录、哪些动作可以撤销？数据是否经过严格治理与事实锚定？以及最关键的——交付成果如何量化评估？当前 AI 行业最大的痛点就是无法交付可衡量的业务结果，而 Zig 在创立之初便坚持以可验证的统计数据为导向，甚至将量化业务指标直接写入客户合同。在早期这或许伴随着风险，但这是让 AI 真正具备可用性的唯一途径。 正如 Steve Ancheta 所强调的："智能体的表现上限，完全取决于它们所立足的地基（Agents are only as good as the ground they walk on）"。正确的落地方式并非让用户直接在一个通用聊天框里提问，而是由企业与 AI 原生团队在短时间内共同构建一个受治理的统一知识层。Zig 采用前向部署（Forward Deployed）模式，直接派遣工程师深入企业内部，梳理散落在各处的碎片化数据，构建一个不断复合增长的组织上下文图谱（Compounding Context Graph），使其成为整套系统的 AI 大脑。随后在其上叠加操作系统，并部署专属的调研、触达及商机管线智能体。这种深度部署虽然需要花费更多前期时间，但自我学习和迭代的速度会呈指数级提升，通常在 3 到 6 个月内，系统对组织运转规律的理解就会超越大多数人类员工。 ```mermaid flowchart TD accTitle: Zig AI 部署与多智能体架构流程图 subgraph DataLayer ["1. 底层数据与知识治理"] A["碎片化企业数据"] --> B["前向部署工程师梳理"] B --> C["复合上下文图谱\n(Compounding Context Graph)"] C --> D["安全与治理准入规则"] end subgraph AgentLayer ["2. 专用多智能体协作层"] D --> E1["调研智能体\n(Research Agent)"] D --> E2["触达智能体\n(Outreach Agent)"] D --> E3["管线管理智能体\n(Pipeline Agent)"] end subgraph ActionLayer ["3. 业务执行与闭环"] E1 & E2 & E3 --> F["一键审批 / 人工确认"] F --> G["自动化业务执行与日志记录"] G --> H["跨部门复用 (RevOps / HR / 运营)"] end ``` --- ### [12:10](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=730s) 智能体落地的四步法与单一任务多智能体分工 无论企业选择与哪家供应商合作，成功实施 AI 智能体都应遵循清晰的四步方法论：首先完成数据结构化，其次确立数据治理红线（明确哪些数据可对外暴露、哪些坚决隔离），第三步围绕新的数据流动路径重构业务工作流，最后通过小范围验证后再全面推广扩展。在验证阶段，企业应组建一个由内部骨干组成的先锋群组（Champion Group），让系统在小规模实战中不断迭代并向人类学习。Zig 甚至会安排深度的员工面对面访谈——这种方式听起来虽然传统，但效果极佳，因为一线销售人员最清楚客户从踏入大门到最终离开的全流程细节。 当企业的数据底座与管线彻底铺设完毕后，智能体就不再只是零散的产品功能点，而是转变为解决实际组织痛点的生产力工具。一旦上下文图谱建立起来，这套系统便能从销售部门顺畅扩展至运营、收入运营（RevOps）、人力资源与薪酬管理等跨职能场景。在架构设计上，系统必须做到合规可控：提供一键式批准机制、所有操作全面可追溯且可逆，并在受治理的数据闭环中实现自我学习与自主驱动。针对业内关于“单一全能智能体与多智能体”的争论，Steve 明确指出，单个智能体虽然理论上可以处理多项任务，但实际效果往往不尽如人意；真正高效的架构应坚持“每个工作流配备一个专属智能体”，由多个各司其职的智能体协同完成特定业务动作。 - **实施四步法**：数据结构化 $\rightarrow$ 建立治理红线 $\rightarrow$ 重塑工作流 $\rightarrow$ 小范围验证并规模化推广。 - **治理与控制设计**：一键授权审批、全流程操作日志记录、支持动作即时撤销、基于受控数据的持续自我强化。 - **架构原则**：反对“单智能体全包”模式，推崇“单工作流单智能体”的多智能体分布式协同。 --- ### [14:22](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=862s) 销售生产力的范式颠覆与“自建还是采购”的陷阱 AI 正在重塑销售团队的生产力格局。数据表明，借助 AI 开展销售的业务代表完成业绩指标（Quota）的概率是普通代表的 3.7 倍；而在数据成熟度上的每 1 美元投入，能够带来高达 8 美元的回报。引入 AI 智能体后，销售人员的一周工作节奏彻底翻转：原本被录入 CRM、搜索资料等繁重事务所占据的时间被完全释放，CRM 系统实现自动记录与更新，销售周期大幅缩短，企业数据资产实现自动复合累积。据行业预测，到 2028 年，AI 智能体的数量将以 10 比 1 的比例超越人类销售人员；在全球 15 万亿美元的 B2B 采购交易中，AI 智能体将深度参与其撮合与流转；而在当下，已有 60% 的品牌在利用代理型 AI 进行一对一的客户互动。 ``` 传统销售模式： [ 关系沟通 ] [ 大量繁琐的手动数据录入与跨系统切换 ] AI 原生销售模式： [ 深度客户关系经营与关键谈判 ] ▲ │ (完全由智能体自主支撑) [ 自动化调研 / 智能触达 / 实时上下文整合 / CRM 自动同步 ] ``` 在这一变革下，未来的客户经理（AE）不再是兼职的数据录入员，而是转变为“以关系和谈判为核心的交易促成者”，其本质是管理多名虚拟数字员工的管理者。针对大企业中普遍存在的“自建还是采购（Build vs. Buy）”之争，Steve 提出了明确警告。许多大型机构凭借雄厚的人力与资金选择自主研发，但现实中往往演变成灾难：销售团队不仅没有被解放，反而不得不花费与以往操作点状工具相同的时间去维护和修补自建的智能体代码。企业必须从底层搭建标准化的管线，才能让智能体真正实现自主运行，避免沦为新的维护负担。 - **销售效能倍增**：AI 赋能代表达标率高出 3.7 倍，数据治理投入产出比达 1:8。 - **2028 年行业预测**：智能体与人类销售比例达 10:1，深度介入 15 万亿美元 B2B 采购市场。 - **自建陷阱**：盲目自建容易让销售人员从“工具使用者”沦为“智能体维护工”，违背降本增效初衷。 --- ### [16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) 现场实操演示：语音驱动的智能体协作与即时执行 在演讲尾声，Steve 在现场网络信号受限的挑战下进行了实时语音交互演示，生动展示了未来销售人员管理虚拟员工的日常场景。Steve 唤醒手机上的语音助手，告知自己在 AI4 大会现场刚偶遇了来自 Sinch 公司的 Carlos，希望能发送一封跟进邮件以便下周建立联系。系统不仅秒级响应，还主动调取了底层知识库中的关键业务情报进行补充提醒：“Tony 在 3 个月前曾联系过 Sinch，而且他们的 Gong 合约将在年底到期，眼下正是切入的最佳时机，我会把针对 Gong 的竞争情报整合进这封跟进邮件中。” 整个交互过程极为流畅自然。智能体在数秒内拟定好邮件草案，并主动确认是否附上了 Steve 的个人会议预约链接。在得到 Steve 的语音确认后，智能体迅速完成发送，并主动询问：“需要我在 Carlos 回复的第一时间直接打电话通知你吗？”在得到肯定答复后，智能体承诺将在收到回信的瞬间通过电话提醒 Steve，并礼貌地祝愿他在 AI4 大会度过愉快的一天。这一全自动、有上下文感知、且完全在授权范围内行动的闭环，赢得了全场热烈的掌声。 - **即时上下文提取**：智能体自动关联历史跟进记录（Tony 于 3 个月前触达）与商业情报（Sinch 的 Gong 竞品合约年底到期）。 - **人机协同闭环**：语音指令下发 $\rightarrow$ 智能草拟与竞品要点整合 $\rightarrow$ 会议链接自动挂载 $\rightarrow$ 人工一键语音授权发送 $\rightarrow$ 智能体承诺回信时主动电话提醒。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=4s) AI 的核心优势在于处理“围绕交易（Surround the sale）”的繁杂辅助事务，但它根本无法、也永远无法取代人与人之间真实的商业情感连接和信任建立。 — Steve Ancheta > [03:15](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=195s) 市面上最大的竞争对手不是竞品，而是铺天盖地的“市场噪音（Noise）”。 — Steve Ancheta > [05:24](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=324s) 通用大模型是基于公开网页和文献训练的，只懂得抽象的“大众标准”，完全不理解具体某家企业内部独特的“优质业务标准”。 — Steve Ancheta > [08:30](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=510s) 让销售人员停下手中与客户的真实交谈，去思考如何编写精确的 Prompt，这本质上是把能用来创造营收的宝贵时间白白浪费掉。 — Steve Ancheta > [10:23](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=623s) 智能体的表现上限，完全取决于它们所立足的地基（Agents are only as good as the ground they walk on）。 — Steve Ancheta > [12:10](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=730s) 真正高效的架构应坚持“每个工作流配备一个专属智能体”，由多个各司其职的智能体协同完成特定业务动作。 — Steve Ancheta > [14:22](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=862s) 未来的客户经理（AE）不再是兼职的数据录入员，而是转变为“以关系和谈判为核心的交易促成者”，其本质是管理多名虚拟数字员工的管理者。 — Steve Ancheta > [16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) Tony 在 3 个月前曾联系过 Sinch，而且他们的 Gong 合约将在年底到期，眼下正是切入的最佳时机，我会把针对 Gong 的竞争情报整合进这封跟进邮件中。 — AI 智能体 > [16:51](https://www.youtube.com/watch?v=7MXJHMxwI3w&t=1011s) 需要我在 Carlos 回复的第一时间直接打电话通知你吗？ — AI 智能体 ## Mentioned Steve Ancheta, Zig AI, Ai4, Siebel, Salesforce, ChatGPT, Claude, Gemini, Sinch, Gong, Carlos, Tony ## Video description Steve Ancheta - Founder and CEO, Zig AI The Sales Software stack as we know it is being replaced. The next generation of sales teams won't toggle between a dozen SaaS tools, they'll operate through a single AI-native system where autonomous agents handle research, outreach, follow-up, and pipeline management while compounding institutional knowledge with every interaction. I'll explore why org-approved agents — ones that act independently but operate within defined boundaries, are the architecture separating high-performing teams from everyone else, and share what we've proven at zig.ai building in this space. If you're a sales leader wondering why your current stack feels like it's working against you, this talk will show you what's already replacing it. [Source](https://www.youtube.com/watch?v=7MXJHMxwI3w)

Source quotes

[0:04] Steve Ancheta: AI 的核心优势在于处理“围绕交易（Surround the sale）”的繁杂辅助事务，但它根本无法、也永远无法取代人与人之间真实的商业情感连接和信任建立。 [3:15] Steve Ancheta: 市面上最大的竞争对手不是竞品，而是铺天盖地的“市场噪音（Noise）”。 [5:24] Steve Ancheta: 通用大模型是基于公开网页和文献训练的，只懂得抽象的“大众标准”，完全不理解具体某家企业内部独特的“优质业务标准”。 [8:30] Steve Ancheta: 让销售人员停下手中与客户的真实交谈，去思考如何编写精确的 Prompt，这本质上是把能用来创造营收的宝贵时间白白浪费掉。 [10:23] Steve Ancheta: 智能体的表现上限，完全取决于它们所立足的地基（Agents are only as good as the ground they walk on）。 [12:10] Steve Ancheta: 真正高效的架构应坚持“每个工作流配备一个专属智能体”，由多个各司其职的智能体协同完成特定业务动作。 [14:22] Steve Ancheta: 未来的客户经理（AE）不再是兼职的数据录入员，而是转变为“以关系和谈判为核心的交易促成者”，其本质是管理多名虚拟数字员工的管理者。 [16:51] AI 智能体: Tony 在 3 个月前曾联系过 Sinch，而且他们的 Gong 合约将在年底到期，眼下正是切入的最佳时机，我会把针对 Gong 的竞争情报整合进这封跟进邮件中。 [16:51] AI 智能体: 需要我在 Carlos 回复的第一时间直接打电话通知你吗？

Source quotes

[8:30] Steve Ancheta: input plus context equals output [8:30] Steve Ancheta: answer when asked [10:23] Steve Ancheta: Agents are only as effective as the ground they walk on. [16:51] Steve Ancheta: Hey, I just ran into Carlos over at Sinch here at AI4. Wanted to send him an email so we can connect next week. [16:51] Zig AI Agent: Tony reached out to Sinch three months back and their Gong renewal is at year-end. That's right around the corner, so timing's perfect. I'll pull our Gong intel into the follow-up. [16:51] Steve Ancheta: Yeah, go ahead and send it [16:51] Zig AI Agent: Sent. Want me to call you the moment Carlos replies? [16:51] Zig AI Agent: Have a good rest of your day at AI4, Steve.

Original transcript

[0:04] Everybody welcome over these next 20 minutes.  We're going to go over a topic that I'm [0:10] extremely passionate about a little bit background  about myself. I've been in B2B sales for 16 years [0:16] held positions as account executive sales engineer  sales manager head of sales and a founded two [0:22] companies are very sales driven and sold both of  those in the last 10 years a lot in a very little [0:27] bit of time, but that's how most sales people  work. So if you're if you're in sales you kind [0:31] of understand. And what we're going to talk about  is going to be the way that we're going to empower [0:36] the sales teams of tomorrow by utilizing AI  agents or approved AI agent specifically. [0:42] So let's talk about the problem and kind of what  it looks like today. On the left what you'll see [0:47] is this is your average sales reps week. The green  is how much time they spend actually selling and [0:54] talking to customers. Everything else is well  everything else. What we found is that you know [1:01] you you probably hear a lot of noise in the market  around you know AI is replacing sales people [1:07] you know companies are starting to do layoffs  because of AI in their sales organization and [1:10] then they end up hiring people back and  it's really because of this AI is really [1:14] really good at things that surround the sale.  What it can't figure out nor will it is how to [1:20] make human connection and understand kind of that  person-to-person relationship building aspect. But [1:26] what the technology market has done up until this  point is they have thrown tools at the problem [1:31] either tools or personnel. And that's caused an  issue that's compounded over the last couple years [1:35] which is 53% of sales reps were hitting quota  in 2012 while 16% were hitting quota across the [1:42] country as of last year. And so as technology has  progressed they have tried to iterate and make the [1:51] sales person's life a little bit easier but it's  really caused more issue. It's created solutions [1:58] in certain aspects and caused issues in others.  And so, if you go way back, right? There was [2:03] the spreadsheet era, right? This was kind of the  first step toward organization. And funny enough, [2:08] the way that a lot of sales people still  work today, which is crazy, but it's true. [2:12] Um it's a very comfortable way of keeping records  from a day-to-day standpoint. [snorts] Then [2:18] the CRM was born, right? With uh solutions like  Siebel, right? Um a lot of on-prem solutions, [2:24] customized solutions. And then Salesforce came  to the market and started taking that market [2:29] share. So, this cloud the cloud CRM. Then what  happened shortly after that was the stack boom. [2:34] The stack boom got really interesting because in  the stack boom, you started seeing these companies [2:39] that solved very small pieces of the problem, but  only those pieces of the problem. And we'll get [2:44] into the reason why that is in a little bit, but  it's a little bit interesting. Something that I [2:48] didn't know until I got really deep into building  products. So, after that, we saw copilots in 2023. [2:54] And then today is where we stand in the agentic  era, right? The agentic era, in my opinion, [2:59] is arguably one of the most fascinating booms  in technology that we will have in our lifetime. [3:04] And the reason why is because agents, when built  correctly, can operate very close to a human [3:10] to make their life easier. But the key is when  they're built correctly. And it all comes down to [3:15] one thing, which is the data. So, right now, there  are roughly 30,000 SaaS products on the market uh [3:22] enterprise-wide. And a thousand of those  run inside of the average enterprise, right? [3:27] So, there's only one integration layer, and  it's a human being. Uh RevOps does this, [3:32] account executives do this, GTM engineers do this,  but they piece these tools together in the way [3:37] that they work at least in an effective fashion  for the organization itself. The problem with that [3:43] is that roughly 1/3 of those tools inside of an  enterprise are actually integrated in a day-to-day [3:48] uh in the day-to-day aspect, right? The  average is about 10 tools that are used [3:53] uh by reps to close a single deal and 70% of the  week goes towards managing those tools. Not really [4:00] effective when sales folks are paid on merit as  well as they're not the cheapest employees in [4:05] the world and it's a very competitive market where  they're able to go find jobs pretty easily, right? [4:10] So, what we found through our research as we were  building Zig is we found that there is four stages [4:15] to AI adoption. This is kind of where everybody  lives. There's stage zero, which is a organization [4:20] is curious around AI. Um this has really no  dollar impact because you're you're you're [4:25] not even testing yet. You're kind of just seeing  what's out there. You're not spending a dollar [4:29] so you're not you don't expect any ROI at this  stage. So there's really a net zero impact there. [4:34] When you start experimenting, that's where  you start losing you start losing ROI. [4:38] So in the experimenting phase, you're trying to  figure out what works and there's so many things [4:42] what what I tell my team at Zig is that our number  one competition is not actually competitors on the [4:47] market. Our number one competition is noise. Um  and the only way that you could be successful in [4:52] today's market is you break through that noise  and have a differentiation that is valuable [4:56] to the organizations that we serve. And so the  experimenting piece is where you're losing money. [5:00] Most organizations live here right now. Stage  one is where you see a lot of organizations [5:07] in in today's market specifically in the  mid-market enterprise space, you see them in [5:10] stage one. Our goal and the goal of AI native  companies that do in my opinion it correctly [5:18] is to get an organization from stage one to stage  two and then eventually to stage three, right? [5:24] So, let's talk about why generative AI. I'm  sure all of you have seen this report that [5:29] came out that 95% of generative AI pilots were  failing, right? Let's figure out why, right? [5:35] So, 95% of gen AI pilots deliver absolutely no  P&L impact and 60% of those initiatives have [5:41] either been abandoned or will be abandoned  before they see their complete maturity, [5:46] right? And the reason is is because right now,  based on studies, only 7% of organizations across [5:51] the country are actually a their data is AI ready.  So, there's there's 100% I wouldn't say 100%, [5:58] but I would say 80 to 90% of companies that are  actually utilizing AI in some way, shape, or form, [6:03] but they have zero AI data readiness. And so, this  is what separates the other 5%. This is my thesis [6:09] at least, and you you could say that I'm wrong,  but so far it's it's proven in the wild successful [6:14] is that what separates the other  5% is deep integration, right? [6:18] You have to put data quality first. Adoption is  owned by the line, and you have to build a system [6:25] that continuously learns, right? And it has to  learn from human behavior and human interaction, [6:29] because every organization runs a little bit  differently, and AI is essentially trained [6:35] on literature as its source material and web  traffic, right? And so, when you're doing that, [6:39] it has a very generalized sense of what good  looks like, but it's very different than from an [6:43] organizational sense of what good looks like. So,  show of hands really quick, how many in the room [6:49] have used some form of AI in their day-to-day  when it comes to doing the their job? Great. [6:56] How many feel like it gives you the right output  every single time that you use it? Show of hands. [7:02] Yeah, there it is. And so, I didn't know if that  was going to work, and I'm really glad it did. Um [7:08] Uh so, 81% of your employees are  already using unapproved AI today. Um [7:14] 71% have fed it sensitive data, right? I I I  have a funny story that I I read. Whether the [7:20] story is true or not, I don't know, but I thought  I read it a couple times, so I'm hoping it's true. [7:24] Um or else you guys are going to fact-check me  after this, and then come up to my booth and [7:28] tell me that I was wrong. But, there  is there was a report that came out, [7:31] and the report said that there was a competitor  who was bidding on a government uh RFP, [7:36] and they were bidding against another competitor  that they were always competing against in RFPs. [7:40] And the sales wrapper, whoever it was, wanted to  see if they can extract the information around the [7:47] RFP. They had a hunch that their competitor  may have fed their RFP into a open model. [7:53] And they did, and they didn't get every piece of  information, but they got enough to kind of know [7:57] strategically where they should be, and they won  the deal. And that story came out. Now, again, [8:01] whether that story is true or not, I don't know.  Um but I thought it was very interesting because [8:05] I've heard kind of I've heard horror stories at  a at a much smaller sense. So, what's happening [8:10] right now is 81% of employees have admitted to  using AI that's not approved. 93% of executives, [8:17] 71% have put sensitive data, and the average  annual cost of insider risk is roughly $17.4 [8:23] million per enterprise, right? Not small numbers.  So, here's the biggest difference between agents [8:30] and co-pilots. Co-pilots will answer when asked.  So, think of your gen your generative model, your [8:36] ChatGPTs, your Clauds, your Geminis, you know,  you choose. Those are Those are more co-pilots. [8:40] You ask it. The problem with sales people is that  sales people don't think the way engineers do. [8:46] AI was built very engineer-friendly. A- A- and AI  is not this magic wand that you that you um you [8:53] wave over something. AI is very simple arithmetic.  It is input plus context equals output, right? [9:01] So, the better the input, the better the context,  the better the output. Engineers can think in [9:04] that fashion. Salespeople can't. And And not  that they can't, I shouldn't say that. It's [9:08] It's not part of their natural workflow. Their  natural workflow is I need to work the way that [9:13] I do because I'm good at what I do and the way  that I do it. I need something that is going to be [9:16] side-by-side with me and learn from my process and  continue to iterate and get better. The problem [9:22] is is when you ask a salesperson to stop having  conversations so that it can now talk to an AI, [9:28] that is minutes and hours that they are spent away  from actually driving revenue. So, when AI that [9:34] works is it owns the goal and the entire workflow  in in the sales uh in the sales conversation, [9:39] it researches, it drafts, it updates, it follows  up, right? I'm a I'm a sales person. I like to [9:44] think of myself as a good sales person. I'm still  pretty bad at follow-ups manually. I I think most [9:49] sales people can admit because we're always so  busy. We're always looking for kind of that next [9:53] relationship, that next deal and you know, the the  pipeline report what whatever is on your plate. [9:58] Um the Here's the biggest thing is a human  should always be in the loop. So a a reviews it [10:04] and they approve and then the value is in the  finished outcomes on a per team basis, right? [10:12] I thought my clicker broke. Um Orga Approve means  least privileged human approval gates and every [10:17] action is logged and reversible. So in order  for you to build Orga Approve AI agents or to [10:23] buy Orga Approve AI agents and I will say this, if  you're procuring some sort of AI native software [10:27] currently, you should ask these questions, right?  You want these answers before you deploy anything [10:32] into your environment, right? What type What is  the least privileged access? What type of human [10:36] approval gates do you provide? What is logged and  what's reversible, right? What is grounded and [10:41] governed data and what is measured, right? What  is measurable outcome look like? Because right [10:46] now the biggest issue that AI companies have is  actually delivering measurable outcomes. When we [10:51] built Zig, we were very adamant that we know what  good looks like. We put statistics behind it. So [10:56] when we deliver, that's our measurable outcomes  and that we that we put in our contracts to our [11:00] customers. And whether risky or not, right? Early  days it probably was, we did it because it's the [11:05] only way that AI actually knows how to operate. So  this is what it looks like, right? This is how you [11:10] do it correctly. Agents are only as good as the  ground they walk on. So you give them one governed [11:15] knowledge layer that the organization and the AI  native company have co-built together in a short [11:20] period of time, have tested, have proven out, have  checked the security postures. Then you deploy [11:25] things like research agents and outreach agents  and pipeline agents. So the way that we do this [11:29] at Zig is we do this in a forward deployed fashion  where we send an engineer into the organization, [11:35] they look at all the fragmented data wherever it  lives, they build a compounding context graph, [11:39] which becomes the AI brain around the entire  foundation, and then we layer our operating [11:44] system and our and then deploy our agents into  that operating system. When done correctly, [11:48] it is a very effective way to adopt AI. Takes  a little bit longer, it's not just, you know, [11:52] a self-serve, jump on a chatbot and, you know,  ask it a question, but it but the outcomes and the [11:58] self-learning aspect become quicker and quicker  and it iterates faster and faster. And by month, [12:04] you know, 3 to 6, the system knows the  organization better than most humans do. [12:10] So, these are the simple four steps, right? And  whether you do it with companies like ours or you [12:14] do it with any other company, these are the four  steps that they should follow. It is structuring [12:18] the data, you set the governance bar, what data  do you want to expose, what data don't you, [12:23] you redesign your workflows around the way that  data now moves, right? And then you prove it, [12:28] and then you scale, right? So, you prove it  with a small group inside of the organization, [12:31] call that the champion group, you iterate and  iterate and iterate, it learns from those humans, [12:35] right? We even do as far as human I'm sorry, human  interviews, which sounds archaic, but it actually [12:40] works really, really well, because they know  your organization at the ground level better than [12:43] anybody else. The sales people who are driving  the revenue understand how the organization runs [12:50] from the first stop of a customer coming in  the door to their last stop before they leave. [12:56] So, once the plumbing is laid, then agents stop  being features and they start solving actual [13:01] organizational problems. The really cool thing  about when data is structured is there are so [13:06] many use cases inside of an organization that  go way past sales, operations, rev ops, HR, [13:12] payroll, things like that can all now  be layered in there slowly but surely, [13:16] because you've now built that foundational context  graph inside of the organization that continues to [13:20] compound and learn from new behavior. And so, it's  org approved by design, so we a one-tap approval. [13:28] Some companies do multiple, you know, uh uh an  input type of approval where you actually tell the [13:33] agent this is approved to go out. Logged, visible,  reversible uh actions, right? Learning from the [13:39] governed data, so you want a system that's going  to be self-learning and self-driving, right? Um [13:44] because as it learns more and more, it's going  to find the best ways and the the least path of [13:49] resistance to move at a faster rate than  any human could. And then, it runs the [13:54] same playbook that you just saw, right? So, the  data foundation first, the governance built in, [13:58] and then one agent per workflow. Right now,  there's this kind of debate inside of the AI [14:03] world is can you deploy a single agent to do  multiple tasks? The answer is yes, the answer [14:07] is also is not well. So, it's best to be able to  do a multi-agent type of um deployment where each [14:14] agent is responsible for one or a small group  of actions that are relative to the org itself. [14:22] So, this is what the future looks like. Rep  selling with AI are right now 3.7 times more [14:26] likely to hit quota, and data maturity goes toward  that $8 per $1 spent, right? And this is what it [14:34] looks like. The week flips, right? The CRM writes  itself. So, all of that green that you saw or all [14:38] the white that you saw now turns green, that green  turns white, right? It it's a much more effective [14:43] type of week for the human seller. Cycles  compress, the data compounds. I it's it it becomes [14:48] this kind of natural progression of the entire  organization when the piping is laid properly. [14:54] So, by 2028 agents will outnumber human sellers  10 to 1. And and intermediate $15 trillion in [15:02] B2B purchases will will be made with AI agents  having some sort of hand in that transaction. [15:08] And 60% of brands right now are using agentic  AI for one-to-one customer interactions. [15:13] That's that's today. That's today, and  that's what the next 2 years looks like. [15:18] So, tomorrow sales teams won't just use AI, [15:21] they'll actually manage them, right? This is This  is my theory. I've been in sales for a long time, [15:26] and I think what the future account executive  looks like is they're not this kind of data [15:30] entry plus relationship. They are relationship  first, conversation closing, and everything else [15:36] that surrounds those conversations will be done  by agents that allow them to move faster, allow [15:41] them to get in front of more customers, have more  meaningful conversations for the organization, [15:45] and drive pipeline. Now, this isn't a very quick  fix. It's not something that you could just spin [15:51] up tomorrow and it's going to work. It's something  that has to be worked through in a very methodical [15:55] format, right? For For it to be successful,  and most importantly for it to be scalable, [16:01] it has to be built at the foundation correctly.  One of the biggest debates that I have heard in [16:05] the enterprise space with enterprise customers  that we've talked to is buy versus build, [16:09] right? A lot of organizations have gone down this  build path because they have the capital, they [16:14] have the resources, they have the relationships to  be able to build these things out internally. What [16:19] they have found though, and we just had I just had  a conversation uh earlier today about that, is um [16:25] this very large organization had all of their  sales teams build their own agents, and now [16:30] they're spending just as much time as they were  in point tools having to maintain their agents so [16:35] that they continue to work in an effective format.  That's not what you want your sales people doing. [16:40] So, if you lay the piping correctly, then it all  becomes autonomous for them, and they become a [16:45] manager of virtual employees, essentially. And  I think Well, I actually did well on time today. [16:51] So, I'm going to show you a very, very quick,  right? Hopefully it works. There's There's a [16:55] little Reception issues are horrible in here, but  I'm going to show you a quick kind of example of [17:01] what managing an agent would look like.  Let's see here. Hopefully this picks it up. [17:10] Hi Steve, good to hear from you. How can  I help? Hey, I just ran into Carlos over [17:15] at Cinch here at AI 4. Um wanted to send  him an email so we can connect next week. [17:21] Carlos, yeah. I can set that up. Oh,  actually one thing worth knowing. [17:26] Tony reached out to Synch 3 months back and  their Gong renewal is at year end. That's [17:32] right around the corner so timing's perfect.  I'll pull our Gong intel into the follow up. [17:40] Awesome. [17:44] Draft ready when you are? [17:45] Yeah, go ahead and send it. Can you make  sure my meeting link's on there, please? [17:50] Yeah, already attached. Want  me to send it? Yeah, please. [17:54] Sent. Want me to call you the moment Carlos  replies? Yeah, if you don't mind that'd be great. [18:01] Perfect. I'll call you the moment he's back. [18:05] Have a good rest of your day at AI4, Steve.  Thank you. Thank you, everybody. [applause]