MXD Process Boosts Sales Conversion 10 Points via Low-Cost AI Tools

Steel and Strategy: A No-Hype Playbook for Selling Smarter with AI With MXD Process — Ai4

Ai4

Industrial equipment maker MXD Process increased lead conversions to 27.4% and cut quoting cycles by treating artificial intelligence as structured workflow automation rather than buying overhyped enterprise software.

AI-assisted summary of Steel and Strategy: A No-Hype Playbook for Selling Smarter with AI With MXD Process — original source. Check the original for full context.

Source published: 2025-08-28

Detailed notes

## Background Published on YouTube by Ai4 on August 28, 2025, this presentation features Eric Seiberling, VP of Sales and Marketing at MXD Process. According to the page description, the session provides a brutally practical, no-hype playbook detailing how a traditional industrial manufacturing company applied a crawl-walk-run framework to integrate AI into sales workflows, shorten deal cycles, and boost pipeline without large IT budgets.

## TL;DR Drawing lessons from the 1999 dot-com crash, Seiberling details how MXD Process resolved severe lead routing and quoting bottlenecks by taking a disciplined, operational approach to AI. The company first optimized its existing HubSpot and CallRail tech stack before deploying custom GPT assistants—such as 'Arlin'—to augment a lone application engineer and streamline complex technical proposals. Over an 18-month rollout, MXD Process reduced quoting turnaround times from five days to 1.5 days, lifted conversion rates from 17.4% to 27.4%, and increased average deal sizes by 20%.

## Atmosphere & style The talk strikes a refreshingly candid, 'savagely pragmatic' tone, driven by Seiberling's grounded delivery as an experienced commercial operator rather than a starry-eyed tech evangelist. The narrative moves engagingly from cautionary historical tales like Pets.com to humorous office lore involving Dungeons & Dragons and 'intern-level' AI blunders. It feels like an authentic, roll-up-your-sleeves briefing from an industrial manufacturing leader, demystifying enterprise AI into practical, bite-sized workflow experiments.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=X0snP9C8vFM&t=4s) The generative AI hype closely mirrors the 1999 dot-com boom, where technology was mistaken for an overnight panacea rather than a tool requiring execution, infrastructure, and operational maturity. - [03:06](https://www.youtube.com/watch?v=X0snP9C8vFM&t=186s) Mid-market and traditional industrial companies need a 'savagely pragmatic' mindset to extract commercial returns without large IT budgets or multi-million-dollar tech investments. - [05:54](https://www.youtube.com/watch?v=X0snP9C8vFM&t=354s) Adopting a disciplined 'crawl, walk, run' deployment framework enables teams to run low-risk operational experiments, discard failures quickly, and scale proven workflows. - [07:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=464s) Upfront business process mapping must precede any technology deployment to expose friction points ('torch points') where automation will generate genuine commercial leverage. - [07:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=464s) AI should fundamentally be understood as workflow automation—computers executing structured processes and deterministic decisions based on clear business rules. - [10:33](https://www.youtube.com/watch?v=X0snP9C8vFM&t=633s) Before buying new standalone AI software, organizations should 'sweat the assets' by exhausting the native automation capabilities of platforms they already pay for. - [10:33](https://www.youtube.com/watch?v=X0snP9C8vFM&t=633s) Drastically improving responsiveness—such as reducing lead routing to 3 minutes and outreach to under 24 hours—provides a decisive competitive advantage in complex B2B sales. - [16:18](https://www.youtube.com/watch?v=X0snP9C8vFM&t=978s) LLMs function like 'the smartest, dumbest interns': highly capable conceptually, but prone to failure in specialized industrial domains unless bound by strict instructions and human supervision. - [17:20](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1040s) Targeted domain-specific assistants, built using simple Custom GPTs and internal engineering documents, can drastically compress quoting bottlenecks without enterprise IT overhead. - [18:20](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1100s) Lighthearted change management, such as playfully naming AI agents after Dungeons & Dragons characters, demystifies the technology and reduces employee adoption anxiety. - [19:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1184s) Commercially hyped AI platforms often fail in niche contexts; processes should be vetted manually end-to-end before connecting them via workflow tools like Make.com with a human in the loop. - [20:28](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1228s) Pragmatic AI implementation delivered measurable bottom-line gains at MXD Process, lifting lead-to-sale conversion from 17.4% to 27.4% and increasing average deal size by 20%. - [20:28](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1228s) Management's core role is shifting from direct task delegation to coaching team members on how to delegate effectively to AI models across progressive maturity stages. - [21:28](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1288s) Generative AI acts as a great equalizer across industries—elevating weak performers to mediocre and turning strong performers into powerhouses—making early, iterative adoption imperative.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=X0snP9C8vFM&t=4s) The 1999 Dot-Com Boom and AI Déjà Vu

Reflecting on the parallels between modern generative AI hype and past tech cycles, the presentation opens with a trip back to the summer of 1999 during the peak of the dot-com boom. Back then, common wisdom dictated that the internet would change everything instantly: physical brick-and-mortar storefronts would disappear overnight, every consumer purchase would move online, and automakers would sell entirely direct-to-consumer. During that period, Eric Seiberling was a brand manager at Procter & Gamble (P&G) launching one of the company's earliest internal web platforms—a direct-to-consumer e-commerce initiative for a juice-dispensing appliance that had failed to secure retail shelf distribution through Walmart. The team was convinced they were pioneering direct one-to-one consumer relationships that would mint endless revenue.

```mermaid flowchart TD accTitle: Technology Hype Cycle vs Operational Reality subgraph Hype["1999 Dot-Com Boom & AI Era"] A["Sky-High Expectations 'Everything changes overnight'"] --> B["Aggressive Marketing Blitz (e.g., Pets.com 70% brand awareness)"] end subgraph Reality["Operational Reality"] B --> C["Execution Bottlenecks (Logistics, personalization, support)"] C --> D["99% Dot-Com Failure Rate"] D --> E["20-Year Maturation & Adaptation (BOPIS, Instacart, Pragmatic Tools)"] end ```

The quintessential symbol of that era's exuberance was Pets.com. Founded in 1998, the company surged into public consciousness with massive ad buys across the Macy's Thanksgiving Day Parade, Super Bowl commercials, and *Good Morning America*, achieving a staggering 70% brand awareness in just four months. Yet, within a short span, Pets.com vanished along with roughly 99% of its contemporary dot-com startups—its former web address surviving today merely as a URL redirect to PetSmart. The failure stemmed from the fact that executing e-commerce—managing complex logistics, running customer call centers, and delivering personalization—was brutally difficult in 1999, requiring over two decades of infrastructure maturation to truly master. While physical retail adapted rather than perished (evolving through curbside pickup, Instacart, and specialized independent shops), the lesson remains that emerging technology is ultimately a tool rather than an overnight panacea. For business leaders witnessing today's generative AI frenzy, the experience feels like pure déjà vu.

---

[03:06](https://www.youtube.com/watch?v=X0snP9C8vFM&t=186s) A Savagely Pragmatic Mindset at MXD Process

Taking what he describes as a "savagely pragmatic perspective," Eric emphasizes that AI is a tool capable of delivering immense organizational capability, but it will not magically overhaul business operations overnight—especially for mid-market companies unable to pour millions into speculative development. Rather than adopting breathless tech-evangelism, non-technical commercial leaders must step back, evaluate operational realities, and identify practical, stepwise methods to implement and scale tools that generate genuine commercial returns.

Eric's perspective is informed by a 25-year career across sales and marketing. After beginning at Procter & Gamble, he held roles at engineering software giant Dassault Systèmes and spent several years at Amazon Web Services (AWS) before transitioning to become the Vice President of Sales and Marketing at MXD Process. Transitioning from enterprise tech giants to an industrial manufacturing business brought immediate operational constraints: there was zero dedicated software development budget, and the internal technical development team effectively consisted of "me, myself, and I." However, the agility of a lean organization meant decisions could be made and executed without bureaucratic friction.

```mermaid flowchart LR accTitle: MXD Process Organizational Profile subgraph Org["MXD Process Profile (83 Total Employees)"] direction TB P["50+ Shop Floor Production"] FO["Front Office / Sales & Marketing"] IT["2 IT Staff Total • 1 ERP Maintenance • 1 Network & PC Support"] end subgraph Business["Core Manufacturing"] M["Custom Process Equipment • Stainless Steel Tanks • Industrial Mixers • Automation Control Systems"] end Org --- Business ```

MXD Process is a specialized industrial equipment manufacturer that engineers and builds stainless steel tanks, mixing systems, and automated controls for liquid processing lines across diverse sectors, ranging from food and beverage (e.g., commercial tomato sauce) to heavy industrial chemicals (e.g., epoxy, caulking, and coatings). The domain demands deep technical expertise in chemical engineering and computational fluid dynamics (CFD). The company operates across three physical facilities with a total headcount of 83 people, over 50 of whom work directly in manufacturing and shop-floor production. The administrative and front-office footprint is intentionally lean: the entire internal IT department comprises just two employees—one dedicated to keeping the Enterprise Resource Planning (ERP) platform running, and the other managing local PC infrastructure and networking.

- **MXD Process Headcount:** 83 employees across 3 sites (>50 in production, 2 in IT). - **Core Product Lines:** Custom industrial mixing equipment, stainless steel tanks, automated process control systems. - **Engineering Requirements:** Sanitary compliance, chemical engineering specifications, and computational fluid dynamics (CFD).

---

[05:54](https://www.youtube.com/watch?v=X0snP9C8vFM&t=354s) Baseline Operational Bottlenecks and the Crawl-Walk-Run Strategy

When Eric joined MXD Process two years prior, the commercial sales pipeline was plagued by severe operational logjams. An initial audit revealed a massive backlog:

- **6,000 Unrouted Leads:** Roughly half of the entire inbound lead volume generated over the prior year was sitting completely untouched in the system. - **7.8-Day Initial Contact Time:** Average speed-to-lead was nearly eight days from inbound form submission to salesperson outreach. - **5-Day Quoting Turnaround:** Generating a customized technical proposal required an average of five working days. - **17.4% Conversion Rate:** Overall lead-to-sale conversion stood at 17.4% across sales cycles that ranged from two weeks to multiple months depending on equipment complexity.

Recognizing the latent potential of AI but lacking the time and technical background to implement it, the company's owner tasked Eric with figuring out how to leverage AI across the commercial organization. In response, Eric instituted a disciplined "crawl, walk, run" deployment strategy across the executive team. Rather than greenlighting sweeping, high-risk digital transformations, the team committed to small, low-risk experiments designed to test hypotheses, kill failures rapidly, double down on initial successes, and gradually scale proven workflows into daily sales operations.

---

[07:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=464s) Process Mapping and Redefining AI as Workflow Automation

Before touching software or writing code, Eric insisted on performing thorough, upfront business process mapping. Through twelve successive iterations on a conference room whiteboard, he mapped the complete customer lifecycle from initial awareness through post-sale delivery. Every operational bottleneck, friction point, and task that employees actively avoided or frequently mismanaged was documented in red ink as a "torch point" bleeding organizational efficiency.

A prime example was the company's credit check procedure for customer financing terms. The historical workflow required mailing out paper credit packets, collecting references, and conducting manual phone verification—a cumbersome three-week process that suffered a 90% abandonment rate. By mapping the full intermediate workflow with all defects exposed rather than an idealized future state, the team could clearly identify exactly where automated intervention would deliver the highest leverage.

```mermaid flowchart TD accTitle: Lead Triage, Routing, and Dynamic Nurturing Architecture subgraph Inbound["Inbound Lead Capture"] A["Web Form Fill / CallRail Inbound"] --> B{"Triage & Qualification"} end

subgraph Qualification["HubSpot Automation Logic"] B -->|"Sales Ready (SQL)"| C["Industry Specialization Routing (3-Minute SLA)"] B -->|"Not Ready (MQL)"| D["4-Week Dynamic Nurture Stream (Automated emails & task triggers)"] end

subgraph Action["Sales Execution"] C --> E["Direct Sales Rep Outreach (<24-Hour SLA)"] D -->|"Engagement / Reply"| E D -->|"21% Meeting Booking Rate"| E end ```

The team then outlined the discrete "jobs to be done" where automation and AI could assist—spanning sales management, revenue forecasting, deal reviews, and pipeline coaching. While many organizations immediately rush to large language models (LLMs) and autonomous agents, Eric reframed AI down to its fundamental principle: a computer executing a structured process and making deterministic decisions on behalf of human operators based on established business rules. Under this pragmatic definition, well-architected automated workflows represent the foundational layer of artificial intelligence.

---

[10:33](https://www.youtube.com/watch?v=X0snP9C8vFM&t=633s) The Crawl Phase: Optimizing HubSpot and CallRail

The "crawl" phase focused exclusively on extracting maximum value from MXD Process's existing tech stack. At the time, the company utilized HubSpot purely as a basic digital rolodex for logging contacts, companies, and deals, ignoring its broader automation engine. Instead of hiring external consultants for an extended multi-month engagement, Eric brought the company's HubSpot agency on-site for a one-week agile sprint. Operating in daily feedback loops—scoping requirements at 8:00 AM, building throughout the day, validating configurations with a sales rep and marketing coordinator, and reviewing progress each evening—they completely overhauled lead routing within five days.

Because MXD Process serves wildly divergent industries with unique technical demands—such as 3-A Sanitary Standards and Good Manufacturing Practice (GMP) requirements for food and beverage versus explosion-proof engineering for industrial paints and coatings—leads were automatically routed based on industry specialization. The results were immediate:

- **Lead Assignment Speed:** Inbound form fills were automatically assigned and routed within 3 minutes. - **Speed to First Contact:** First human sales contact dropped from 7.8 days to under 24 hours.

Next, dynamic lead nurturing workflows were built within HubSpot to bifurcate inbound interest into Sales Qualified Leads (SQLs) and Marketing Qualified Leads (MQLs). Prospects requesting immediate quotes were fast-tracked to specialized reps, while earlier-stage prospects entered a 4-week automated nurture sequence integrating behavioral branching, targeted email touchpoints, and scheduled rep check-ins. This dynamic nurture stream achieved an extraordinary **21% meeting booking rate**, dramatically outperforming the typical 2% to 4% B2B email engagement benchmark. In industrial equipment sales, where the first vendor to respond with an accurate technical quote overwhelmingly wins the deal, this rapid responsiveness provided a decisive competitive edge.

```mermaid flowchart LR accTitle: CallRail ConvertAssist Workflow Integration subgraph CallRail["CallRail Call Tracking"] CR["Inbound Phone Call (Dynamic 1-800 Number)"] --> CA["ConvertAssist AI Engine"] CA --> P1["Action Plan Generation"] CA --> P2["Sales Rep Coaching Notes"] CA --> P3["Draft Response Email"] end subgraph CRM["HubSpot Integration"] P1 & P2 & P3 --> CO["Custom Call Object"] CO --> SR["Sales Rep Review & Send"] end ```

The company applied the same optimization mindset to CallRail, a platform originally deployed only for dynamic 1-800 phone number tracking to attribute pay-per-click ad spend. The team activated an underutilized native feature called **ConvertAssist**, which analyzes recorded inbound customer calls using AI to automatically produce three valuable outputs:

1. **Structured Action Plans:** Itemized technical requirements and commercial next steps derived directly from the caller's conversation. 2. **Sales Coaching Prompts:** Tactical guidance highlighting opportunities for the sales rep to improve objection handling and discovery. 3. **Draft Follow-Up Emails:** Pre-written customer correspondence summarizing the call.

Rep reps are required to review and refine the generated drafts rather than sending them raw. To ensure end-to-end visibility, MXD Process engineered an integration passing ConvertAssist conversation intelligence directly into HubSpot via a custom call object.

---

[14:55](https://www.youtube.com/watch?v=X0snP9C8vFM&t=895s) Takeaway #1: Sweat Existing Tech and the "Smartest, Dumbest Interns"

The primary takeaway from the crawl phase is straightforward: **"Use the technology you already pay for."** Rather than succumbing to vendor pitches touting standalone AI solutions, businesses must audit their existing software subscriptions and exhaust their native capabilities first. Generating immediate operational wins with in-house tools creates crucial internal momentum and executive buy-in, paving the way for targeted investments down the line. Crucially, organizational momentum is achieved by framing initiatives around tangible operational and business value (such as faster contact times and higher conversion rates) rather than abstract IT cost-avoidance metrics, which rarely prompt small, budget-conscious companies to take action.

Moving from rule-based workflow automation to LLM implementations revealed another foundational insight: **"AIs are the smartest, dumbest interns you ever had."** While LLMs possess vast conceptual intelligence and rapid processing speed, their practical utility in highly specialized, highly technical B2B manufacturing environments requires intense structure, clear boundary constraints, and vigilant human oversight.

[16:18](https://www.youtube.com/watch?v=X0snP9C8vFM&t=978s) The Intern Analogy: Why LLMs Require Close Supervision

Reflecting on early experiments with generative models, Seiberling recalls asking AI to conduct technical research on lithium refinement processes for battery manufacturing. Early iterations like GPT-2 and GPT-3 produced answers that were downright "dumb." While newer iterations such as GPT-4 and o3 delivered solid raw information, they still missed the mark on practical industrial reality. The models lacked an organic understanding of physical manufacturing plant environments, real-world operational constraints, and technical disciplines like computational fluid dynamics (CFD).

To bridge this gap, teams must actively train and oversee their tools. Seiberling illustrates this using the analogy of a brilliant intern: if a company hired an intern with a PhD, leadership still would not blindly hand them the keys to run the business. Any manager who has worked with interns knows they require continuous guidance, guardrails, and supervision. Generative models demand the exact same mindset. Smaller organizations face significant risk if they get overexcited, assume the model's output is infallible, and deploy it prematurely without verifying what has actually been produced.

- Early AI models failed on nuanced industrial queries, lacking context around plant constraints and computational fluid dynamics. - Advanced reasoning models possess strong data but still require foundational domain instruction and continuous supervision. - Treating an LLM like an unsupervised "PhD intern" can lead businesses into costly errors.

---

[17:20](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1040s) Solving the Quoting Bottleneck: Building Custom GPT "Arlin"

MXD Process took its first targeted step into large language models to address a crippling operational bottleneck in quoting. At the time, the company had only a single application engineer who was perpetually swamped. Because MXD Process serves an enormous breadth of industrial sectors, expecting one engineer to maintain deep, encyclopedic mastery of every niche application was unrealistic. The business desperately needed a way to augment his capacity with on-demand technical resources.

Instead of waiting for an enterprise-wide rollout, Seiberling found instructions via a QR code to build a Custom GPT. The team uploaded 20 to 30 internal engineering documents and product resources to establish the core knowledge base. Seiberling worked intimately with the lone application engineer to author tailored system instructions and refine prompts. Because only one person sat in the role, driving user adoption was straightforward, allowing them to focus heavily on practical change management—specifically how the tool would integrate into the engineer's daily workflow rather than disrupting it.

```mermaid flowchart LR accTitle: Custom GPT Deployment Workflow for Quoting subgraph INGEST["Knowledge Base"] direction TB Docs["20-30 Engineering & Product Specs"] Prompts["Custom Instructions & Prompts"] end

subgraph ENGINE["Custom GPT ('Arlin')"] direction TB Agent["Domain-Trained Agent"] end

subgraph OUTCOME["Sales Execution"] direction TB AE["Application Engineer"] Quote["Accelerated Quote Turnaround"] end

Docs --> Agent Prompts --> Agent Agent <--> AE AE --> Quote ```

---

[18:20](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1100s) Change Management, D&D Lore, and Expanding the LLM Fleet

A critical element of organizational change management is demystifying AI and stripping away employee anxiety. The team came up with a disarming name for their primary assistant: Arlin. Over weekly lunch breaks, the sales and engineering teams bonded over Dungeons & Dragons sessions. The game's Dungeon Master constantly introduced an idiosyncratic non-player character (NPC) named Arlin—a character who was eager to help, genuinely nice, but fundamentally an idiot. Recognizing that their custom assistant shared the same well-meaning, slightly oblivious nature, they adopted the moniker. Seiberling joked that he later concocted a retrofitted corporate backronym—the "Agent for Revenue"—to justify it, but the lighthearted D&D origin succeeded in making the tool approachable across the department.

Building on Arlin's success, MXD Process developed two additional purpose-built models: Aspen and Acosta, the latter dedicated specifically to supporting sales workflows. The concrete operational impact across these specialized assistants was immediate and dramatic:

- Quoting turnaround times for complex lithium battery applications plunged from 5 full days down to just 1.5 days. - Market positioning and product research for the company's ShearPRO product line dropped from over 10 days down to a single hour. - The company embraced a rapid experimentation philosophy: build quickly, test rigorously, iterate constantly, or kill what does not perform.

---

[19:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1184s) Tool Selection and Workflow Automation: "All That Glitters Is Not Gold"

Seiberling cautions that the market is saturated with overhyped software, noting that "all that glitters is not gold." MXD Process tested commercially hyped AI platforms like Jasper and 11x, only to watch them fail terribly inside their industrial context. The experience led the company to adopt a disciplined, work-backwards framework for developing AI capabilities: 1. Define crystal-clear expectations for the final output or decision required. 2. Carefully map out the exact human workflow. 3. Build a role-based LLM tailored strictly to that workflow. 4. Continuously refine and iterate based on practical user feedback.

For marketing automation, Seiberling avoided expensive enterprise point solutions by stringing together bespoke workflows using Make.com, keeping a human agent in the loop for final reviews. The team adheres to a strict protocol where every prompt and multi-step process is first executed manually. Only after the logic and responses prove consistent do they transition the flow into automated pipelines using Make.com or Zapier.

---

[20:28](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1228s) Business Impact and the AI Maturity Framework

The transition to augmented workflows generated substantial top- and bottom-line revenue gains. Over an 18-month span, MXD Process increased its lead conversion rate from 17.4% to 27.4%—a 10-percentage-point increase—while simultaneously expanding average deal sizes by 20%. This newfound leverage validated the practical approach to AI-assisted selling.

Seiberling emphasizes that management roles are undergoing a fundamental transformation. While managers today cannot simply delegate responsibilities blindly without active direction, their broader challenge is teaching team members how to delegate tasks effectively to AI agents. He highlights the AI maturity framework popularized by Allie K. Miller ("AI with Allie"), which tracks an organization's progression through distinct stages:

1. **Multitasker:** AI handles disconnected, routine micro-tasks. 2. **Copilot:** AI works side-by-side with employees to accelerate existing tasks. 3. **Delegation:** Users assign complex chunks of work directly to specialized models. 4. **Teammate:** AI operates autonomously as a trusted peer; MXD Process envisions future agents functioning as project managers overseeing complex, multi-stakeholder enterprise deals.

---

[21:28](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1288s) The Playbook for Implementation and Leveling the Playing Field

In closing, Seiberling offers an urgent call to action for traditional businesses: begin experimenting immediately. Organizations cannot afford to wait for the technology to settle. Teams should start small, test multiple low-risk initiatives, fail quickly, and celebrate incremental operational wins before eventually hiring external experts to scale mature workflows.

The fundamental reality of generative AI is that it serves as a great equalizer. It elevates bottom-tier, terrible performers into mediocre ones, while transforming already great performers into hyper-efficient powerhouses. Companies that delay adoption risk losing their edge, as even the technological laggards among their competitors will inevitably adopt these tools to close the gap. Emphasizing that he is a business operator rather than a technologist, Seiberling invites attendees to connect on LinkedIn to share experiences and continue the conversation.

## Notable quotes > [03:06](https://www.youtube.com/watch?v=X0snP9C8vFM&t=186s) savagely pragmatic perspective — Eric Seiberling

> [03:06](https://www.youtube.com/watch?v=X0snP9C8vFM&t=186s) me, myself, and I — Eric Seiberling

> [05:54](https://www.youtube.com/watch?v=X0snP9C8vFM&t=354s) crawl, walk, run — Eric Seiberling

> [07:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=464s) torch point — Eric Seiberling

> [14:55](https://www.youtube.com/watch?v=X0snP9C8vFM&t=895s) Use the technology you already pay for. — Eric Seiberling

> [14:55](https://www.youtube.com/watch?v=X0snP9C8vFM&t=895s) AIs are the smartest, dumbest interns you ever had. — Eric Seiberling

> [18:20](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1100s) Agent for Revenue — Eric Seiberling

> [19:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1184s) all that glitters is not gold — Eric Seiberling

## Mentioned Eric Seiberling, MXD Process, Procter & Gamble, Walmart, Pets.com, PetSmart, Macy's, Good Morning America, Instacart, Dassault Systèmes, Amazon Web Services (AWS), HubSpot, CallRail, ConvertAssist, GPT-2, GPT-3, GPT-4, OpenAI o3, Arlin, Aspen, Acosta, ShearPRO, Dungeons & Dragons, Jasper, 11x, Make.com, Zapier, Allie K. Miller, AI with Allie

## Video description Eric Seiberling, VP of Sales and Marketing, MXD Process

Abstracts

In this talk, I’ll share how an old-school manufacturing company—built on stainless steel and handshakes—is using AI to transform how we sell, without buying into the hype.

This isn’t a tech demo or a futurist vision. It’s a brutally practical playbook built from real-world experiments inside a traditional industrial sales org. I’ll walk through how we started small with AI-powered research and follow-up, then scaled what worked into our CRM and sales workflows—and what we learned (and unlearned) along the way.

The talk will follow a simple crawl-walk-run framework to show how any sales team, regardless of industry or tech background, can start using AI to shorten sales cycles, improve quoting, and grow pipeline—without waiting on IT or a massive digital transformation project.

If you’re tired of buzzwords and want a straight path to real results, this session is for you.

[Source](https://www.youtube.com/watch?v=X0snP9C8vFM)

中文简报

MXD Process借助低成本AI工具将销售转化率提升10个百分点工业设备制造商MXD Process将人工智能视为结构化的工作流自动化，而非购买被过度炒作的企业软件，从而将线索转化率提升至27.4%并缩短了报价周期。借鉴1999年互联网泡沫破裂的教训，Seiberling详细阐述了MXD Process如何通过严谨、重实操的AI应用路径，解决严重的线索分配与报价瓶颈。该公司首先优化了现有的HubSpot与CallRail技术栈，随后部署了如“Arlin”等定制GPT助手，以辅助仅有的一名应用工程师并精简复杂技术方案的制定流程。在为期18个月的推行中，MXD Process将报价周转时间从5天缩短至1.5天，转化率从17.4%提升至27.4%，平均订单金额也增加了20%。当前生成式 AI 的炒作与 1999 年互联网泡沫高度相似；企业不能将其视作万能灵药，而应秉持极致务实的态度，通过梳理业务流程瓶颈，将 AI 本质化地落实为基于明确规则的自动化工作流。 (0:04) 中小型制造企业应优先充分挖掘现有软件的原生功能，例如 MXD Process 仅凭优化现有 HubSpot 配置，就将销售线索路由时间缩短至 3 分钟以内，首次联络时间压缩至 24 小时之内。 (10:33) 大语言模型如同“最聪明但也最笨的实习生”，缺乏工业现场常识与流体动力学等专业背景，必须配合严格的边界约束与人工监督才能在复杂 B2B 领域发挥实效。 (14:55) MXD Process 打造了代号为“Arlin”等定制 GPT 助手辅助唯一的应用工程师，将锂电池复杂报价周期从 5 天压缩到 1.5 天，并使产品调研用时从 10 天降至 1 小时。 (17:20) 在 18 个月的务实落地中，MXD Process 将线索转化率从 17.4% 提升至 27.4%，平均客单价提高 20%，证明传统工业团队通过低风险渐进式实验即可显著拉开竞争差距。 (20:28)

Reading summary

Drawing parallels to the 1999 dot-com crash, MXD Process Vice President of Sales and Marketing Eric Seiberling argues that mid-market businesses should avoid speculative AI software and instead deploy pragmatic, human-supervised automations within existing tech stacks to drive measurable revenue. Pragmatic Automation Over Generative AI Hype (0:04) Recalling the dot-com bubble when early e-commerce pioneers like Pets.com failed despite massive brand awareness, Seiberling views current generative AI enthusiasm with skepticism. He contends that emerging technologies require decades of operational infrastructure to mature, warning that advanced tools cannot instantly fix broken operational realities, particularly for lean companies lacking enterprise budgets. Process Mapping and In-House Tech Optimization (5:54) Operating with an 83-person headcount and a two-person IT department, MXD Process initially faced severe pipeline bottlenecks: 6,000 unrouted leads, a 7.8-day response time, a 5-day quoting turnaround, and a 17.4% conversion rate. Rather than investing in bespoke code, the company implemented a 'crawl, walk, run' strategy. Seiberling mapped customer journeys across twelve whiteboard iterations to identify operational friction points, such as a paper-based credit check that had a 90% abandonment rate. Sweating Existing Software: HubSpot and CallRail (10:33) During the crawl phase, MXD Process maximized platforms it already paid for rather than purchasing new tools. Following a one-week agile sprint with their HubSpot agency, the team built industry-specific lead-routing logic that cut lead assignment times to 3 minutes and compressed initial contact times from 7.8 days to under 24 hours. A four-week dynamic nurture sequence achieved a 21% meeting booking rate, substantially outperforming standard B2B benchmarks of 2% to 4%. Supervising Custom GPTs as 'Smart Interns' (16:18) Seiberling characterizes large language models as 'the smartest, dumbest interns you ever had,' noting that while advanced reasoning models offer strong baseline data, they lack native comprehension of industrial realities like plant constraints and computational fluid dynamics. To alleviate work on a lone application engineer, the company built a Custom GPT called 'Arlin' by ingesting 20 to 30 engineering documents, reducing complex lithium battery quoting times from 5 days to 1.5 days. Disciplined Workflows and Commercial Returns (19:44) After proprietary tools like Jasper and 11x failed in MXD Process's manufacturing setting, the company adopted a strict rule to execute all multi-step prompts manually before automating them through Make.com or Zapier. Over 18 months, these augmented workflows raised lead-to-sale conversions from 17.4% to 27.4% and boosted average deal size by 20%. Seiberling notes management's role is shifting toward training staff to delegate discrete tasks to specialized models across Allie K. Miller's AI maturity spectrum. Seiberling views AI as a great equalizer that can elevate lower-performing staff and accelerate high performers, warning that businesses delaying low-risk experimentation risk falling behind competitors who adopt these tools to close capability gaps.

精炼阅读摘要

MXD Process销售与营销副总裁Eric Seiberling指出，面对当前生成式AI的狂热炒作，资源有限的传统制造企业切忌盲目跟风，而应秉持极其务实的“爬、走、跑”策略，通过梳理底层业务流程、深挖现有软件潜力，并在严密监督下部署定向大语言模型，以此推动销售效能与业绩的实质性增长。警惕技术泡沫与务实的精益制造定位 (0:04) Seiberling以1999年互联网泡沫时期的Pets.com为例警示当下的AI热潮：Pets.com曾在四个月内借助超级碗和梅西百货游行广告将品牌知名度推升至70%，但由于当年履约物流、客户支持与个性化服务的基础设施极其不成熟，公司迅速破产，同批倒闭的初创企业高达99%。实体商业最终依靠路边自提和专业化演进存活，这表明新兴技术绝非一蹴而就的灵丹妙药，基础设施的真正成熟往往需要数十年的逐步磨合。作为中型工业设备制造商，MXD Process生产用于食品、饮料及重度化工流体处理的不锈钢储罐、混合系统与自动化控制柜，工程要求涉及计算流体动力学（CFD）及卫生合规标准。企业在3处设施拥有83名员工，其中50多人驻扎车间一线，而IT部门仅有2人分别负责ERP维护与PC网络，且没有专门的软件研发预算。Seiberling强调，非技术背景的管理者必须摆脱技术狂热，立足资源极其有限的现实，走低风险、循序渐进的实用主义路线。重构流程并榨取现有技术栈原生能力 (5:54) 在引入复杂模型前，MXD Process首先梳理了销售运营的严重瓶颈：上一年度约半数（多达6000条）入站线索在系统内完全未被分配，首联平均耗时达7.8天，定制技术报价耗时5天，全流程成单转化率仅17.4%。团队在白板上经过12轮推演完整绘制了客户全生命周期，标出诸如耗时三周、流失率达90%的纸质信用审查等“灼痛点”，并将AI重新定义为“计算机依据明确规则执行结构化流程并进行确定性决策”。在“爬行”阶段，团队没有采购昂贵的外部独立AI系统，而是集中挖掘已付费工具的原生功能。通过与代理机构进行为期一周的敏捷冲刺，MXD Process重构了原本仅用作数字通讯录的HubSpot系统，依据3-A卫生标准或防爆要求等行业属性实现自动分流，将线索分配压缩至3分钟内，销售人员首联时间降至24小时以内；同时配置4周动态培育序列，使会议预订率达到21%，远高于B2B行业2%至4%的平均水平。此外，团队启用了呼叫追踪工具CallRail的ConvertAssist原生AI功能，将客户通话自动转录生成行动方案、销售辅导建议和跟进邮件草稿，再经人工审核后同步至CRM系统。将LLM视作实习生并定向破解报价瓶颈 (16:18) 针对大语言模型，Seiberling提出了“最聪明也最笨的博士实习生”比喻：通用模型虽然拥有庞大知识储备与强大推理能力，但对物理工厂的现场环境、工程规范及流体动力学等缺乏实际感知，若脱离持续引导与明确边界便会产生严重偏差。MXD Process在测试Jasper和11x等高知名度AI营销工具遭遇失败后，确立了严格的反向构建工作流：先明确决策目标，再手动执行验证提示词与步骤逻辑，最后才通过Make.com或Zapier串联自动化，并始终保留人工审核环节。针对全公司仅有一名应用工程师因业务覆盖过宽而陷入报价积压的痛点，团队导入20至30份内部工程文档与技术规范，构建了名为“Arlin”的定制GPT（该名称源自团队跑团桌游《龙与地下城》中热心却笨拙的非玩家角色，以此化解员工抵触情绪）。在人机协作模式下，锂电池等复杂工业场景的报价周期从5天大幅压缩至1.5天，新产品线的市场定位与竞品调研耗时从10天缩短至1小时。随后，团队又针对销售等细分场景扩充了Aspen和Acosta等专用助手。商业增长验证与AI成熟度演进 (20:28) 通过将结构化自动化与定制模型嵌入销售流程，MXD Process在18个月内取得了显著的商业成果：线索成单转化率从17.4%跃升至27.4%（增长10个百分点），平均交易规模同步提升了20%。这些客观数据证明了基于流程梳理的实用化技术落地路径对企业收入和利润的实际支撑作用。 Seiberling引用Allie K. Miller的AI成熟度框架指出，组织的应用形态通常经历多任务工具（Multitasker）、人机协同副驾驶（Copilot）、任务委托（Delegation）到具备自主性的团队伙伴（Teammate）四个阶段。他认为，AI本质上是一个拉平竞争差距的“平权器”：它能将表现底层的员工拉升至中等水平，更能让高绩效者成倍放大效能；管理者的核心职责正从单纯指导员工，转变为训练员工如何高效向AI代理委派复杂任务。 Seiberling呼吁传统企业无需等待技术完全成熟，应立刻以低风险的小规模试验建立反馈闭环并淘汰无效方案。不过需要指出的是，该案例主要建立在销售前端与单一人工工程瓶颈的局部优化上，这种由销售主管直接驱动、依赖轻量低代码工具搭建的受控工作流，在面对更加庞大复杂、跨系统整合壁垒高筑且合规容错率极低的工业场景时，能否具备同等的敏捷性与泛化能力仍有待进一步验证。

Detailed notes — zh

## Background 本内容由 YouTube 频道 Ai4 于 2025 年 8 月 28 日发布，由 MXD Process 销售与营销副总裁 Eric Seiberling 主讲。视频分享了一家深耕不锈钢制造的传统工业企业如何在不盲从科技炒作的前提下，通过务实且可落地的 AI 战术重塑销售流程并驱动业绩增长。 ## TL;DR 演讲者以 1999 年互联网泡沫破裂的历史为鉴，提出面对 AI 浪潮应秉持“野蛮实用主义”，切忌盲目相信炒作。针对 MXD Process 资源匮乏、线索积压与报价周期长等痛点，团队采取了“爬、走、跑”的渐进式战略：首先盘活 HubSpot 与 Call Rail 等现有工具实现线索极速分发与智能跟进；随后将 LLM 视作需要严格指导与监督的“博士实习生”，打造了名为 Arlin 的定制工程报价助手；最终通过逆向设计工作流与 Make.com 低代码自动化，将线索转化率从 17.4% 显著提升至 27.4%，平均客单价提高 20%。 ## Atmosphere & style 整场演讲洋溢着制造业实干家的硬核务实风格与幽默自嘲，摒弃了空洞的技术布道。演讲者通过《龙与地下城》NPC 命名 AI 助手等生动故事，将复杂的变革管理讲得妙趣横生且极具亲和力。现场节奏明快、案例真实，既有对技术盲目放权的尖锐警示，又有落地见效后的从容底气，带给观众强烈且清晰的实战启发与行动紧迫感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=X0snP9C8vFM&t=4s) 1999 年的互联网泡沫（如 Pets.com 在 4 个月内知名度达 70% 却迅速倒闭）表明电商落地花费了近 20 年，当前的 AI 浪潮同样需要警惕盲目炒作，保持‘极其务实’的态度。 - [03:56](https://www.youtube.com/watch?v=X0snP9C8vFM&t=236s) 面对仅有两名 IT 人员且预算受限的现实，中小型传统制造企业不应追求庞大的 IT 变革，而应采取‘爬、走、跑’（Crawl, Walk, Run）的小步快跑策略。 - [04:46](https://www.youtube.com/watch?v=X0snP9C8vFM&t=286s) 在引入技术前，MXD Process 先经历了 12 轮白板梳理，精准定位了 6000 条积压线索、7.8 天跟进延迟等核心业务痛点。 - [10:00](https://www.youtube.com/watch?v=X0snP9C8vFM&t=600s) 落地第一原则是‘优先使用已付费的技术’：通过挖掘现有系统（HubSpot 和 Call Rail）实现线索 3 分钟内自动分发，将首次跟进时间压缩至 24 小时以内，邮件会议预约率提升至 21%。 - [16:18](https://www.youtube.com/watch?v=X0snP9C8vFM&t=978s) 大语言模型（LLM）如同‘拥有博士学位的实习生’，虽知识丰富但缺乏特定工业场景常识，必须保持人工监督（Human-in-the-loop）与精准指令指导。 - [17:20](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1040s) 通过打造融入团队文化的定制 GPT（如命名为 Arlin 的助手），成功将复杂工业报价调研周期从 5 天压缩到 1.5 天，产品定位研究时间从 10 天以上缩短至 1 小时。 - [19:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1184s) 采用‘逆向设计’构建自动化：先明确业务输出目标、定义人工流程，再手动测试 Prompt，最后通过 Make.com 等低代码工具实现自动化闭环。 - [20:46](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1246s) 经过一年半的务实推进，公司整体线索转化率从 17.4% 提升至 27.4%，平均客单价提升了 20%。 - [20:46](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1246s) AI 正在拉平行业竞技场，让平庸者合格、让优秀者更高效；企业应尽早从小切口启动实验，把持续的渐进式改进作为核心竞争力。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=X0snP9C8vFM&t=4s) 互联网泡沫的教训与 AI 时代的反思演讲者 Eric Seiberling 开场借用 1999 年夏天的互联网狂热（.com Boom）拉开序幕。他回忆道，当年所有人都在宣扬互联网将彻底颠覆一切——实体店会倒闭，所有人都会在线上购物，汽车厂商会直接向消费者直销。作为当时宝洁公司（Procter & Gamble）的品牌经理，Seiberling 亲自参与了宝洁早期电商网站的搭建，当时项目因为无法打入沃尔玛（Walmart）等传统渠道，转而通过自建电商系统建立与消费者的直接联系，梦想着借此掀起销售变革。他特别举了 **Pets.com** 的经典案例：该公司成立于 1998 年，凭借极其出彩的营销手段（如梅西百货感恩节大游行、超级碗广告、登上《早安美国》）迅速崛起，在短短四个月内将品牌知名度提升到了 70%。然而最终，Pets.com 与当时近 99% 的互联网初创公司一样破产倒闭，其域名如今沦为了 PetSmart 的重定向链接。Seiberling 分析指出，原因在于“做电商真的很难”——特别是在 1999 年，供应链物流、呼叫中心、个性化推荐等配套极度落后，行业整整花了 20 年才摸索出可行门道。尽管如此，传统行业并未彻底消失，超市结合 Instacart 等技术发展出了线上下单自提与送货上门，实体书店与咖啡店也重新复苏。面对当下铺天盖地的 AI 浪潮，Seiberling 直言这种狂热感令人产生“既视感”（déjà vu）。作为一个拥有 25 年销售与营销经验的实干家而非纯粹的技术专家，他的核心主张是“不要盲目相信炒作”。AI 固然是一项能力巨大的工具，但绝不可能一夜之间颠覆一切，尤其是对于无法立即进行巨额投资的企业而言。企业必须保持极其务实的态度，思考如何将其落地到具体业务中。 - **Pets.com 品牌知名度**：仅用 4 个月便达到 70%。 - **电商成熟周期**：从 1999 年泡沫破裂到如今形成成熟模式，经历了约 20 年的演进。 - **核心观点**：AI 是一种工具而非万能神药，应用时需秉持“极其务实”（savagely pragmatic）的态度。 --- ### [03:56](https://www.youtube.com/watch?v=X0snP9C8vFM&t=236s) MXD Process 的现实困境与低成本破局思路 Seiberling 介绍了自己的职业背景，他曾任职于宝洁、达索系统（Dassault Systèmes）以及亚马逊云科技（AWS），两年前加入 MXD Process 担任销售与营销副总裁。MXD Process 是一家专注于工业搅拌设备、罐体及控制系统制造的小型工业设备企业，业务高度依赖化工、计算流体力学等专业工程技术，涉及食品饮料（如符合 3-A 卫生标准与 GMP 规范）、涂料防爆等高度技术化的定制需求。公司全员仅有 83 人，分布在三个厂区，其中 50 多人属于一线生产人员，前端管理部门极其精简。公司面临的技术资源短缺十分严峻：全公司仅有两名 IT 运维人员，一人负责维持 ERP 系统，另一人负责维护网络和办公电脑，根本没有多余的人力或预算去组建大型治理团队或规划宏大的数字转型战略。Seiberling 加入公司时，面临着一堆棘手的销售痛点： ```mermaid flowchart LR accTitle: MXD Process 原始销售流程痛点 subgraph Leads["线索处理泥潭"] A["6,000 条未分配线索"] --> B["响应延迟: 7.8 天"] end subgraph Sales["转化瓶颈"] B --> C["报价周期: 5 天"] C --> D["成交转化率: 17.4%"] end ``` 在资源极度受限的前提下，公司企业主赋予了 Seiberling 充分的决策权去探索 AI 的可能性。Seiberling 提出了贯穿其职业生涯的 **“爬、走、跑”（Crawl, Walk, Run）** 渐进式战略：先从小规模实验切入，快速测试，不行就立刻止损；一旦看到效果### 互联网泡沫的历史镜鉴与“野蛮实用主义” 演讲者 Eric Seiberling（MXD Process 销售与营销副总裁）开场以轻松的语气感谢观众穿越展厅的远道而来，随后将时光拉回 1999 年夏天的互联网泡沫（Dot-com boom）鼎盛时期。当时他还在宝洁（P&G）担任品牌经理，由于产品无法进入沃尔玛（Walmart）等传统渠道，便顺应当时的“电商将颠覆一切”的浪潮，搭建了宝洁最早的内部电子商务网站之一。他借用 Pets.com 这一经典案例——在 1998 年推出后通过超级碗广告和梅西百货游行获得了极高的知名度，却在短短数月内轰然倒塌——指出“电子商务其实非常艰难”，供应链、呼叫中心和个性化服务在当时极其复杂，行业整整花了 20 年才真正掌握其门道。结合这段历史，Eric 提出并强调了他对当前人工智能（AI）热潮的核心观点：**野蛮实用主义（Savagely Pragmatic）**。他认为当前的 AI 狂热与 1999 年的互联网泡沫极其相似（似曾相识感），但技术终究只是工具，不可能在一夜之间改变一切。对于无法立即投入巨资的企业来说，切忌盲目相信炒作，而必须冷静思考技术如何落地于自身业务，并采取务实、渐进的步骤去实践和推广。 * **Pets.com 品牌知名度**：在短短 4 个月内达到了 70% 的品牌知名度，但最终破产，其域名如今重定向至 PetSmart。 * **演讲者背景**：拥有 25 年销售与营销经验，曾任职于宝洁（P&G）、达索系统（Dassault Systèmes）及亚马逊云科技（AWS）。 --- ### [04:46](https://www.youtube.com/watch?v=X0snP9C8vFM&t=286s) 传统制造企业的痛点与“爬-走-跑”战略 MXD Process 是一家典型的传统工业设备制造企业，共有 83 名员工分布于三个厂区，其中超过 50 人为一线生产人员。公司主营工业搅拌机、储罐和控制系统等高度定制化的过程设备，涉及化学工程与计算流体力学等极高门槛的技术领域。由于团队精简，全公司仅有两名 IT 人员（一人维护 ERP 系统，一人维护网络和个人电脑），既没有预算，也没有人力去搭建复杂的 AI 治理团队或大型 IT 战略架构。面对老板“利用 AI 提升业绩，你自己看着办”的要求，Eric 提出了**“爬、走、跑”（Crawl, Walk, Run）**的敏捷推进战略。 ```mermaid flowchart LR accTitle: MXD Process 业务流程重塑与 AI 落地路径 subgraph Phase1["业务梳理 (Crawl)"] A["白板梳理业务流程"] --> B["标识红色痛点与效率漏洞"] B --> C["定义 AI 角色与任务 (Jobs-to-be-Done)"] end subgraph Phase2["挖掘现有堆栈 (Walk)"] D["HubSpot 自动化流程"] --> E["Call Rail 智能语音分析"] E --> F["线索 3 分钟内精准分发"] end subgraph Phase3["规模化扩展 (Run)"] G["LLM 深入辅助"] --> H["缩短销售周期与提升转化"] end Phase1 --> Phase2 --> Phase3 ``` 在正式引入技术之前，Eric 经历了约 12 轮白板讨论，将公司从潜在客户感知到售后服务的整个销售流程全面梳理，用红色标记出所有效率流失的“痛点”（如冗长的信用检查流程原先需耗时 3-4 周，90% 最终未能完成）。完成流程梳理后，团队进一步拆解了 AI 可以在销售管理、预测和交易审查中扮演的具体角色。 * **MXD Process 初始痛点数据**： * **积压线索**：入职时发现 6,000 条未分发的线索（占全年线索总量的一半）。 * **响应延迟**：从首次接触到跟进平均耗时 **7.8 天**。 * **报价耗时**：完成一份报价需要 **5 天**。 * **转化率**：从线索到最终成交的转化率仅为 **17.4%**。 --- ### [10:00](https://www.youtube.com/watch?v=X0snP9C8vFM&t=600s) 爬行阶段：盘活现有技术堆栈（HubSpot 与 Call Rail） Eric 强调，许多人在谈论 AI 时会立刻联想到大语言模型（LLM）或检索增强生成（RAG）等前沿概念，但广义上的 AI 本质上是“计算机根据规则代表我们运行流程并做出决策”。因此，MXD Process 的“爬行”阶段并没有购买昂贵的新软件，而是直接挖掘现有软件（HubSpot）的潜力： 1. **基于行业分发的自动化规则**：取消原先混乱的分配机制，根据销售人员的专业领域（如熟悉食品饮料卫生标准的销售员对接食品客户；熟悉防爆要求的对接涂料客户）进行自动路由。Eric 邀请代理机构开展为期一天的敏捷开发（Agile Sprints），迅速落地该功能。 2. **动态培育流（Dynamic Nurturing）**：在 HubSpot 中建立为期 4 周的自动化培育流程，包含多轮邮件与电话提示，根据客户行为动态调整后续路径。 3. **挖掘语音工具（Call Rail）的隐藏功能**：利用 Call Rail 现成的 Convert Assist 功能，自动监听电话通话，生成行动计划、销售跟进建议及邮件草稿，并将其无缝接入 HubSpot CRM。 * **线索路由时效**：表单填写后 **3 分钟内**完成自动分发。 * **首次跟进时间**：销售人员首次联系客户的时间从 **7.8 天缩短至 24 小时以内**。 * **邮件培育效果**：搅拌机培育流获得了高达 **21% 的会议预约率**（行业平均水平仅为 2%-4%）。 * **核心认知 1**：“请优先使用你已经在付费的技术（Use the technology you already pay for）。”不要盲目追求炫酷的新产品，先从现有工具中挖掘价值，以此建立内部信任和业务动能。 --- ### [16:00](https://www.youtube.com/watch?v=X0snP9C8vFM&t=960s) 人工智能的本质：最聪明又最愚蠢的实习生在进入大语言模型（LLM）应用阶段后，Eric 给出了一个形象的界定：**AI 就像是你招过的“最聪明但也最愚蠢的实习生”**（Smartest, dumbest intern you ever had）。它拥有海量的知识储备和极高的处理速度，但缺乏对特定复杂工业领域的实际常识，必须接受明确的指令、严谨的规则和人工的严格审核（Human-in-the-loop），才能真正为高技术门槛的业务场景创造价值。 ### [16:18](https://www.youtube.com/watch?v=X0snP9C8vFM&t=978s) LLM 就像“博士实习生”：需要指导而非盲目放权在探讨制造工艺中的锂提炼（用于电池制造）调研时，演讲者回顾了模型的演进历程。在 GPT-2 到 GPT-3 的早期阶段，模型的表现非常愚蠢；等到 GPT-4 或 GPT-4o（o3）时，模型虽然能提供优质信息，但依然会“错失核心要点”（missed the boat）。原因在于模型缺乏对真实制造工厂运作、物理限制以及计算流体力学（CFD）等复杂工程维度的深度理解。因此，企业必须主动去“教导”这些模型。演讲者提出了一个生动的类比：如果公司来了一位拥有博士学位的实习生，你会直接把业务钥匙交给他全权管理吗？任何管理过实习生的人都会回答“绝不可能”。实习生再聪明，也需要严密监督和持续指导——对待 LLM 亦是如此。特别是对于中小企业而言，如果仅仅觉得“这太棒了”便盲目放任模型运行，不理解潜在的边界和漏洞，极容易给业务带来巨大麻烦。 ```mermaid flowchart TD accTitle: LLM Deployment Comparison subgraph Wrong["错误方式：直接放权"] A1["大模型 / 博士实习生"] --> B1["直接接管关键业务"] B1 --> C1["忽视工厂约束与物理现实 / 产生严重错误"] end subgraph Right["正确方式：指导与监督"] A2["大模型 / 博士实习生"] --> B2["定制工程提示词 & 知识库输入"] B2 --> C2["资深工程师严密审核 / 流程闭环"] end ``` --- ### [17:20](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1040s) 打造专属定制 GPT：命名为 Arlin 的变革管理故事 MXD Process 构建的第一个定制 LLM 重点攻克的是“报价环节”。当时团队仅有唯一一名应用工程师（Application Engineer），每天都被海量报价需求淹没；加上公司业务覆盖的工业领域极其广泛，单凭一人根本不可能掌握所有行业的技术细节。为了给他提供技术支援，团队通过扫描 Custom GPT 的二维码教程搭建了专用工具，上传了 20 到 30 份工程与产品资源文档，并撰写了专门的指令与提示词。由于当时只有一位工程师，推广阻力小，演讲者与他紧密配合，深入探讨了该工具具体如何嵌入日常工作流中。在组织变革管理（Organizational Change Management）中，消除员工对新技术的恐惧至关重要。团队赋予了这个 AI 助手一个极具人情味的名字——“Arlin”。工程团队平时在午休时会一起玩《龙与地下城》（Dungeons & Dragons），地下城主经常提到游戏里一个名叫 Arlin 的非玩家角色（NPC）：他虽然是个“呆子”，但为人热心、极度渴望提供帮助。团队顺理成章地将这个 GPT 命名为 Arlin（演讲者打趣称后来为了合理化，硬编了个全称叫“收入代理” Agent for Revenue）。赋予它名字后，所有人都能轻松、亲切地称呼它，新工具的推行也变得极其顺畅。此后，团队又陆续开发了名为 Aspen 和 Acosta（负责销售业务）的大模型。 - **报价与技术调研提速**：在锂电池相关项目中，应用工程师使用定制提示词，将工作周期从 5 天大幅缩短至 1.5 天。 - **产品研究与定位**：在 ShearPro 生产线的产品研究中，市场定位分析时间从 10 天以上直接压缩到仅需 1 小时。 - **核心迭代原则**：坚守“构建、测试、快速迭代，不行就果断放弃”（Build, test, iterate or kill it）的敏捷态度。 --- ### [19:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1184s) 逆向设计工作流与基于 Make.com 的自动化实践在踩坑经验方面，演讲者直言“发光的并不都是金子”（All that glitters is not gold）。MXD Process 曾尝试采购 Jasper、11x 等市面上的现成 AI 营销与销售工具，但结果均以惨败告终。团队转而采用了一套自主可控的“逆向推导流程”： 1. **明确决策产出**：首先清晰界定业务最终需要什么输出与决策； 2. **定义工作流**：将具体的人工业务流程彻底理顺并标准化； 3. **构建基于角色的 LLM**：针对特定角色定制专属的 LLM 提示词与知识库； 4. **手动测试再自动化**：先手动运行所有 Prompt 验证效果，确认无误后再利用 Make.com 或 Zapier 等轻量工具进行自动化串联，并在节点末端保留人工/Agent 最终审查。 ```mermaid flowchart LR accTitle: Workflow Automation Design A["明确决策与产出目标"] --> B["梳理并定义业务工作流"] B --> C["构建基于角色的 LLM"] C --> D["全流程手动运行与测试"] D --> E["通过 Make.com / Zapier 自动化并保留人工审查"] ``` 由于定制商业软件费用高昂，团队选择用低代码工具自己搭建。面对当前 AI 工具的爆发，演讲者强调当今职场角色正在发生本质转变：“现阶段你必须进行明确指挥（Direct），而不能完全甩手委托（Delegate）；但同时，我们必须学会如何将任务拆解并委托给 AI，而不是什么都亲力亲为。” --- ### [20:46](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1246s) 业务成效、进化阶梯与行动建议通过这套务实的落地打法，MXD Process 在一年半的时间内取得了极为显著的商业成果： - **线索转化率**（Lead Conversion Rate）：从原本的 **17.4%** 飙升至 **27.4%**。 - **平均客单价**（Deal Size）：整体提升了 **20%**。演讲者引用了行业博主 "AI with Ally" 提出的成熟度演进框架，指出企业应用 AI 往往经历四个阶段： 1. **多任务处理者**（Multitasker）：处理零散琐碎事务； 2. **副驾驶**（Co-pilot）：协同员工处理专业场景； 3. **委托执行者**（Delegate）：承担独立的工作流环节； 4. **正式团队成员**（Teammate）：深度融入团队（公司未来计划打造一个专门负责大项目管理的 AI 队友）。演讲者最后留给全场的实战秘诀是：**立即行动，绝不要等待；从小处着手，快速尝试并迅速淘汰失败方案；将持续的渐进式改进视为胜利；在必要时寻求专家帮助**。 > “AI 正在拉平竞技场。它让表现糟糕的人变得平庸，让表现优秀的人变得更加高效。你的竞争对手中哪怕是最落后的跟跑者也正在迎头赶上，所以你必须立刻行动起来。” ## Notable quotes > [00:04](https://www.youtube.com/watch?v=X0snP9C8vFM&t=4s) 不要盲目相信炒作。 — Eric Seiberling > [00:04](https://www.youtube.com/watch?v=X0snP9C8vFM&t=4s) 极其务实（savagely pragmatic） — Eric Seiberling > [04:46](https://www.youtube.com/watch?v=X0snP9C8vFM&t=286s) 利用 AI 提升业绩，你自己看着办 — Eric Seiberling > [10:00](https://www.youtube.com/watch?v=X0snP9C8vFM&t=600s) 请优先使用你已经在付费的技术（Use the technology you already pay for）。 — Eric Seiberling > [16:00](https://www.youtube.com/watch?v=X0snP9C8vFM&t=960s) 最聪明但也最愚蠢的实习生（Smartest, dumbest intern you ever had） — Eric Seiberling > [16:18](https://www.youtube.com/watch?v=X0snP9C8vFM&t=978s) 如果公司来了一位拥有博士学位的实习生，你会直接把业务钥匙交给他全权管理吗？任何管理过实习生的人都会回答“绝不可能”。 — Eric Seiberling > [17:20](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1040s) 构建、测试、快速迭代，不行就果断放弃（Build, test, iterate or kill it） — Eric Seiberling > [19:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1184s) 发光的并不都是金子（All that glitters is not gold） — Eric Seiberling > [19:44](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1184s) 现阶段你必须进行明确指挥（Direct），而不能完全甩手委托（Delegate）；但同时，我们必须学会如何将任务拆解并委托给 AI，而不是什么都亲力亲为。 — Eric Seiberling > [20:46](https://www.youtube.com/watch?v=X0snP9C8vFM&t=1246s) AI 正在拉平竞技场。它让表现糟糕的人变得平庸，让表现优秀的人变得更加高效。你的竞争对手中哪怕是最落后的跟跑者也正在迎头赶上，所以你必须立刻行动起来。 — Eric Seiberling ## Mentioned Eric Seiberling, MXD Process, Pets.com, PetSmart, 宝洁 (Procter & Gamble), 沃尔玛 (Walmart), Instacart, 达索系统 (Dassault Systèmes), 亚马逊云科技 (AWS), HubSpot, Call Rail, Convert Assist, GPT-2, GPT-3, GPT-4, GPT-4o, Custom GPT, Arlin, Aspen, Acosta, ShearPro, Jasper, 11x, Make.com, Zapier, AI with Ally, 龙与地下城 (Dungeons & Dragons) ## Video description Eric Seiberling, VP of Sales and Marketing, MXD Process Abstracts In this talk, I’ll share how an old-school manufacturing company—built on stainless steel and handshakes—is using AI to transform how we sell, without buying into the hype. This isn’t a tech demo or a futurist vision. It’s a brutally practical playbook built from real-world experiments inside a traditional industrial sales org. I’ll walk through how we started small with AI-powered research and follow-up, then scaled what worked into our CRM and sales workflows—and what we learned (and unlearned) along the way. The talk will follow a simple crawl-walk-run framework to show how any sales team, regardless of industry or tech background, can start using AI to shorten sales cycles, improve quoting, and grow pipeline—without waiting on IT or a massive digital transformation project. If you’re tired of buzzwords and want a straight path to real results, this session is for you. [Source](https://www.youtube.com/watch?v=X0snP9C8vFM)

Source quotes

[0:04] Eric Seiberling: 不要盲目相信炒作。 [0:04] Eric Seiberling: 极其务实（savagely pragmatic） [4:46] Eric Seiberling: 利用 AI 提升业绩，你自己看着办 [10:00] Eric Seiberling: 请优先使用你已经在付费的技术（Use the technology you already pay for）。 [16:00] Eric Seiberling: 最聪明但也最愚蠢的实习生（Smartest, dumbest intern you ever had） [16:18] Eric Seiberling: 如果公司来了一位拥有博士学位的实习生，你会直接把业务钥匙交给他全权管理吗？任何管理过实习生的人都会回答“绝不可能”。 [17:20] Eric Seiberling: 构建、测试、快速迭代，不行就果断放弃（Build, test, iterate or kill it） [19:44] Eric Seiberling: 发光的并不都是金子（All that glitters is not gold） [19:44] Eric Seiberling: 现阶段你必须进行明确指挥（Direct），而不能完全甩手委托（Delegate）；但同时，我们必须学会如何将任务拆解并委托给 AI，而不是什么都亲力亲为。 [20:46] Eric Seiberling: AI 正在拉平竞技场。它让表现糟糕的人变得平庸，让表现优秀的人变得更加高效。你的竞争对手中哪怕是最落后的跟跑者也正在迎头赶上，所以你必须立刻行动起来。

Source quotes

[3:06] Eric Seiberling: savagely pragmatic perspective [3:06] Eric Seiberling: me, myself, and I [5:54] Eric Seiberling: crawl, walk, run [7:44] Eric Seiberling: torch point [14:55] Eric Seiberling: Use the technology you already pay for. [14:55] Eric Seiberling: AIs are the smartest, dumbest interns you ever had. [18:20] Eric Seiberling: Agent for Revenue [19:44] Eric Seiberling: all that glitters is not gold

Original transcript

[0:04] Thank you all very much and thank you [0:05] for making the longest journey you can [0:07] possibly take from the Garden Arena to [0:09] here. So I appreciate your both your [0:11] commitment and your hiking skills in [0:13] order to be able to make it back. [0:16] So I don't know if anybody remembers the [0:18] summer of 1999. I'm old enough to [0:20] remember and it was the height of the [0:23] dot boom. It was the internet was going [0:26] to change absolutely everything. stores [0:29] were going to close. Everybody's gonna [0:30] buy everything online. [0:33] Car companies would sell direct and it [0:35] would just create this whole new era and [0:37] it was going to change absolutely [0:39] everything. Does anybody remember that? [0:41] Yeah. And I the older of us remember [0:43] that. But it was like one of these [0:44] things. I was a brand manager at PNG at [0:47] the time and I was launching one of the [0:48] first internal.coms. [0:50] We were launching a new appliance and we [0:52] were making the juice that went inside [0:54] it and so we we couldn't get [0:55] distribution through Walmart. So, what [0:57] we had to do is we set up an e-commerce. [0:59] So, it's like great, we can have a [1:00] onetoone relationship with our [1:02] consumers. We're going to be able to do [1:03] this. We're going to make lots of money. [1:05] It's just going to be awesome. And then, [1:07] you know, we're just going to grow and [1:08] grow and really build this new [1:10] capability for PNG. [1:12] But my personal favorite, [1:15] I don't know if anybody recognizes this [1:17] guy, but it's from Pets.com. Now, [1:19] Pets.com launched in 1998, and it like [1:23] immediately took off. Their advertising [1:25] was absolutely incredible. They were in [1:27] the Macy's Day Parade. They had a great [1:29] super set of Super Bowl ads. They were [1:31] on Good Morning America. And they [1:32] reached a 70% brand awareness in only [1:35] four months. [1:37] They were going to deliver everything [1:38] you needed for your pet at home. And so [1:41] it was just absolutely fabulous. And in [1:43] between that and Amazon was looking like [1:45] this is perfect. This is what I can do. [1:49] What happened? [1:51] They're dead. [1:53] just like almost 99% of the other [1:55] dotcoms that all launched in that time. [1:58] But what happened? Why? It's because [2:01] e-commerce is hard. Anybody been [2:03] involved in e-commerce? [2:06] Logistics, your call center, [2:08] personalization, all those things, [2:10] particularly in 1999, were horrible. It [2:12] was really, really hard to do. And it's [2:15] taken us 20 years to start to be able to [2:17] figure out how to do it and how to do it [2:20] well. [2:22] This is where pets.com is today. It's [2:25] the URL redirect for PetSmart. [2:29] Grocery stores did not die, but they [2:31] have changed. So, if we look through [2:32] COVID and those kinds of things, they [2:34] were able to adapt to technology. You've [2:36] got Instacart now for home delivery. You [2:38] got buy online, pick up at store on the [2:40] curb, those kinds of things. It adapted, [2:43] but it didn't totally disrupt and change [2:45] everything. Now, the bookstores [2:48] mainly were killed, but now you see [2:50] independent bookstores and coffee shops [2:52] starting to come back. And so, it's one [2:54] of these things when we think about the [2:56] technology, when we think about the [2:57] whole dot era, we end up learning a [3:01] lesson that it's a tool and how do we [3:03] end up leveraging that? [3:06] But now we're in this realm of AI. And [3:08] how many times, what does it feel like [3:10] to people? [3:12] I don't know to you, but it feels like [3:13] to me it's deja vu. [3:15] And I just have a very simple, this is [3:18] from my savagely pragmatic perspective. [3:21] I'm not a technologist. I've been in [3:23] sales and marketing for 25 years. And my [3:26] initial inclination is don't believe the [3:28] hype. [3:29] It is a tool. It can provide tremendous [3:31] capability for us, but it's not going to [3:34] change everything overnight. [3:36] And particularly for companies that [3:38] can't make the investment straight away. [3:40] But we what we have to do is we got to [3:42] take a step back and we really have to [3:43] think about it. We have to think about [3:45] how does this apply to our business and [3:47] then what are the savagely pragmatic [3:49] steps I can take in order to go from [3:51] beginning and start to use it and then [3:53] keep rolling it out and to be able to [3:54] take advantage of it. So my name is Eric [3:57] Cyberling. I started at Proctor and [3:59] Gamble but I've I've worked at um tech [4:03] companies. I worked at Disso Systems if [4:04] anybody knows that uh engineering [4:06] software. I spent years a couple years [4:08] at AWS, but now I'm the VP of sales and [4:11] marketing for a small industrial [4:13] equipment company. And it was really [4:15] interesting to go from a really big [4:16] company to a small one. [4:19] The problem is I didn't have any budget. [4:22] Any development team that I have [4:23] consists of me, myself, and I. And it's [4:26] like how do we end up doing this? How do [4:28] we end up moving forward? But on the [4:30] other hand, the great part about my job [4:32] is I get to make the decision. So it's [4:35] like how do we end up moving forward? [4:36] how do we end up going? And so what I'm [4:38] going to do today is I'm just going to [4:39] tell you a little bit about our AI [4:41] story, what we've done, what we've [4:43] learned, and what we've been able to [4:44] achieve. [4:46] So MXD Process is a small industrial [4:48] equipment company. We make mixers and [4:51] tanks and control systems. Basically [4:54] everything you need to make a production [4:55] line. So if you whether you're making [4:57] tomato sauce or I'm creating um coaul, [5:01] epoxy or anything liquid, if I'm making [5:04] a manufacturing plant, we make the [5:05] equipment to be able to do that. It's a [5:08] very specialized industry. It's about [5:10] process equipment. It's about chemical [5:12] engineering, computational fluid [5:13] dynamics. It's extraordinarily [5:15] technical. We have 83 people in the [5:18] company across three sites. Over 50 of [5:20] them are in production. So we have a [5:22] very small front office. I only have two [5:24] IT people. One basically is trying to [5:26] maintain our ERP system and the other [5:28] one is trying to maintain our network [5:30] and our personal computers. [5:34] We really don't have much bandwidth in [5:36] order to be able to do the work that [5:38] normally I hear people going okay you [5:40] need to build a governance team and we [5:42] need to get this IT strategy and build [5:45] all these things. [5:47] We don't have the people or the money to [5:48] do that. So how do you go forward? How [5:51] do you how do you think about it? [5:53] So, I ended up being hiring hired almost [5:56] two it was two years ago this month. And [5:59] when I showed up, they presented me some [6:01] problems. [6:04] We had 6,000 unouted leads when I showed [6:07] up. That was half of the leads they they [6:10] got for the year. They just sat there. [6:13] No one touched them. We ended up having [6:16] 7.8 days from the first contact. Not [6:18] only because of just the unouted leads, [6:20] but also it was just how did it get [6:22] there? Making sure that we had good [6:24] processes for salespeople to follow up. [6:27] It took us 5 days to complete a quote [6:29] and we had a 17.4 conversion rate when [6:31] you talk about from lead to um sale. Now [6:36] our sales are big. So it's you know our [6:38] sales sales length depending on what the [6:40] cost is can range anywhere from two [6:42] weeks to multiple months. But you know [6:45] that close rate it's what can we do [6:46] better? And so I was challenged by our [6:49] owner to say what can you do? How can we [6:52] leverage AI? So the advantage I had is I [6:54] had an owner who was very forward [6:56] thinking of saying let's use AI. I can [6:58] see the potential in it. I just don't [7:00] know how and I don't have time. You [7:01] figure it out. Anybody in that position? [7:08] This is a strategy I've taken across my [7:10] entire career, which is kind of a crawl, [7:12] walk, run. And it's this is the the [7:15] mantra I even used with our executive [7:16] team of we're going to we're going to [7:18] crawl, we're going to walk, and then [7:20] eventually we'll run. But we can't just [7:22] start getting up and running. And so, [7:24] how do we think about that? What are the [7:25] small experiments we can run? How do we [7:27] end up testing it? How do we kill things [7:29] quickly if it doesn't work? But then, [7:32] how do we when we see an inkling of [7:33] something that works well, what do we [7:35] do? How do we end up moving forward in [7:37] that? And then ultimately, how do we [7:39] scale it? And so that was the approach [7:41] that that we took. And so the first [7:44] thing and then most people want to skip [7:45] to the technology, but what I forced [7:48] myself to do um really you see the [7:50] whiteboard that was about the 12th [7:52] iteration of the business process of me [7:54] walking through it. And what you see in [7:56] the red are all the problems. Those are [7:58] the torch points. Those were the things [7:59] that people didn't want to do. it took [8:02] too much time or they just skipped it or [8:04] they did it wrong and all these things [8:06] that were bleeding efficiency from our [8:08] business and it was getting into some of [8:10] the details about you know whether it [8:13] was leads or whether it was even just [8:14] trying to when someone wanted terms for [8:16] their contract. How did we do the credit [8:19] checks? What it was originally was it [8:21] took four three weeks because we sent [8:23] out give us your credit packet, give us [8:25] references, we'll call you up, we'll go [8:27] back out and it was that whole process [8:29] and 90% of the time it was never [8:30] completed. But it was looking across all [8:33] of that and working through and doing [8:35] that hard work up front that allowed us [8:37] to move much faster in the technology [8:38] space. And so what you end up seeing on [8:41] that big after the big squirrels is um [8:45] that's our new sales process and it goes [8:47] from the initial kind of awareness all [8:49] the way through to post sale. And so we [8:51] have the whole thing documented. But the [8:53] other nice thing about that we did it [8:54] was it wasn't the end tobe process. It [8:57] was kind of the middle with all the [8:59] problems highlighted. [9:01] So then what what I ended up doing is we [9:05] took and we said okay what are the jobs [9:07] to be what are the roles that AI can [9:09] play and what are the jobs to be done. [9:12] Oh by the way the slides will be [9:13] available for download as well. So now I [9:17] get it there there's going to be some [9:18] small things on here because you'll see [9:19] like prompts and stuff but it just [9:22] they'll be available. So but it was like [9:24] sales manager or about forecasting or [9:27] about how do I do deal reviews? the [9:29] simple things of how do I go through and [9:31] how you know what are the things that I [9:32] can teach an agent to do now I don't [9:35] know if it's even possible this was just [9:36] early brainstorming because I'm no [9:38] expert at this point in time I knew very [9:41] little about LLMs and all those kinds of [9:44] things I was just figuring it out as I [9:45] went because I'm smart it's more because [9:48] I'm stubborn [9:50] so we ended up taking these roles and [9:52] processes and say okay how do we think [9:55] about this how do we go at this and it [9:57] was like immediately a lot of people [9:58] like, oh, go to the LLM, build an agent, [10:01] do all those things. But instead, I stop [10:04] and say, what technology do we already [10:06] have? [10:08] Cuz the issue is AI. Every time we think [10:10] AI, we think about LLMs, think about [10:12] RAG, we think about all this cool stuff, [10:14] right? [10:17] What is really artificial intelligence? [10:20] It's the computer running a process and [10:22] making decisions on our behalf based on [10:25] rules. So, what's an automated workflow? [10:29] it is artificial intelligence and so [10:31] that's where we started. So what we did [10:33] is you know in our crawl is we automated [10:35] the simple stuff and so we we had [10:37] HubSpot and the only thing we used [10:38] HubSpot for was basically to do contacts [10:41] and companies and deals but we really [10:44] used any other tracking for it. So what [10:47] we instead is we said okay lead [10:49] assignment you saw the 6,000 leads how [10:51] do we assign them? How do we make them [10:52] faster? Well, what makes the most sense [10:54] in our business was to actually um route [10:57] those leads to the different salespeople [10:58] by industry. Because if you do that by [11:01] industry, because we cover so many [11:03] industries, if I have a salesperson that [11:05] really understands food and beverage and [11:07] understands the 3A requirements and good [11:10] manufacturing practices and sanitary [11:12] things, they're going to do much better [11:14] making that sale. Whereas somebody who [11:16] understands paint and like explosion [11:18] proof and those kinds of things, it's [11:20] all very different. But what we did is [11:21] we had our sales people specialize and [11:24] then we routed the right leads to them. [11:26] And so the way that we ended up doing [11:27] this was not me going into HubSpot. We [11:29] had an agency, but instead of let's go [11:32] back and forth via email for 4 months. I [11:35] brought them to on site and for a week [11:36] we did agile development. They showed up [11:38] at 8 a.m. told them what we wanted. They [11:41] ran it through my salesperson and my [11:43] marketing person, my my my uh one sales [11:46] rep and my marketing person, I only have [11:48] one, reviewed it with them. that I [11:50] reviewed at the end of the day. We reset [11:52] the strategy and we kept doing that for [11:54] a week. And so we walked out at the end [11:57] very quickly of being able to route [11:59] leads within 3 minutes of a form fill [12:01] and being able to have the initial [12:02] contact of a salesperson actually [12:03] reaching an individual under 24 hours [12:07] from what was that 7.8 days. [12:11] That was just using what I had. [12:15] And how many times that we sit there and [12:16] we think about it when we talk about AI [12:18] that we want to jump to the shiny object [12:20] instead of saying this is what I have in [12:22] house. How do I move forward? So the [12:25] second thing we ended up doing is we [12:26] started doing dynamic nurturing inside [12:28] HubSpot and so this is a newer [12:30] capability that came out with HubSpot [12:32] but basically you start to be able to do [12:34] various paths based on their behavior to [12:36] be able to make those changes. So, we [12:38] sorted leads whether are they a [12:40] marketing qualified lead or a sales [12:41] qualified lead based on the activity [12:43] that they took. And then it's the MQL, [12:46] so the marketing leads that aren't quite [12:48] ready. They're not raising their hand [12:50] going, "Give me a quote." We had a we [12:52] had basically a four-week nurturing [12:54] stream that hit them with multiple [12:56] emails, plus there were calls embedded [12:57] in it. And if they responded back or [12:59] what, depending on what actions they [13:00] took, changed how it routed, changed [13:03] what we did, and then it told the [13:05] salesperson how to move forward next. [13:07] And so, you know, we ended up going to I [13:11] think it's I know it's really hard to [13:12] see, but when we talk about D kind of [13:15] like mixer nurturing, we ended up having [13:17] 21% meetings based on that email nurture [13:20] stream. Now, those of you in marketing, [13:22] that's huge. [13:24] That that is absolutely huge because [13:26] most of the time email response rate is [13:28] about two two to 4%. So being able to [13:31] get that level of meeting rep is is big [13:33] in our business because in our business [13:36] the company that reaches out to you [13:38] first and gets you a quote first usually [13:40] wins. And so that was really important [13:42] to our business. So we we kind of we [13:45] went through HubSpot and we said here [13:47] are all the things we can do and then we [13:48] asked the question I wonder if there's [13:50] any other things we can use inside our [13:52] tech stack. It was pretty simple tech [13:54] stack but we looked and so the other one [13:56] I don't know do we have any call rail [13:57] users here? [13:59] Call rail is it rotates your 1-800 [14:01] numbers on your website that you can [14:03] attach to cost per click. Well, they [14:06] launched a little hidden feature in [14:07] there called convert assist. And so this [14:10] is what it does. And so it listens to [14:12] the phone call. So someone calls you and [14:14] it goes through and it says, "Okay, [14:17] here's your action plan based on your [14:18] phone call. [14:20] Here is kind of the coaching that it [14:22] ends up doing for both the sales rep." [14:24] And then finally, here's the draft email [14:26] for you to be able to to to respond [14:28] right away. Now, you don't take the [14:30] draft email by itself and send it. You [14:33] put in the review. You do the work. You [14:34] do those things. But just even that was [14:37] a big win. And we already had it. The [14:40] toughest part for us was to get it from [14:43] Call Rail into our HubSpot instance [14:45] because we used that as kind of [14:46] everything. We just drew everything back [14:48] into that. We've been able to do that in [14:50] a custom call object and it's really [14:52] made a difference. [14:55] So big takeaway number one is please use [14:58] the technology you already pay for. You [15:01] will come to conferences like this and [15:03] the and sorry people that are AI [15:05] providers, they're going to say, "Oh, [15:08] buy my stuff. It's better. It'll do [15:09] these great things." But use what you [15:11] have first. Get the value for that. This [15:13] created a lot of moment momentum for us [15:16] inside our company to be able to allow [15:18] me to start spending a little bit of [15:19] money. And that's part of when you talk [15:21] about the the real way of how do I get [15:24] an IT project moving is you start always [15:27] focusing on what is the operational and [15:29] what is the business value. The IT value [15:32] or cost savings or cost avoidance ends [15:34] up [15:36] it's nice but it's not going to get a [15:39] company to move off the dime especially [15:41] smaller companies that are watching [15:42] every penny. And so go through your tech [15:46] stack, really understand what's there, [15:49] and then be able to move forward. I know [15:50] it sounds like duh, right? But the thing [15:54] is, a lot of times we forget that. And [15:57] so it's just a good reminder. [16:00] The second thing is we moved to LLM, we [16:02] found out that AIS are the smartest, [16:04] dumbest interns you ever had. [16:08] It's just true. And yes, the image was [16:10] generated by Chad GPT. [16:12] And this is what we end up finding. Now [16:14] I told you we have a very complex [16:15] industry. So like we've got to do [16:18] research on on lithium refinement to go [16:20] into batteries in a manufacturing [16:22] process. And what we got back especially [16:25] when we're looking at JT GBT2 going into [16:28] three is it they were it was dumb. Then [16:32] by the time it got up to like four or [16:34] 03, it had good information, but it [16:36] really it missed the boat because it [16:39] didn't understand, oh, this is what a [16:40] manufacturing plant is and all the [16:42] constraints and then understanding [16:44] computational fluid dynamics and [16:46] understanding all these different [16:47] elements. And so you got to teach, you [16:49] have to teach your intern. You just, [16:51] it's kind of like when you and the [16:52] reason I use that analogy is this. If [16:54] you got a PhD intern that came to work [16:56] in your business, would you hand the [16:58] keys over to him and let him run it? If [17:00] anybody's ever managed interns, that [17:02] would be no. They require supervision. [17:05] They require guidance. They require And [17:07] that's the way you have to think about [17:08] your LLMs. [17:10] And if you don't, you could get yourself [17:12] in a lot of trouble, especially smaller [17:13] companies, cuz it's like, "Oh, yeah, [17:15] this is great." And they just move on [17:16] too quickly without realizing what [17:18] they've done. [17:20] So, we started with our first LLM was [17:23] really about quoting. We had a huge [17:25] issue there. I only had one application [17:26] engineer at the time and he was he was [17:29] he was swamped [17:31] and so it's like what can we do to help [17:33] him and the other problem was we covered [17:35] such a breadth of industries it's [17:36] impossible for one person to know and so [17:39] how do we end up providing the technical [17:40] resources that he needed to do that and [17:43] so we we I ended up finding the [17:45] instructions that's QR code for a custom [17:47] chat GPT and so that's what we created [17:50] so we uploaded you know 20 30 product [17:53] you know engineering and product resour [17:55] sources to build Arlin. I'll tell you [17:57] about the name in a minute. We wrote [17:58] custom instructions and prompts and then [18:00] we really worked with my application [18:03] engineer because there was only one at [18:04] the time so adoption was easy but I [18:06] worked with him a lot and built it to [18:08] really help him and then how here's how [18:10] it fits into we talked a lot about how [18:13] will you use this how will this work in [18:14] your work process those kinds of things. [18:16] So the organizational change management [18:17] elements in order to be able to do it [18:20] the other name and I'll talk about [18:22] Arland for a minute. So the other part [18:24] about organizational change management [18:25] is how do you take the fear out of a new [18:27] tool? And so Arlin, my team along with [18:31] the engineers and all those things, they [18:32] have a weekly Dungeons and Dragons game [18:35] over lunch. I know it sounds weird, but [18:36] it's great. It's a great team building [18:38] exercise, but the dungeon master kept [18:40] telling me about this dumb non-player [18:43] character named Arlin. He was super [18:45] helpful, but he was an idiot. And so [18:48] it's like, okay, but he but he was he [18:50] was nice. He wanted to help. He wanted [18:52] to do those things. And so we just [18:53] decided that's what we're naming them [18:55] Zarlin. So and they're like, "Why do you [18:57] call it Arland?" Well, it's the agent [18:59] for revenue, but it was more just me [19:01] trying to make it up to make it fit. But [19:03] no, they so everybody calls our LLM by [19:06] name. We also created two more LLMs for [19:10] Aspen and uh Aosta. So Aosta is for [19:13] sales. So here's an application [19:16] engineer. This is the prompt we were [19:17] able to use with it for lithium [19:19] batteries. And so we ended up going from [19:21] 5 days to one and a half. [19:24] We we also use it for product research. [19:26] We were able to it this is for our sheer [19:28] pro line. We were able to go from [19:32] was it 10 plus days to an hour to [19:34] actually do our market positioning. [19:37] So the big thing is build, test, iterate [19:39] or kill it. So I'm sorry I got to move a [19:41] little fast. So [19:44] the third thing and this is the third [19:45] lesson. All that glitters is not gold. [19:47] So we tried Jasper, we tried 11x, we [19:49] tried those things and they failed [19:51] terribly. But so this is the process [19:53] we're using and working backwards of be [19:55] clear on the output of the decision, [19:57] define the workflow and then build the [19:59] rolebased LLM and then just keep [20:01] refining. [20:03] And so this is how we're doing for [20:04] marketing now. We're using make.com in [20:06] order to thread it through with the [20:08] final review of the agent. This is still [20:09] under development right now, but I can't [20:11] afford any of the custom tools that are [20:14] out there. So this is me just kind of [20:16] working on it. And so we test it [20:18] manually first. We run all the prompts [20:20] manually. And then what we're doing is [20:21] we're just going to automate it using [20:22] make or zapier or one of those. I'll [20:24] figure out which one's easiest. [20:28] The biggest thing is this. Our jobs are [20:30] changing. [20:32] You have to direct. You cannot delegate [20:34] today. But the other thing is our jobs [20:37] are we have to teach people not to do [20:38] but to delegate. And so how do we end up [20:41] doing that? And how do we think about [20:43] that? [20:46] So this was the result. We actually [20:48] improved lead conversion scores from [20:49] 17.4 to 27.4%. [20:52] It's huge. We also increased deal sizes [20:55] by 20%. [20:56] And so that was in about a year and a [20:58] half. And so it's given me a lot of [21:00] leverage to be able to move forward. [21:03] Looking back there's a framework existed [21:05] if especially operators if you don't [21:08] know if you don't follow AI with Ally [21:10] you should. But this is kind of her [21:12] framework that she has around you start [21:14] with AI as a multitasker, you move to a [21:16] co-pilot, you then delegate, and then [21:18] eventually some at some point you'll [21:20] move to a teammate. We're thinking about [21:21] one as one as a project manager for big [21:23] deals, but that's kind of in the future. [21:28] Here's what I want to leave you with the [21:30] recipe. Get started now. Do not wait. [21:32] It's never going to be you can't start [21:34] early enough. Start small. Try a bunch [21:36] of things and fail fast. [21:39] Incremental improvement is a win. And [21:40] then ultimately you have to call in the [21:42] experts. And that's kind of where I'm [21:43] starting to get to. [21:45] But the fundamental, and this is the big [21:47] takeaway at the end, AI levels the [21:49] playing field. It makes terrible [21:51] performers mediocre. [21:54] It makes great performers more [21:56] efficient. [21:57] And so you need to get started now [22:00] because even your lagards that you [22:02] compete with are going to start catching [22:04] up. And so just move forward and think [22:06] about that. This is what our journey [22:08] was. Every journey is a little bit [22:10] different, but I'm not a technologist [22:13] and so there's a lot of room for [22:14] improvement. It could have been done, [22:15] but this has really worked for us and [22:17] hopefully it'll work for you, too. [22:20] If you want to get a hold of me, feel [22:22] free. Connect on LinkedIn. I'll answer [22:24] any questions outside or if you hit me [22:26] on LinkedIn if you don't have time. Love [22:28] to do that. [22:30] [Applause]