Presentation by Michael Chorey, Founder of FreshAi, Wendy's

Ai4

AI-assisted summary of Presentation by Michael Chorey, Founder of FreshAi, Wendy's — original source. Check the original for full context.

Source published: 2025-08-29

Detailed notes

## Background Published on YouTube by the channel Ai4 on August 29, 2025, this presentation features Michael Chorey, Founder of FreshAI at Wendy's. According to the video description, Chorey shares real-world lessons and a leadership playbook from building and scaling the world's first generative AI drive-thru agent, discussing what worked, what caused friction, and how to build customer-facing AI with soul and emotional resonance.

## TL;DR Michael Chorey recounts the journey of creating FreshAI at Wendy's, transitioning from a scrappy, duct-taped garage pilot into an enterprise-scale generative AI ordering system operating across hundreds of drive-thrus nationwide. Rooted in Wendy's core values of hospitality and dignity, the initiative overcame extreme menu customization complexities and outdoor acoustic challenges by partnering with Google's LLM research team. Chorey details key operational lessons, noting that if starting today, he would accelerate rollouts via autonomous agent workflows and lock down franchisee unit economics upfront rather than deferring them. Addressing concerns about automation, he emphasizes that voice AI does not replace labor, but instead reallocates crew members to deliver warmer, unhurried hospitality at the pickup window.

## Atmosphere & style The presentation carries a warm, reflective, and candidly pragmatic tone, balancing corporate innovation with deeply human storytelling. Chorey establishes an emotional foundation early by recalling his childhood memories of his hardworking mother finding dignity at the drive-thru window, seamlessly connecting heartfelt sentiment to the gritty realities of running extension cords through parking lots. The session takes on the gravity of an emotional milestone when Chorey reveals it is his final day at Wendy's and his first time presenting without his trademark red company jacket, giving his technical advice the resonance of a parting manifesto.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=6j790ra1r-M&t=4s) Customer-facing automated AI must remain grounded in hospitality, empathy, and human dignity rather than solely optimizing quantitative business KPIs. - [02:50](https://www.youtube.com/watch?v=6j790ra1r-M&t=170s) High-tech product innovation requires balancing commercial objectives with emotional resonance, ensuring automated systems support both guests and kitchen crew. - [04:42](https://www.youtube.com/watch?v=6j790ra1r-M&t=282s) A single menu item like Wendy's Dave's Double has over 200 billion customization combinations, creating conversational complexity that breaks rigid rule-based bots and necessitates frontier LLMs. - [06:01](https://www.youtube.com/watch?v=6j790ra1r-M&t=361s) Project Kittyhawk began as an agile, garage-built pilot and became the first external enterprise deployment permitted to run Google's proprietary frontier LLMs in the field. - [08:34](https://www.youtube.com/watch?v=6j790ra1r-M&t=514s) Project Armstrong scaled FreshAI into the quick-service industry's first enterprise GenAI voice agent, serving tens of millions of customers across hundreds of locations nationwide. - [10:17](https://www.youtube.com/watch?v=6j790ra1r-M&t=617s) Transitioning from a 4-person startup team to enterprise-wide adoption introduced organizational drag, stakeholder governance, and approval cycles that slowed deployment velocity. - [11:43](https://www.youtube.com/watch?v=6j790ra1r-M&t=703s) Modern AI rollouts should leverage autonomous agents across every business function and adopt rapid 'vibe coding' to compress deployment cycles from years to months. - [13:08](https://www.youtube.com/watch?v=6j790ra1r-M&t=788s) Deferring franchise fee structures and financial models created significant friction; sustainable enterprise AI adoption requires resolving unit economics from day one. - [15:02](https://www.youtube.com/watch?v=6j790ra1r-M&t=902s) Generative AI initiatives should directly target the company's core revenue engine—such as Wendy's drive-thru representing 80%–85% of sales—rather than peripheral experiments. - [17:08](https://www.youtube.com/watch?v=6j790ra1r-M&t=1028s) FreshAI was designed for labor reallocation rather than workforce reduction, shifting staff away from headset order-taking to focus on order accuracy and genuine hospitality at the pickup window. - [19:13](https://www.youtube.com/watch?v=6j790ra1r-M&t=1153s) AI ordering delivers conversational consistency and eliminates dead air, enabling customers to complete their orders in 55 seconds without feeling rushed or interrupted. - [20:22](https://www.youtube.com/watch?v=6j790ra1r-M&t=1222s) Customer sentiment surveys show a clear adoption tipping point by the second visit, as drivers realize the AI agent offers a patient, unhurried ordering interaction.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=6j790ra1r-M&t=4s) Finding Soul and Dignity in the Drive-Thru

Michael Chorey introduces his talk on "the soul in the drive-thru," tracing the origins of his perspective back to his childhood in the late 1980s and early 1990s rather than starting in a corporate boardroom. Back then, his mother worked double shifts on the hard concrete and sweltering heat of his grandparents' convenience store drive-thru. At the end of exhausting days, when she was too tired to cook dinner, she would pick up seven- or eight-year-old Chorey and say, "Let's go out." Because money was tight, going out meant pulling through a McDonald's or Wendy's drive-thru to split an inexpensive meal or get a kids' meal and a basic burger.

Beyond the special bonding time with his mother, one vivid detail stayed with Chorey: a drive-thru worker who consistently greeted his mother as "mama" or "mom." While the young Chorey puzzled over why a stranger would call his mother that, growing into adulthood revealed the true meaning of the gesture. After spending all day serving others on her feet, the drive-thru became a rare moment where someone took the time to serve, acknowledge, and respect her. The worker didn't merely thrust a bag through the window or aggressively upsell expensive items she couldn't afford; they treated her as a human being.

Chorey emphasizes that every interaction in quick-service restaurants (QSR) or retail holds this potential for human connection. While modern enterprises obsess over quantitative key performance indicators and return on investment, Chorey argues that the heart of any customer-facing technology must remain grounded in hospitality and human dignity.

[02:50](https://www.youtube.com/watch?v=6j790ra1r-M&t=170s) Grounding Fast Food Innovation in Hospitality and Core Values

Approaching high-tech product design requires understanding the fundamental "why" behind human behavior rather than rushing to bare-minimum automated outcomes. Chorey notes that ordinary people lead busy, demanding lives and do not spend their days deeply contemplating fast-food visits. When a customer decides to stop by a drive-thru or a retail store, the last thing technology should do is disrupt their routine or make their day harder. The goal of product development must be to uncover and protect the most meaningful micro-moments of hospitality within that brief encounter.

Drawing on his early product career at L Brands and Victoria's Secret, as well as his tenure at Wendy's, Chorey explains that successful innovation requires balancing commercial objectives with emotional resonance. Delivering strong business ROI should never come at the expense of customer and employee experience. Guided by Wendy's founder Dave Thomas and his foundational core values, the guiding principle for deploying automated technology must always be to "do it with dignity" for both the guest at the speaker post and the crew member inside the kitchen.

[04:42](https://www.youtube.com/watch?v=6j790ra1r-M&t=282s) The 200-Billion-Way Dave's Double: Why Drive-Thrus Defy Traditional AI

Three years prior, Chorey's team set out to answer a radical question: *Could a large language model (LLM) take drive-thru orders in real time?* At that time, nearly every existing automated voice solution on the market relied on rigid, traditional rule-based trees and deterministic dialog flows. While ordering fast food might seem deceptively simple to an individual consumer who orders the same item every visit, the mathematical and linguistic reality behind the menu is staggering.

Taking Wendy's signature Dave's Double as an example, there are over 200 billion possible ways to customize the sandwich. When factoring in the messy nuances of natural human speech—such as expressing the same preference through "no lettuce," "remove the lettuce," or "I don't want any of the green stuff"—traditional deterministic flowcharts break down exponentially. Recognizing these limitations long before ChatGPT and GPT-3 entered the mainstream public consciousness, Chorey's team partnered directly with Google's LLM research teams to explore whether frontier language models could handle the chaotic domain of restaurant ordering.

```mermaid flowchart TD accTitle: Evolution of Drive-Thru Ordering Intelligence subgraph Traditional["Traditional Voice Bots"] A["Rigid Rule-Based Trees"] --> B["Fails on Conversational Nuance"] B --> C["Breaks under Menu Permutations"] end

subgraph GenAI["Wendy's FreshAI LLM Approach"] D["Frontier Large Language Models"] --> E["Natural Language Understanding\ne.g., 'No green stuff'"] E --> F["Dynamic Customization Handling\n>200B Dave's Double Variants"] end ```

[06:01](https://www.youtube.com/watch?v=6j790ra1r-M&t=361s) Project Kittyhawk: Duct Tape, Garage Prototypes, and Google Research

The initial pilot was dubbed "Kittyhawk"—named after the site of the Wright brothers' pioneering flight in North Carolina, honoring both Ohio aviation history and the audacious, fragile nature of building something never attempted before. At the time, Wendy's became the first external enterprise permitted to run Google's proprietary LLMs outside Google's own walls. Because LLM infrastructure was not yet standardized, the team had to solve countless baseline challenges from scratch: live audio streaming, speech-to-text accuracy in noisy outdoor conditions, prompt orchestration, transcript governance, and legal compliance.

Operated like an agile, scrappy startup, the original FreshAI team consisted of Chorey and a handful of colleagues building hardware and software prototypes in Chorey's home garage, located just two miles from the first test restaurant. The team physically ran extension cords across the restaurant parking lot to power their equipment, observing live ordering sessions for a few hours at a time before retreating to iterate on the software overnight. Crucially, the engineers lived inside the restaurant alongside the crew members, gathering immediate feedback to ensure the system removed stress from kitchen workers rather than imposing friction on their shifts.

- **Codename:** "Kittyhawk", signifying the fragile, high-risk first flight of generative voice AI. - **Proximity:** Prototyped in a residential garage two miles from the initial pilot store. - **Pioneering Access:** First third-party enterprise implementation of Google's frontier LLMs in the field.

[08:34](https://www.youtube.com/watch?v=6j790ra1r-M&t=514s) Project Armstrong: Scaling from a Garage Pilot to Tens of Millions of Orders

Following the fragile breakthroughs of Kittyhawk, the initiative entered its scale phase, codenamed "Armstrong"—symbolizing the giant leap from a single rickety test site to mission-grade nationwide deployment. Moving from store one to store two and three presented intense operational hurdles, but the compounding learnings allowed deployment velocity to accelerate rapidly from bespoke experiments to dozens of store rollouts per month.

FreshAI evolved into the world's first enterprise-scale generative AI ordering agent in the quick-service industry. Operating across the majority of US states in hundreds of restaurants—with hundreds more queued for installation—the agent handles conversational ordering for hundreds of thousands of diverse real-world customers daily, spanning Gen Z, elderly guests, rural dialects, and urban slang. What began as a duct-taped garage experiment three years earlier transformed into a foundational pillar of Wendy's public enterprise growth strategy.

- **Scale:** Expanded from a single test store to rolling out across dozens of locations per month. - **Footprint:** Deployed across the majority of US states in hundreds of locations, with hundreds more planned. - **Volume:** Serves tens of millions of live customers overall, processing hundreds of thousands of voice orders daily across diverse demographics.

[10:17](https://www.youtube.com/watch?v=6j790ra1r-M&t=617s) The Friction of Scale: From Lean Innovation to Enterprise Operations

As FreshAI transitioned from a scrappy pilot into a live enterprise system, the team faced the inevitable growing pains of organizational scaling. In the earliest days, a small core team of just three or four people could handle everything directly, from ideation to restaurant-level training. However, expanding across the footprint required training tens of thousands of crew members and serving hundreds of thousands of live drive-thru customers every single day. With that explosion in footprint came a surge of new stakeholders, expanding operational layers, and an influx of meetings, approval cycles, and governance hurdles.

Reflecting on that journey, Chorey acknowledges that navigating this enterprise bureaucracy noticeably slowed their momentum. Instead of purely dreaming big and tackling exciting engineering challenges, the team became consumed with operational alignment and consensus-building. Yet, despite the organizational friction that scale introduces—whether in bleeding-edge generative AI or a traditional product rollout—the system did not collapse. The platform pushed through, continuing to expand and operate reliably under intense real-world demand.

- **Scale footprint**: Scaled from single-store pilots to tens of thousands of trained crew members serving hundreds of thousands of drive-thru customers daily. - **Enterprise drag**: Transitioned from a 3-to-4-person rapid deployment team to complex multi-stakeholder governance and approval loops.

---

[11:43](https://www.youtube.com/watch?v=6j790ra1r-M&t=703s) Rewriting the Playbook: Agents Everywhere, Vibe Coding, and Velocity

Looking back at the three-year journey to bring FreshAI to market, Chorey admits that if he were to start over today, he would discard the original playbook entirely: "What we did at the beginning of this story, I would not do again." While Wendy’s remains the industry leader in generative AI drive-thru technology, a three-year timeline is "far too slow" in the current technological landscape. In the early days, modern acceleration tools did not exist—there was no ChatGPT to resolve queries instantly, nor could developers "vibe-code" lean prototypes and live demos overnight.

If building from scratch today, Chorey would prioritize speed, bold intentionality, and end-to-end agentic workflows across every business function. Rather than relying solely on human-heavy manual processes for operational rollouts, autonomous agents should handle early research, operational logistics, and employee onboarding from day one.

```mermaid flowchart TD accTitle: FreshAI Playbook Evolution subgraph OldApproach["Original 3-Year Linear Rollout"] A["Lean 4-Person Pilot"] --> B["Manual Ops & Training"] B --> C["Delayed Economic Modeling"] C --> D["Multi-Year Enterprise Scale"] end subgraph AgenticApproach["Modern Accelerated Blueprint"] E["Agentic Research & Vibe Coding"] --> F["AI-Driven Ops Training"] F --> G["Upfront Unit Economics"] G --> H["Rapid Multi-Unit Deployment"] end ```

---

[13:08](https://www.youtube.com/watch?v=6j790ra1r-M&t=788s) Confronting the Economics Early and Preserving Product "Vibe"

A major hard-learned lesson from the rollout was the danger of deferring tough business decisions. Although franchisees and corporate stakeholders were energized by the vision and consumer story, franchise adoption ultimately hinges on bottom-line financial performance. The team initially operated with only a rough estimate of the fee structure and franchise pricing model, adopting a mindset of "we'll figure it out at the end." Chorey warns that pinning down the economic model late in the lifecycle proved to be one of the most difficult hurdles of the entire program, making early financial viability a mandatory first-order priority for any future deployment.

Simultaneously, engineering leaders must obsess over preserving the product's emotional core—its personality, warmth, and feel—as it scales into standard operational procedure. Scaling must not dilute the "vibe" or soul that made the product resonate with customers in early trials.

- **Franchise realities**: High-level vision excites stakeholders, but sustainable scale requires immediate, proven bottom-line ROI. - **Economic timing**: Deferring franchisee fee structures and financial models created significant friction; unit economics must be resolved upfront.

---

[13:38](https://www.youtube.com/watch?v=6j790ra1r-M&t=818s) A Milestone Departure and the Future of Hospitality AI

Chorey marks a deeply personal turning point in his career during the presentation. Known across the industry for his signature red Wendy’s jacket, he points out that this talk is his first time appearing without it: "My journey at Wendy's actually ends today. Today is my last day with Wendy's." After five total years at the company and three years spent nurturing FreshAI "like a child," he is stepping down to pursue his next venture within the hospitality sector.

Looking across the broader food service landscape, Chorey predicts that generative AI drive-thrus and hospitality agents will rapidly transition from novel, headline-grabbing innovations into an everyday standard across the sector. As he embarks on his next chapter, he plans to practice the very lessons shared in this talk—delivering not merely operational outcomes, but true "presence at scale" that treats customer and employee interactions with dignity.

---

[15:02](https://www.youtube.com/watch?v=6j790ra1r-M&t=902s) Starting Bold: Transforming the 85% Drive-Thru Core

Chorey urges leaders deploying AI never to relegate generative models to peripheral, low-stakes experiments. At Wendy's, the team deliberately targeted the core engine of the business: the drive-thru, which accounts for 80% to 85% of total sales.

Rather than settling for small, safe pilots, innovators should pursue transformative, high-impact bets with robust guardrails that create tangible value for customers, frontline crew, and shareholders alike.

- **High-stakes deployment**: Rather than testing in low-traffic channels, FreshAI targeted the 80%–85% of company revenue running through the drive-thru. - **Impact formula**: True enterprise AI must simultaneously elevate three groups: guests, crew members, and bottom-line shareholders.

---

[16:12](https://www.youtube.com/watch?v=6j790ra1r-M&t=972s) Q&A: The GCP Partnership and Hybrid Architecture

During the audience Q&A, Tomas from Triabs AI consultancy asked how Wendy’s navigated the technical challenges of building and adapting their generative AI product on Google Cloud Platform (GCP).

Chorey explained that Wendy's pursued a unique hybrid model, entering a high-profile strategic partnership with Google from the earliest days of enterprise LLMs as one of their flagship generative AI implementations. While they occasionally encountered architectural headaches and platform limitations as early adopters, Google worked closely with them to refine and mature the platform capabilities over time.

---

[17:08](https://www.youtube.com/watch?v=6j790ra1r-M&t=1028s) Q&A: Rooted in Dave Thomas’s 1969 Legacy, Not Workforce Reduction

An audience member asked whether the decision to use LLMs was primarily driven by labor reduction or simply chasing the "cool factor" of new technology. Chorey firmly pushed back, explaining that FreshAI was never intended as a workforce reduction strategy. In fact, Wendy's founder Dave Thomas invented the modern drive-thru in 1969, establishing a long heritage of service innovation that the team felt obligated to honor.

Rather than merely swapping out a human voice for a synthetic bot, the project leveraged large language models because natural language understanding excels at parsing complex conversational data. By letting AI handle conversational order capture, restaurant labor is reallocated toward order assembly and the pickup window. This eliminates the hurried, impersonal handoff where a bag is shoved through a window, freeing crew members to make eye contact, smile, and offer a genuine human thank-you.

- **Drive-thru history**: Dave Thomas created the modern drive-thru concept in 1969; FreshAI was built to modernize that legacy rather than strip out human connection. - **Labor reallocation**: Voice ordering shifts labor from wearing a headset to improving food preparation accuracy and enhancing guest hospitality at the pickup window.

---

[19:13](https://www.youtube.com/watch?v=6j790ra1r-M&t=1153s) Q&A: Overcoming Machine Resistance with Warmth and Consistency

A participant raised the concern that talking to an automated voice feels impersonal, noting that they had previously driven away from an AI-powered drive-thru specifically to avoid speaking to a machine. Chorey acknowledged that conversational quality varies widely across the industry, but defended a well-crafted system as vastly superior in consistency and patience.

To illustrate the difference, Chorey referenced two real customer videos recorded during Wendy's recent promotional campaign for the Netflix series *Wednesday*:

- **Traditional lane without FreshAI**: A customer waited in total silence at the speaker post for 55 seconds before an overwhelmed worker finally greeted them with a rushed "Hello." - **Lane powered by FreshAI**: The system instantly and warmly greeted the vehicle the moment it reached the detection loop: *"Hello, welcome to Wendy's. What can I get for you today?"*

Chorey highlighted that a properly tuned AI never rushes the customer, provides absolute conversational consistency, and eliminates painful dead air, delivering a consistently welcoming brand experience.

[20:22](https://www.youtube.com/watch?v=6j790ra1r-M&t=1222s) Consistency, Human Connection, and Customer Adoption

Illustrating the tangible difference in the drive-thru experience, Michael Chorey highlights a side-by-side scenario: within 55 seconds, a customer placing their order with the AI agent is completely finished and pulling forward toward the pickup window, while a driver in the adjacent traditional lane is left sitting at the speaker box, still waiting for a crew member to acknowledge them. This stark contrast highlights the core value proposition of deploying conversational AI in quick-service restaurants: unmatched consistency and immediate customer confidence that an order will be taken without delay.

Far from stripping away the human element, automating the initial order-taking at the speaker post shifts human labor to where genuine connection actually matters—at the physical pickup window. Relieved from the stress of juggling headsets and multi-tasking under pressure, crew members have the breathing room to deliver authentic hospitality: offering a warm greeting, saying thank you, and checking in on regular customers with a personalized *"how have you been?"* By recovering those critical seconds, the face-to-face handoff transforms into a far more meaningful, relaxed interaction.

This operational shift is backed by hard data and nationwide customer sentiment surveys, which demonstrate that guests broadly prefer interacting with the AI drive-thru agents over traditional ordering channels. This appreciation is especially pronounced by a customer's second visit; once someone experiences the system a second time, the value proposition immediately clicks. Customers realize they are no longer rushed through their menu choices or interrupted mid-sentence, enjoying a seamless, unhurried ordering experience.

```mermaid flowchart TD accTitle: Drive-Thru AI Ordering Flow vs Human Window Connection subgraph OrderCapture["Speaker Post (AI Agent)"] A["Immediate Greeting & Order"] --> B["55s Completed Order"] end

subgraph WindowService["Pickup Window (Human Crew)"] B --> C["Driver Pulls Forward"] C --> D["Meaningful Interaction: 'Hello', 'Thank you', Regular Check-in"] end ```

- **55-Second Completion:** Customers using the AI agent are fully wrapped up and moving to the window within 55 seconds, eliminating the friction of waiting unacknowledged at the speaker. - **Enhanced Human Touchpoints:** Crew members freed from headset-juggling can invest quality time into window interactions, greetings, and building rapport with regulars. - **The "Second Visit" Tipping Point:** Survey data reveals customer preference spikes after the second visit, as drivers realize they won't be rushed, interrupted, or bounced around while ordering.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=6j790ra1r-M&t=4s) Let's go out. — Michael Chorey's mother

> [02:50](https://www.youtube.com/watch?v=6j790ra1r-M&t=170s) do it with dignity — Michael Chorey

> [04:42](https://www.youtube.com/watch?v=6j790ra1r-M&t=282s) Could a large language model (LLM) take drive-thru orders in real time? — Michael Chorey

> [11:43](https://www.youtube.com/watch?v=6j790ra1r-M&t=703s) What we did at the beginning of this story, I would not do again. — Michael Chorey

> [11:43](https://www.youtube.com/watch?v=6j790ra1r-M&t=703s) far too slow — Michael Chorey

> [13:08](https://www.youtube.com/watch?v=6j790ra1r-M&t=788s) we'll figure it out at the end. — Michael Chorey

> [13:38](https://www.youtube.com/watch?v=6j790ra1r-M&t=818s) My journey at Wendy's actually ends today. Today is my last day with Wendy's. — Michael Chorey

> [13:38](https://www.youtube.com/watch?v=6j790ra1r-M&t=818s) presence at scale — Michael Chorey

> [19:13](https://www.youtube.com/watch?v=6j790ra1r-M&t=1153s) Hello, welcome to Wendy's. What can I get for you today? — FreshAI Agent

> [20:22](https://www.youtube.com/watch?v=6j790ra1r-M&t=1222s) how have you been? — Michael Chorey

## Mentioned Michael Chorey, FreshAi, Wendy's, Dave Thomas, Dave's Double, Google, Google Cloud Platform (GCP), ChatGPT, GPT-3, Project Kittyhawk, Project Armstrong, L Brands, Victoria's Secret, McDonald's, Tomas, Triabs AI, Netflix, Wednesday

## Video description Michael Chorey, Founder, Wendy's

I built and scaled the world’s first generative AI drive-thru agent—serving tens of millions of live customers in one of the most complex, high-pressure environments imaginable. What started as fragile, duct-taped pilots became a platform that redefined how fast food operates. But here’s the truth: if I had to do it again, I’d do it differently.

In this talk, I share what I learned from scaling GenAI in the real world—what worked, what slowed us down, and what I’d rebuild from day one. From using agents across every part of the business to designing for emotional resonance ("vibe-code"), I’ll walk through how to move faster, scale smarter, and build with soul. This isn’t a product story—it’s a leadership playbook for anyone deploying AI that has to feel as good as it functions.

[Source](https://www.youtube.com/watch?v=6j790ra1r-M)

中文简报

Wendy's FreshAI 创始人 Michael Chorey 演讲 Michael Chorey 回顾了在 Wendy's 打造 FreshAI 的历程，讲述了它如何从一个简陋拼凑的车库试验项目，演进为在全美数百家汽车穿梭餐厅运行的企业级生成式 AI 点餐系统。该项目根植于 Wendy's 待客之道与尊重员工的核心价值观，通过与 Google 的 LLM 研究团队合作，克服了极高的菜单定制复杂性以及严苛的户外声学挑战。Chorey 详细阐述了关键的运营经验，指出如果放在今天启动，他会通过自主智能体工作流来加快推广节奏，并在前期就确定加盟商的单店经济模型，而非延后处理。针对关于自动化的担忧，他强调语音 AI 并非取代人力，而是将员工重新分配至取餐窗口，从而提供更温暖、更从容的待客服务。面对 Wendy's 单款汉堡 Dave's Double 超过 2000 亿种定制组合与复杂的口语表达，传统规则引擎彻底失效，团队由此携手 Google 探索前沿大模型在穿梭餐厅的实时点餐落地。 (4:42) FreshAI 从车库原型“Kittyhawk 项目”起步，成为业内首个获准运行 Google 专有前沿大模型的外部企业部署，并随后通过“Armstrong 项目”扩展为覆盖全美数百家门店、服务数千万顾客的企业级生成式 AI 语音智能体。 (6:01) 从 4 人敏捷团队扩展到企业级推广带来了繁复的跨部门治理与审批摩擦；创始人 Chorey 总结称，若在当下重新启动该项目，应在所有业务环节全面采用自主 Agent 与“vibe coding”，将交付周期从数年大幅压缩至数月。 (11:43) 推迟制定特许经营费结构与财务模型给后期规模化带来了显著阻力，企业级 AI 部署必须自第一天起明确单店经济模型，同时在扩张中守住产品的情感核心与人情味（vibe）。 (13:08) 部署穿梭餐厅生成式 AI 并非为了削减人工，而是将员工从繁重的耳麦点餐中解脱出来，将人力重新配置到取餐窗口以提供更从容、真诚的人性化服务与餐品制作。 (17:08)

Reading summary

Michael Chorey argues that automating Wendy's drive-thru with generative AI succeeds only when high-stakes technical innovation is anchored in customer dignity, labor reallocation, and upfront economic viability rather than simple workforce reduction. Overcoming Menu Complexity with Generative Models (4:42) Traditional quick-service voice automation relied on deterministic, rule-based decision trees that consistently failed in real-world drive-thru environments. While fast-food ordering appears straightforward, menu permutations create vast linguistic complexity. For example, Wendy's signature Dave's Double burger has more than 200 billion possible customization combinations, compounded by unpredictable natural phrasing such as substituting 'no lettuce' with 'remove the green stuff.' Recognizing that rigid flowcharts could not accommodate conversational nuance, Chorey’s team partnered directly with Google’s research division before widespread commercial adoption of large language models. Operating as the first external enterprise running Google's frontier models outside the company's research environment, the project tested whether natural language processing could handle live audio streaming, ambient noise, and dynamic menu customization in real time. Scaling from Garage Prototype to National Deployment (6:01) The pilot, codenamed 'Kittyhawk,' originated as an agile hardware and software operation run from Chorey's garage two miles from the pilot store. Engineers ran extension cords across restaurant parking lots to test prototype systems in live environments, working directly alongside kitchen staff to ensure the technology reduced worker strain instead of introducing new operational bottlenecks. Following initial testing, the rollout shifted to Project Armstrong, expanding deployment velocity from bespoke single-store setups to dozens of locations per month. FreshAI evolved into an enterprise system operating across hundreds of Wendy's restaurants in the majority of US states, processing hundreds of thousands of daily voice transactions and serving tens of millions of customers across diverse demographics and regional speech patterns. Reallocating Drive-Thru Labor to Human Hospitality (17:08) Chorey emphasizes that FreshAI was deliberately targeted at Wendy's primary commercial engine—the drive-thru channel, which generates 80% to 85% of total sales—rather than low-risk peripheral operations. Grounded in founder Dave Thomas’s 1969 legacy, the initiative was explicitly framed not as a headcount reduction mechanism, but as a labor reallocation strategy. Removing the burden of continuous headset monitoring allows crew members to focus on order preparation accuracy and direct human engagement at the pickup window. Operational comparisons demonstrated that an AI lane could complete an entire order and move the vehicle forward within 55 seconds, whereas overwhelmed staff in conventional lanes sometimes left drivers waiting unacknowledged at the speaker post for the same duration. Chorey reported that customer sentiment surveys showed an overall preference for the automated system, with satisfaction climbing sharply on second visits once guests realized the conversational agent would not interrupt or rush their decisions. Enterprise Drag and Upfront Unit Economics (13:08) The transition from a lean, three-to-four-person garage team to an enterprise-wide rollout exposed significant organizational friction. Training tens of thousands of restaurant crew members and navigating corporate governance, multi-stakeholder approvals, and consensus-building noticeably slowed deployment momentum. Chorey noted that in the current development landscape, a three-year development cycle is obsolete, advocating instead for agentic workflows and rapid prototyping from inception. A primary operational mistake was deferring franchise economic modeling. The development team initially prioritized technical execution while postponing the definition of franchise fee structures and financial terms. Chorey cautioned that leaving unit economics to the end generated acute commercial friction, making immediate, proven financial viability a critical requirement for any future franchise technology rollout. Marking his departure from Wendy's to launch a new venture, Chorey concluded that conversational agents will quickly transition from novel experiments to standard industry infrastructure. However, the operational lessons from FreshAI indicate that scaling enterprise AI depends as much on resolving franchise unit economics and organizational inertia as on model performance.

精炼阅读摘要

FreshAI 创始人 Michael Chorey 阐述了 Wendy's 利用大语言模型重塑汽车穿梭餐厅（Drive-Thru）点餐的历程，主张面向消费者的 AI 部署应根植于服务温度与人文尊严，而非单纯追求削减人力成本或追踪量化指标。技术初心：从个人记忆到尊重人性 (0:04) Chorey 回顾了童年时期母亲在便利店汽车穿梭窗口长时间轮班的经历。每次疲惫下班后带他去快餐店时，一位员工亲切称呼其母亲为“妈妈”，这份来自陌生人的尊重让他深刻体会到，快餐服务窗口不仅是商业交易场所，更是展现人文关怀的微小契机。他认为现代企业往往过度沉迷于投资回报率和关键绩效指标，但客户面对的技术核心必须始终建立在待客之道与人格尊严之上。顾客在日常忙碌的生活中不应被生硬的技术打乱节奏，自动化产品的设计目标应是发掘并守护这些短暂互动中的温暖微时刻，遵循 Wendy's 创始人 Dave Thomas 的核心价值观，让顾客与后厨员工都获得体面对待。技术破局：语言模型征服海量定制 (4:42) 在 ChatGPT 和 GPT-3 广泛普及的三年前，Chorey 团队便着手探索大语言模型在点餐场景的实时落地。当时市场上的语音方案多依赖死板的规则树与确定性对话流，但在快餐场景中，单是 Wendy's 经典的 Dave's Double 汉堡就存在超过 2000 亿种可能的定制组合。面对自然口语中五花八门的表达方式（例如将“不加生菜”说成“不要任何绿色的东西”），传统流程图会彻底失效。为此，团队与谷歌的大语言模型研究团队达成合作，成为首个获准在谷歌外部运行其专有模型的第三方企业，从零攻克户外嘈杂环境下的语音转文字、流式音频传输以及合规治理等底层工程挑战。工程演进：从车库原型到全美规模部署 (6:01) 该项目最初代号为“Kittyhawk”，Chorey 与几名成员在距测试门店仅两英里的自家车库里搭建软硬件原型，甚至将延长线穿过停车场为设备供电。工程师与店内员工并肩工作，以确保系统减轻后厨压力。随后项目进入“Armstrong”规模化阶段， FreshAI 发展为快餐业首个企业级生成式 AI 点餐代理。该系统已覆盖美国大部分州的数百家门店，每月拓展数十家，累计服务数千万顾客，每天处理数十万涵盖各年龄段与不同方言的语音订单。随着规模扩张，早期 3 至 4 人的精简团队转变为涉及数万名员工培训和复杂多层级审批的企业架构，虽然组织流程拖慢了推进速度，但系统在高负荷真实需求下保持了稳定运行。重构策略：早期经济模型与敏捷开发 (11:43) 回顾三年历程，Chorey 直言如果今天重新起步，绝不会沿用原有的漫长路径。在当前技术环境下，三年周期过于缓慢，如今可以利用“vibe coding”和全链路自主代理来加速早期的运营培训与逻辑构建。更关键的教训在于商业模式的尽早确立：团队初期因推迟明确加盟商定价与收费结构，抱着“最后再解决”的心态，导致后期推广面临极高阻力。他强调，任何新技术的大规模落地都必须在一开始就厘清单元经济效益。此外，Chorey 在演讲中宣布这是他代表 Wendy's 的最后一天，结束了五年的公司生涯及三年的 FreshAI 孵化工作，但他预计生成式 AI 穿梭餐厅将很快从行业创新转变为全行业的日常标配。服务重构：释放双手以回归人工温度 (15:02) Chorey 反驳了将 AI 视为削减工时或单纯追逐噱头的观点，指出项目直面贡献全店 80% 至 85% 营收的穿梭餐厅核心场景。AI 自动化点餐的核心价值在于劳动力重新分配：员工无需长时间佩戴耳机在嘈杂中分心记录订单，从而能将精力集中在餐品精准制作以及取餐窗口的面对面服务上。现场对比数据显示，在传统车道顾客于对讲机前死等 55 秒才听到仓促问候时，AI 车道的顾客已在 55 秒内完成点餐驶向取餐窗口。这一转变让员工有充裕时间微笑、致谢并与熟客寒暄。调研表明，顾客在第二次体验后对 AI 穿梭餐厅的偏好度显著提升，普遍认可其稳定、从容且不被打断的点餐体验。 Chorey 总结认为，企业级生成式 AI 必须同时为顾客、一线员工与股东创造价值。尽管 FreshAI 证明了前沿模型在高并发交易中的可行性，但如何在规模化扩张中平衡加盟商的底线经济诉求，并在技术渗透全行业的同时守住最初的人文温度，仍是快餐智能化转型中需持续面对的命题。

Detailed notes — zh

## Background 该内容由 Ai4 于 2025 年 8 月 29 日发布在 YouTube 平台。页面简介显示，温蒂汉堡（Wendy's）FreshAI 创始人迈克尔·乔里（Michael Chorey）在演讲中分享了打造并规模化全球首个生成式 AI 汽车穿梭餐厅智能体的实战经历，复盘了其中的成功经验与推进瓶颈，并提出了一套兼顾业务赋能与情感共鸣（vibe-code）的 AI 领导力落地指南。 ## TL;DR 迈克尔·乔里从童年时在汽车穿梭窗口感受到的“尊严与款待”出发，阐述了将生成式 AI 引入温蒂汉堡点餐场景的产品哲学。团队从车库手搭原型的“小鹰号”起步，携手 Google 攻克了户外降噪与海量个性化点餐语言理解难题，最终实现了全美数百家门店的大规模落地与数千万人次的服务。迈克尔深入反思了规模化带来的组织流程阵痛，主张未来 AI 项目应推行端到端全面 Agent 化并尽早确立财务模型；同时他在演讲当天宣布从温蒂离职，强调 AI 的核心价值在于解放一线员工精力、重塑服务业的人性温度。 ## Atmosphere & style 整场演讲兼具深情的人文底色与硬核务实的极客精神，开篇关于母亲与尊严的童年往事真挚动人，随后迅速切入高压工程实战与车库拓荒的敏捷历程。迈克尔的分享坦诚犀利，毫无保留地剖析了大公司官僚流程与规模化推进中的真实阵痛。当他现场脱去标志性的红夹克、宣布当天是自己在温蒂的最后一天并展望下一代 AI 变革时，现场氛围达到了兼具释怀告别与雄心远瞻的高潮。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=6j790ra1r-M&t=4s) 商业交互的核心是尊重与尊严：汽车穿梭窗口的微小交互本质上是对顾客尊严的确认，技术与商业追求 ROI 必须与人性化关怀和款待精神深度结合。 - [02:50](https://www.youtube.com/watch?v=6j790ra1r-M&t=170s) 以人为本的产品设计哲学：顺应温蒂汉堡创始人 Dave Thomas 的价值观，AI 系统研发不应只满足于最低交付，而要切实解决用户痛点并带着尊严去服务。 - [04:42](https://www.youtube.com/watch?v=6j790ra1r-M&t=282s) 打破规则系统的技术破局：面对单款汉堡超过 2000 亿种配置组合及千奇百怪的自然口语长尾表达，温蒂汉堡携手 Google 成为全球首个将 LLM 落地于外部真实复杂业务的企业。 - [06:01](https://www.youtube.com/watch?v=6j790ra1r-M&t=361s) “小鹰号”车库精益拓荒：团队在车库手工组装设备并拉线至单店实测，攻克户外降噪收音与调用合规难题，通过与一线店员深度共创确保系统能实质减轻员工负担。 - [08:34](https://www.youtube.com/watch?v=6j790ra1r-M&t=514s) “阿姆斯特朗”全美规模化：FreshAI 从单店 POC 跃升为覆盖全美数百家门店、日常服务数十万人次的成熟系统，成为推动温蒂汉堡整体业务增长的核心引擎。 - [10:17](https://www.youtube.com/watch?v=6j790ra1r-M&t=617s) 规模化带来的组织扩张阵痛：随着业务触达数万名一线员工并接待海量顾客，跨部门官僚流程与繁琐审批在后期明显拖慢了原本敏捷的极客开发节奏。 - [11:38](https://www.youtube.com/watch?v=6j790ra1r-M&t=698s) 全面 Agent 化与早期锁定商业模式：复盘 3 年历程认为推进依然偏慢，下一代 AI 产品应将 Agent 渗透至研发与运营全流程，且必须在项目初期尽早解决加盟商经济模型与定价难题。 - [13:38](https://www.youtube.com/watch?v=6j790ra1r-M&t=818s) 直面核心业务命脉：汽车穿梭餐厅贡献了温蒂汉堡 80% 至 85% 的营收，打造颠覆性 AI 体验绝不能从小修小补入手，而要敢于直接攻坚高并发、高压力的核心腹地。 - [16:12](https://www.youtube.com/watch?v=6j790ra1r-M&t=972s) 引入 LLM 的初衷并非削减用工成本：技术创新的核心目的在于重塑服务体验，将劳动力从机械的点餐沟通中解放出来，重新分配至餐品制作与出餐环节。 - [19:13](https://www.youtube.com/watch?v=6j790ra1r-M&t=1153s) 秒级响应与消除催促感：相比高峰期传统门店让顾客在麦克风前长时间空等，FreshAI 能提供零延迟、标准且热情的迎宾交互，彻底消除了人工催促带来的压迫感。 - [20:22](https://www.youtube.com/watch?v=6j790ra1r-M&t=1222s) 赋能取餐窗口的人际温度：AI 接管重复点餐后，释放了员工在取餐窗口的时间与精力，让一线员工能从容地与顾客真诚交流寒暄，重塑真实的人情关怀。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=6j790ra1r-M&t=4s) 汽车餐厅里的“灵魂”与童年关于尊严的记忆迈克尔·乔里（Michael Chorey）以“汽车餐厅里的灵魂”（the soul in the drive-thru）为主题展开了演讲。作为温蒂汉堡（Wendy's）FreshAI 的主导者，他带领团队打造了全球首个应用于汽车穿梭餐厅（Drive-Thru）的生成式 AI 点餐智能体。在剖析技术与系统之前，他将时间拨回到了 20 世纪 80 年代末至 90 年代初，分享了一段塑造他人生底层价值观的童年往事。当时只有七八岁的迈克尔，母亲在祖父母经营的一家便利店汽车穿梭窗口打工。母亲经常顶着酷暑在坚硬的水泥地面上连续上双班，工作极其辛苦。下班接他回家时，疲惫不堪的母亲常常无力做晚饭，便会带他拐进麦当劳或温蒂的汽车点餐道。因为囊中羞涩，母子俩通常只能买一份最便宜的餐食平分，或者买一份儿童餐外加一个普通汉堡。让迈克尔终生难忘的是餐厅员工对待母亲的态度——有位店员每次都会亲切地称呼他的母亲为 "mama" 或 "mom"。年幼的他当时感到困惑，甚至心想“她又不是你妈，干嘛这么叫”，但成年后他才深刻领悟到，那一刻给予了他母亲莫大的尊严。一个为生计劳碌整日、一直在服侍他人的底层劳动者，在那个窗口被当成尊贵的客人认真对待、热情问候，而不是被粗暴地塞过来一个餐袋、或是被强行推销根本负担不起的高价套餐。迈克尔指出，这正是他理解商业与人性连接的起点：无论是快餐业还是零售业，商业追求投资回报率（ROI）固然重要，但每一次看似微不足道的交互，其核心都在于“款待精神”（Hospitality）与对人的尊重。 - **时代背景**：20 世纪 80 年代末至 90 年代初的家庭记忆 - **核心体验**：低收入家庭在汽车穿梭窗口感受到的温情与平等待遇 - **底层洞察**：每一次微小交互本质上都是对顾客尊严的确认（"a moment of dignity"） --- ### [02:50](https://www.youtube.com/watch?v=6j790ra1r-M&t=170s) 以人为本的产品哲学与温蒂创始人的核心价值观迈克尔强调，他在产品研发与技术落地中从不满足于“只做最低限度的交付”或片面追求快速产出，他始终深究工作背后的“根本原因”（the why）。现代人的生活节奏极度忙碌，几乎没有人会在一天的日常中花心思去深思熟虑“今天去哪家快餐店点餐”，顾客之所以光顾，图的就是便捷顺畅。如果在这一刻系统发生故障或体验受挫，就会直接破坏顾客原本就紧绷的一天。从早年在维多利亚的秘密（Victoria's Secret）及其母公司 L Brands 的从业经历，到后来在温蒂汉堡掌舵 AI 创新，迈克尔始终坚持将商业 ROI 与有意义的人性化体验相结合。这与温蒂汉堡创始人戴夫·托马斯（Dave Thomas）所倡导的企业价值观不谋而合——无论是面向在窗口点单的顾客，还是在后厨高压环境下挥汗如雨的员工，设计产品的第一准则永远是“带着尊严去服务”（do it with dignity）。 --- ### [04:42](https://www.youtube.com/watch?v=6j790ra1r-M&t=282s) 汽车点餐的极致复杂度与大语言模型的破局探索大约三年前，迈克尔的团队开启了一段探索之旅，率先提出了一个在当时极具颠覆性的设想：“大语言模型（LLM）能否直接用来做汽车餐厅点餐？”在那个时期，市面上所有的语音交互方案几乎全都是传统的基于硬编码规则与流程分支（rule- and flow-based）的旧式 AI 系统。外界通常认为快餐点单非常简单，但实际场景的工程与语义复杂度极其惊人。以温蒂汉堡的招牌汉堡“Dave's Double”为例，仅这一款汉堡在后厨系统里就存在超过 2000 亿种可能的配置组合方式（包括肉饼熟度、酱料、加减配料等）。再加上真实人类口语表达中千奇百怪的自然语言变体——例如“不要生菜”（No lettuce）、“去掉生菜”（Remove the lettuce）、“别放任何绿色的玩意儿”（I don't want any of the green stuff）——传统规则系统在面对此类长尾表达时极易崩溃。在 ChatGPT 爆火出圈、甚至在 GPT-3 全面普及之前，温蒂团队就主动找到了战略合作伙伴 Google 的 LLM 研究团队探讨合作。温蒂汉堡也因此成为了全球首个在 Google 围墙之外、将大语言模型落地到实际外部业务场景中的企业。 ```mermaid flowchart TD accTitle: FreshAI 发展阶段演进图 subgraph P1["1. 探索破局（2021–2022）"] A["提出设想：LLM能否胜任复杂点餐？"] --> B["携手 Google 引入前沿大模型能力"] end subgraph P2["2. 小鹰号原型（Kittyhawk）"] C["车库手工搭建初代软硬件"] --> D["单店实地拉电线测试与店员共创"] end subgraph P3["3. 阿姆斯特朗规模化（Armstrong）"] E["多门店快速复制与工程调优"] --> F["全美数百家门店投产 / 日常服务数十万人"] end B --> C D --> E ``` --- ### [06:01](https://www.youtube.com/watch?v=6j790ra1r-M&t=361s) 车库创业与初代原型代号“小鹰号”（Kittyhawk）作为业内首创项目，团队将其初代试点项目秘密命名为“小鹰号”（Kittyhawk）——以此致敬莱特兄弟在北卡罗来纳州完成的人类首次动力飞行。这一命名不仅因为迈克尔与莱特兄弟同为俄亥俄州老乡，更象征着一项前途未卜但雄心勃勃的拓荒尝试。当时大模型的基础设施完全没有标准化，团队必须从零摸索一系列棘手难题：如何在嘈杂的户外环境中采集清晰音频？如何实现高精度的实时语音转录？如何向企业级系统开放大模型调用接口？转录文本该如何存储合规？在法律合规层面是否可行？团队完全以精益初创公司的方式运作。迈克尔和几位核心成员在自家车库里亲手组装了初代原型设备，车库距离第一家测试的温蒂门店仅有两英里路程。他们把设备搬到餐厅现场，从停车场另一头拉起长长的接线板插上电源，让真实顾客进行几个小时的点单实测。测试结束后团队迅速回到车库复盘迭代，第二天再带回餐厅继续验证。迈克尔和团队几乎“住在了餐厅里”，每天与一线店员深入交流。他们清醒地认识到，虽然技术人员擅长打造数字化体验，但真正常年累月在高压快餐环境中工作的是餐厅员工。AI 系统不仅要为顾客提供顺畅一致的品牌体验，更必须切实减轻一线员工的操作负担，而不是给他们添乱。 - **代号寓意**：Kittyhawk，致敬俄亥俄州莱特兄弟的首飞精神与从零拓荒的勇气 - **开发模式**：车库手搭原型、拉延长线在单店实测、每日与一线店员高频迭代反馈 - **攻关难点**：户外降噪收音、大模型调用接入、转录处理合规性与法律授权 --- ### [08:34](https://www.youtube.com/watch?v=6j790ra1r-M&t=514s) 从“阿姆斯特朗”到全美规模化落地的战略支柱在“小鹰号”验证成功之后，项目迎来了实现飞跃的规模化阶段，代号被命名为“阿姆斯特朗原型”（Armstrong Prototype），象征着从莱特兄弟的离地首飞迈向阿姆斯特朗登月般的系统化跃迁。迈克尔坦言，虽然跨过首家门店极其艰难，但从第一家店复制到第二家、第三家门店的工程难度呈指数级上升；与此同时，团队的迭代效率和知识沉淀也随着数据积累而迅速加速。如今，FreshAI 已经从单店试点拓展至全美范围的大规模企业级落地，覆盖了美国绝大多数州，入驻了数百家温蒂餐厅，且仍有数百家门店正在排队上线。系统已经累计服务了数千万名真实顾客，每天稳定接待数十万名不同背景的用户——从 Z 世代年轻人到老年长辈，从乡村居民到都市人群，均能自然流畅地交互。这项在三年前尚不存在的技术实验，如今已经从一个脆弱的概念验证（POC）蜕变为温蒂汉堡整体业务增长蓝图中的核心驱动引擎，并在公司财报与公开战略中被高频提及。迈克尔对团队在极端严苛环境下敏捷突进的成果深感自豪。 - **规模化数据**：覆盖全美绝大部分州，已上线数百家餐厅，累计服务数千万真实顾客，单日服务量达数十万人次 - **用户普适性**：覆盖从老年人群到年轻一代（Gen Z）、从乡村到都市的全谱系消费群体 - **商业地位转变**：从车库边缘创新跃升为驱动全公司未来战略增长的关键核心业务 ### [10:17](https://www.youtube.com/watch?v=6j790ra1r-M&t=617s) 规模化带来的组织阵痛与运营挑战随着试点取得初步成果，FreshAI 像所有初创产品一样进入了严苛的运营优化阶段。团队规模必须从最初精悍的 3 到 4 人小组向外拓展，因为原班人马已经无法承受全美门店的培训负荷。此时，系统已经触达数以万计的门店一线员工（crew members），每天在汽车穿梭餐厅（drive-thru）接待数十万名顾客。随着组织规模急剧膨胀，跨部门的利益相关者纷至沓来，曾经“天马行空、专注解决最酷问题”的极客氛围，逐渐被冗长低效的官僚流程所取代——团队不得不把精力消耗在“到底要开多少次会才能推进”以及“逐级审批流程该怎么走”等繁琐事务上。 Michael 坦言，回看整个历程，组织扩张确实让团队的推进速度变慢了。无论是在大会走廊里与同行交流，还是在行业活动中探讨“从规模化中习得的教训”，这都是几乎所有企业在从尖端生成式 AI 探索迈向实际产品落地时必然遭遇的经典阵痛。不过，这些组织层面的阻力并未击垮团队，FreshAI 顶住了重重压力，在规模化扩展的轨道上平稳运行至今。 - **组织培训量级**：累计培训数万名门店一线员工，使其掌握人机协作与优质服务流程。 - **日常吞吐负荷**：全美部署门店每天稳定承载数十万人次的真实进店与点餐交互。 --- ### [11:38](https://www.youtube.com/watch?v=6j790ra1r-M&t=698s) 抛弃旧玩法：全面 Agent 化与尽早确立商业模式站在当下的时间节点复盘，Michael 直言如果重新打造一款 AI 产品，过去的旧打法必须全盘推翻。当年 FreshAI 起步时，行业中甚至还没有 ChatGPT 可以用来随时解答开发疑问，团队也无法通过“氛围编码”（vibe-code）快速生成 Demo 或进行极简敏捷迭代。尽管 FreshAI 历经 3 年依然保持着行业领头羊的地位，但 Michael 认为“3年实在太慢了”。面对新一代 AI 技术浪潮，团队必须以更大的胆识、更激进的速度和更明确的目的性去推进每一步决策。如果要重新设计落地路径，首要原则是将 AI Agent 渗透到业务的每一个毛细血管，包括前期的技术研究、运营调度以及员工培训等全链路。此外，团队必须以最快速度验证商业价值。虽然温蒂汉堡很早就向加盟商和利益相关者描绘了激动人心的愿景，但加盟商最终最在意的永远是底层利润（bottom line）的实质改善。在当年的开发中，团队把最棘手的难题推迟到了最后——即面向加盟商的收费标准与整体经济模型（economic model）。Michael 强调，未来任何 AI 项目都必须在初期就优先确立财务上可持续的商业模式，绝不能留到后期再亡羊补牢。 ```mermaid flowchart TD accTitle: FreshAI 迭代策略演进对比 subgraph Legacy["传统早期路径（耗时 3 年 / 推进迟缓）"] L1["单点技术验证"] --> L2["人工主导培训与运营"] L2 --> L3["后期硬啃加盟商经济模型"] end subgraph Future["下一代 Agent 敏捷路径（极速落地 / 价值优先）"] F1["端到端全面 Agent 化\n(研发/培训/运营)"] --> F2["尽早锁定财务与盈利模型"] F2 --> F3["快速规模化并改善底层利润"] end ``` - **核心反思**：3 年的产品迭代周期在生成式 AI 时代过于冗长，必须追求更激进的落地节奏。 - **关键卡点**：加盟商定价与财务模型是规模化过程中最艰难的硬骨头，必须在项目早期确立。 --- ### [13:38](https://www.youtube.com/watch?v=6j790ra1r-M&t=818s) 告别红夹克：从温蒂汉堡离职与未来的大胆愿景在追求规模化的同时，如何守护产品的“灵魂与调性”（product personality / vibe），确保技术不失温度，是 Michael 始终坚守的核心。在演讲现场，Michael 罕见地没有穿他标志性的红色夹克（red jacket），借此向全场宣布了一个重磅消息：当天正是他在温蒂汉堡任职 5 年、全身心孕育 FreshAI 3 年的最后一天（"Today is my last day with Wendy's"）。尽管 FreshAI 就像自己的孩子一样难以割舍，但他决定正式翻篇，投身于酒店餐饮（Hospitality）领域下一场更为宏大的前沿变革中。 Michael 预判，AI Agent 在汽车穿梭餐厅及整个服务行业的普及将迅速变成“家常便饭”，未来顾客在任何服务场景中遇到 AI 都不再会感到惊讶。他向现场所有创新者发出号召：打造产品切忌仅仅执着于冰冷的投资回报率（ROI）或商业指标，更要关注产品在大规模部署时所展现的“存在感与尊严”。正如温蒂汉堡有 80% 到 85% 的业务发生在那条狭窄的穿梭车道上，团队一开始就选择啃下最难的硬骨头，打造颠覆性体验。入局 AI 绝不能从小修小补入手，而要以绝对的勇气（"Be bold"），在保障安全的前提下，为顾客、员工和股东带来真正有意义的深远变革。 - **职业生涯转折**：结束在温蒂汉堡的 5 年任期（其中 3 年主导 FreshAI），开启下一代酒店餐饮 AI 创业项目。 - **业务基本盘**：温蒂汉堡 80% 至 85% 的营收来自汽车穿梭餐厅，属于高压、高并发的核心命脉。 --- ### [16:12](https://www.youtube.com/watch?v=6j790ra1r-M&t=972s) 问答：GCP 云原生挑战与采用大语言模型的真正初衷在观众问答环节，来自 Triabs 的咨询顾问 Tomas 率先提问，探讨基于谷歌云平台（GCP）构建与适配 FreshAI 所面临的技术挑战。Michael 回应称，温蒂汉堡走了一条独特的混合架构路线。作为谷歌早期在生成式 AI 领域的旗舰级战略合作伙伴，双方共同攻克了大量底层平台瓶颈；虽然日常开发中免不了遇到各类技术痛点，但谷歌团队始终保持着极高的协作热情，持续推动平台能力的快速迭代。紧接着，现场观众追问温蒂汉堡最初决定引入大语言模型（LLM）的底层动机，询问这究竟是为了“削减用工成本”，还是单纯觉得技术很酷。Michael 明确反驳了裁员论调，并深情回顾了温蒂汉堡创始人戴夫·托马斯（Dave Thomas）在 1969 年发明现代汽车穿梭餐厅的历史底蕴。他指出，温蒂汉堡探索 LLM 的初衷绝不是为了“用机器声音替换喇叭里的真人”，而是为了重塑汽车穿梭餐厅的服务体验。通过将繁琐的点餐沟通工作交给擅长理解自然语言的 AI，员工劳动力得以重新分配到餐品制作与出餐环节。 - **技术平台合作**：深度绑定 Google Cloud 生态，作为旗舰案例推动生成式 AI 在严苛低延迟场景的落地。 - **出餐窗口体验重塑**：减少以往由于忙乱而“直接把餐袋从窗口塞给顾客”的冷漠交互，让一线员工有从容的时间在取餐口与顾客真诚微笑致谢。 --- ### [19:13](https://www.youtube.com/watch?v=6j790ra1r-M&t=1153s) 问答：AI 是否让点餐失去人情味？针对观众提出“看到 AI 点餐就想绕开、担心机器夺走人情味”的疑虑，Michael 坚定地给出了否定回答。他坦言目前市面上的语音点餐产品良莠不齐，但他坚信温蒂汉堡打造的 FreshAI 是目前市场上最出色、最稳定的系统。FreshAI 不仅能始终保持友善、一致的沟通态度，更关键的是它“永远不会催促顾客思考”，从而消除了顾客在人工点餐时常有的压迫感。为了证明这一点，Michael 分享了一个极具戏剧性的真实对比案例：在温蒂汉堡近期联动 Netflix 热播剧《星期三》（Wednesday）推出大型营销活动当天，社交媒体上流传出两条鲜明的实测视频。在一家尚未部署 FreshAI 的传统门店里，顾客开车停在麦克风前足足等待了 55 秒，在不得不主动大喊“有人吗？”之后才有疲惫的店员上线接单；而在部署了 FreshAI 的门店，车辆刚一驶入，系统便立即以极致饱满的热情主动问候："Hello, welcome to Wendy's. What can I get for you today?"（您好，欢迎光临温蒂汉堡，今天想吃点什么？），整个过程既专业又直奔主题，完美化解了传统点餐中的尴尬与等待。 - **服务响应对比**：无 AI 门店在促销高峰期让顾客在窗口空等长达 55 秒；FreshAI 门店实现秒级零延迟、标准化且热情的主动迎宾。 - **交互核心优势**：消除人工点餐时的催促感与情绪波动，提供极具耐心、始终如一的友好交互。 ### [20:22](https://www.youtube.com/watch?v=6j790ra1r-M&t=1222s) AI Agent 带来的服务一致性与窗口人际温度在得来速车道的实际运营中，AI Agent 展现出的效率与稳定性尤为显著。当使用 AI 点餐的车道在短短 55 秒内便顺利完成整个点餐流程、车辆径直开往前方的取餐窗口时，另一条车道上的传统人工服务顾客却往往还停在原地，苦苦等待店员开口应答。这正体现了切换到 AI 系统后带来的核心价值：极致的“一致性”（Consistency）与“确定性”（Confidence）——顾客从驶入车道那一刻起就能确信，自己的点餐需求会立即得到响应，绝不会被晾在一旁。更深层的价值在于对门店员工的解放。AI 接管了繁琐、重复的点餐流程后，成功释放了取餐窗口处员工的时间与精力，而窗口恰恰是“更有意义的人际互动发生的地方”。员工不再被点餐耳机的催促所撕扯，从而能够从容地向顾客送上一声真诚的问候与感谢，甚至能微笑着对熟客寒暄一句“最近怎么样？”。省下来的短短几秒钟，让原本冰冷机械的取餐流程重新焕发出真正的人性温度。全美的运营数据与客户调研结果均证实，广大顾客整体上其实更偏好与 AI Agent 互动。这种体验通常在“第二次体验”时迎来爆发式的认知转变——只要顾客尝试过一次并再次光顾，他们往往就会彻底想通并爱上这种模式。因为在与 AI 的交流中，他们真切地体会到“自己不会被催促，也不会被打断或手忙脚乱”（"I don't get rushed. I don't get jumped around"），从而获得了更从容、更受尊重的点餐体验。 - **效率对比**：AI 车道在 55 秒内即可完成完整点餐并让顾客驶向取餐窗口，彻底消除了人工点餐常见的初期等待与静默。 - **赋能人际互动**：将员工从点餐台解绑，使取餐窗口成为传递温度的阵地（如向熟客问候、真诚致谢）。 - **心智转变契机**：顾客在第二次使用后满意度大幅提升，核心痛点在于消除了被人工催促、打断的不适感。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=6j790ra1r-M&t=4s) the soul in the drive-thru — Michael Chorey > [00:04](https://www.youtube.com/watch?v=6j790ra1r-M&t=4s) a moment of dignity — Michael Chorey > [02:50](https://www.youtube.com/watch?v=6j790ra1r-M&t=170s) the why — Michael Chorey > [02:50](https://www.youtube.com/watch?v=6j790ra1r-M&t=170s) do it with dignity — Michael Chorey > [04:42](https://www.youtube.com/watch?v=6j790ra1r-M&t=282s) I don't want any of the green stuff — Michael Chorey > [13:38](https://www.youtube.com/watch?v=6j790ra1r-M&t=818s) Today is my last day with Wendy's — Michael Chorey > [13:38](https://www.youtube.com/watch?v=6j790ra1r-M&t=818s) Be bold — Michael Chorey > [19:13](https://www.youtube.com/watch?v=6j790ra1r-M&t=1153s) Hello, welcome to Wendy's. What can I get for you today? — Michael Chorey > [20:22](https://www.youtube.com/watch?v=6j790ra1r-M&t=1222s) I don't get rushed. I don't get jumped around — Michael Chorey ## Mentioned Michael Chorey, FreshAI, 温蒂汉堡 (Wendy's), 戴夫·托马斯 (Dave Thomas), 麦当劳 (McDonald's), 维多利亚的秘密 (Victoria's Secret), L Brands, Dave's Double, Google, Google Cloud Platform (GCP), 莱特兄弟 (Wright brothers), Kittyhawk, Armstrong Prototype, ChatGPT, GPT-3, Tomas, Triabs, Netflix, 《星期三》(Wednesday) ## Video description Michael Chorey, Founder, Wendy's I built and scaled the world’s first generative AI drive-thru agent—serving tens of millions of live customers in one of the most complex, high-pressure environments imaginable. What started as fragile, duct-taped pilots became a platform that redefined how fast food operates. But here’s the truth: if I had to do it again, I’d do it differently. In this talk, I share what I learned from scaling GenAI in the real world—what worked, what slowed us down, and what I’d rebuild from day one. From using agents across every part of the business to designing for emotional resonance ("vibe-code"), I’ll walk through how to move faster, scale smarter, and build with soul. This isn’t a product story—it’s a leadership playbook for anyone deploying AI that has to feel as good as it functions. [Source](https://www.youtube.com/watch?v=6j790ra1r-M)

Source quotes

[0:04] Michael Chorey: the soul in the drive-thru [0:04] Michael Chorey: a moment of dignity [2:50] Michael Chorey: the why [2:50] Michael Chorey: do it with dignity [4:42] Michael Chorey: I don't want any of the green stuff [13:38] Michael Chorey: Today is my last day with Wendy's [13:38] Michael Chorey: Be bold [19:13] Michael Chorey: Hello, welcome to Wendy's. What can I get for you today? [20:22] Michael Chorey: I don't get rushed. I don't get jumped around

Source quotes

[0:04] Michael Chorey's mother: Let's go out. [2:50] Michael Chorey: do it with dignity [4:42] Michael Chorey: Could a large language model (LLM) take drive-thru orders in real time? [11:43] Michael Chorey: What we did at the beginning of this story, I would not do again. [11:43] Michael Chorey: far too slow [13:08] Michael Chorey: we'll figure it out at the end. [13:38] Michael Chorey: My journey at Wendy's actually ends today. Today is my last day with Wendy's. [13:38] Michael Chorey: presence at scale [19:13] FreshAI Agent: Hello, welcome to Wendy's. What can I get for you today? [20:22] Michael Chorey: how have you been?

Original transcript

[0:04] So today I'm going to be talking about [0:07] the soul in the drive-thru. Um, so I am [0:10] from Wendy's. We led the first [0:12] generative AI agent in the world in the [0:15] drive-thru. And so today I'm going to [0:17] talk a little bit about that story. [0:19] Before I get into that, I want to go all [0:22] the way back. Imagine me instead of up [0:25] here as the the man I am today. seven, [0:28] eight-year-old boy, late 90s, I'm sorry, [0:30] late 80s, early 90s. And my story really [0:33] kind of starts there. And imagine, you [0:36] know, I had my mom and she worked back [0:39] then for my grandparents uh convenience [0:42] store drive-thru. She'd work doubles. [0:44] She'd work all day. And at the end of [0:46] the day when she was really tired, she'd [0:48] come home and uh pick me up from my [0:50] grandparents. On the way home, uh [0:52] sometimes she was too tired to go home [0:54] and, you know, make dinner. Um, and so [0:57] she'd be like, "Let's go out." And we'd [1:00] go through a drive-thru, whether it was [1:01] McDonald's or or Wendy's most of the [1:04] time. And back then we didn't have a lot [1:07] of money. And so we'd often get, you [1:09] know, a very inexpensive meal and split [1:11] it or I'd get a kids meal and she'd get, [1:12] you know, just a burger, something [1:14] pretty basic and straightforward. [1:17] And I have these fond memories because [1:18] they were always really special to me, [1:20] just me and my mom spending time [1:22] together. Um, but I also remember how [1:25] the crew members would treat her. This [1:28] one lady specifically always called her [1:30] mama or mom, which as a kid I was like, [1:32] "She's not your mom. Why are you calling [1:34] her that?" Um, but it kind of always [1:37] stuck with me. And this is the moment [1:39] that I didn't know as a kid, but as an [1:42] adult and I grew up and I began to [1:45] realize that like that moment was a [1:47] moment of dignity for my mom. She worked [1:49] all day. She served people, you know, [1:52] walking through the drive-thru on the [1:53] hard concrete and the hot, maybe not as [1:55] hot as it is here, but, you know, it [1:57] would get really hot in that drive-thru. [1:59] And instead of serving somebody else, [2:01] somebody was there to serve her. They [2:02] took a moment to um to acknowledge her, [2:05] to say thank you, to greet her. And, you [2:08] know, they could have just put a hand [2:09] out the window and given her a bag. They [2:12] could have rushed her along and tried to [2:14] get her to buy something more expensive [2:15] that she probably couldn't afford, but [2:17] they didn't. They treated her with [2:19] respect. And looking back, this is [2:21] really a moment of a moment of dignity. [2:24] This is one of the first times that I [2:25] really started to understand how [2:27] important every little interaction is [2:30] for a person that's coming to your [2:31] business. Whether it's in the fast food [2:33] industry, the retail industry, those [2:35] moments, those unique moments are what [2:37] is so important and critical. We often [2:40] talk a lot about our objectives, our [2:42] ROIs, but at the end of the day, I'm a [2:44] huge believer in these moments about [2:47] hospitality. [2:50] And so for me a little bit, how I [2:52] approach the work I've done has been [2:54] grounded in that. I don't believe in [2:57] just doing the bare minimum. I don't [2:59] believe in rushing to an outcome. For [3:03] me, it's deeply important to understand [3:05] the why behind what I'm doing. And so I [3:08] think you also heard a little bit from [3:09] the panel before. It was interesting. [3:10] They were talking about the human [3:12] aspect. Everything we do, whether it's a [3:15] retail company, a finance company, [3:17] health or QSR, quicks service restaurant [3:19] or fast food, like those moments of [3:21] interaction are really important. All of [3:23] us are super busy. Let me ask a [3:26] question. Who thinks very hard about [3:29] going to fast food throughout your day? [3:33] Not a single person. I think that's the [3:35] key. Like we have very busy lives and [3:39] for you to show up to a fast food [3:40] restaurant or to show up to your [3:42] favorite retail. If that moment is [3:45] disruptive, why? Like the last thing on [3:49] earth I would want is to have a a [3:51] drive-thru experience I create make your [3:53] day worse. And so a lot of what I do is [3:56] try to find the maximum and the most [3:58] meaningful moments within that [4:00] interaction. Whether it was my time [4:02] early in my career with Victoria Secrets [4:04] and Elbrands or more recently with my [4:07] time at Wendy's, it is always about how [4:10] can I not only achieve the ROI and the [4:13] objectives of the business because I'm [4:15] also a good product person. I have a a [4:17] goal I do need to achieve, but I also [4:19] want to understand how I can do it and [4:21] make it a meaningful experience for you [4:23] as a customer or the crew member within [4:25] the restaurant. And most importantly, [4:27] and this is a throughine that I've come [4:28] to realize with my time at Wendy's and [4:31] our founder Dave Thomas and his core [4:33] values is do it with dignity. Whether [4:36] it's the the customer or the crew [4:37] member, that has to be the first thing [4:39] we care about. [4:42] So, I kind of alluded at the very [4:43] beginning, my team built the world's [4:46] first generative AI agent for taking [4:49] orders in the drive-thru. This is a [4:51] journey that we have been on for about 3 [4:53] years. And we started with a really bold [4:56] question and that was could an LLM take [4:59] orders? Up until then, basically all of [5:03] the other products out in the market [5:04] were traditional AI rule and flow-based. [5:07] And I like to talk a lot about how [5:09] complex fast food ordering is. Most of [5:11] you show up and probably are like, "It's [5:14] not that hard." It's because you [5:15] probably get the same thing most of the [5:17] times. But at Wendy's, the classic [5:19] Dave's Double, there is over 200 billion [5:21] ways to configure that sandwich. And [5:24] when you look across all the markets [5:25] within the QSR space or the retail [5:27] space, complexity is everywhere. Then [5:30] you add in the nuances of how people [5:32] talk. No lettuce. Remove the lettuce. I [5:34] don't want any of the green stuff. It [5:36] gets exponentially harder to make a good [5:38] experience. [5:41] And so my team started this journey long [5:44] before Chat GBT exploded, before GPT3 [5:48] came out. And we went to Google who are [5:51] our good partners and we found um we [5:53] knew some that worked in the LLM [5:55] research space and we asked them this [5:56] question. And so this is what started [5:58] our long journey. And this little bit [6:01] right here I'm going to talk about is a [6:03] story I've never really talked about in [6:04] public. Our very first pilot, we gave it [6:08] a code name just like every startup and [6:10] tech team loves to do. Come up with a [6:12] cool name and we called it Kittyhawk. [6:15] after the very first flight, the human [6:18] flight in a plane uh in North Carolina. [6:21] Uh it was meaningful for a couple [6:23] reasons. One, because I'm from Ohio and [6:26] the Wright brothers were also from Ohio. [6:29] But it was also this idea, something [6:32] bold and big. We didn't know how to to [6:36] create an agent to order in the [6:38] drive-thru, especially with an LLM. [6:40] Early in the journey when we were [6:42] working with Google, Wendy's was the [6:44] first company to ever use an LLM outside [6:47] of their walls. It's common place now. [6:49] I'm sure half of you right now probably [6:50] have an LLM up, whether it's ChateBT, [6:53] Claude, or or Gemini. They didn't know [6:55] how to make it accessible. And so early [6:58] in this journey, we had to figure out [7:00] things that weren't just can I do it. We [7:03] had to figure out countless things that [7:05] you would have never thought about. How [7:07] do you get audio? How do you uh [7:10] understand what a customer is saying? [7:11] How do you get access to the LLMs? What [7:13] do you do with the transcripts? Can I [7:15] even legally do it? We were every day [7:18] asking and trying to answer questions [7:20] that we never thought we would. And so I [7:22] also like this idea of Kittyhawk is it [7:25] was very fragile in the beginning. We [7:27] were listening to the team up here [7:29] talking about startups. The the Wendy's [7:31] team with Fresh AI, we treated it as a [7:34] startup. It was me and a few other [7:36] people. We literally built the first [7:38] prototype out of my garage at my house [7:41] two miles from the very first Wendy's to [7:44] get Fresh AI or this voice AI a ordering [7:46] agent. And we'd work on it. We'd take it [7:49] up to the restaurant. We'd plug it in. [7:51] We'd bring an extension cord across the [7:53] parking lot to the outlet and we'd have [7:55] customers come through for a few hours [7:58] um and get feedback. And then we go and [8:00] we iterate and we show up the next day. [8:02] We lived in the restaurant working to [8:04] make this experience better, more [8:07] effective, and enjoyable for everyone. [8:10] We talked to crew members every single [8:12] day. We got their feedback. Yes, we [8:14] maybe knew how to create something that [8:16] was meaningful from an experience [8:17] perspective, but I didn't work and [8:20] breathe and live in the fast food [8:21] restaurant every day. Those crew members [8:23] did. And so, this also had to work for [8:25] them. had to make their job easier, not [8:27] just a consistent onbrand frictionless [8:30] experience that's meaningful for the [8:31] customer. And so after Kittyhawk came [8:34] our scale moment where it really all [8:37] started to click and we called it the [8:39] Armstrong prototype. [8:42] And this also embodies the spirit and [8:45] the soul of going from this rickety [8:47] small like hum humong hum very large [8:53] sorry very large uh successful and hard [8:55] thing to do with the first flight or the [8:57] first restaurant. But it was [8:58] exponentially harder to go to site two [9:00] to site three for a lot of reasons. But [9:02] a lot of other things also got easier [9:04] along that journey. We got faster. We [9:07] learned more. There was more [9:08] information. And so we went from one [9:11] store to dozens every month. And now the [9:14] product is across the United States. [9:17] It's not only the first generative AI [9:19] ordering agent, it's the first one to go [9:21] scale across uh an enterprise [9:24] organization uh within our industry. [9:27] We're in the majority of the states. [9:29] We're in hundreds of restaurants with [9:30] hundreds more coming. And we have had [9:33] over tens of millions of real everyday [9:36] people talking to this agent. hundreds [9:39] of thousands every single day. Grandmas, [9:42] Gen Z, everyday people, country, people [9:45] from the country, people from the city. [9:48] And it works. This agent works well. And [9:51] we went from a proof of concept to this [9:54] is now a critical part of the growth [9:56] strategy of that company. [9:59] Something that didn't exist three years [10:00] ago is now the fundamental driver for [10:02] their growth plan. It's something that [10:04] they talk about openly. And I'm deeply [10:07] proud of the ability to do that. But the [10:10] one thing I will talk about as part of [10:12] this story and as we talk about scale is [10:15] we were really really fast and nimble in [10:17] the beginning. But over time, just like [10:20] any new product or or or startup, we had [10:24] to optimize for operations. More people [10:28] got involved. Our three or four person [10:30] team couldn't do training for all the [10:32] restaurants. At this point in time, [10:33] we've trained tens of thousands of crew [10:36] members to understand how to create a [10:38] great experience. And I said hundreds of [10:39] thousands of customers come through [10:41] every single day. And so more [10:43] stakeholders got involved, more team [10:44] members got involved. And instead of [10:46] just dreaming big and thinking about [10:49] exciting problems, we had to start [10:51] thinking about well how many meetings do [10:53] we have to have to solve this? Um how [10:55] what approvals do we need to get? And so [10:58] this is a part as I reflect on the [11:00] journey where we did slow down a little [11:01] bit. As I sit up here talking to all of [11:04] you and and I talk to folks out, you [11:07] know, in the hallways and at other [11:08] events and we talk about what did we [11:10] learn from scale. This is a classic item [11:13] and I think it's not unheard of whether [11:15] it's the most cutting edge generative AI [11:17] or launching a very basic product is [11:20] scale has its own problems to to bring [11:23] from an organization perspective. But [11:25] the other thing I would add though is [11:27] for this story, it didn't stop us. It [11:29] didn't cause us to collapse. This is a [11:33] is a product that is continuing to [11:34] scale. [11:37] And so here's a little bit of what I've [11:38] learned and what I want to give back to [11:39] all of you on this call. [11:43] Sorry, is that the old playbooks don't [11:46] work. Even what we did at the beginning [11:51] of this story, [11:54] sorry. What we did at the beginning of [11:56] this story, I would not do again. So, as [11:58] I sit here today and I've been [11:59] reflecting a lot, if I were to start [12:01] something new again today, what would I [12:03] do? I wouldn't run the same playbook I [12:05] did in the beginning. Back then, we [12:07] didn't have chatbt to answer a lot of [12:09] our questions. We couldn't vibe code [12:11] through um how to develop and and to to [12:14] show off demos quickly or to work [12:16] through in a lean way. We had to think [12:18] through as time went on those became [12:20] available to us. But I look back at our [12:22] journey 3 years that's way too slow. 3 [12:25] years later we're still the leader [12:27] within the space on this product. But [12:29] it's far too slow. I would want to ask [12:31] how can we go faster? How can I be even [12:33] more bold and be more intentional about [12:35] every decision I make along the path? [12:38] And so the first thing is is I would use [12:40] agents everywhere like the research I [12:43] would use an agent for um for operations [12:47] training all of those things I would [12:48] start using agents [12:50] the other thing I would look at is how [12:52] do I get to value faster like I said [12:55] we've scaled to one restaurant we've [12:57] been in many restaurants for almost two [12:58] years now but that early scale was too [13:01] slow we told a very powerful story to [13:04] our uh franchises our stakeholders they [13:07] were excited [13:08] But at the end of the day, they want to [13:09] see a bottom line improvement. And so we [13:11] needed scale faster. We also put some of [13:13] the really hard decision-m things off [13:15] till the end. The economic model, we had [13:17] a rough idea of what it would look like, [13:19] the fee we would charge to franchises, [13:22] but we said, "We'll figure it out at the [13:23] end." And I'll tell you, it's been one [13:25] of the hardest things to to nail down. [13:28] And so as I look at a new journey ahead [13:30] of me, getting to a financially viable [13:33] model early would be one of the first [13:35] things we spend more time on. [13:38] And I also started with at the very [13:40] beginning, what is the feel of the [13:42] product? And how do you make sure that [13:44] as you go to scale, you don't risk [13:46] losing that product personality, that [13:48] feel, the vibe of what your product [13:50] brings? And so for those of you who may [13:53] have like looked me up before this, I'm [13:55] like historically in a red jacket. It's [13:57] like my signature look. This is my first [13:59] talk not in red and it's because my [14:03] journey at Wendy's actually ends today. [14:05] Today is my last day with Wendy's. And [14:08] so after 5 years and three exciting uh [14:13] years working on this fresh AI product, [14:15] I've decided to move on to the next big [14:17] exciting thing. I can't share what that [14:19] is yet. I think that news will come out [14:21] later this week. Um, but I'm really [14:23] excited for [14:26] the journey that I've been on and this [14:28] product that is like a child to mine. [14:30] And so as I look across this industry, I [14:33] think this is going to be a commonplace [14:34] thing. It's not this exciting [14:36] innovation. Oh my gosh, what was that? [14:39] Going through a drive-thru or going to, [14:41] you know, a hospitality place where [14:42] there is an agent is going to be [14:45] commonplace in no time at all. And I'm [14:47] really excited to look forward to the [14:49] next big thing. the next thing in [14:51] hospitality is where my attention is [14:53] going. And so as I go forward, I'm going [14:56] to take my own advice that I just talked [14:57] to you. These are things that I'm [14:58] looking at as I begin my next journey. [15:02] And it's not just about outcomes. It's [15:04] about present at presence at scale. So [15:07] as you think about how your product is [15:09] going to make an impact, don't get [15:11] caught up on just the ROI. Don't get [15:13] caught up on just the business needs. [15:15] Think about [15:17] I I went too far. Think about, you know, [15:19] my mom and I when I was small and how [15:21] important that interaction was and the [15:24] the dignity that you can bring to your [15:26] product. And so my call to all of you [15:28] is, you know, be bold. With Wendy's, we [15:31] started with something that was [15:33] revolutionary. 80 to 85% of our business [15:36] went through the drive-thru. Don't start [15:38] with something small. Do something big [15:40] and bold. do it in a safe way, but [15:42] ensure that you can drive meaningful [15:44] impact to your customers, to your crew [15:46] members, and to your um shareholders. [15:51] And so, that's really it. I think we're [15:53] almost out of time. So, if there's time [15:54] for a couple questions, I'll take them. [15:56] But for anyone in the room that wants to [15:58] connect and hear more, you can find find [16:00] me on LinkedIn and connect. I love [16:02] chatting with innovative people. [16:12] Hi, thank you. My name is Tomas. We are [16:15] I'm from Triabs II consultancy. So my [16:18] question is how do you feel about facing [16:21] the challenge of adapting your product [16:23] on GCP? [16:26] >> So can you say it one more time? How do [16:27] I feel about my [16:28] >> the challenge of making your product on [16:31] for Wendy's on GCP using GCP? [16:35] So for uh how do I what about the [16:38] challenge of making my product on GCP [16:39] for for Wendy's? So I would say Wendy's [16:41] had a unique path. We did a hybrid [16:44] approach. We had a strategic partnership [16:46] with Google from the very early days. We [16:48] were one of their like flagship [16:50] generative AI capabilities. But today [16:52] we're kind of probably just like every [16:54] other customer. And so they're [16:56] continuing to work and improve the [16:57] capabilities of those platforms. We run [16:59] into headaches at times, but they're [17:01] very good and passionate about [17:03] continuing to make it better all the [17:05] time. So, yep. [17:08] >> Hi. Um, yes. Uh, what can you say more [17:11] about the initial decision to decide to [17:13] use LLMs for this use case? So, was it [17:16] primarily about workforce reduction? How [17:18] what was it? [17:19] >> Yeah. [17:19] >> Was it that you were or was it just more [17:21] like the cool factor? Can can we do it? [17:23] >> Yeah. So, why did we use LLMs? So [17:26] actually our sto story started a little [17:28] bit before that. We were going to go [17:30] down the path of using traditional AI. [17:32] But the thing that I would say is this [17:34] was never about reducing workforce or [17:36] any of that for Wendy's. Um a lot of [17:39] people don't know that our founder Dave [17:40] Thomas actually invented the modern [17:42] drive-thru in 1969. So every restaurant, [17:45] fast food restaurant you've ever been to [17:47] that has a drive-thru, Wendy's is the [17:49] reason that exists. And so when we [17:52] started to go down this path, I did a [17:55] lot of research about the drive-thru. I [17:57] understood where there were problem [17:58] points. And we didn't just want to take [18:00] a human voice out of the speaker and [18:02] replace it with a robot. We wanted to [18:04] make the thing that we created [18:06] meaningful meaningfully better for the [18:08] future. And so going down this path, we [18:10] really started to understand how can we [18:12] make it better for our customers. That [18:14] story about me and my mom or crew [18:17] members interacting with customers, that [18:18] was the fundamental driver for this [18:21] product in the drive-thru. [18:23] Yes, labor is reallocated to other [18:26] things to enable your food to be more [18:28] accurate. The thing I'm super excited [18:30] about with voice AI at restaurants, [18:33] whether it's Wendy's or other products [18:34] out there, is that you can see [18:36] meaningfully more interaction happening [18:38] up that pickup window. Crew members are [18:40] more free to spend a second or two with [18:42] you to say thank you. No more classic [18:44] like hand out the window with the bag as [18:46] you're driving up if you're lucky for it [18:47] to be ready. And so for us, why did we [18:50] go down this path? It's to create a more [18:53] meaningful experience to go back to what [18:55] our core values were of having a very [18:57] human connection. Uh this technology [19:00] helps us do that. Um it also is really [19:02] good at understanding human language. [19:04] And so at the end of the day, that's all [19:06] ordering is is a communication of data [19:08] back and forth. [19:13] Um, do you feel like it's taken some of [19:16] the personal element out of it? Um, I've [19:20] driven up and I've seen one of these AI [19:22] ordering things and I just drove through [19:24] the drive-thru because I'm like, the [19:25] last thing I want to do is talk to a [19:26] machine. [19:28] Yeah. So, does it take away some of the [19:30] personal aspect, the the AI in the [19:32] drive-thru? Um, I don't think so. I [19:36] think at that exact moment you're going [19:38] to find and and I will say it varies. [19:40] Some of them are drastically better than [19:42] others. I personally believe the one I [19:44] made with Wendy's is the best one on the [19:46] market. But it is consistent. It is [19:49] friendly. It doesn't rush you. I [19:52] recently saw a video a customer pulled [19:54] up at a drive-thru. Wendy's had just [19:56] launched a new campaign for Wednesday, [19:59] their Netflix deal. and at a site with [20:02] no fresh AI. A customer waited for 55 [20:07] seconds and then they were like, [20:08] "Hello." And then the person jumped on [20:10] and took their order. We saw another [20:12] video from social media on that same [20:14] day. A customer pulled up and the agent [20:16] greeted them, "Hello, welcome to [20:18] Wendy's. What can I get for you today?" [20:20] Super friendly, straight to the point. [20:22] And by the time that second order got to [20:25] 55 seconds, they were done and heading [20:27] on up to the window where that other [20:29] customer was still sitting waiting for [20:31] somebody to acknowledge them. I think [20:32] that's a really core example of what you [20:35] get when you switch to this place is you [20:38] get consistency. You get um [20:41] confidence of somebody's going to take [20:43] your order. And I would reinforce that [20:45] it actually frees up the person at the [20:47] pickup window, which is where a more [20:48] meaningful human interaction happens. [20:51] You get a hello, you get a thank you, [20:53] you ask for regulars, how have you been? [20:56] We free up a few seconds for that [20:58] interaction to become more meaningful. [20:59] We've seen it in the data. We've also [21:01] seen through customer surveys that [21:03] across the country, people actually [21:06] prefer using the agents overall. And [21:08] most of them too, it's like the second [21:11] time you go. So if you've I would I [21:13] would say try it once and then go back [21:15] again. And usually for most people like [21:17] they it just clicks for them that like I [21:20] don't get rushed. I don't get jumped [21:21] around.