威尔·博德韦斯，电话访谈系列联合创始人兼首席执行官

面谈

威尔·博德韦斯，电话访谈系列联合创始人兼首席执行官

在 Google 上的首选来源中添加 Unite.AI

威尔·博德韦斯(Will Bodewes)是Phonely公司(Phonely)的共同创始人兼首席执行官，一家语音AI公司专注于帮助企业接听电话，改善客户对话的结果。 在创办公司之前，他在墨尔本大学攻读博士学位，研究应用人工智能和系统，可以在自然交流的同时遵循指令。 他之前创办了音频技术创业公司，并在NCAA的越野滑雪和超耐力自行车比赛中竞争。 在Phonely，他带来研究了部署语音agents的实际挑战的背景，包括对话可靠性，耐久性，集成性，以及持续优化预订，销售，和客户服务工作流程。

Phonely是一个AI语音-agent平台，使企业能够将电话对话自动化，安排预约，合格线索，路由呼叫，并与现有软件连接客户互动。 其工具将自定义的谈话工作流程与调用历史，分析，A/B测试相结合，以便团队可以测量性能，并随着时间的推移改进结果。 该公司为各种组织服务，从小企业到大型企业，称公司支持10 000多个企业，每天处理12万多个电话。 电话还开发了自己的对话模式，包括阿尔玛，重点是让自动通话更快，更可靠，更适合现实世界的商业对话。

在墨尔本大学攻读AI博士学位的同时，是什么让你相信声音AI已经到了可以支持一个独立公司的地步，你在市场上看到了什么，其他人失踪了？

当时没有证据证明语音AI能够支持一个独立的公司。 有几家公司在做这项工作，大约有五十个用户，所以，当我们开始电话，这是一个赌注。 当时，一切都集中在聊天上，所以声音是新媒介。 这个问题没有解决，所以我们想，“为什么不去追？为什么不设法解决这个问题？”

您的博士研究一开始的目标是建立一个“像机器人一样遵循指示，像人一样说话”的AI。 哪些技术突破使你意识到这一研究可以成为商业语音AI平台的基础？

这项研究对了解所发生的事情的基本技术是宝贵和重要的。 它让我可以看到曲线周围，看看管道里会有什么。 我对这些模型是什么以及它们是如何工作的有技术理解，这使我有能力下注。 具体地说，我想建立一个公司并产生一个影响，这是与我所知道和能够预见到的事物最相关的领域。 我想成为其中的一部分。

你将每通电话描述为企业历来无法系统优化的漏斗。 优化电话交谈实际是什么样子，AI如何随着时间的推移改善转换，解析率，或客户体验？

假设你计划约一个水管工。 电话上的人被训练来收集你的细节并安排预约时间，但视他们如何提问，所以，你可能会安排一个约会关于你的管道问题，它证明你也需要修复你的水槽。 agent训练有素，这可以提升新接待员没有受过训练的产品，一个简单的ChatGPT声音并不知道如何做。 您需要一种为它而建造和训练的产品。

除此之外，人类可能会错过电话，可能不会接听电话，或者可能会发送电话到语音邮件。 这杀死了美国的小企业。 他们没有民众接听电话，如果一个电话得不到企业想要的结果，因为AI的速度，我们可以改变对话

现在，每个人都看AI 节省费用。 但是我们认为人们应该把AI看成是增长。 企业实际上可以通过采用人工智能来发展，使他们能够比以前表现更好。

Voice AI必须解决基于文本的AI经常可以避免的挑战，包括耐久，中断，口音，背景噪音，情感提示，以及无法预测的对话。 在企业规模中，其中哪些问题最难可靠解决？

懒惰是最大的我们现在解决了这个问题，但在早期，由于两个原因，这是一个巨大的头痛。

首先，我们在建设一些不可靠的创业企业。 OpenAI的API请求有很大一部分会失败，如果你在电话上，那和挂断是一回事，所以我们必须围绕它做一些回顾。

第二个是能力。 当你向 OpenAI 发出请求时，你会被困在数以百计或数以千计的其他请求后面的排队里，可能要30秒才能回来为了绕过它，我们审视了OpenAI在世界各地的终点位置，并切换了我们所击中的终点，这些终点是各国睡觉和需求低的。 我们必须建造所有这些基础设施，以便找到当时需求最低的终点。 这是一个非常有趣的技术问题。 现在，我们训练自己的模型，包括阿尔玛，所以我们不必担心，因为我们管理计算和需求。

电话现在每天处理超过12万个电话，涉及数千家企业。 agents操作时

百分百提高时间非常重要。 我们所解决的最严峻的工程挑战之一是可靠性和速度。 人工智能空间的移动速度非常快，Y Combinator有一句谚语：动作要快，但当客户每天拨打12万个电话时，所以，我们的许多过程正在想出如何以我们所需要的速度运输，保持适切性，在产品上发展，同时保持企业客户的幸福和稳定，因为我们不断构建新的特征，以在市场上保持领先。 可靠性可能是我们在建立公司时遇到的最大挑战之一。

Generative AI系统可以不可预测的行为，但客户端的手机agent可能正在处理支付，预订，保健问题，或者其他高价互动。 你如何设计agents的声音保持灵活和对话而不产生幻觉或在他们授权的边界外行动？

大约两年前，我们发现了这个难题，而现在，它被广泛应用于整个行业。 我们意识到LLMs最大的问题是背景。 当你把太多在一个迅速，像几页的指示，事情出错的机会增加。

所以我们设计了一个8-agent的管弦系统，从一个agent转移到另一个，每个agent都有本地化的背景。 这与人类如何在呼叫中遵循剧本相似：你得到一些信息，你进入下一阶段等等。 这样我们就可以在代码层面分开步骤。 如果我需要收集你的名字，电子邮件，电话号码和出生日期在给你退款之前，我可以保证，信息收集在代码级别，然后继续。 如果你只依赖LLM，通过构建这个 agent 管弦乐层，可以使可靠性达到100%，因为所有步骤在转移到新的 agent 之前都会被捕获，这个agent 仍然有上下文，但被隔离在自己的指示下。

据报道，一个电话部署在短短三个月内为一名客户带来了1 000万美元的净新收入。 是什么让AI创造了这种程度的可衡量的商业影响，你从哪里看到今天的声音AI产生了最强的ROI?

我们可以做到这一点，因为我们有客户的衡量标准。 他们正在优化销售，因为他们是入境呼叫，他们是合格呼叫者，所以我们可以跟踪产出。 视当天情况而定，他们的呼叫中心销售率约为35%至42%。 在从客户那里得到信息后，我们带走了他们的流量并优化了agent. 我们以20%的销售率开始在接下来的几个月里，我们调整了东西，优化了以至于在多数情况下，我们比他们的呼叫中心做得更好。 我们的销售率不断提高，因为我们了解了问题，确保了经验始终如一和完美无缺。 我们能超越呼叫中心

随着agents的语音能力增强，企业应该从哪里划清可以完全自动化的对话与人类应该继续参与的情况之间的界限？

企业应该能够实现任何它们想要遵循的可重复的过程的自动化，而这种过程的定义是足够的。

但自动化在几个领域变得具有挑战性。 一个是没有API的集成：假设你需要集成，但软件没有API. 虽然有些浏览器自动化技术可以做，另一个是当有多个交割从agent到一个人类，然后回到agent时。 第三是背景差距。 说我把钱包忘在餐馆了也无法检查钱包是否在餐厅。

有些情况总是需要人类，同时需要跨越多个系统的更为复杂的工作。 避免使用 agent 进行这些操作。

电话上说融资在发展到10 000多家企业之后，公司的下一阶段是什么样子，你相信哪些技术能力可以区别下一代语音AI平台？

下一阶段是建设团队并继续投资。 我们正全力以赴。 许多企业都冒昧进入了语音AI，但对我们来说，我们正在建立一个优化平台，我们用声音作为频道。 对于关心线索和手机创收的企业来说，我们的整个重点就是优化AI agents，以实现客户想要的结果。 将Phonely视为一个优化公司，在那里人们想要优化自己的agents以做得更好。

你们的学术研究也探索了利用AI和智能计技术来扩大发展中国家的能源获取和清洁烹饪。 超越Phone，你如何看待AI被应用于诸如能源获取等大型物理世界问题，而你最终希望返回的领域呢？

AI可以做很多好事，但很多问题都是由人造成的，所有问题都无法用技术来解决。 但让技术和人类的变革同时发生，对世界来说是净积极的。

谢谢你的好采访希望多学点的读者应该去拜访范

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Interviews

Will Bodewes, Co-Founder and CEO of Phonely – Interview Series

Add Unite.AI to your preferred sources on Google

Will Bodewes is the Co-Founder and CEO of Phonely, a voice AI company focused on helping businesses answer calls and improve the outcomes of customer conversations. Before starting the company, he pursued a PhD at the University of Melbourne, researching applied artificial intelligence and systems that could follow instructions while communicating naturally. He previously founded an audio-technology startup and competed in NCAA cross-country skiing and ultra-endurance cycling. At Phonely, he brings that research background to the practical challenges of deploying voice agents, including conversational reliability, latency, integrations, and the continuous optimization of bookings, sales, and customer-service workflows.

Phonely is an AI voice-agent platform that enables businesses to automate phone conversations, schedule appointments, qualify leads, route calls, and connect customer interactions with existing software. Its tools combine customizable conversation workflows with call histories, analytics, and A/B testing so teams can measure performance and improve results over time. The company serves organizations ranging from small businesses to large enterprises and says it supports more than 10,000 businesses and handles over 120,000 calls per day. Phonely also develops its own conversational models, including Alma, with a focus on making automated calls faster, more reliable, and better suited to real-world business conversations.

You began building Phonely while pursuing a PhD in AI at the University of Melbourne, with an early focus on creating systems that could follow instructions while communicating naturally. What convinced you that voice AI had reached the point where it could support a standalone company, and what did you see in the market that others were missing?

There was no proof of voice AI being able to support a standalone company at the time. There were a couple of companies working on it that had around five or 10 users, but no one had successfully deployed voice AI that wasn’t an IVR system. So when we started Phonely, it was a bet. At the time, everything was focused on chat, so voice was a new medium. It was a problem that wasn’t solved, so we thought, “Why not go after it? Why not try to solve it?”

Your PhD research began with the goal of creating an AI that could “follow instructions like a robot and talk like a person.” What technical breakthroughs made you realize that this research could become the foundation for a commercial voice AI platform?

The research was valuable and important for understanding the fundamental technology of what was happening. It allowed me to see around the curve and see what was coming up in the pipeline. I had the technical understanding of what these models were and how they worked, and that gave me the ability to make a bet. Specifically, I wanted to build a company and make an impact, and this was the area most related to what I knew about and could see coming. I wanted to be a part of that.

You’ve described every phone call as a funnel that businesses historically haven’t been able to systematically optimize. What does optimizing a phone conversation actually look like, and how can AI improve conversion, resolution rates, or customer experience over time?

Let’s say you’re scheduling an appointment with a plumber. The person on the phone is trained to collect your details and schedule the appointment, but depending on how they ask the questions, they may be able to teach you about things you didn’t know were part of that business. So you might schedule an appointment for your plumbing issue, and it turns out you also needed to fix your sink. A voice AI agent that’s trained really well does the same thing and helps the business upsell customers and generate more money. It could upsell a product that a new receptionist wasn’t trained on and add another ticket to the order, making the customer happier and creating a better business outcome. A simple ChatGPT voice doesn’t know how to do that. You need a product that’s built and trained for it.

On top of that, humans can miss calls, may not answer the phone or may send calls to voicemail. That kills small businesses in America. They don’t have people answering the phone, lack consistent quality and have no way of optimizing it. With AI, if a call doesn’t get the outcome a business wants, we can optimize it with the data we have. Because of AI’s speed, we can change the conversation, which couldn’t happen nearly as fast with humans.

Right now, everyone looks at AI for cost savings. But we think people should look at AI as growth. Businesses can actually grow by adopting AI, enabling them to perform better than they did before.

Voice AI has to solve challenges that text-based AI can often avoid, including latency, interruptions, accents, background noise, emotional cues, and unpredictable conversations. Which of these problems has been the hardest to solve reliably at enterprise scale?

Latency is the biggest one. We solved the problem now, but in the early days, it was a huge headache for two reasons.

First, we were building on top of startups that weren’t reliable. A significant portion of API requests to OpenAI would fail, and we wouldn’t get a response back. If you’re on the phone, that’s the same thing as hanging up, so we had to build retries around that.

The second was capacity. When you sent a request to OpenAI, you’d get stuck in a queue behind hundreds or thousands of other requests, and you didn’t know where you’d be in that queue. It could take 30 seconds to come back. To get around it, we looked at where OpenAI had endpoints around the world and switched which endpoints we hit based on which countries were sleeping and had low demand. We had to build all this infrastructure just to find the endpoints with the lowest demand at that moment. It was a pretty interesting technical problem. Now, we train our own models, including Alma, so we don’t have to worry about that, because we manage the compute and the demand.

Phonely now handles more than 120,000 calls per day across thousands of businesses. What have you learned from operating voice agents at that scale that wasn’t obvious when you were building the first version of the platform?

It’s really important to be up 100% of the time. One of the toughest engineering challenges we’ve been solving is reliability and speed. The AI space is moving incredibly fast, and we’re moving fast even for an AI company. There’s a saying at Y Combinator: Move fast and break things. But when you have customers making 120,000 calls a day, you can’t afford to break things. So a lot of our process was figuring out how to ship with the velocity we need to stay relevant and iterate on the product, while keeping enterprise customers happy and stable as we keep building new features to stay ahead in the market. The reliability piece was probably one of the biggest challenges we dealt with when building the company.

Generative AI systems can behave unpredictably, but a customer-facing phone agent may be handling payments, bookings, healthcare questions, or other high-stakes interactions. How do you design voice agents that remain flexible and conversational without hallucinating or acting outside their authorized boundaries?

We figured this out the hard way about two years ago, and now, it’s pretty widely adopted across the industry. We realized the biggest issue with LLMs is context. When you put too much into a prompt, like pages of instructions, the chances of something going wrong increases.

So we designed an eight-agent orchestration system that transfers from one agent to another, with each agent having localized context. It’s similar to how a human follows a script on a call: You get some information, you move on to the next stage and so forth. That lets us separate the steps at the code level. If I needed to gather your name, email, phone number and date of birth before giving you a refund, I could guarantee that information was gathered at the code level before moving on. You can’t guarantee that if you just rely on an LLM. By building this agent orchestration layer, you can get reliability to 100%, because all the steps are captured before transferring to a new agent that still has context but is isolated to its own instructions.

One Phonely deployment reportedly generated $10 million in net-new revenue for a customer in just three months. What allowed AI to create that level of measurable business impact, and where do you see voice AI producing the strongest ROI today?

We could do that because we had the customer’s metrics. They were optimizing for sales since they were inbound calls, and they were qualifying callers, so we could track the output. Depending on the day, their call center was doing about a 35% to 42% sell rate. After getting that information from the customer, we took their flows and optimized the agent. We started at about a 20% sell rate. The following months, we iterated, tweaked things and optimized to the point where we were outperforming their call center in the majority of cases. Our sell rate kept increasing because we figured out the issues and made sure the experience was consistent and perfect every single time. We were able to outperform a call center.

As voice agents become more capable, where should businesses draw the line between conversations that can be fully automated and situations where a human should remain involved?

Businesses should be able to automate anything that’s a repeatable process they want followed, and that’s defined well enough.

But automation becomes challenging in a few areas. One is integrations without APIs: Let’s say you need an integration, but that software doesn’t have an API. That’s difficult to get around, though there are some browser automations you can do. Another is when there are multiple handoffs that go from an agent to a human and back to an agent. The third is context gaps. Say I left my wallet at a restaurant. That’s something AI is never going to be able to solve, because it doesn’t know if you have your wallet, and it can’t go check whether the wallet is at the restaurant.

Some situations will always need a human, along with more complicated work that threads across multiple systems. Avoid using an agent for those.

Phonely recently raised a $16 million Series A after growing to more than 10,000 businesses. What does the next stage of the company look like, and which technical capabilities do you believe will differentiate the next generation of voice AI platforms?

The next stage is building out the team and continuing to invest in it. We’re going all in on this idea. A lot of businesses are venturing into voice AI, and they’ll probably do it decently in different areas. But for us, we’re not just building a voice agent. We’re building an optimization platform, and we’re using voice as a channel. For businesses that care about leads and their phones generating revenue, our whole focus is optimizing AI agents to accomplish the outcomes customers want. Think of Phonely as an optimization company, where people want to optimize their agents to do better.

Your academic research has also explored using AI and smart-meter technology to expand energy access and clean cooking in developing economies. Looking beyond Phonely, how do you see AI being applied to large physical-world problems such as energy access, and is that an area you ultimately hope to return to?

There’s a lot of good that can be done with AI, but a lot of problems that exist are people-inflicted, so there needs to be change from that side. All problems can’t be solved with technology. But allowing technology and human change to happen simultaneously is what will be net positive for the world.

Thank you for the great interview, readers who wish to learn more should visit Phon

Auto-fetched and translated by AI Atlas. All rights belong to the original author.