一种新型计算，而非新工具
当生成式 AI 到来时，Localiza 采取了一种定义后续一切的观点：这不仅仅是工具包中的又一款软件；这是公司运营方式的改变。
公司没有追求孤立的用例，而是围绕三大战略支柱组织与 OpenAI 的合作：通过大规模实验转型业务、创造新的客户界面，以及构建 AI 原生产品。连接这三者的，是围绕智能、上下文和速度重新设计流程的决定。
“我们将其视为一种强大的工具，帮助我们的团队做好他们本已擅长的事情。他们践行我们的文化，理解客户的需求，并知道如何提供优质服务。”
—André Petenussi，Localiza 首席技术官
支柱 1：规模赋能学习
指导内部采用的原则很直接：规模赋能学习，学习创造更好的产品。
在实践中，这意味着将 ChatGPT 许可证作为实验和生产力的资源进行分发，而不仅仅是作为个人福利。Localiza 各部门的团队，从财务到日常运营，都在持续进行自定义 GPT、智能体和 Codex 的活跃实验。
Codex 最初进入技术部门——软件工程师、产品经理、设计师和数据科学家。但采用并未止步于此：分析师、经理、PMO、律师和其他业务专家开始构建自己的原型和自动化，以在其专业领域内检验假设。
这是公司未曾预料到的效果——重新定义了开发之前的阶段，多学科团队在解决方案进入开发队列之前就进行原型设计。
他表示，瓶颈在于人而非技术。
“最大的挑战是帮助一个庞大的团队学习新的工作方式并建立新技能，同时确保我们所有的 19 个部落和 130 个小队高效且一致地工作。”
—André Petenussi，Localiza 首席技术官
支柱 2：客户用自然语言描述他们的需求
第二个领域始于对行为的观察：人们已经开始用自然语言描述他们需要什么，在表达意图的那一刻，为出行服务创造了一个新的入口。
Localiza 的内部数据证实了这一转变：自 2024 年以来，来自 LLM 的流量一直在快速增长，ChatGPT 成为主要渠道。
公司选择尽早在这一领域建立存在。2026 年 3 月，它成为巴西首个——也是拉丁美洲首批之一——提供直接集成到 ChatGPT 中的应用的出行平台。客户可以探索租赁、订阅和二手车选项，并在对话中开始他们的旅程，然后在 Localiza 的安全环境中完成交易。
“我们相信我们能够以这些新方式服务客户，是因为我们早早接受了这一挑战，并在每一步都持续学习。”
—André Petenussi，Localiza 首席技术官
在内部，Localiza 在 ChatGPT 中的应用被视为其本来面目：一个旨在让公司贴近客户行为前沿、加速从真实互动中学习，并探索 AI 如何补充现有生态系统的实验。
支柱 3：Liza，一款围绕对话设计的产品
该公司 AI 原生产品最清晰的例子是 Liza，一个对话式助手，已成为其运营中既定的成功案例。
与传统聊天机器人的区别在于架构，而非界面。Liza 不是遵循固定的决策树，而是围绕对话、意图和上下文构建的：它结合了 OpenAI 模型、Localiza 服务和专业工具，以支持体验的关键部分，从理解初始请求到协助租赁旅程的关键阶段。
其目标是高质量地处理重复性旅程，释放团队进行复杂的、更高价值的互动。
但成功不是以自动化量来衡量的。
Localiza 在获得此类技术之前就已经在映射和衡量其客户服务旅程——这一基础使得能够清晰地决定智能体应在何处介入，以及人类需要在何处干预。
“我们一直致力于让客户旅程尽可能无缝，并且我们开发了衡量该体验质量的指标。我们当然也跟踪转化率，但我们的主要关注点是客户体验。”
—André Petenussi，Localiza 首席技术官
Liza 还充当持续学习的循环：来自互动的聚合、受监控的信号有助于识别需求、摩擦点和服务机会。结合语音能力——语音转文本和文本转语音——这些信号支持其从交易型聊天机器人演变为能够跨不同出行旅程学习和扩展的 AI 原生客户界面。
“这释放了团队的大部分精力，使他们能够更有信心和轻松地处理更有意义的问题，帮助我们为客户履行使命。”
—André Petenussi，Localiza 首席技术官
自 GPT‑3.5 起共同学习
与 OpenAI 的合作开始得很早：Liza 的第一个版本仍运行在 GPT‑3.5 上。
对于 Petenussi 来说，这段关系的核心价值在于两家并行发展的公司之间的主动学习，而非特定功能——合作伙伴理解 Localiza 的挑战及其能力的不同成熟度，无论是在技术内部还是外部。
这是他经常向市场上其他公司提出的论点：对于一项仍在成熟的技术，试错和紧密合作具有相当大的价值。
“我们相信我们仍处于一项正在改变市场、我们的工作方式以及我们使用产品和服务方式的技术的早期阶段。它将在未来 5 年、10 年和 15 年帮助塑造我们的世界。”
—André Petenussi，Localiza 首席技术官
成果概览
2026 年 3 月，Localiza 成为巴西首家、也是拉丁美洲首批之一推出直接集成到 ChatGPT 中的应用的公司；最近，它作为广告商加入了该平台在巴西的广告试点。
自 2024 年以来，来自 LLM 的流量一直在快速增长，ChatGPT 是主要渠道。
Liza 助手拥有 60 多个专业智能体，基于推理模型、MCP 集成和编排租赁旅程的专业工具构建。
公司各部门的团队——从财务到运营——都在持续进行自定义 GPT、智能体和 Codex 的活跃实验。
Codex 进入技术部门并扩展到业务领域，分析师、经理、PMO 和律师构建自己的原型和自动化。
技术运营分为 19 个部落和 130 个小队。
给领导者的经验教训
将 AI 视为一种新的运营模式，而非生产力层。最大的收益来自于围绕智能、上下文和速度重新设计流程。
利用规模来产生学习。同时进行大规模实验比完美的试点能产生更多的组织知识。
将实验扩展到技术之外。当律师、PMO 和分析师在他们自己的领域内进行原型设计时，上游工作的性质就会改变。
在自动化之前绘制并衡量你的旅程。拥有质量指标可以让你决定代理应在何处介入，以及何处人类不可或缺。
采用的瓶颈是再培训，而不是模型。改变数百人的工作方式比整合技术更难——也更具决定性。
提示
从客户的问题和旅程开始，而不是从模型功能开始。
将许可证分配为实验预算，而不是个人福利。
在扩展对话代理之前建立质量指标。
将出现在新的对话渠道中视为一次学习实验，而不是一次获客活动。
闭环：利用每次交互产生的信号来改进产品本身。
后续步骤
Localiza 的雄心是将 AI 置于其下一次体验和流程的中心——无论是在工作如何产生、为客户创造的体验，还是运营问题的解决方面。
已经在与 OpenAI 一起开发的下一批领域包括 Liza 和销售导向代理的演进、将对话界面建立为客户入口点，以及代理式软件开发的新周期。
“主要关注点是我们如何创造更好的体验。无论是为了提高我们业务的盈利能力，确保其健康，还是最重要的是，为我们的客户和员工创造愉悦的体验。”
—André Petenussi，Localiza 首席技术官
A new kind of computing, not a new tool
When generative AI arrived, Localiza took a view that defined everything that followed: this was more than another piece of software in the toolkit; it was a change in the way the company operates.
Rather than pursuing isolated use cases, the company organized its collaboration with OpenAI around three strategic pillars: transforming the business through experimentation at scale, creating new customer interfaces, and building AI-native products. What connects all three is the decision to redesign processes around intelligence, context, and speed.
“We see this as a powerful tool that helps our teams do what they already do well. They live our culture, understand our customers’ needs, and know how to deliver great service.”
—André Petenussi, CTO, Localiza
Pillar 1: scale enables learning
The principle guiding internal adoption is straightforward: scale enables learning, and learning creates better products.
In practice, that meant distributing ChatGPT licenses as a resource for experimentation and productivity, rather than simply as an individual benefit. Teams across Localiza, from finance to day-to-day operations, maintain active experiments with custom GPTs, agents, and Codex.
Codex first arrived in the technology department—software engineers, product managers, designers, and data scientists. But adoption did not stop there: analysts, managers, PMOs, lawyers, and other business specialists began building their own prototypes and automations to test hypotheses within their areas of expertise.
This was an effect the company had not anticipated—a redefinition of the stages before development, with multidisciplinary teams prototyping solutions before they enter a development queue.
The bottleneck, he says, is human rather than technical.
“The big challenge is helping a large team learn new ways of working and build new skills, while ensuring that all 19 of our tribes and 130 squads work efficiently and consistently.”
—André Petenussi, CTO, Localiza
Pillar 2: customers describe their needs in natural language
The second area starts with an observation about behavior: people have begun describing what they need in natural language, creating a new entry point for mobility services—at the very moment they express their intent.
Localiza’s internal data confirms this shift: traffic originating from LLMs has been growing rapidly since 2024, with ChatGPT emerging as the main channel.
The company chose to establish a presence in this space early. In March 2026, it became Brazil’s first mobility platform—and one of the first in Latin America—to offer an app integrated directly into ChatGPT. Customers can explore rental, subscription, and used-car options and begin their journey in the conversation, before completing the transaction in Localiza’s secure environment.
“We believe we’ve been able to serve our customers in these new ways because we took on this challenge early and have kept learning at every step.”
—André Petenussi, CTO, Localiza
Internally, Localiza’s app in ChatGPT is treated as what it is: an experiment designed to keep the company close to the frontier of customer behavior, accelerate learning from real interactions, and explore how AI complements the existing ecosystem.
Pillar 3: Liza, a product designed around conversation
The company’s clearest example of an AI-native product is Liza, a conversational assistant that has become an established success story in its operations.
The difference from a traditional chatbot lies in the architecture, not the interface. Rather than following a fixed decision tree, Liza was built around conversation, intent, and context: it combines OpenAI models, Localiza services, and specialized tools to support key parts of the experience, from understanding the initial request to assisting with key stages of the rental journey.
The ambition is to handle repetitive journeys with quality and free teams for complex, higher-value interactions.
But success is not measured by the volume of automation.
Localiza was already mapping and measuring its customer service journeys before it had access to this type of technology—and that foundation made it possible to decide clearly where an agent should step in and where a human needs to intervene.
“We’ve always been committed to making the customer journey as seamless as possible, and we’ve developed metrics to measure the quality of that experience. We track conversion too, of course, but our main focus is the customer experience.”
—André Petenussi, CTO, Localiza
Liza also serves as a continuous learning loop: aggregated, monitored signals from interactions help identify needs, points of friction, and service opportunities. Combined with voice capabilities—voice-to-text and text-to-voice—these signals support its evolution from a transactional chatbot into an AI-native customer interface that can learn and scale across different mobility journeys.
“This frees up much of the team to tackle more meaningful problems with greater confidence and ease, helping us deliver on our mission for our customers.”
—André Petenussi, CTO, Localiza
Learning together, since GPT‑3.5
The partnership with OpenAI began early: the first version of Liza still ran on GPT‑3.5.
For Petenussi, the central value of the relationship lies in active learning between two companies evolving in parallel, rather than in a specific feature—with the partner understanding Localiza’s challenges and the different maturity levels of its capabilities, both within and outside technology.
It is an argument he often makes to other companies in the market: with a technology that is still maturing, trial and error and close collaboration have considerable value.
“We believe we’re still in the early years of a technology that is transforming markets, how we work, and how we use products and services. It will help shape our world over the next 5, 10, and 15 years.”
—André Petenussi, CTO, Localiza
Results overview
In March 2026, Localiza became the first company in Brazil, and one of the first in Latin America, to launch an app integrated directly into ChatGPT; more recently, it joined the platform’s advertising pilot in Brazil as an advertiser.
Traffic from LLMs has been growing rapidly since 2024, with ChatGPT as the main channel.
The Liza assistant operates with more than 60 specialized agents, built on reasoning models, MCP integration, and specialized tools that orchestrate the rental journey.
Teams across the company—from finance to operations—maintain active experiments with custom GPTs, agents, and Codex.
Codex reached the technology department and spread to business areas, with analysts, managers, PMOs, and lawyers building their own prototypes and automations.
The technology operation is organized into 19 tribes and 130 squads.
Lessons for leaders
Treat AI as a new operating model, rather than a productivity layer. The greatest gains come from redesigning processes around intelligence, context, and speed.
Use scale to generate learning. Simultaneous experiments at scale produce more organizational knowledge than a perfect pilot.
Take experimentation beyond technology. When lawyers, PMOs, and analysts prototype within their own areas, the nature of upstream work changes.
Map and measure your journeys before automating them. Having quality metrics in place enables you to decide where an agent should intervene and where a human is indispensable.
The adoption bottleneck is reskilling, not the model. Changing how hundreds of people work is harder—and more decisive—than integrating the technology.
Tips
Start with the customer’s problem and journey, rather than a model feature.
Allocate licenses as an experimentation budget, rather than an individual benefit.
Establish quality metrics before scaling conversational agents.
Treat a presence in new conversational channels as a learning experiment, rather than an acquisition campaign.
Close the loop: use the signals generated by each interaction to improve the product itself.
Next steps
Localiza’s ambition is to put AI at the center of its next experiences and processes—whether in how work is produced, the experience created for customers, or the resolution of an operational problem.
The next areas already being developed with OpenAI include the evolution of Liza and sales-focused agents, the establishment of conversational interfaces as a customer entry point, and the new cycle of agentic software development.
“The main focus is how we create better experiences. Whether that is to improve the profitability of our business, to ensure its health, or, above all, to create delightful experiences for our customers and our employees.”
—André Petenussi, CTO, Localiza
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文