75%。
这是OpenAI自家英文客服热线1-888-GPT-0090目前由AI独立解决的来电比例。7月22日，OpenAI正式发布企业级AI Agent平台OpenAI Presence——一个经过多年企业级部署打磨的产品，能让企业部署可信的语音和聊天Agent来处理客户支持、外呼销售和高风险内部工作流。BBVA在墨西哥探索银行业务语音支持，软银在测试自然日语客户对话，IAG在探索极端天气等高峰期的保险理赔支持。这不是一个API更新，这是OpenAI的第二次战略转型：从卖token变成卖软件。
上一次类似的转型发生在2023年3月，ChatGPT从研究预览变成商业化产品，OpenAI从实验室变成了API公司。三年后，Presence的发布意味着OpenAI要更进一步——它不再满足于做模型供应商，等客户自己把API缝成可用的系统；它要自己下场，把模型、策略、护栏、评估工具、改进闭环打包成一个完整的企业软件产品，直接卖给CIO和CTO。
Presence到底是什么？不是ChatGPT Enterprise 2.0
很多人看到“OpenAI发布企业Agent”的第一反应是：这不就是ChatGPT Enterprise加了Agent功能吗？不是。
ChatGPT Enterprise解决的是“让员工安全地用大模型”的问题——一个聊天界面，加上企业级安全和数据隔离。Presence解决的是一个完全不同的问题：让AI Agent在生产环境中可靠地替公司干活。这两者之间的距离，比大多数人想象的要大得多。
根据OpenAI官方博客的描述，Presence的每个部署从一个具体的“活儿”开始：解决账单问题、处理保险理赔、响应员工IT服务请求。Agent只获得完成这项工作所需的知识和系统访问权限。公司设定策略：Agent能做什么、什么时候需要审批、什么时候必须转人工。上线后，生产会话和转接案例暴露系统短板，Codex自动提出更新建议，团队测试审批后再发布。
这段话听起来简单，但里面包含了企业软件最核心的几个要素：权限最小化、策略引擎、护栏系统、仿真测试、持续改进闭环。这不是给一个大模型加个system prompt就能解决的问题。这是Salesforce、ServiceNow、SAP们花了二十年搭建的东西。
OpenAI举了自家客服线的数据：Presence上线后几周内就达到甚至超过了人工客服的质量基准，现在独立解决75%的来电问题；更关键的是，利用Codex驱动的改进循环，在短短10天内人工转接率下降了15个百分点。10天、15个百分点——这个改进速度是传统企业软件迭代周期的零头。
为什么是现在？三个信号表明时机已到
OpenAI不是今天才想做企业软件。Presence的发布时机背后，是三个关键信号的交汇。
第一个信号：Agent可靠性终于跨过了生产可用门槛。2024年的Agent还停留在demo阶段——能订个餐厅、能查个天气，但一到真实业务场景就掉链子：记错用户信息、违反公司政策、在关键步骤卡住不知道转人工。2025年到2026年，随着推理模型成熟、工具调用可靠性提升、评估体系完善，Agent在特定垂直场景的表现开始接近甚至超过人类一线员工。NTT DATA集团用ChatGPT Enterprise加Codex后，事故分析时间从数小时缩短到30分钟，9000名员工在使用——这不是试点，这是规模化部署。
第二个信号：企业不再问“AI能不能用”，而是问“怎么让AI可靠地干活”。OpenAI在博客开篇就点明了这个转变：“企业面临的挑战不再是证明AI Agent能工作，而是让它们足够可靠以在生产中完成高价值工作。”过去两年，几乎所有大企业都做过AI试点，但真正跑到生产环境、产生可量化ROI的比例很低。瓶颈不在模型能力，而在把模型、系统、流程、权限、监控缝合成一个可靠系统的工程能力。大多数企业没有这个能力。Presence就是OpenAI给这个问题交出的答卷。
“企业面临的挑战不再是证明AI Agent能工作，而是让它们足够可靠以在生产中完成高价值工作。”—— OpenAI官方博客
第三个信号：竞争压力倒逼OpenAI从平台走向应用。Anthropic在企业市场的攻势越来越猛——Claude for Work已经在很多企业与ChatGPT Enterprise正面竞争；Google的Gemini Enterprise Agent Platform也在同一天发布新模型；微软Copilot虽然生态绑定强，但OpenAI需要保持自己独立于微软的企业客户关系。更重要的是，ServiceNow、Salesforce这些传统企业软件巨头正在把AI Agent深度嵌入自己的产品——如果OpenAI只做底层模型，它就会被这些应用层公司管道化。Presence是OpenAI说：“我不只是你的模型供应商，我也是你的应用层竞争对手。”
Presence的产品哲学：控制环比模型更重要
仔细读Presence的产品设计，你会发现一个有意思的重心转移：OpenAI不再把模型本身当作核心卖点，而是把“控制环”（control loop）当作核心卖点。
什么是控制环？就是围绕模型的一整套系统：策略定义（Agent能做什么不能做什么）、护栏（实时检测和拦截违规行为）、审批流程（哪些动作需要人确认）、仿真测试（上线前模拟各种场景）、评估工具（量化Agent表现）、持续改进（生产数据反馈驱动模型和策略更新）。这些组件单独看都不新鲜，但把它们整合成一个开箱即用的产品，OpenAI是第一家。
这种设计哲学反映了一个深刻的行业认知转变：在企业场景中，模型能力只是基础，可靠性和可控性才是决定能否投产的关键。一个准确率99%但会在1%的情况下闯祸的Agent，在客服场景中可能意味着错误退款、违规承诺、隐私泄露——这些风险的代价远大于1%的效率提升。Presence的核心价值主张不是“我的模型最聪明”，而是“我的Agent不会把你的生意搞砸，而且越用越好”。
Codex驱动的改进闭环尤其值得关注。传统企业软件的更新周期是月度甚至季度级别的——业务部门提需求，IT排期开发，测试，上线。Presence的做法是：生产环境中的每次转接、每次差评、每次人工干预都是信号，Codex自动分析这些信号、定位问题根因、生成策略或配置更新建议，团队审批后灰度发布。这个循环把Agent的迭代速度从“季度”压缩到了“天”。前面提到的10天内人工转接率下降15个百分点，就是这个闭环的威力。
企业AI市场的格局重塑
Presence的发布将对企业AI市场产生深远影响，首当其冲的是三类公司。
第一类是传统企业软件巨头：Salesforce、ServiceNow、SAP、Oracle。这些公司过去一年都在拼命往自己的产品里加AI功能，但本质上是在旧架构上贴AI补丁。Presence是从Agent原生的角度重新设计的产品——策略引擎、护栏、评估、改进闭环都是围绕Agent工作流设计的。如果OpenAI在垂直场景（客服、IT支持、HR、财务）做到足够深的产品化，这些传统厂商面临的就不只是“加个AI功能”的竞争，而是架构层面的代差。不过传统巨头的优势是数据和工作流深度——Salesforce里有你全部的客户数据，ServiceNow里跑着你全部的IT流程，这些不是一个新Agent平台短期内能替代的。
第二类是AI Agent创业公司。过去两年涌现了大量企业Agent初创公司——做客服Agent的、做IT支持Agent的、做销售Agent的。Presence的发布直接杀入了它们的赛道。OpenAI的优势是模型能力、品牌认知、资金储备；创业公司的优势是垂直场景深度、灵活性、客户响应速度。接下来12个月会是一场残酷的淘汰赛——就像AWS的崛起没有灭掉所有企业SaaS，但确实把基础设施层的创业公司清洗了一遍。
第三类是微软。这可能是最微妙的关系。OpenAI和微软的关系在过去一年经历了多次紧张时刻——从Anthropic合作到Copilot竞争。Presence作为OpenAI独立推出的企业软件产品，在某种程度上和Microsoft 365 Copilot形成了竞争关系。但微软的企业渠道、Office生态、Azure基础设施又是OpenAI短期内无法替代的。这种“既合作又竞争”的关系会继续持续，双方都在小心翼翼地维持平衡。
对中国市场而言，Presence短期内不会直接进入，但它的产品形态给国内大模型公司指明了方向：只卖API和模型授权是不够的，企业需要的是完整的、可信赖的、能持续改进的Agent系统。飞书aily、钉钉悟空、百度文心智能体都在往这个方向走，但在策略引擎、护栏系统、持续改进闭环这些“无聊但关键”的企业级组件上，和Presence比还有明显差距。
终局判断：AI的“Oracle时刻”
回顾企业软件的历史，1977年Oracle的成立是一个标志性事件。在Oracle之前，数据库只是IBM大型机上的一个组件；Oracle把关系型数据库做成了独立的企业软件产品，开启了一个价值数千亿美元的市场。
Presence可能是AI Agent领域的“Oracle时刻”——把大模型从一个技术组件变成了一个独立的企业软件品类。当然，OpenAI不是唯一在做这件事的公司，但它是第一个把“模型+策略+护栏+评估+改进”完整打包成产品、并且有大规模生产数据（自家客服75%解决率）背书的公司。
这条路并不容易。企业软件的护城河从来不是技术领先，而是深度嵌入客户业务流程后的切换成本、合规认证、行业Know-how、服务网络。OpenAI在这些方面几乎是从零开始。Presence目前还是Limited GA，只对合格企业客户开放，由OpenAI的Forward Deployed Engineers和系统集成商主导部署——这意味着它还不是一个可以自助购买的SaaS产品，服务交付能力是瓶颈。
但方向是明确的。当AI Agent的可靠性跨过生产门槛，当企业有强烈的降本增效压力，当OpenAI这样的公司愿意把模型能力封装成解决方案而不是只卖API——企业软件的AI重构就开始了。Salesforce用了20年做到年营收300亿美元，ServiceNow用了15年做到年营收100亿美元。AI原生的企业软件公司会用多长时间？Presence给出了一个值得关注的起点。
明天见。
75%.
That's the share of incoming calls now handled independently by AI on OpenAI's own English-language support line, 1-888-GPT-0090. On July 22, OpenAI officially launched OpenAI Presence — a battle-tested enterprise AI agent platform forged through years of production deployments, enabling companies to deploy trusted voice and chat agents for customer support, outbound sales, and high-risk internal workflows. BBVA is exploring AI-powered voice banking in Mexico; SoftBank is testing natural Japanese customer conversations; IAG is exploring insurance claims support during peak events like severe weather. This is not an API update. This is OpenAI's second strategic transformation: from selling tokens to selling software.
The last comparable pivot happened in March 2023, when ChatGPT went from research preview to commercial product and OpenAI went from lab to API company. Three years later, Presence means OpenAI is going further — it's no longer content to be a model vendor waiting for customers to stitch APIs into usable systems; it's going straight to market itself, packaging models, policies, guardrails, evaluation tools, and an improvement loop into a complete enterprise software product sold directly to CIOs and CTOs.
What Exactly Is Presence? It's Not ChatGPT Enterprise 2.0
Many people's first reaction to “OpenAI launches enterprise agents” is: isn't this just ChatGPT Enterprise with agent features bolted on? It isn't.
ChatGPT Enterprise solves the problem of “letting employees use LLMs safely” — a chat interface with enterprise-grade security and data isolation. Presence solves a fundamentally different problem: making AI agents reliably do real work in production. The gap between those two is wider than most people realize.
According to OpenAI's official blog, every Presence deployment starts with a specific “job”: resolving billing issues, handling insurance claims, responding to employee IT service requests. The agent receives only the knowledge and system access required for that job. The company sets policies: what the agent can do, when it needs approval, when it must escalate to a human. After launch, production sessions and escalations expose gaps; Codex proposes updates; teams test and approve before rollout.
That sounds simple, but it packs in the core elements of enterprise software: least-privilege access, policy engines, guardrail systems, simulation testing, continuous improvement loops. You can't solve this by slapping a system prompt on a frontier model. This is the stuff Salesforce, ServiceNow, and SAP spent two decades building.
OpenAI shared data from its own support line: within weeks of deployment, Presence met or exceeded human support quality benchmarks and now independently resolves 75% of incoming calls; crucially, using the Codex-powered improvement loop, human handoffs dropped by 15 percentage points in just 10 days. Ten days, 15 points — that iteration speed is a fraction of traditional enterprise software cycles.
Why Now? Three Signals the Timing Is Right
OpenAI didn't wake up yesterday wanting to build enterprise software. Three key signals converged to make Presence's launch inevitable.
Signal one: Agent reliability has finally crossed the production threshold. Agents in 2024 were still demo-stage — they could book a restaurant or check the weather, but fell apart in real business scenarios: misremembering user data, violating company policy, getting stuck at critical steps without knowing to escalate. From 2025 into 2026, with reasoning models maturing, tool-calling reliability improving, and evaluation frameworks solidifying, agent performance in specific verticals began approaching or even surpassing frontline human workers. NTT DATA Group cut incident analysis from hours to 30 minutes using ChatGPT Enterprise and Codex across 9,000 employees — that's not a pilot; that's scaled deployment.
Signal two: Enterprises stopped asking “can AI work?” and started asking “how do we make AI work reliably?” OpenAI states this shift plainly in the blog's opening: “The challenge for enterprises is no longer proving that AI agents can work, it's making them reliable enough to do high-value work in production.” Over the past two years, nearly every large enterprise ran AI pilots, but the share that made it to production with measurable ROI was low. The bottleneck wasn't model capability — it was the engineering work of stitching models, systems, processes, permissions, and monitoring into a reliable system. Most enterprises don't have that capability in-house. Presence is OpenAI's answer.
“The challenge for enterprises is no longer proving that AI agents can work, it's making them reliable enough to do high-value work in production.”— OpenAI Official Blog
Signal three: Competitive pressure is forcing OpenAI up the stack from platform to application. Anthropic's enterprise offensive is intensifying — Claude for Work is already competing head-to-head with ChatGPT Enterprise in many accounts; Google's Gemini Enterprise Agent Platform launched new models the same day; Microsoft Copilot has strong ecosystem lock-in, but OpenAI needs to maintain enterprise customer relationships independent of Microsoft. Crucially, traditional enterprise software giants like ServiceNow and Salesforce are embedding AI agents deep into their products — if OpenAI only sells base models, it risks being piped by these application-layer companies. Presence is OpenAI saying: “I'm not just your model vendor; I'm also your application-layer competitor.”
Presence's Product Philosophy: The Control Loop Matters More Than the Model
Read Presence's design carefully, and you'll notice an interesting center of gravity shift: OpenAI is no longer selling the model as the core value proposition; it's selling the “control loop” as the core value proposition.
What's a control loop? It's the entire system around the model: policy definition (what the agent can and can't do), guardrails (real-time detection and interception of violations), approval workflows (which actions require human confirmation), simulation testing (running scenarios before launch), evaluation tooling (quantifying agent performance), and continuous improvement (production data feeding back into model and policy updates). None of these components is new in isolation, but packaging them into an out-of-the-box product? OpenAI is the first.
This design philosophy reflects a deep industry realization: in enterprise settings, model capability is table stakes; reliability and controllability are what determine whether something ships to production. An agent that's 99% accurate but causes havoc in the 1% — wrong refunds, policy violations, privacy leaks — carries risks whose costs far outweigh that 1% efficiency gain. Presence's core promise isn't “my model is the smartest”; it's “my agent won't screw up your business, and it gets better over time.”
The Codex-powered improvement loop deserves special attention. Traditional enterprise software updates on monthly or even quarterly cycles — business submits requirements, IT schedules development, tests, ships. Presence's approach: every escalation, every negative rating, every human intervention in production is a signal; Codex automatically analyzes these signals, diagnoses root causes, generates policy or configuration updates; teams approve and gradually roll out. This loop compresses agent iteration from “quarters” to “days.” The 15-point handoff reduction in 10 days is exactly this loop at work.
Reshaping the Enterprise AI Landscape
Presence's launch will have ripple effects across enterprise AI, and three categories of companies stand in the blast zone.
First, traditional enterprise software giants: Salesforce, ServiceNow, SAP, Oracle. All of these have been furiously adding AI features to their products over the past year, but they're essentially bolting AI onto legacy architectures. Presence is designed from an agent-native perspective — policy engines, guardrails, evaluation, improvement loops all built around agent workflows. If OpenAI achieves sufficient product depth in verticals (customer service, IT support, HR, finance), these incumbents face not just an “add AI features” competition but an architectural generation gap. That said, the incumbents' advantage is data and workflow depth — Salesforce holds all your customer data; ServiceNow runs all your IT processes — things a new agent platform can't displace overnight.
Second, AI agent startups. The past two years saw a flood of enterprise agent startups — customer service agents, IT support agents, sales agents. Presence enters their lanes directly. OpenAI's advantages are model capability, brand recognition, and capital; startups' advantages are vertical depth, agility, and customer responsiveness. The next 12 months will be a brutal shakeout — just as AWS's rise didn't kill all enterprise SaaS, but it certainly cleaned out infrastructure-layer startups.
Third, Microsoft. This is the most delicate relationship. OpenAI and Microsoft's partnership has seen multiple tensions over the past year — from the Anthropic partnership to Copilot competition. Presence, as an independently launched enterprise software product from OpenAI, competes to some extent with Microsoft 365 Copilot. But Microsoft's enterprise channels, Office ecosystem, and Azure infrastructure are irreplaceable for OpenAI in the near term. This “frenemy” dynamic will continue, with both sides carefully maintaining balance.
For the Chinese market, Presence won't enter directly any time soon, but its product shape points the way for domestic LLM companies: selling APIs and model licenses alone isn't enough; enterprises need complete, trustworthy, continuously improving agent systems. Feishu aily, DingTalk Wukong, and Baidu Wenxin Agents are all heading in this direction, but on the “boring but critical” enterprise components — policy engines, guardrail systems, continuous improvement loops — there remains a clear gap versus Presence.
Endgame: AI's “Oracle Moment”
Look back at enterprise software history, and Oracle's founding in 1977 stands as a landmark. Before Oracle, databases were just a component on IBM mainframes; Oracle turned the relational database into an independent enterprise software product, opening a market worth hundreds of billions of dollars.
Presence may be AI agents' “Oracle moment” — turning LLMs from a technical component into an independent enterprise software category. OpenAI isn't the only company working on this, of course, but it's the first to package “model + policy + guardrails + evaluation + improvement” into a complete product backed by large-scale production data (75% resolution on its own support line).
This road won't be easy. Enterprise software moats were never built on technical leadership alone — they're built on switching costs from deep embedding in customer workflows, compliance certifications, industry know-how, and service networks. OpenAI is starting from near-zero on all of these. Presence is still limited GA, available only to eligible enterprise customers through Forward Deployed Engineers and select system integrators — meaning it's not yet a self-serve SaaS product; service delivery capacity is the bottleneck.
But the direction is clear. When AI agent reliability crosses the production threshold, when enterprises face intense cost pressure, when a company like OpenAI is willing to package model capabilities into solutions rather than just selling APIs — the AI reconstruction of enterprise software has begun. Salesforce took 20 years to reach $30 billion in annual revenue; ServiceNow took 15 years to hit $10 billion. How fast will AI-native enterprise software companies move? Presence is a noteworthy starting line.
See you tomorrow.
OpenAI Presence · 企业级Agent · 75%自动解决率 · BBVA · 软银 · IAG · 策略引擎 · 护栏系统 · Codex改进闭环 · 企业软件 · AI商业化
OpenAI Presence · enterprise agents · 75% auto-resolution · BBVA · SoftBank · IAG · policy engine · guardrails · Codex improvement loop · enterprise software · AI commercialization
