1亿用户。日均1000万次咨询。这不是某个C端聊天机器人的数据，是一个健康AI Agent的成绩单。
WAIC 2026上，蚂蚁集团首次完整展示了面向智能体商业时代的三层AI布局：AI应用层、智能体商业生态层、技术基座层。其中最亮眼的数字来自应用层——健康AI「阿福」用户数已突破1亿，日均处理超过1000万次健康咨询。
1亿用户是什么概念？很多大模型公司全产品加起来都没这么多用户。而且这1亿不是「注册了但不用」的僵尸用户，是每天都在问健康问题的活跃用户。这可能是目前国内用户量最大的垂直AI Agent。
三层布局：从技术到生态到应用，蚂蚁想做智能体时代的「操作系统」
蚂蚁的三层布局，思路很清晰，野心也很大。
最上层是AI应用层——自己下场做标杆产品。健康AI「阿福」、AI版支付宝「阿宝」，这些都是蚂蚁自己做的Agent。为什么自己做？因为要打样——告诉合作伙伴和开发者，「智能体应该这么做，能做到这个效果」。同时也积累用户、数据和场景理解。
中间层是智能体商业生态层——这是蚂蚁真正的野心所在。AI支付已支持3亿笔智能体支付，适配了95%的通用智能体框架。什么意思？就是不管你用什么框架做的Agent，只要接入了支付宝的AI支付能力，就能自动完成支付。这就像智能体时代的「支付基础设施」——就像当年移动支付时代的支付宝一样，只不过这次服务的对象从人变成了Agent。
最底层是技术基座层——百灵大模型、灵波具身智能、OceanBase AI数据库、安全可信能力。这些是支撑上面两层的基础。蚂蚁不跟别人比单模型跑分，比的是「模型+数据+场景+支付」的综合能力。
三层叠在一起，你会发现蚂蚁想做的不是某一个爆款Agent，而是智能体时代的「操作系统」——你做Agent可以用我的模型、接我的支付、跑在我的生态里。就像Windows时代，微软自己也做Office，但更重要的是所有软件都跑在Windows上。
为什么健康AI先跑出来？场景选对了，事半功倍
阿福能跑到1亿用户，不是偶然的。健康这个场景，天生适合AI Agent。
第一，需求高频且普遍。每个人都会有健康疑问——「这个症状要不要去医院？」「这个药有没有副作用？」「体检报告上这个指标偏高是什么意思？」这些问题不大不小，去医院嫌麻烦，上网搜怕不靠谱，问AI正好。
第二，AI能真正帮上忙。健康咨询不是什么高难度的医疗诊断，就是信息整合和科普——把专业的医学知识用普通人能听懂的话讲出来。这件事AI特别擅长——知识量比任何一个医生都大，而且24小时在线，永远耐心。
第三，有支付宝的流量入口。很多AI产品的最大难题是获客——你做得再好，没人知道也白搭。但阿福不一样，它长在支付宝里，支付宝有几亿用户。只要产品体验过得去，用户获取成本几乎为零。
"Agent创业最大的误区：总想着做通用Agent，却忽略了垂直场景的真实需求。先在一个场景跑通1亿用户，比做一个什么都能做但没人用的通用Agent有价值得多。"—— Dawn Vision编辑部
阿福的1亿用户，给整个行业提了个醒：别总想着做「下一个ChatGPT」，先找一个真实的、高频的、有流量入口的垂直场景，把它做深做透。用户量上来了，数据就有了；数据有了，模型就更强了；模型更强了，用户体验就更好了——这是一个正向循环。
智能体的竞争，从来不比谁更「通用」，比谁更「有用」。
明天见。
100 million users. 10 million daily consultations. This isn't data from some consumer chatbot — it's a report card for a health AI agent.
At WAIC 2026, Ant Group for the first time fully showcased its three-layer AI architecture for the agent business era: AI application layer, agent business ecosystem layer, and technology foundation layer. The most eye-catching number comes from the application layer — health AI "A Fu" has surpassed 100 million users, handling over 10 million health consultations daily.
What does 100 million users mean? Many LLM companies don't have that many users across all their products combined. And these 100 million aren't zombie users who "signed up but never use" — they're active users asking health questions every day. This is likely China's largest vertical AI agent by user count.
Three Layers: From Tech to Ecosystem to Apps — Ant Wants to Be the 'OS' of the Agent Era
Ant's three-layer strategy is clear in its thinking — and ambitious in its scope.
Top layer: AI application layer — building flagship products themselves. Health AI "A Fu," AI-powered Alipay "A Bao" — these are agents Ant built itself. Why build your own? To set the example — showing partners and developers "this is how agents should be built, and this is what they can achieve." It also accumulates users, data, and scenario understanding.
Middle layer: agent business ecosystem layer — this is where Ant's real ambition lies. AI payment already supports 300 million agent transactions, adapted to 95% of common agent frameworks. What does that mean? No matter what framework you use to build your agent, as long as it integrates Alipay's AI payment capabilities, it can automatically complete payments. It's like "payment infrastructure" for the agent era — just like Alipay in the mobile payment era, except this time it's serving agents instead of people.
Bottom layer: technology foundation — Bailin LLM, Lingbo embodied AI, OceanBase AI database, trusted security capabilities. These are the foundation supporting the two layers above. Ant doesn't compete on single-model benchmarks — it competes on the combined capability of "model + data + scenario + payment."
Stack all three layers together and you see what Ant really wants to be — not just another hit agent, but the "operating system" of the agent era. You build agents using my models, connecting to my payment, running in my ecosystem. Just like in the Windows era — Microsoft made Office too, but what mattered more was that all software ran on Windows.
Why Did Health AI Break Out First? Pick the Right Scenario and Everything Gets Easier
A Fu reaching 100 million users isn't an accident. The health scenario is天生 (naturally) suited for AI agents.
First: high-frequency, universal demand. Everyone has health questions — "should I go to the hospital for this symptom?" "does this medicine have side effects?" "what does this high indicator on my physical exam mean?" These questions aren't serious enough for a hospital visit, but searching online feels unreliable — asking AI is perfect.
Second: AI can actually help. Health consultation isn't high-stakes medical diagnosis — it's information synthesis and科普 (health education): translating professional medical knowledge into language ordinary people understand. This is something AI excels at — it has more knowledge than any single doctor, it's available 24/7, and it's always patient.
Third: Alipay's traffic entry point. The biggest challenge for many AI products is user acquisition — however good your product is, it's useless if nobody knows about it. But A Fu is different — it lives inside Alipay, and Alipay has hundreds of millions of users. As long as the product experience is decent, user acquisition cost is essentially zero.
"The biggest mistake in agent startups: always trying to build general agents while ignoring real demand in vertical scenarios. Getting 100M users in one scenario first is way more valuable than building a general agent that does everything but nobody uses."— Dawn Vision Editorial
A Fu's 100 million users is a wake-up call for the entire industry: stop trying to build "the next ChatGPT." First find a real, high-frequency vertical scenario with a traffic entry point, and go deep on it. When user volume goes up, you get data; with data, your model gets better; with a better model, user experience improves — it's a virtuous cycle.
The competition in agents is never about who's more "general." It's about who's more "useful."
See you tomorrow.
Agent创业最大的误区：总想着做通用Agent，却忽略了垂直场景的真实需求。先在一个场景跑通1亿用户，比做一个什么都能做但没人用的通用Agent有价值得多。
—— Dawn Vision编辑部
The biggest mistake in agent startups: always trying to build general agents while ignoring real demand in vertical scenarios. Getting 100M users in one scenario first is way more valuable than building a general agent that does everything but nobody uses.
— Dawn Vision Editorial
Ant Group · three-layer agent business · health AI A Fu · 100M users · 300M AI payments · agent OS · vertical scenarios
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 9 个源信号生成，经编辑部人工审核。素材来源：36氪、WAIC现场报道。
This article was generated by the Dawn Vision cognitive engine processing 9 source signals, with editorial review. Sources: 36Kr, WAIC on-site reports.
