Why Ant's Health App Is Betting on the Family Care Manager

When Ant Group rebranded its health AI app to Ant Afu in late 2025, it came with some big numbers. Monthly active users jumped from 15 million to 30 million in a month, daily health questions hit 10 million, and by early 2026, total users crossed 100 million. But third-party tracking told a different story: around 29 million MAU in June, with users opening the app about 18 times a month and spending just 13 minutes total on it.

Those numbers can be spun two ways. Optimists see a breakout for health AI. Skeptics see a classic case of red-packet-fueled downloads that don't stick. Both takes miss something crucial: they treat total users, MAU, and retention as if they're directly comparable. They're not. The company's own metrics and third-party data are measured differently, and no public data can prove how many people actually got healthier.

The real question for product people is simpler: How does an app positioned as an 'AI health friend' turn a one-off health question into a long-term, trustworthy, actionable relationship?

After digging through team interviews, product updates, public data, and user feedback, my take is this: Afu's most valuable unit isn't the individual patient. It's the 'family health manager'—the person who handles health tasks for parents, kids, spouse, and themselves. The product tries to bundle scattered, low-frequency medical needs into an ongoing family duty using family profiles, long-term records, health behaviors, and real-world services. Whether that works depends not on daily chats but on whether the product can safely move someone to a clear next step, and then use the data from that step to make the next one better.

Let me be clear: 'family health manager' is a model I'm proposing based on public materials, not a proven conclusion. What the team wants, what the product currently does, and whether users actually benefit are three separate questions. I'll keep them apart.

Beyond the chatbot surface

On the surface, Afu looks like a medical chatbot. You describe symptoms, upload a lab report or snap a pill bottle, and the AI asks follow-ups and gives advice. Reviews usually focus on answer accuracy or how it stacks up against DeepSeek or Doubao.

But Afu isn't content to stay in Q&A. In 2024, Ant launched an AI health assistant inside Alipay. By June 2025, it had a standalone app called AQ, renamed Afu that December, with the product structure reorganized into health Q&A, health companionship, and health services. In 2026, Ant's CEO summarized the direction as '1+3+X': professionalism as the base, plus anthropomorphism, personalization, and action-taking, expanding into specialty areas.

This evolution reflects a different view of health needs. Hospitals handle diagnosis and treatment. Daily life covers eating, sleeping, and exercise. But there's a big gray zone in between: Should you rush a feverish kid to the ER? What does that nodule on your dad's report mean? Can Mom's new medication be taken with her old one? Is that health article she forwarded trustworthy? Which department should your brother see for a second opinion, and what documents does he need?

These questions don't always require a hospital visit, but they shape what a family does next. Old search engines dump information. Generic AI chatbots generate an explanation. But the user still has to decide and act. Afu tries to go two steps further: fill in the health context, then connect to doctors, hospitals, insurance, or health management services.

So, a more accurate framing is: Q&A is the most visible entry point, but Afu wants to extend value beyond the answer.

The hidden primary user: the family health manager

In a 2025 interview, Ant's health division president admitted the team initially expected young, AI-savvy users to adopt AQ first. Instead, they found a wave of users aged 40–60. These folks are starting to deal with their own abnormal lab results, sleep issues, and chronic conditions—while also managing health tasks for their aging parents and school-age kids.

For a healthy young adult, serious medical needs are rare. One physical a year, maybe a cold. That doesn't sustain daily app usage. But family health responsibilities aren't evenly distributed. In most families, one person ends up as the de facto coordinator: gathering information, assessing risk, booking appointments, accompanying to visits, reminding about meds. They may not be the patient, but they're the decision-maker.

I call this person the 'family health manager.' Their tasks fall into four buckets:

• Monitoring: tracking vitals, meds, and symptoms for multiple people.
• Interpreting: understanding lab reports, doctor notes, and treatment plans.
• Coordinating: scheduling visits, arranging referrals, and handling insurance.
• Supporting: providing emotional backup and daily health coaching.

From this angle, family health frequency is the sum of multiple members, multiple life stages, and ongoing management tasks. A kid's fever, a parent's blood pressure, a spouse's cholesterol, your own insomnia—none alone is high-frequency. But together, they create steady demand for one person.

That's why Afu added large-text mode, dialect support, voice input, and photo recognition. A blank chat box works for someone comfortable with prompts. But for a middle-aged person asking about a parent's condition, with limited medical vocabulary, the interface itself is the first barrier. The closer input is to everyday conversation, the more likely the product can serve people who'd otherwise never touch AI.

Still, not every 40–60-year-old user is managing a whole family. Some are mostly dealing with their own chronic issues. Younger users might focus on weight loss or sleep. And seniors might use the app directly, not through a proxy. So 'family health manager' is a hypothesis to test, not a label for everyone. To prove it, Afu would need to show data on multi-member accounts, reuse of family profiles, share of proxy questions, and differences in long-term engagement between single- and multi-member accounts.

There's also the privacy elephant. Creating a health file for a parent requires consent. What if a report gets filed under the wrong family member? How much health info can an adult child see? How do you revoke proxy access? China's Personal Information Protection Law treats medical data as sensitive, requiring specific purpose, strict necessity, and separate consent. A family can be a unit of value, but not automatically a unit of data authorization. The deeper the family integration, the more careful permission design has to be.

The real goal: getting to a safe next state

Health users don't just want knowledge. They want to know 'What should I do right now?' That could mean: keep observing at home, get more info, stop self-medicating, book a regular appointment, talk to a human doctor soon, or head to the ER. A good health AI shouldn't push everyone toward a transaction, nor should it give definitive diagnoses to seem useful. It needs to land people in a clear, safe next state.

Here's how Afu's ideal flow breaks down:

1. Lab report interpretation is the clearest first-value moment

Compared to open-ended symptom chats, report interpretation has a concrete input: the user has a document with results they don't understand. Afu can read it, flag abnormal values, prioritize them, and suggest next steps—recheck, see a doctor, or lifestyle changes. Report reading lowers three costs at once: medical jargon, anomaly screening, and decision-making. If the user confirms the interpretation and it's saved to their file, the product gets its first piece of long-term context.

But recognition accuracy isn't the same as medical accuracy. Reference ranges vary by lab, and a single value depends on age, sex, history, meds, and other results. An AI can read every character perfectly and still misinterpret because it lacks context.

2. Proactive follow-up questions help, but can create false completeness

The AI clinic asks a series of questions to gather details like location, duration, associated symptoms, and meds. That's better than expecting users to write a full prompt. It shifts the burden of questioning from user to product. The risk: as the progress bar fills, users may believe the system has everything needed for a diagnosis. Medical completeness isn't a checklist; it depends on the condition and risk. The product should clearly state what it's basing its advice on, what's missing, and the limits of its confidence.

3. 'See a doctor' isn't failure—vague referrals are

Users often dismiss an AI that says 'go to the hospital.' But caution isn't a flaw. The real issue is whether the AI explains why, when, which department, what red flags to watch for, and what to prepare beforehand. A generic 'seek medical attention' is just a disclaimer. A risk-stratified recommendation with trigger conditions, urgency levels, and next steps is a useful output.

4. Connecting a service isn't the same as completing it

Afu says it connects to thousands of hospitals, real doctors, and services like cloud accompaniment and insurance payments. But 'connecting' can mean anything from info lookup to full integration. For users, tapping 'book appointment' isn't task completion. Did they find the right specialist? Did the booking go through? Did the AI's summary reach the doctor? Did results flow back into the file? What happens if it fails? That's why a better metric than 'question completion rate' is 'safe next-state arrival rate': in tasks that require action, how many users completed an appropriate observation, record, human consult, booking, ER escalation, or follow-up within a target time—and for non-action scenarios, how many were correctly told to just watch and wait, avoiding unnecessary panic or visits. This metric doesn't incentivize transactions and counts 'no action needed' as a win.

Features must pass five checks

Afu has lots of features: AI clinic, report reading, skin photo analysis, specialist avatars, health files, mini-goals, reminders, device sync, appointments, consults, insurance, and accompaniment. Listing them in order makes for a superficial review. A feature needs to do more than fit into the flow. It must pass five tests: Does it reduce real cost? Does it improve decisions? Does it drive the next step? How does it recover from failure? And what new safety or privacy risks does it introduce?

Long-term relationships require accurate memory

Afu wants to personalize advice using health files and conversation memory. The CEO calls it 'knows you and is professional'—the more you use it, the more complete your file, the harder to switch. That's a sound product logic, but it's also the riskiest claim. Two detailed App Store reviews mention trouble retrieving past info and inconsistent labeling of the same metric. Two reviews aren't proof of systemic amnesia, but they highlight a key danger: a single session can tolerate missing context, but in long-term management, a wrong object, time, or value can amplify errors in every subsequent recommendation.

Medical memory shouldn't mean the model silently hoarding as much as possible. It needs at least five attributes: clear subject (self, father, mother, child), timestamp (current vs. historical), source (report, device, self-report, AI inference), error correction (users can fix misrecognitions), and control (users choose what history to use, can delete or export). When memory informs advice, 'view memory' becomes a core feature, not a privacy footnote.

The smart scale is an activation experiment

In June 2026, Afu launched a 'lose 100 million jin' campaign with cheap smart scales, daily check-ins, and AI coaching. A 21-day challenge followed, then Ant invested in Mint Health, a food database, to power photo-based meal logging. These moves aren't just about acquiring users. The scale forces a critical activation sequence: download app, pair device, get first body data, receive AI interpretation, set a goal, re-measure. That's closer to product value than a red packet—users at least complete a measurement and get an interpretation. Weight is also intuitive, changes relatively fast, and is easy to grasp, making it ideal for building a 'measure-interpret-act-remeasure' loop.

But a 21-day challenge isn't a habit. The real metrics are: first measurement completion after pairing, number of valid measurement days at 30/60/90, retention after incentives end, whether diet logging continues, and whether weight changes come from sensible behaviors, not water weight. The meal photo feature is just an estimate—Chinese cuisine is complex. Allowing users to correct names and portions isn't a weakness; it's a necessary error-correction mechanism for a serious product.

Commercial neutrality needs to be explained

In late 2025, Afu publicly emphasized that health answers contain no ads and aren't influenced by commercial factors. By June 2026, it launched an insurance AI agent and partnered with a health insurer on a product. These aren't necessarily contradictory—one is about content, the other about services. But it raises a governance question: when the same product understands your health anxiety and can also recommend doctors, drugs, insurance, and services, how do you keep editorial and commercial separate? Users need to know what's health advice, what's a product, why something is recommended, whether the platform earns money, and whether there's a free alternative. The more complete the health data, the more targeted the sales—and the more critical the trust firewall.

Stickiness isn't in the chat box

When people talk about Afu's stickiness, they often lump together different mechanisms: events drive recall, family aggregation drives demand, data creates continuity and switching costs, behaviors drive revisits, and services close the loop. These can reinforce each other, but they can also fail independently.

First, separate acquisition, activation, retention, and outcomes. Red packets and ads get downloads. Cheap scales push activation. Report reading or a good consult provides first value. Retention only happens when the next real health event brings the user back, or when they keep using files, devices, or services. A better metric set would include: activation rate (e.g., completed a report interpretation), repeat usage rate over 30/90 days, completion rate of specific health tasks (e.g., scheduled a follow-up), and outcome measures like improved lab values or reduced unnecessary ER visits. Different tasks need different observation windows. Weight management can be weekly or monthly; chronic disease management is monthly or quarterly; a physical exam might trigger a visit a year later. Using one daily active number for all health tasks just pushes the product toward meaningless reminders.

Second, a health product shouldn't aim to trap users in the app. For entertainment, longer sessions are good. For health, if a problem is solved and the user doesn't open the app for a while, that's success. If the app keeps pinging users with anxiety-inducing alerts, high engagement might mean the product is making things worse. So I'd define Afu's north star as: the number of family health tasks safely and correctly moved to a clear next state within a reasonable time window—including booking, consulting, logging, follow-up, or confidently waiting. This metric rewards execution without pushing transactions.

Third, the data flywheel can spin backward. The optimistic loop: more data → better advice → more action → more results → more trust. But there's a dark loop: a misread report gets filed under the wrong person, subsequent answers reference that error, the user notices inconsistencies, trust erodes, they stop sharing data or leave. Health data is only a moat if it's accurate, sourced, user-controllable, and improves task outcomes. Otherwise, it's just an increasingly unwieldy liability.

Five things we still can't see from outside

Public data shows two things clearly: people are willing to ask AI health questions, and Ant can reach a massive audience quickly. But scale isn't the finish line. We still lack proof on these five fronts:

• Reliability of long-term memory. No public data on file completeness, reuse rates, cross-session contradictions, family member misattributions, or user corrections. Privacy policies allow storing some inputs as structured memory, but can users easily review, edit, and turn it off?
• Clear handoff between AI and humans. Marketing uses phrases like 'AI health friend' and 'doctor AI avatars,' while fine print says AI isn't a substitute for professional diagnosis. These can coexist, but the interface must make clear who's answering: a general AI, an AI trained by doctors, a human-supervised AI, or a real online consult. Different levels carry different responsibilities. The more anthropomorphic the AI, the clearer the responsibility label must be.
• Real service completion. Hospital count and doctor count only show capacity. The real test is booking success rate, wait times, how well the AI's info transfers to the doctor, whether results flow back, and what happens if something fails. Regional differences in hospital systems and partnerships will cause big gaps in experience.
• Commercial offerings without eroding content trust. The insurance agent is just the start. If Afu adds drug sales, paid consults, or checkup packages, it needs auditable disclosure rules: are health recommendations separate from ads? Does platform revenue affect rankings? Can users see non-commercial options? In an interview, Ant's CMO said the platform technically could infer diseases from booking patterns and target ads, but that's a path health products should resist. 'Can we monetize?' isn't the hard question; 'How to monetize without burning trust?' is.
• Real health outcomes. We see questions asked, resources linked, activities joined. We don't see goal achievement rates, follow-up adherence, lab improvements, reduced unnecessary visits, or whether anxiety went up or down. This doesn't mean Afu isn't tracking these internally, nor should we demand a clinical trial from a consumer product at this stage. But the strength of conclusions must match the evidence. Right now, we can say Afu has built a path toward long-term family health management. We can't say that path has already produced stable health improvements.

The real test is whether next time is better

So why does Ant build a standalone app for something as low-frequency as medical care? Because it's betting not on a person getting sick every day, but on family health responsibility being a constant. The health events of parents, kids, spouses, and yourself are coordinated by one person. Reports, device data, and behavior accumulate. AI explains, triages, reminds, and connects to humans when needed. If that loop holds, low-frequency illness doesn't mean low-value product.

For now, the accurate conclusion is: Afu has identified a promising product unit and built a path from Q&A to family health management. It hasn't yet proven the long-term relationship with public data.

For product managers working on AI, Afu leaves three questions more useful than any feature list:

• After this task, will the next service be better because of the history?
• Does the AI's answer bring the user to a clear, safe, recoverable next state?
• When the model is unsure, a service fails, or risk escalates, does the product know when to stop and hand off responsibility?

If these aren't answered, a rich feature set is just a more complex chat box. Remembering accurately, being able to hand over, and knowing when not to continue—that's the real threshold for AI health tools to move from utility to lasting relationship.