Original · Unique Research · 2026-08-08
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening narrative, themed sections, afterword, and complete panel transcript. Customer-case figures and company metrics are speaker self-reports attributed to the named founders, not independently verified findings. This article describes product strategy and deployment logic; it is not medical advice and makes no clinical-efficacy claims. A sensitive real-person health anecdote in the source is rendered generically per Source Record editorial guidance. Company and person names are preserved as source attributions.
Original · Unique Research · 2026-08-08 20:00 Shanghai
AI Industry Observation
"Once you believe it, you explore in that direction. If you don't believe it, you can't even find the door."
Siemens interviews 50,000 people a year. Not "wants to" — "must": campus-recruiting candidates pour in like a flood; people alone couldn't handle it.
A nursing-home director is willing to pay high prices for a top-tier hospital doctor to come see patients; the doctor won't come, not even for the money.
A sovereign fund in Singapore has a 30-person legal team serving five to six markets; per-person load is exploding.
What these scenarios share: extremely high professional barriers, not enough people, and money spent without getting the job done. AI looks like the cure, right? Four founders — recruiting, legal, medical general practice, medical interaction — sat and talked, and I found something counterintuitive: what really stumps them was never technology.
The first cut is belief.
"Do you believe large models can handle your industry?"
Fang Xiaolei has done AI interviewing for nine years, from small models of a few hundred million parameters to today. He says the hardest thing is one translation: translating the physical world's hiring standards into something the AI world understands. Behind it is the industrial-psychology competency model, a pile of hidden rules, each dug out through scientific modeling. Many people get stuck here — a batch comes in, a batch leaves.
Max Ye does legal AI, contract review. He says the first mile has three things: data, scenario understanding, and expert oversight. Data must be structured — lawyers' track changes, highlights, comments on edited files are all treasures, organized into a playbook. Then scenario understanding, different contracts different rules. Then traceability — telling the lawyer why a change was made. Document engineering takes enormous time.
Sounds all right, all solid. But Deng Jiang punctured the paper screen directly.
"He said: 'The first hurdle is belief. Do you believe large models can do full-domain, full-modality general computation? Do you believe in three months it can?' This is sharp. The guests talked about data, scenarios, workflows; Deng Jiang says all that is the second gate. The first gate is whether the founder himself believes."
He used Musk as an example: five years ago everyone thought rocket recovery was fantasy, pure-vision self-driving even more absurd, only lidar could solve it. But Musk believed. Once you believe, you explore in that direction. If you don't, the technical path is wrong from the start.
Put simply, many companies start businesses based on "today's technical limitations." That's extremely rational in most eras, but not in the AI era. Silicon-based evolution is too fast; design a product around today's boundaries, and in three months the ground shifts.
So not enough data? You can fill it. Don't understand the scenario? You can learn. But if the founder doesn't believe, the thing is fundamentally impossible.
Sun Junwei's phrasing is more down-to-earth. Hengfang Health has done doctor-patient interaction for five years; his first mile is "product power." It's not enough that AI tech is impressive — it's deep coupling with the scenario. Top experts at Fudan Cancer Hospital in Shanghai, Peking University Cancer Hospital, Zhongshan, Ruijin get visited by large-model companies every day; they try it for two or three days then drop it. Why? Because AI gives a "useful" answer but not a "usable" experience.
In the end, in AI healthcare, technology isn't the barrier; scenario coupling is. Sun Junwei says it plainly: doctors have the ability to pay, but not the willingness, and still less the habit. Early Hengfang Health 1.0 charged doctors directly by driving the product into the folds of those three minutes in the clinic.
Taming Large Models: No Silver Bullet, Only "Cutting"
A general large model is like an intern with huge talent but a quirky temper — can chat about anything, but drops the ball in professional scenarios. How do you tame it?
Fang Xiaolei's path is the most pragmatic. When they started AI interviewing there were no large models; in the small-model era they trained everything themselves. Now with large models, their strategy is to use different ones in different places. Stages with low information-security requirements call APIs directly; core judgment stages use self-developed models. In autumn recruiting, hundreds of thousands of candidates pour in simultaneously; security first, stability first, cost first, effectiveness can be compromised a little.
He said one thing I remember: everything is constantly in a process from imbalance to balance. In plain language, no once-and-for-all solution; always dynamically tuning.
Max Ye's approach is more "cutting." In legal scenarios the model is uncontrollable, especially contract review: you tell it "review from the buyer's side," and it doesn't understand. So ThinkSpace AI cut one review scenario into 20-plus workflows, completely imperceptible to the user. Each stage tests different models' strengths; summarization uses light models, complex judgment uses heavy models. Then a cross-check layer: changed clauses must tell the lawyer why, which rule, which playbook.
"Cut into 20-plus workflows, and the model's hallucination is locked in a cage. But Deng Jiang's approach is the most aggressive. He said: 'Our company requires all employees to give up thinking.' Only set goals and make decisions; all the middle thinking goes to AI. Their company has no UI, no designers, no product managers, no programmers. The former programmers only give Cursor goals and confirm; they're not allowed to touch code."
This sounds arrogant, but his logic is self-consistent: if your product's goal is "make users give up thinking," you yourself must first learn to give up thinking. The future human role is decision, not thinking.
Sun Junwei's strategy is "hold the orthodox, strike the unexpected." Hold the orthodox means on the strategic side he doesn't obsess over self-developed vs. base model, but focuses on whether he's an excellent medical product manager who can find the "folds" in expert diagnosis. Strike the unexpected means on the tactical side, each doctor gets a private RAG, using the doctor's historical data to constrain model output. When the doctor is in the room, never hand the patient wholly to AI.
Who Pays? Cost Reduction Is the Admission Ticket; Efficiency Gain Is the Premium
No matter how strong the tech, you have to make money. The four guests' paying-logics fit together into one complete map.
Fang Xiaolei plays pure cost reduction. Siemens interviews 50,000 people a year; before it took piles of HR and business-leader time, travel, and organizing cost. Put an AI interviewer on, and campus-recruiting season cost drops directly 80%-90%. He says customers pay for a simple reason: compare it directly to management expense before AI, and the ledger is clear.
But he added a sharper line: using one human to judge another is itself flawed. The AI interviewer's real value is "one ruler measuring all the way" — a unified standard.
Max Ye plays efficiency. A Singapore sovereign fund's 30-person legal team serves five to six markets; even a two-person small fund covers multiple markets. Repetitive contract review eats huge time; AI does the pass first, flags key clauses, flags risks, gives edit suggestions, and hands back a "semi-finished product." Legal moves from executor to decision-maker. All revision data consolidates into enterprise know-how; new hires inherit it directly, no fear of people leaving.
Their product essentially helps the legal department build a self-iterating memory bank.
"Deng Jiang's play is the most counterintuitive. He says he's not selling an AI system; he's selling a 'digital employee.' For 10 million, I give you 1,000 excellent general-practice doctors, each costing only 10,000 a year — do you buy? The analogy is sharp. He prices not by software but by labor cost. Out-of-hospital scenarios have the strongest willingness to pay, because nursing homes and primary clinics want top-tier hospital doctors and can't hire them even paying. Then how cost-effective the digital doctor is is visible to the eye."
Sun Junwei's commercial understanding is the coolest. He says if you only position as cost reduction, you easily become a project-based vendor — "today he's the AI guy, comes in to run a project, becomes pure beast of burden." To be a "farming people," you must accumulate, must take the efficiency route, because efficiency gain has premium room. Hengfang Health's payers split two ways: doctors and enterprises. The doctor side solves the concrete folds of "three minutes in the clinic, desire exceeds capacity"; the enterprise side (pharma, insurance) solves the strategic question of "where are new patients, how do existing patients keep taking their medicine."
Wang Chaochao CC added: you can't simply summarize as cost reduction or revenue increase; the core is ROI. Investment is limited, revenue may be unlimited; the key is the return ratio.
Future: A-to-A Chats, Robots Watching You Take Medicine at Home
After the present, the future. Three to five years out, what do these professional scenarios become?
Fang Xiaolei's description is most like sci-fi, but already happening. He built an AI group-recruiting Agent that helps HR fully automatically chat with candidates on recruiting sites and ask for resumes. Not enough — outside products already build "candidate doubles": your recruiting Agent chats with the other side's candidate Agent, A-to-A. Once it's about right, move to WeChat or WhatsApp, where a human closes. Recruiting used to happen only in daytime; soon 24/7. Both sides' efficiency rises, and matching quality may even improve, because AI has no emotional swings and doesn't get tired and misjudge.
Max Ye used a self-driving analogy: legal document processing from L1 to L5. Basic documents can be fully automated; complex documents are semi-automated, give suggestions, humans make final decisions. The future extends from the legal department to whole corporate governance — enterprise document management, compliance, know-how consolidating, more and more automated.
"But Deng Jiang's imagination jumps furthest. He says every new technology brings changes in production relations. In 10 or 15 years, every household has a robot washing clothes and cooking. When this robot has a medical brain, it's your full-time family doctor. You develop a red rash on your back you haven't even noticed; it sees it and diagnoses on the spot. The medical order isn't given to you, it's given to the robot: 'why didn't you take medicine at noon? Remember tonight.' Medical supply moves from centered on the hospital to centered on the home."
Sun Junwei strongly agrees on "out-of-hospital" but adds a layer: the essence of future medicine is trust. No matter how correct today's AI advice, patients still don't believe it, because no one is liable. But as you get positive feedback again and again, trust slowly builds. He even envisions a scene: a sudden death tragedy is heartbreaking, but if from the start, every physical exam showing high LDL, an Agent would remind you or even "warn" you. Insurers will change too: a 40-year-old, one eats meat every day and doesn't exercise, another stays active, same premium doesn't make sense. The future must be patient-specific, thousand people thousand faces.
Aftertaste
Talking this through, one image stays in my head. Maybe ten years later, the robot at your home reminds you you missed today's medicine. Not because you forgot, but because the medical order went to it, not you.
By then, the center of medical computation may no longer be the hospital. Just as battery makers became the core of the auto chain because of EVs, AI healthcare, through reconstructing production relations, is turning "home" into the new health hub.
Four founders, four completely different deployment paths. But the underlying logic is strikingly consistent: AI isn't replacing people; it's solving "there simply aren't enough people." And cost reduction, efficiency gain, infinite supply are just different facets of this production-relation reconstruction.
"In the end, it's still that line: do you believe it? Once you believe it, you explore in that direction. If you don't believe it, you can't even find the door."
More Conversation Details
Speakers
Jinyu Intelligence CEO — Fang Xiaolei (方小雷)
ThinkSpace AI founder & CEO — Max Ye
Yuanqi Wisdom founder & CEO — Deng Jiang (邓江)
Hengfang Health CMO — Sun Junwei (孙君韡)
Host
Unique Capital partner — Wang Chaochao CC (王朝超)
Wang Chaochao CC: Today's keyword is professional scenarios. We have four guests from different fields — legal, medical, recruiting, enterprise process; four scenarios, four different deployment paths. Please introduce yourselves and your companies.
Fang Xiaolei: Hello everyone, I'm Fang Xiaolei, CEO of Jinyu Intelligence. We started as an AI-interviewing company, from 2017-2018, nine years now. Now we're more of an AI Talent System company — AI talent development services, focused on AI engineers and full-stack AI engineers; plus providing enterprises with AI interviewers, AI group-recruiting Agents, and AI learning-interaction platforms, talent-deployment system services.
Max Ye: Hello everyone, I'm Max, CEO and founder of ThinkSpace AI. We're a legal AI company, mainly building an Agent OS for legal work — Agents in legal scenarios, mainly serving in-house legal teams. Currently based in Singapore, glad to meet everyone.
Deng Jiang: Hello everyone, I'm Deng Jiang, CEO of Yuanqi Wisdom. We're a Beijing AI healthcare company doing general-practice, full-modality general computation — positioned to use large models and Agent technology to achieve general computation across all medical scenarios and modalities. We already have many deployments with leading medical institutions, including primary-care institutions. I believe in the future, most people's many medical services will be delivered with AI backing. Thank you.
Sun Junwei: Hello everyone, good afternoon, I'm Sun Junwei, CMO of Hengfang Health. Coincidentally, Hengfang was also founded in 2020, dedicated to using AI to solve doctor-patient interaction. We're an AI healthcare company; our customers are pharma or insurance companies, and our users are doctors and patients. The problem we solve is the clinic's limited three minutes — giving patients a better experience, doctors a better experience, and the enterprise (the pharma payer) a better experience. That's Hengfang's business scenario.
Where Does the First Cut of Agent Deployment Land?
Wang Chaochao CC: Everyone says general AI capability is improving fast, but professional scenarios have their own barriers — law relies on statutes and evidence, medicine on evidence-based practice, recruiting on matching. In your professional fields, what's the first hard-to-cross obstacle for Agent deployment? Maybe data, scenario understanding, user trust, or something else?
Fang Xiaolei: Our earliest AI product was the AI interviewer, now very popular; many leading enterprises across industries use it. Looking back, the hardest part was a translation — translating physical-world standards into AI-world standards. For recruiting, behind it is IO psychology, the competency model, now competency-plus-skill model. Through a whole scientific modeling method, translate the enterprise's hiring requirements, many hidden rules, into something AI understands. We did this for years. Further, obtain data, label corresponding data, run back-to-back human-AI comparison experiments with clients, validating that AI indeed understands the criteria. This is hard. Many companies entered the industry and many left, basically stuck at this point. That's our industry.
Max Ye: Our first mile had three places. First is data, the foundation; second is scenario understanding, the core; third is expert oversight of results. Starting with data, we mainly do non-litigation contracts; legal contract scenarios have many annotations, mainly three metadata: track changes, highlights, comments. Lawyers make many adjustments when editing files; this data is often precious. We need to structure it, help lawyers build a playbook reusable on future repetitive contracts — first. Second, data must be traceable. Lawyers reviewing contracts need to know why they reviewed this way, why written this way; the writing must be trained to the lawyer's own preferences. We spent a lot of time here — document engineering. We have lots of technology in parsing, mapping, and understanding document tree structures. Then scenario understanding; different scenarios need different rules. Finally, traceability back to the lawyer and oversight of results. Roughly like that.
Deng Jiang: The first hurdle — I think the first hurdle is belief. As a large-model startup, do you believe large models can solve problems? In medicine, we see many companies doing so-called vertical polishing; do you believe current AI can achieve full-scenario autonomous driving, full-scenario programming, full-scenario video generation? Do you believe it can do full-scenario general medical computation — unified general-practice full-modality computation, not specialist computation? Belief is important. Second, the silicon-based evolution speed large models represent is very fast; when it can't do it today, do you believe it can in three months? This is actually the first hurdle AI entrepreneurship must cross. We see many companies starting businesses based on current tech and limitations; that's extremely buy-in in most eras, because it evolves too fast. So I think the first hurdle is this.
Wang Chaochao CC: Let me press: is the first hurdle the founder, the founding team, or the customer?
Deng Jiang: Actually the founder and founding team's belief. Like Musk believing rockets are recoverable — that's a belief. Five years ago there were many technical gaps, including pure-image self-driving; in that era everyone thought only lidar could solve it, but he believed pure images would solve it as tech advanced. So this choice is far more important than data-layer problems. From another angle, the supply side must believe first before moving the demand side. Or only if you believe will you explore in that direction and possibly find it. If you don't believe, technical exploration goes wrong from the start.
Sun Junwei: Let me share my shallow understanding. The core of the first mile is that a tech company doesn't live in grand narratives but in microscopic folds. I believe AI-scenario monetization is a common problem, so we greatly value user experience. Among doctors, patients, and enterprises, we focus on doctors, because doctors are the starting point of today's Chinese healthcare and big-health system. This group has the ability to pay but not the willingness, still less the habit. So early Hengfang 1.0 charged doctors. The first mile is always product power. How to define product power? Healthcare is a scenario every large-model company heavily bets on. From a useful product to a usable product, besides impressive AI tech, something easily overlooked is deep coupling with the scenario. Many experts — Fudan Cancer, Peking Cancer, Zhongshan Ruijin — get visited by too many large-model companies pitching AI healthcare; they try it for two or three days then stop. Hengfang pursues deep scenario coupling plus a bit of AI tech to make doctors pay. This is what Hengfang has done, and I think a relatively important point in AI-plus-medical scenarios.
How to Tame General Large Models into Professional Agents?
Wang Chaochao CC: Everyone knows large models are general engines, but professional scenarios need professional tools or Agents, not just prompt engineering. How do you tame the general model's capability into an Agent? Technically how?
Fang Xiaolei: When we started AI interviewing there were no general large models; it was the small-model era, all trained ourselves. The earliest models might have only a few hundred million parameters, low cost, few cards needed, before the export restrictions. Starting with our own models achieved good results, but unlike large models it couldn't follow up or do deep analysis. It could judge well — whether a candidate's answer was good, average, or poor — reaching good human-consistency standards. With large models, everything started very green. First-generation large models still had limited parameters and intelligence, so SFT wasn't enough; incremental pre-training was needed in '23, '24. Now it's changed a lot; the product is complex, not an early product, complexity high, many places use different models. Some models, for cost or information-security reasons not important, may use other companies' APIs; but core key judgment points may still use self-developed models. So it's not a choice for us; use the right one in each place. More consideration is cost. Take the AI interviewer — there are very busy seasons, every campus-recruiting autumn maybe hundreds of thousands or even hundreds of thousands of candidates a day. Then security first, stability first, cost first, maybe even compromise effectiveness a bit. Different stages have different KPIs and priorities. Everything is constantly in imbalance-to-balance.
Max Ye: In our view the model itself is very uncontrollable, especially in legal scenarios. I started in May '23, half a year after GPT came out; at 3.5 and 4 we did many different trials and pre-training. But for contract review, it can't identify stance. Even if we tell it to review as buyer, seller, or some stance, it can't precisely judge. But after O3 came out, we found a big shift — much of the training or control we did before wasn't needed; the model itself already understands certain stances. Now each model's strengths differ. Some models don't need reasoning for simple tasks, like summarization; we test different models' strengths in the first stage. Very extensive tests on what each model suits. Also called multimodel; in the product we design different models in different workflows. Beyond the model end, after testing we cut workflows very finely. In one review scenario we actually cut 20-plus workflows, imperceptible to users. Only cut finely enough to minimize model hallucination — second means. Third, beyond workflow cutting, model-strength testing and selection, finally cross-check, cross-reference. When it comes out, it needs strong traceability — tell the user why a clause changed this way, which rule or playbook. Based on these three, the final edit right goes to the lawyer — accept or not, maximum user choice. They can rely on their own judgment again whether to execute. Of course they can rely directly on AI, but we keep this right with them. That's our work.
Deng Jiang: From the source, our company wants to build an AI Native company to make AI Native products. So we have a basic requirement: all employees give up thinking. Giving up thinking means you only set goals and make decisions; all middle thinking goes to AI. For example, we want to build a general-practice full-modality Agent — that's the goal. Where do skills come from? All options should come from the best model — give 5, 6, 7 options, I decide. So what skills, tools, workflows, technical architecture it needs — if your goal is a product that makes users give up thinking, you yourself must first learn to give up thinking. This is the role change we think future AI Native, or AI deeply fused into work and life, requires — we decide, not so-called thinking. For example, an app — most people tell AI "I want a blue-style app, generate one." We don't; we say we want an app, for which users, you produce. Let me give a medical example. We always align goals with clients, not requirements. He says my goal is that all critical patients get a warning within one second — that's his goal. How to achieve it? The whole solution goes to AI — make an app, a system, an Agent, a robot, or what? To achieve the goal, which 12345 things to do? The final demo? It may produce 10 or even 20 demos; I only decide. Goals from the client, tasks aggregated by me, solutions all from AI, I only decide. Like today's tattoo-image generation — actually let it generate 10, 100, 1000 images, you pick one. You tell it image details, the goal — like "I want a headshot for a visa" — it generates a pile with different hairstyles and backgrounds, you pick. For AI in all scenarios, especially complex knowledge computation, cooperating with humans, humans set goals and decide. PPT is the same.
Wang Chaochao CC: Hearing this, Yuanqi's degree of AI-ification is quite high.
Deng Jiang: Our company has no traditional white-collar roles — no UI, no designer, no product manager, no programmer.
Wang Chaochao CC: Then what do the leaders do?
Deng Jiang: We have former programmers, but they only decide, aren't allowed to write code. Give Cursor a goal, let Cursor write code, finally just confirm. They aren't allowed to modify or touch code.
Sun Junwei: My answer to this is hold the orthodox, strike the unexpected. Hold the orthodox comes from strategy; strike the unexpected from tactics. Mr. Deng's scenarios are very vertical and deep. Hengfang Health is relatively high fault-tolerance, doing doctor-patient interaction, relatively shallow, high fault-tolerance. Responding to my earlier point, we strive to make a useful product not stay on theory or a launch event but actually be used by every expert, even paid for. So from useful to usable, beyond tech, more important is deep scenario coupling ability. The so-called hold-the-orthodox: we don't care that much whether self-developed or base model; what matters is whether you're an excellent medical product manager, whether you know where today's expert diagnostic folds are, and know today's model iteration boundary — this iteration serves the useful-to-usable path above. The so-called strike-the-unexpected: everyone knows today's large-model stage necessarily has probabilistic inference — "the weather is great" it'll probably infer "hot" or "bad." But in medical interaction, fault tolerance is low. So we definitely follow the "when doctor, when model" principle; each doctor has a private RAG to constrain their model as much as possible. That's the hold-the-orthodox and strike-the-unexpected I mentioned.
Wang Chaochao CC: How far have you taken doctor-patient communication?
Sun Junwei: It's subjective. My definition lies with the patient, because everything comes back to the doctor's KPIs. At least patients' gain is strong; they spontaneously write thank-you notes, take medicine on time, even high activity. The precise standard of doctor-patient interaction is patient activity; patients' perceived gain isn't lip service but their behavior. In serving thousands of doctors and nearly a million patients nationwide, backend activity is basically 60%-70%.
Who Is the Customer? Why Pay?
Wang Chaochao CC: There's an old saying: no matter how strong the tech, you still have to make money; no money, nothing to talk. So we all face a problem — who's the customer? Why pay? What's their core paying reason? Cost reduction, revenue increase, doing what was impossible, or something else?
Fang Xiaolei: OK, the core HR problem is still cost reduction. Many complex factors, but the original driving force is cost reduction. For example, our earliest scenario was campus recruitment. Client Siemens has at least 200,000 candidates a year, interviews 50,000. Previously interviewing 50,000 by humans was impossible — no time, very high cost. Now we can interview 50,000 every year, because an AI interviewer session is cheap. The vast majority of clients prioritize efficiency and accuracy, but first you must be cheap enough. Previously so many people traveling to many places for campus recruiting. Of course now there's Tencent Meeting, but it consumes so much business-leader time and HR time organizing and interviewing each person — that's itself a time and cost to the business. AI can knock out 80%-90% of campus-recruiting cost. So the driving force is cost reduction. But only Fortune 500, tier-one giants, relatively leading companies have such budget or cost outlay.
Wang Chaochao CC: Super-large and large enterprises yes, but actually any company growing from 100 to 1,000-plus — recruiting isn't hiring 500 from 100 and ending at 600. Often there are large volumes of repetitive, high-turnover roles; these projects also need AI interviewing. Extending to companies above 100 people, in China nearly 4 million companies, so the market is large enough.
Fang Xiaolei: Understood. The customer's paying reason is direct comparison to the whole management and recruiting-management expense before.
Wang Chaochao CC: Yes, everyone thinks that at first, then finds a more serious problem — using one human to judge another.
Fang Xiaolei: Right, so an AI interviewer can deeply solve many problems, unify or relatively standardize standards. We call it "one ruler measuring all the way."
Max Ye: On willingness to pay, I think mainly two points. First, efficiency gain. For a legal team, usually not huge — like a Singapore sovereign fund, the whole legal team is only 30 to 40 people, but a very large fund. Many clients, like smaller funds, two people serving five to six markets, heavy workload. Our first product, Syndoc, core ability is to replace repetitive work with AI, based on structuring their past data — first, data consolidation, enterprise know-how and knowledge; second, when someone leaves or trains a junior, they can get up fast, avoiding stance loss or re-teaching time. Second, the second product does whole relationship management. In past scenarios, I need an NDA, send it from business to legal; it may sit in many files for a while then come back. We want to rebuild the whole flow — on receiving email, the Agent does first-stage analysis: what contract, key clauses, what to note, how to change. When it comes back to legal, it's already been edited once, not just received. Everything flagged, prompted, edited; what they do is decide, saving the repetitive time. After the flow changes, re-consolidate all know-how, forming a closed loop. Enterprise structure and work-style change is also important, moving from today's efficiency to future reduction of repetitive work. So we target the legal department inside enterprises.
Wang Chaochao CC: In other words, compete with law firms.
Max Ye: Law firms can use our product too, but first ICP is in-house legal.
Wang Chaochao CC: OK, Mr. Deng, why would your customers pay?
Deng Jiang: Just last week I talked with a founder of a fairly large medical listed company. I said, for 10 million I sell you an AI system, would you buy? He said I'd evaluate whether the 10 million is worth it. I said this: I supply you 1,000 excellent general-practice doctors, each costing only 10,000 a year, would you? He said very willing. So OK, our cooperation — treat me as a labor-outsourcing company; I supply excellent digital employeesmastering general-practice medical knowledge, and I can supply infinitely, as many as you want. Because inside such medical companies there are many roles doing medical knowledge computation — hiring medically-backed people for clinical data analysis, case organization, diagnosis and medication. We must see what's actually being done now — knowledge computation: legal knowledge computation, medical knowledge computation. What we really replace is the person in that original role. But in mental work, especially complex mental work, supply is far short; it's not replacement, not that enterprise people are enough, but insufficient — hence high salaries to poach. We want to build a new supply relationship, infinitely supplying high-quality medical-knowledge computation. What's actually landing now isn't inside the hospital but outside. Many out-of-hospital clients are building cooperation with us, because outside, to get quality medical service you must hire doctors at high prices and still can't. A nursing home wants a top-tier hospital doctor to visit once; he won't come, not even for money. So in many under-supplied scenarios we have bigger opportunity; willingness to pay is strong, because anyway it's cheaper than hiring people.
Wang Chaochao CC: So you're playing the revenue-increase,revenue generation logic? Grow GMV, grow revenue?
Deng Jiang: The previous two, Mr. Fang and Max, played cost reduction, management-expense reduction; this side is revenue increase.
Wang Chaochao CC: Belief is priceless. You believe, then I book it for you — how? I can use unlimited digital employees to create for you.
Deng Jiang: Theoretically, and I don't even need belief, because many scenarios are easy to test — one test, one calculation and it's clear.
Wang Chaochao CC: That's the revenue logic. Mr. Sun, your view.
Sun Junwei: I think commercialization is very important for AI Agent companies. In my shallow understanding, when seeking customers, we should hold the cost-reduction floor to break through the efficiency-gain ceiling. If positioned as cost reduction, you easily have no premium room and become project-based — today he's the AI guy, comes in to run a project, becomes pure vendor, pure beast of burden; lots of people, resources, and energy dissipate. I've always called this nomadic people, fighting wars to stop wars. To be a farming people, to accumulate, it must be efficiency gain, because efficiency has more premium room. From Hengfang's scenario, payers are partly doctors, partly enterprises. Back to deep scenario coupling: doctors today have limited three minutes in clinic, want to serve patients well but desire exceeds capacity; Hengfang cuts into these concrete sub-scenarios where capacity falls short. For enterprises, I came from a pharma company; the enterprise first-principle is always: today a new innovative drug comes out, where are new patients? How do existing patients keep taking medicine? How to find evidence-based medicine and monitor adverse reactions? To meet strategic goals, traditionally the enterprise hires lots of people — medical reps, regional managers, familiar to all. The May 1 national anti-corruption campaign actually helped Hengfang a lot; because it's labor-intensive, the enterprise logic is "I don't know if you'll succeed, but I really have no more road; willing to try with you," which put me on the efficiency path. This is a path Hengfang Health began exploring late last year or this year, and my shallow understanding of Agent commercialization.
Wang Chaochao CC: After Mr. Sun's sharing, I think a correction is needed — the cost-reduction and efficiency gain earlier; Mr. Deng and Mr. Sun can't be summarized by revenue alone; it should be ROI. For device or client side, investment is limited, revenue may break through infinitely, so the core is ROI; relatively accurate.
What Will the Industry Look Like in Three to Five Years?
Wang Chaochao CC: Today's roundtable keyword is deployment, but deployment isn't the end; it's a new start. Next, in three to five years, what paradigm change might Agents bring to the industries you serve? Deeper human-machine collaboration, transfer of decision rights, or reconstruction of the whole service model?
Fang Xiaolei: I've also been thinking about something recently. We have a product called AI group-recruiting Agent that helps recruiting HR fully automatically chat with candidates on many recruiting sites and ask for resumes. Before, this consumed over half a recruiting colleague's time — browsing recruiting sites, tedious. At first the eyes could handle it; after a while it got tiring.
Wang Chaochao CC: Before going on stage, did you use eye drops because of this?
Fang Xiaolei: I long ago let AI watch; I don't watch myself. For recruiting colleagues this product is a huge blessing, replacing simple work. But candidates, in the next three to five years, may only be chatting with AI on recruiting sites, key is they don't know it's AI. Meanwhile outside some products do the reverse — help candidates apply, the candidate's double. Candidate job-seeking is our HR recruiting double; many outside companies made candidate doubles; finally your recruiting Agent chats with someone else's candidate Agent, A-to-A. After chatting, maybe move the session to WeChat or WhatsApp, then a human interacts. For both sides efficiency rises; previously this business only happened in daytime, soon maybe 24/7, completing good matching. Future many tasks will first get an AI double pass, doing initial screening and matching, then humans step in. In process terms, the recruiting Agent goes first — this should happen first.
Wang Chaochao CC: Unknown, because many uncertainties.
Fang Xiaolei: Actually the deeper we go, the more we find product difficulty rises geometrically. Serving many clients, clients have many deep requirements; we find the large model's boundary is also very limited, not that it can do anything. We don't do human-AI collaboration — AI does first then human takes over; no, we do 100% AI then humans do. So it gets harder. Future maybe only certain roles, certain level, certain category will do better; it should be a gradual unfolding. Demand is strong, model growth is rapid.
Wang Chaochao CC: The hard problem is with you.
Fang Xiaolei: We look forward to models getting stronger, meeting halfway.
Max Ye: The model itself, as Mr. Fang said, will definitely get stronger. Through model progress, how will the industry change? First, many basic and repetitive tasks will definitely be replaced by AI. Second, in judgment, from L1 to L5, like a car — from file understanding, to generating content, to making some judgments, to fairly automated direct execution; the future will definitely go more automated.
Wang Chaochao CC: What state do you think we're in now in legal work? What L can the product reach?
Max Ye: We think for basic, relatively junior documents, fully automatable. But for complex documents, not yet, because complex documents need many roles. We don't want AI to decide everything; legal itself wouldn't want that, because if what they look at still needs another pass, not much time is saved. So in complex documents, we're semi-automatic — through past edit styles and suggestions, provide edit suggestions, you judge feasibility. On this basis, the whole workflow from legal work to corporate governance — enterprise document management, compliance — much know-how and tasks are similar. Under this framework, enterprises get more and more automated on the legal side.
Deng Jiang: I'm thinking about this. Actually every new technology brings changes in production relations; this is something every enterprise with a long view must watch. For example, new energy vehicles — battery makers weren't in the auto chain before, but because of NEVs, battery makers became an important part. Medicine's old supply production relations relied on people — medical schools training doctors into hospitals, with the hospital as core expanding many things — data collection in-hospital, disease screening, diagnosis, medication management in the hospital, because the people who can provide service are in the hospital. Now out-of-hospital detection is richer, smart bands and watches as collection ends are richer, but computation is still in the hospital. No matter how many bands sold outside, finding an abnormality means registering, going to the hospital — that's the traditional production relation. The cost reduction and efficiency gain mentioned is actually a business doable under current production relations. But if you build a company hoping for the next 5, 10, 20 years, you must actively embrace new production relations. What's a new production relation? Let me imagine: 10 or 15 years, every household has a robot washing and cooking — do you believe it? When this robot has a medical brain, it's your full-time family doctor. A red rash on your back you haven't noticed; it sees it, diagnoses on the spot, says there's a rash, suggests ointment. Ten years is long; I'm being very conservative. Future medical supply, from data collection to computation, may center on the home, not the hospital. Health management will be more at home. In hospital you get a disease, the doctor orders medicine morning, noon, night; in future this order goes not to you but to the home robot, which watches you every day — why no noon medicine, remember tonight. Production-relation change is what excites entrepreneurs most — a chance to find an important position in new production relations. This is our thinking; we're very certain it will happen. Whether Yuanqi is in it then takes a lot of effort.
Sun Junwei: On this I strongly agree with Mr. Deng — the future scenario is out-of-hospital.
Wang Chaochao CC: Mr. Deng thinks out-of-hospital is more the home; where do you think?
Sun Junwei: It's a philosophical question, and a real one. I think it comes from within — as the general secretary says, everyone is their own first health responsible person. Why do I strongly agree with Mr. Deng's out-of-hospital? Because Hengfang has always explored the future. In doctor-patient interaction, we fiercely avoid becoming a hunting company, nomadic people; we want to be farming people. So we've consolidated lots of out-of-hospital data in doctor-patient interaction, including behavioral data. We're developing a Hengfang patient-journey model that knows your present and past, can predict your future, stands in a god's-eye view, intervening from dimensions like health management. The carrier may not be an app or WeChat group, maybe phone, band, Google Glass — that's another story. In this scenario, we empower everyone. For example, a sudden-death tragedy is heartbreaking; how did that person reach sudden death? It must be from the start — every physical exam shows high LDL; many people comfort themselves "it's fine, eat less meat, more vegetables." But in fact genes decide that before vegetables and meat, you'd unhesitatingly pick meat. Then an Agent would come remind you, even warn you. We serve many tumor patients in pharma; big guys say they don't smoke but often hope to trade health for immediate pleasure. Many young female breast-cancer patients, including people my parents' age, are anxious, over-anxious. Future Agents face patients who are thousand-person-thousand-face. Let me summarize: the essence of medicine is trust. It's not there yet because it's still bound to the doctor; in China, seeking care, as Mr. Deng said, we won't believe AI's pile of correct platitudes; best to go to hospital, because the doctor is liable. That's the current stage. But the patient journey — when patients get positive feedback again and again, as Mr. Deng said, patients will slowly trust out-of-hospital Agent help, like many now starting online medical services. The core factor constraining in-hospital to out-of-hospital is that the diagnostic bottom logic is screening, diagnosis, treatment, management; screening and diagnosis must go to hospital. The doctor orders a test, but no one gives it to you; at home no one comes to draw blood. After screening, I believe today using AI to upload many reports, answers won't differ much. But the hard part is today it's useless; AI inaccuracy is because many neglect one point — did you give enough evidence? Abstractly, AI is already a top-tier Zhuge Liang; many people's differing understanding of AI ignores one point: are you Liu Bei yourself, can you three-times invite Zhuge out? Many people lack the virtue to station an army, so they can't keep this thing. Hengfang wants to level this; we find many patients, in interaction, their cognitive and language expression constrain the final result. That's what Hengfang wants to flatten. Further imagining, when we have an out-of-hospital system for patient interaction, beyond god's-eye health management, when a patient finds a problem, innovative drugs and devices will timely notice you. My old employer is Novartis; one drug comes out and everyone, thousand-face, uses one drug; future must be patient-specific, and I believe China's supply chain can do it. Including future insurers — now insurers are all "40 years old" — like me and CC, 40? I'm not yet 40; I am 40, don't look it. For example, I might eat meat every day and not exercise; you might stay active, older than me, but same premium — doesn't make sense. In this system, when knowing enough of the patient's present and future data, you can predict present, past, future. Insurers must be patient-specific. Then patients do health archiving by choice, even intervene in the future. Today medicine has healthy and sick populations, with many sub-healthy people in between ignored. Once healthy people use Agents, there's a new economy — the longevity economy, a new picture. Of course, what I want to say is, Hengfang Health today has already explored the patient-journey model and gained recognition from many patients and doctors.