---
title: "From Zero to $40M ARR in Six Months: Four Agent Founders on How to Actually Make Money"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-08-11T10:31:17+00:00"
canonical: "https://ffcap.cn/en/research/src-20260811-01html"
source: "https://uniqueresearch.substack.com/p/src-20260811-01html"
language: "en"
---

# From Zero to $40M ARR in Six Months: Four Agent Founders on How to Actually Make Money

[![](https://substackcdn.com/image/fetch/$s_!OOMw!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae7361a4-93bd-4c68-9b7b-dbcd1f9e1e2c_2048x1152.png)](https://substackcdn.com/image/fetch/$s_!OOMw!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae7361a4-93bd-4c68-9b7b-dbcd1f9e1e2c_2048x1152.png)

_Original · Unique Research · 2026-08-11_

_Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, four insight sections, closing reflection, and the full roundtable transcript. All named companies, products, and people are preserved. ARR figures, customer counts, and revenue claims are speaker statements, not independently verified findings._

AI Industry Observer

"Money comes first, then the Agent, then the technology."

Six months, from zero to $40M ARR.

The person saying this is Michael Guan, CEO of Hellyeah AI, which makes digital employees. Its first landing scenario was growth marketing. Note: he's not reporting valuation or vision — he's reporting annualized revenue.

At the roundtable, the other three also came with track records. 94AI, doing voice outbound calls, has several hundred million RMB in domestic plus overseas revenue across 30+ countries. QuickCEP, doing brand going-global, has 100M RMB ARR and serves 70 of China's top 100 global brands. upsello.ai, doing Shopify smart shopping assistants, brings in $1M a year, with five large sellers doing over $100M annual GMV among its clients.

In plain terms: this roundtable doesn't discuss model parameters or AGI faith. It discusses one thing: did your Agent actually help the customer make money, and why should the customer cut you in?

Host Wu Wei's questions got more direct with each one. I've extracted the most valuable segments for you.

Why These Four Areas Specifically

Start with the topic selection logic — these four people chose their tracks based on the same calculation.

Michael's reasoning was the cleanest: growth marketing is "where the reinforcement learning signal is strongest." A digital employee actually runs budget — money made and money lost are crystal clear, and that's training data. His company itself does enterprise internal knowledge distillation and digital employee deployment; marketing growth was just the first vertical, and it hit the bullseye.

94AI's Liu Siping (刘嗣平) did it in reverse. They started doing voice models in 2018 — speech recognition, synthesis, interaction, the whole stack — and then discovered a harsh reality: "After building voice models, many application scenarios are actually very hard to charge for." They finally landed on the phone scenario because financial clients' calls are an essential demand — callbacks, marketing, collections, they must be made.

upsello.ai's Gao Zhou (高舟) spent nearly a decade in quant. This year, clients came directly to him: "You do AI, you talk about data-driven — can your chat actually make me money?" So they built an offensive shopping-assistant live chat plugin on the Shopify ecosystem. When Wu Wei pressed him on revenue, Gao Zhou said about $1M a year. Wu Wei was startled: "$1M for a custom Agent?" Gao Zhou quickly clarified: "No — individual client is 50-60K RMB, and subscription is a few dozen dollars."

QuickCEP's Chen Guang (陈光) was the most ruthless — serving only the brand going-global industry. The reasoning was practical: dozens of e-commerce platforms and social media channels overseas each fight their own war, with all data as islands — brand owners are in agony integrating them. There are no shortcuts; over the years they integrated one by one. Some platforms are so closed you must become an officially certified partner to get API access, and after integrating, you must continuously follow policy changes to keep the Agent's common sense aligned with the platform.

"Do you see it? None of the four said 'because this track is sexy.' Every choice was forced by clients, by data, by money."

The Attribution Hurdle: How Do You Prove the Money You Made Was Yours?

The first death point of revenue-generating Agents is attribution. The client wants the final result; you only did one part of the chain. Why should credit for the growth go to you?

Gao Zhou's solution is more quantitative. He works backward: first figure out how to make money for the client, then reverse-engineer what to build. His attribution logic is clean: a user was going to buy A; the Agent pushed B through upsell or bundling — only B's GMV counts, and A gets zero. All touchpoints happen in live chat; whatever closes in live chat is his credit. This works thanks to Shopify's open ecosystem data.

Michael's solution is simple and brutal: the digital employee goes directly into the client's code repository, and conveniently retrofit a data telemetry system to ensure attribution fairness and transparency. The cost: many clients can't stand this intrusive approach, feeling like you planted something in their source code. Michael's attitude is also hardline: "Understand that a digital employee should be sitting in their office — first walk through their door." They turned down many clients who couldn't accept this. Not every client must be onboarded, he says; clients who can't accept such a forward-looking concept get decisively rejected.

"If you can't pass the attribution hurdle, every subsequent discussion about revenue sharing is a castle in the air. One relies on ecosystem data; the other goes directly into the code repository. Fundamentally, both answer the same question: do you dare put your contribution in the sunlight for the client's finance department to audit?"

Pay for Performance? First Ask If You Deserve It

Pay-for-performance sounds like the politically correct position in the Agent industry, but the veterans' attitudes were surprisingly cool.

Liu Siping put it plainly. Pay-for-performance is of course better — it means clients recognize your core capability. But there are two prerequisites: the client scenario must be suitable, and user quality must be stable. Top-tier banks and large platforms have large user volumes and stable quality, so pay-for-results works. Emerging enterprises acquire through different channels today and tomorrow — "when you set the pay-for-results standard, they might suddenly be very profitable or suddenly lose money," and the deal can't be signed. As a fallback, charge by call volume.

He also exposed something many people don't want to say: charging by call volume distorts you. Pay-for-performance makes you iterate the model for effect-oriented outcomes; pay-by-call-volume makes you care more about increasing consumption and call volume, and effectiveness isn't the top priority. "Clients have a ruler in their hearts" — when user quality is stable, that ruler is stable.

Chen Guang added three more practical issues, all landmines from real operations.

First, mid-to-large enterprises' procurement processes and budget systems simply don't support it — clients themselves can't say how much they'll spend this year. Second, what about sales incentives? Originally the subscription fee is collected once a year, the salesperson gets commission and leaves; now they sign a contract but the client doesn't pay — the salesperson only gets paid after the bill is calculated by results, and sales aren't happy either. Third and most essential: in the client's cognition, only two roles can take a cut — those who control traffic, or those who control transactions. Meta, Google, TikTok control traffic and can take a commission; acquiring platforms and store-building platforms control transactions and can take a GMV cut. Middle links? Clients used to think you're just an auxiliary tool, and giving some token money for subscription is already generous.

"So why can QuickCEP now start taking a revenue cut? Chen Guang gave an example from the Japanese market. In some countries, the client's own team simply can't hold the operation, so QuickCEP uses the entire Agent team to take over consumer operations for that country: EDM, WhatsApp mass messaging, lapsed member reactivation, loyalty operations — the client doesn't need to assign a single person. Complete takeover — he doesn't manage it; it was originally a piece nobody was doing, and you did an extra piece for him."

This sentence is worth underlining three times. To upgrade from charging tool fees to taking a revenue cut, there's a wall between you and the upgrade, and the wall is called: did you take over the entire function? Assist a department and you're a cost; replace a department and you qualify to sit at the revenue-sharing table.

By the way, Chen Guang also rode a wave earlier this year. When OpenClaw blew up, they released PollyReach — a skill for Agents that provides US virtual phone numbers. The origin was very relatable: he lives in Japan, eating yakitori requires reservations, his English isn't good for booking haircuts, and he kept troubling Japanese colleagues to make calls. He thought: can't the Agent do this for me? The Agent can write PPTs and research reports, but can't interact with the physical world. Just install a calling skill on it. Now this skill ranks #1 on Claw Hub with over 100,000 installs. Laying a phone line from the Agent to the atomic world — I think that's more imaginative than many grand narratives.

The Next Function to Be Taken Over Might Be Your Company's Finance

The last topic was the broadest: which positions in enterprises will be taken over by AI? Note: taken over — the 24/7 automatically running kind.

Chen Guang has already applied this internally. Their R&D is landing a "loop engineer" system with three loops: the first loop is Agent self-feedback, iterating at minute-level; the second loop is humans in the loop, programmers giving feedback, cycling hourly; the third loop is external user feedback, cycling daily or weekly. Even recruiting is done this way: OpenClaw chats with candidates on Zhaopin BOSS, scores them by standard, the HR recruiter gives feedback, the sales VP on the business side gives feedback, and three rounds later it auto-adjusts. "Even if the HR team headcount triples, we won't need to hire." The only bottleneck is that Zhaopin BOSS doesn't let Agents chat — "this is annoying" — he says they'll look into using RPA later, but computer-use continuously operating the PC is token-intensive and "a bit wasteful."

But HR people still need to be there. What for? "He has to feed that OpenClaw, continuously tune its skills." You see — the position didn't disappear, but the work content was entirely replaced.

Liu Siping saw it more coldly: what they replace is the external layer of client enterprises. Customer service, telemarketing — these industries used to have massive human labor, plus massive software and hardware companies serving that labor; now those companies' growth space is shrinking. "If what you're serving isn't humans, why do you still need these software?" He named names: traditional outbound call management, soft switches; US communication software companies like Avaya and Genesys. Call centers used to invest millions or tens of millions in systems — "many of these systems won't be needed in the future."

Michael's next stop is finance and tax. The logic is still reinforcement learning: in recruiting, judging whether a person succeeds takes five years — the feedback loop is too long; finance and tax have the shortest feedback loop, you know immediately when something's wrong, and this is a huge human burden in US companies. In his view, all SaaS can be converted by AI: "Buying expensive software and then hiring even more expensive people to operate that software is fundamentally the same logic as buying a more expensive Agent."

Gao Zhou offered a higher-level framework at the end. He said he doesn't want to create anxiety — at this stage it might not be that fast, and in the future it might not be "replacement" but "takeover." What gets taken over first? Positions needing rational support and data support go first. He gave his old profession as an example: quant teams used to hire five or six graduates specifically to write code and strategies, updating massive strategies daily; now those people aren't needed — AI coding handles it all. Where's the human position? At the decision layer. After AI gives you plans A, B, C, D, the aesthetic, taste, and感性 part of judgment is a very big boundary between humans and AI.

The Side That Does the Math First Usually Lives Longest

After this roundtable, I had a strong feeling: these four people talk about Agents more and more like old businesspeople talking about business.

Attribution must be fair; billing depends on client quality; revenue sharing requires taking over the function first. Be decisive when rejecting clients; cost management is always online. These words would hold on any traditional boss's desk — it's just that their production tool has changed from humans to Agents.

"Gao Zhou comes from quant, used to reverse-engineer everything from outcomes. He says when managing client funds, first set this year's target return rate, then reverse-engineer trading strategy and asset allocation. Doing Agents is the same — first think about what money to make for the client, then reverse-engineer what product to build. This order is probably the entire secret behind the words 'revenue-generating Agent': money comes first, then the Agent, then the technology. AI gives A, B, C, D; the human makes the final call. In this division of labor, the side that does the math first usually lives the longest."

More Conversation Details

Speakers

upsello.ai Founder & CEO Gao Zhou (高舟)

Hellyeah AI CEO Michael Guan

94AI Founder & CEO Liu Siping (刘嗣平)

QuickCEP & PollyReach Founder & CEO Chen Guang (陈光)

Host

Unique Research Founder & CEO Wu Wei (吴畏)

Wu Wei: Please each introduce yourself in one sentence, and why you chose the field you're in now, and why this field is better than the other three.

Gao Zhou: Good afternoon everyone, I'm Gao Zhou, founder of upsello.ai. We started AI entrepreneurship in 2023; before that I did quant for nearly ten years. Back then it was mainly influencer marketing, did it for over a year, then stopped, and eventually moved to the application layer. In 2024 we made a customer service product, which is still on the market, mainly serving DHgate. This year we went further because many clients told us: since you do AI and talk about data-driven, can your customer service or external chat directly generate revenue, help me make money? This year we did a lot of research and chose the Shopify track because I think Shopify's entire data interface and ecosystem is currently very complete. We built an Upsell live chat shop plugin based on Shopify, but added more underlying data governance and data components. Because offensive shopping needs user data upfront and a data infrastructure underneath, otherwise it's hard to do proactive actions. This year we did this. We currently serve basically Shopify sellers, with about 5 large sellers above $100M GMV and the rest in the $10M-$20M range. Currently also running some data.

Wu Wei: What's the average deal size?

Gao Zhou: We split into two parts: one is traditional SaaS subscription; the other is Agent customization for Plus users, because they need customized events and workflows.

Wu Wei: How much are you collecting now?

Gao Zhou: We're at about $1M now.

Wu Wei: $1M for a custom Agent?

Gao Zhou: No, you mean per one? Per client it's 50-60K RMB. Subscription is a few dozen dollars.

Wu Wei: Michael, introduce yourself, then explain why you're in the field you're in.

Michael: I'm Michael, co-founder and CEO of Hellyeah AI. We're a US-based enterprise internal knowledge distillation and digital employee platform. The first landing scenario is growth marketing. Why? Because it has the strongest reinforcement learning signal. We deploy growth marketing digital employees into real combat scenarios, let them actually run budget — money made and money lost are crystal clear, and that's the most precious data in the entire reinforcement learning process. We started six months ago, currently at about $40M annualized revenue, ARR $40M, growing very fast. We started from zero six months ago — this is the first vertical, the first digital employee. Our company itself does knowledge distillation and digital employee deployment; this business I understand actually started from the Agent line. Correct. Previously we were helping enterprises build internal reinforcement learning environments, helping them train internal digital employees, and then found the marketing growth entry point and quickly reached this scale.

Wu Wei: Very good. Mr. Liu.

Liu Siping: Hello everyone, we're 94AI. We originally started with voice models — speech recognition, speech synthesis, speech interaction. We've been entrepreneurs for a while, started in 2018, early on experiencing the small-model application stage, and these past two years pushing Agent-type applications. We do voice models and voice interaction; early on we were also looking for scenarios and found the phone scenario more practical — like telemarketing, bank collections, customer service. We've been doing this for years, expanding business in about 30+ countries domestically and overseas, mainly serving finance, e-commerce, retail, education, etc. We have local sales and local people in about ten countries. The largest markets are Indonesia, Philippines, Nigeria, Mexico, Japan — relatively balanced. The core point of our business is: after building voice models, many application scenarios are actually very hard to charge for, hard to get results — that's what we want to discuss. We started in the financial industry because financial client calls are a rigid demand, whether callbacks, marketing, or collections.

Wu Wei: Give us the numbers.

Liu Siping: Currently about a year, domestic plus overseas together, several hundred million RMB in revenue.

Chen Guang: Hello everyone, I'm Chen Guang, our company is called QuickCEP. We focus on one vertical industry called brand going-global — you can think of it as an upgraded version of cross-border e-commerce. We only serve this one industry, going deep. We're a full-lifecycle Agent product, mainly serving mid-to-large big brands. We already serve 70 of China's top 100 going-global brands. When you buy Chinese brand products overseas — like branded phones, branded toys — all the touchpoints overseas are operated on our platform. Our Agent platform helps brands integrate all overseas channels, encompassing all consumer engagement data, and then Agent-driven marketing, after-sales service, and shopping guidance — that's the solution.

Wu Wei: Numbers please.

Chen Guang: Currently about 100M RMB ARMB.

Wu Wei: It looks like everyone is getting results with revenue-generating Agents, which is why we chose this track. I'll start pressing, starting with Mr. Chen. I noticed you keep emphasizing you're a full customer lifecycle management platform — I understand you may have started with CDP, customer service, and also help growth, wanting to manage all customer data on one unified platform, not just sign for one specific scenario. Why do it this way? Why an all-in-one full-lifecycle operations platform?

Chen Guang: Because we found that brand going-global, compared to doing consumer products or brands domestically, faces a more fragmented software ecosystem, IT components, and e-commerce platforms. So we particularly wanted a solution that aggregates all data and is Agent-driven. Early clients pushed us in this direction continuously, with a lot of positive feedback. Originally we only did a few segments, then kept adding — client needs fed us into this shape. Then we found this set works particularly well overseas. A few examples: overseas e-commerce — because the world is big, outside China each country and region has its own e-commerce platforms, social media, traffic channels. Chinese brands going global typically go worldwide — Europe, America, Middle East, Latin America, Southeast Asia. Faced with dozens of platforms and social media to integrate, it's painful. After integrating, they fight their own wars between departments — some are Chinese teams operating overseas, some have local teams, data doesn't connect. Plus marketing data doesn't connect with customer service and after-sales data. Wanting to do very refined, one-to-one operations and marketing is actually very hard. This is a big pain point. How to break these data silos? Different platforms and channels, internal and external — no shortcuts, we integrated one by one over the years. Some platforms are particularly closed, requiring official certified partner status to get API access. After integrating, you continuously follow their policies, continuously research, and keep your Agent's common sense and skills aligned with platform policies — accumulated one by one.

Wu Wei: Let me ask a question I care about. In the future software may all be headless, SaaS interaction less important, all used by Agents. Theoretically, connecting various application data should be easier — they might directly provide CLI or API rather than being so closed. Have you felt changes in software usage interaction in the past six months?

Chen Guang: Actually, overseas e-commerce platforms — whether store-building SaaS or e-commerce marketplaces — have always been open. There have always been APIs, unlike some domestic e-commerce platforms that killed their APIs. So they've always been headless. But during integration, we found how to make these APIs more Agent-friendly. We did a layer in the middle; clients still call our layer. We built a layer on top of the API because the original APIs aren't Agent-friendly schemas or MCP.

Wu Wei: Good. Next question: why did you build PollyReach? What's the relationship between PollyReach and QuickCEP? I find this very interesting.

Chen Guang: Earlier this year OpenClaw blew up; we wanted to capture the opportunity of the times, ride the wave — the effect was decent, so we released it. We'd always been doing ToB and thought, can we try a to-Agent product for the future, providing Infra for Agents? PollyReach is a skill — it's not to C users or ToB enterprises. We provide users with a US virtual phone number, which in a sense is very important Infra for Agents. When I was building a team in Japan, life was inconvenient for a long time — eating yakitori required reservations, and my English isn't good for booking haircuts, so I kept asking Japanese colleagues to call and reserve. I thought: can't OpenClaw help me book? It can't interact with the physical world, the atomic world — it can write PPTs and research reports, but how do you book a Japanese barber shop? It can't. So if it can make calls... just give it a skill to call.

Wu Wei: This new business's revenue or commercialization?

Chen Guang: Commercialization-wise, we're currently ranked #1 for this skill on Claw Hub; this type of skill has over 100,000 Claw installs.

Wu Wei: Mr. Liu, I just listened — you're mainly focused on voice outbound calls, initially focusing on financial industry scenarios like collections. The topic we want to discuss is revenue-generating Agents. Outbound call volume and directly generating revenue — I understand collections or user recall ultimately point to revenue, but call volume doesn't equal revenue. In this process, how do enterprises measure contribution to revenue? To what extent must the product be built before enterprises feel they can calculate by results or revenue?

Liu Siping: Clients care most about results — conversion rates at each step. They don't care how the feature is implemented; they care about the outcome. A good comparison is benchmarking against human teams: is employee conversion rate about the same as humans, and is the revenue generated about the same? Testing one with humans and one with AI directly compared — this is a very direct point. AI might not fully exceed humans in effect, but AI costs much less, so net returns are higher and ROI is better. After clients adopt it, the original human team will be significantly reduced; additionally, many scenarios that previously couldn't be served by humans are now covered. For example, using a credit card — you don't get a bank call reminding you to repay, you don't even get notified when overdue because they're too busy. Now some banks are using it — AI calls you at least before overdue or before credit reporting — actually making some scenario operations more granular.

Wu Wei: Based on the business just mentioned, some follow-up questions. Originally you were helping enterprises do knowledge distillation and provide platforms, entering marketing growth. I understand marketing isn't just ads or influencer marketing — it's the entire stage from user research, market research, strategy generation, delivery, attribution, optimization, and reinforcement learning. After enterprises adopt your marketing growth Agent, what things haven't they handed over to you? What might be handed over in the future, and what might never be handed over?

Michael: Great question. The product form is actually one that can integrate with CI. We believe the best form of digital employee is to provide it directly. Facing clients who need marketing — they're actually very technical — they have concerns when choosing different solutions, including compliance and data security angles; this is the big part. When doing North American B2B sales, if you can solve this, one foot is already in the industry. Back to the question: what can be authorized? Digital employees are like real employees — certain behaviors have autonomy, certain situations need boss approval. Designing digital employees follows this logic. Enterprises have P0 to P10 priority levels — like code debug telling you the priority. Items needing client approval are usually P0 to P3; the rest we proactively complete and present the results for satisfaction review.

Wu Wei: Is this rule internalized, or can it also be continuously trained through reinforcement learning?

Michael: Currently internalized. We're also exploring with clients where their boundaries are. Because real humans' attention is very limited, we want to avoid wasting the attention window — only give the most important, highest-ROI decisions to him. The rest with lower return but very low token cost and easy to implement are done directly.

Wu Wei: Many of your clients are also AI companies, right?

Michael: Correct, currently mainly AI companies in North America, plus many traditional gaming and entertainment companies.

Wu Wei: If you provide a CLI, I understand it calls your CLI within its Agent to complete actions?

Michael: Correct, not just CLI form but also a loop agent form that can continuously do things 24/7.

Wu Wei: Does it worry about internal enterprise data security?

Michael: Very worried, so all data compliance is done very finely.

Wu Wei: There are many details we can expand later. Mr. Gao, you've also done customer service before, now entering the independent site ecosystem, hoping to drive conversion through smart shopping guidance. From the shopping guidance perspective, when is the trigger moment? At what point do users not feel disturbed and actually accept this proactive intervention?

Gao Zhou: When designing the entire early event setup, we had many discussions with clients: which user behaviors to monitor? After monitoring, how to trigger so users aren't particularly annoyed? Now we've built many events, thanks to Shopify's ecosystem data openness — after client authorization, we can get a lot of data. Based on users entering from ad landing pages to the self-built site, all mouse movements, time spent on each interface, add-to-cart and abandonment actions — all monitorable. Based on these actions we form many events, using behavioral psychology and consumer psychology to create events, also customizing events with clients. Through popups — but not large popups; initially a very small bubble, with brand feel and brand language popping up, guiding users to converse with the bubble. We've been continuously iterating and optimizing this — making users feel comfortable while feeling genuinely cared for and guided in the independent site. At the same time, we hope traditional independent sites that are just landing pages, images, text, and video become warmer and more branded through shopping-assistant live chat, while helping users do things.

Wu Wei: Basically through the discussion just now, everyone has more understanding of the products and services your companies provide. Later I designed some general topics everyone can discuss. First topic: we define revenue-generating Agents. When truly creating revenue for clients, even if system capability is good enough, the client's own product must also be strong. There will be cases where the client wants the final result but we only did one part — requirements may be too high. The industry is involuted to the point of requiring full sales revenue sharing or profit calculation, but in reality it's only one segment. How do you balance this?

Gao Zhou: From day one of doing AI, I've been thinking: how does an Agent create value for clients and service partners? Coming from quant, my thinking is more reverse-engineered. For example, when managing client funds, first consider this year's target return rate, then match trading strategy, trading configuration, asset allocation — ultimately it's result-oriented. Doing AI Agents, I also use first principles and result orientation. Since we want to make money for clients and help them, I even considered whether to build an automated ad delivery system, but later found ad traffic is quite hard. Pushed back one step: I'd rather, after user traffic comes in, help improve user GMV even a little through guidance and event touchpoints. The traffic and ad sides aren't open; even if you spend 100K a month on ads, without introduction you might have $100K daily sales, and after introduction, bounce rate and upsell show very obvious improvement. Attribution is simple: all events, touchpoints, and behaviors happen in live chat; monitoring what's completed in live chat is our attribution. For example, determining a user was going to buy A, through upsell, cross-sell, or bundling we pushed B — only attribute the B part, A doesn't count as our GMV. This part is easy to attribute, thanks to Shopify's ecosystem data being all connected.

Wu Wei: Three of you can also discuss this topic. Are there clients you can't do and must give up? The client says let's test first, find it can't run out — not entirely our problem to solve.

Liu Siping: Being able to pay by performance is better for this type of company. Why charge subscription? If you can truly achieve pay-for-performance in many scenarios, it means your control over the client and core capabilities are recognized. If you can't, on one hand it's capability, on the other hand some client scenarios are different — like poor user quality, low conversion rate, can't work through. You have to separate: if the client scenario is suitable without these problems, you should try to do pay-for-results. As a fallback, charge by call volume. During early client admission analysis there's judgment. For top-tier clients, internet platforms or banks, user volume is large enough and quality stable enough for pay-for-results. But emerging enterprises acquire through different channels today and tomorrow — when you set the pay-for-results standard, they might suddenly be very profitable or suddenly lose money. Clients willing to pay by results or by call volume have a ruler in their hearts about how much cost they're willing to bear. When user quality is relatively stable, that ruler is relatively stable, ultimately with ROI measurement, feeding back that it's acceptable within budget. After pay-for-performance, your own发挥 space is larger, and you iterate models better for effect orientation. If charging by call volume, the important thing is increasing consumption and call volume, not necessarily pushing effect to the extreme.

Wu Wei: What do the two of you think? First, how do you attribute what you do; second, what result commitment or billing approach is better?

Michael: The product is simple and brutal. The digital employee goes directly into the code repository, and we customize a unique attribution system including data telemetry and all event tracking. We modify his code repository to ensure the data telemetry system is consistent with ours, ensuring attribution fairness and transparency. Second, pay-for-performance: the simplest is not locking into long-term contracts. Now many enterprises sign annual framework agreements, paying three years at once. Can pay-for-performance be paid month by month? If results are good, they stay; if not, they leave — it's actually mutual selection. We've also rejected many clients who can't accept such an intrusive digital employee solution. Putting code in his source code repository and retrofitting the system, he might feel you planted something. Of course these are very transparent — what code does what. Understand that a digital employee should be sitting in their office — first walk through their door. This is also a mutual selection process. Not every client must be onboarded; if the client can't accept such a forward-looking concept, decisively reject.

Chen Guang: Our track is more complex; let's sort it from the top down. Mr. Wu said pay-by-revenue and pay-by-results aren't quite the same thing. Sometimes the segment you serve can't directly lead to revenue —阶段性成果 can also be a unit of measurement. Currently over 90% of ours is still subscription plus AI Token top-up; we've started trying some pay-by-revenue or pay-by-results. Is it us pushing clients to accept, or clients proactively asking? We've always had this intention to push, but mid-to-large enterprise procurement processes and budget systems are hard to support — they can't say how much they'll spend this year. Another challenge is sales incentive challenges: originally subscription fees collected once a year, sales get commission. Now sign a contract but the client doesn't pay; the salesperson only gets paid after the bill is calculated by results — sales aren't happy either. Additionally, in our industry, the ones who can naturally take a commission are either those who control traffic or those who control transactions. Traffic is Meta, Google; TikTok ad delivery can take a commission; or acquiring platforms and store-building platforms can take a GMV cut. For other segments, clients in the past were unwilling to take a GMV cut — thinking it's just auxiliary, a tool, efficiency improvement — subscription or some token payment is fine. Why can we now start taking revenue cuts in local scenarios? We found some countries where the client's own team can't hold it or do it, so we use the entire Agent team to take over national marketing automation. EDM, WhatsApp mass messaging, lapsed member reactivation — the revenue created doesn't need management, no employees assigned, all employees plus Agents operating old-customer EDM member loyalty year-round, taking a cut of the GMV generated — essentially replacing a national consumer operations department. He accepted it. Complete takeover — he doesn't manage it; it was originally an extra piece nobody was doing, and you did an extra piece for him. For example, Japan generated an extra 10M through member operations, and we negotiate 1-5 points annually — these have all been signed.

Wu Wei: That's a good practical experience. Time permitting, last question. Starting from Mr. Chen. This question is a bit broader: whether you do shopping guidance, marketing growth, customer service, or operations Agents, many directly point to helping clients grow and increase revenue. Which positions or functions in enterprises do you think will continue to be taken over by AI? Or which software or SaaS will continue to be taken over by Agents? By "takeover" I mean 24/7 fully closed-loop automatic operation.

Chen Guang: I'll start. Our R&D is recently landing loop engineer — applying this concept to internal operations governance and product building, following the three latest loops. The first loop is Agent self-feedback, minute-level or hour-level feedback — very sensitive, iterating once. The second loop is humans in the loop — maybe a programmer running a task, programmer feedback on whether it's right, cycling hourly. Another loop is external users, external collaborators, software end-users giving feedback — bugs or experience issues by day, week, or month — these are the three loops. Our R&D now does it this way. For example, internal support departments, HR departments — always one or two people — can we add people using OpenClaw? Chatting on Zhaopin BOSS is annoying and not very Agent-friendly, but later it can all be Agent-driven. Invitations pass first, OpenClaw scores by standard, then HR recruiter feedback, OpenClaw has its own feedback, recruiter feedback, external user is the business — like sales VP, feedback on the resumes and interviews you found, how to adjust next time — that's how it loops. Even if the HR team headcount triples, we won't need to hire — that's how we do it. The only work that grows marginally is the part chatting with Zhaopin BOSS — they don't let Agents chat, which is annoying. We'll think about using RPA later. Because what I've used recently feels like its computer-use rate is much stronger than the previous version, but it's expensive — continuously operating or replaying PC behavior has high token consumption, a bit wasteful.

Wu Wei: You just gave the recruiting example as something that can be taken over by an Agent. But HR still needs to be there?

Chen Guang: HR still needs to be there — he has to feed that OpenClaw, continuously tune its skills.

Wu Wei: Don't worry too much — people in enterprises can self-learn. Mr. Liu, which segments, positions, or functions do you think might be taken over by AI?

Liu Siping: Inside enterprises, as Mr. Chen said, many positions are now being replaced. Focusing more on what we do — replacing the external layer of client enterprises. Customer service, telemarketing — these industries used to have massive human labor and massive hardware and software companies serving them; now those companies' growth space is shrinking. If what you're serving isn't humans, why do you still need these software? Lots of traditional software will be replaced by Agents. For example, in our field, traditional outbound call management, soft switches; US communication software companies like Avaya and Genesys — many are no longer needed. Call centers used to invest in a system — first hardware then software — costing millions or tens of millions; many of these systems won't be needed in the future.

Michael: I think all SaaS can actually be converted by AI. Traditional SaaS, especially in North America — buying expensive software and then hiring even more expensive people to operate that software is fundamentally the same logic as buying a more expensive Agent. Internally we're also doing many other AI directions of distillation and digital employee deployment. Recruiting is a suitable landing scenario, but above it there will be richer signals from a reinforcement learning perspective. Judging whether a recruiting candidate succeeds takes a very long time — five years without leaving and truly creating value before feedback. The shortest feedback loop is finance and tax; the second selection direction is finance and tax. Especially US finance and tax is a big internal burden in companies, with lots of human labor on it.

Gao Zhou: I want to share two points. Based on Mr. Wu's question about what will be replaced, I've also been thinking about this, but maybe don't want to create anxiety. At this stage it might not be that fast, and in the future it might not be "replacement" but "takeover." The thinking dimension might be higher — the more rational or more data-driven layer will be increasingly easy to take over. For example, the coding part — the macro analysis and quant analysis I used to do are very easily taken over. A simple example: we used to hire graduates specifically to write code and strategies — five or six specifically writing strategies, updating a huge number daily; now those people aren't needed, entirely handled by AI coding. Currently positions needing more rational support and data support will gradually be taken over; those needing decision-making and more感性 support will gradually form differentiation. In the future, finding where the boundary between human and AI lies is what to do. The data part and rational part — AI will definitely do better than humans, no need to doubt. What to do is at the decision layer — after AI gives plans A, B, C, D or different strategies, how does the human make the感性 decision — aesthetic, taste — this is a very big boundary with AI.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260811-01html
On-site reading page: https://ffcap.cn/en/research/src-20260811-01html