---
title: "Can AI Actually Hold Up Enterprise Profit Margins?"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-08-03T09:02:01+00:00"
canonical: "https://ffcap.cn/en/research/src-20260803-01html"
source: "https://uniqueresearch.substack.com/p/src-20260803-01html"
language: "en"
---

# Can AI Actually Hold Up Enterprise Profit Margins?

[![](https://substackcdn.com/image/fetch/$s_!oo6P!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F96cc911d-a0fe-4d24-9322-8dcb87e9b099_2048x1152.png)](https://substackcdn.com/image/fetch/$s_!oo6P!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F96cc911d-a0fe-4d24-9322-8dcb87e9b099_2048x1152.png)

_Original · Unique Research · 2026-08-03_

_Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening narrative, five themed sections, and the complete panel transcript. Pricing, cost, user-count, and market figures are speaker self-reports attributed to the named panelists, not independently verified findings. Company and person names are preserved as source attributions._

AI Industry Observation

"Not Selling Tokens, Not Doing Subscriptions, Even Going Out to Buy Companies": An AI Roundtable Tore Off the Industry's Fig Leaf

"The faster you run, the more you lose. AI application companies sit near the bottom of the domestic funding pecking order."

A counterintuitive phenomenon is happening in AI entrepreneurship: the faster you run, the harder you lose.

These are the words of Kuse AI founder Wu Xiankun. At a Unique Bloom roundtable, he and practitioners from Seekee, SenseTime, Google, and WeMeet laid bare several things this industry least likes to say out loud.

Three questions mainly: where the money comes from, how to control cost, and what happens in the second half of the year. Below are the truly valuable parts of that conversation.

Subscriptions Are Being Collectively Abandoned

Everyone building an Agent's first instinct is to copy SaaS homework — subscriptions. Wu Xiankun says this path doesn't work, and the reason is simple: AI is an open-ended solution, not out-of-the-box SaaS; deployment cost is high, and token cost is a marginal cost that won't trend toward zero like software, but rises linearly. The result is that whether growth is fast or slow, gross margin is low or even negative — the faster you run, the more you lose.

AI application companies, in the domestic funding environment, basically sit at the "bottom of the pecking order." Finding investors to subsidize losses by storytelling isn't a sustainable path.

So what to do?

Wu Xiankun chose revenue sharing plus buying companies. They partner with overseas education publishers to build AI workbenches for teachers — the partner brings data (on the scale of China's People's Education Press) and channel credentials, Kuse brings the product, and they sell together to schools. The same logic went into a Shanghai insurance AI project: they only do the design proposal, explicitly refusing implementation and coding, on the grounds that "this part's value isn't that great." More aggressively, they've started doing "AI roll-ups" in broad finance — directly buying companies, embedding AI into core business, and profiting by raising EBITDA and enterprise value, something almost no one in Asia is doing.

Gu Xuebin chose a completely different pricing logic. WeMeet's business-conference AI simultaneous-interpreter Agent is priced at 99 RMB per hour, capped at 100 people per language, with each extra person charged 1 RMB — 99% below existing market AI simultaneous-interpretation pricing. His judgment: this category is essentially an "AI-native service," not selling tokens or subscriptions, but charging by scenario and by the hour.

Xu Leyang runs a hybrid ads-plus-subscription route but admits the company isn't profitable yet. Seekee builds an AI browser for the Latin American market; cash flow is held up first by search ads, and subscription revenue needs time to accumulate — because ordinary users have no interest in understanding what an Agent is; you can only walk them one step first, and once it's good, they'll pay for a single-point feature.

"Behind the three paths is one consensus: subscriptions assume marginal cost is zero, but AI as a business never was, from day one."

What Really Kills AI Companies Isn't Tokens, It's Customer Acquisition Cost

Wu Xiankun summarizes this as a "triangle loop": to-C acquisition cost is high, token cost rises linearly with scale, and after pivoting to to-B, deployment/implementation cost becomes extremely high — three sides blocked, margin can't move.

But Google's Ian Li gave a more piercing data view: tokens aren't the life-or-death factor; customer acquisition cost is. He gave an example: searching "AI tools for free" on Google reaches hundreds of thousands of monthly actives, but this traffic is likely free-riders — no matter how thoughtfully you run them, they won't pay, just consume tokens. Whereas a B2B keyword like "AI agent for marketing" only has 4,000-plus monthly searches and a per-click cost of tens of US dollars, but converts to real paying users. Large traffic doesn't mean valuable; small search volume may be real demand.

He also mentioned an insight very useful for going-global teams: US users care about AI coding and automation; France cares about compliance; Germany cares about API integration and access standards; UK users want to solve SMB efficiency and job interviews — the same product, with copy aimed at the wrong country, directly decides success or failure.

Xu Leyang shared a practical detail for handling token cost: they use best models and tiered models separately. Pure conversation scenarios feel little difference with a next-tier model, but image generation tasks show visible model gaps, so the priciest model is a must — and when model vendors launch the next generation, the previous generation's cost drops directly 80%; that's their key lever for controlling token cost.

Ian Li also mentioned a counterintuitive acquisition trick: certain enterprise clients first charge a $50 consulting fee to screen out invalid traffic — the money is later refunded, but it's enough to quickly filter for genuinely interested paying users.

What Stalls Agents Inside Enterprises Isn't the Model, It's "People"

SenseTime's Cameron Wang says: the biggest obstacle to enterprise Agent deployment isn't at the model layer, but at the engineering and organizational layers.

Engineering-wise, an Agent must call the enterprise's existing business systems, but many enterprises simply don't have standardized APIs; engineers configuring on-site often take 1-2 weeks — completely at odds with the Agent's "plug and play" expectation. Organization is messier: an Agent runs only if the enterprise has clear, documented business processes, but many enterprises' experience is "passed mouth-to-mouth by veteran employees," undocumented, so the Agent can't step in. This is no longer an IT-department problem; it forces the enterprise to first digitize and standardize the whole organizational process, and business owners must personally step in, otherwise the Agent project ends as a stunning demo no one actually uses.

This judgment connects to the token problem itself. Cameron Wang raises that ToB clients have three hard-to-articulate pain points using tokens: cost can't be forecast (the same open task might spend a few thousand tokens or spike to 100,000+); efficiency can't be measured (no one can say what really differs between spending 10 yuan and 1,000 yuan); value can't be confirmed (if inference fails, all spend goes to zero; clients only want to pay for business results, not for uncertainty). She believes a token platform's real value isn't selling tokens but helping clients "manage" them — turning uncertain cost into a determined investment, with the endgame being clients pay only for results while the platform absorbs process volatility.

This Year and Last Year Are Two Species

Ian Li gave a very concrete time slice: last year's going-global AI apps were still in the "muddly fists beating the master" stage — simply packaging big-vendor capabilities like Veo 3 or Sora into a single tool earned traffic. But by June-July this year, such projects had almost vanished; what survived and did better were products that turned AI capability into solutions for specific workflows — like the complete workflow of helping overseas top creators grow followers, rather than an isolated feature point.

Vertical is the second keyword for survival this year. He gave a persuasive case: an AI-scanning 3D-modeling company previously sold image-to-digital-3D-model/portrait as a product, with only a few US dollars monthly; later it focused on the pet track, turning 3D modeling into physical figurines and pet pendants, and average order value jumped from a few dollars to several hundred — the model didn't get stronger; productization and scenario focus solved pricing power.

Gu Xuebin's approach confirms the same logic: WeMeet wins not with a stronger model but with compliance moats (domestic "dual filing" for large models and algorithms) and extreme scenario pricing (99 yuan/hour), converting the past year's sharp token-cost decline into 60%+ gross margin space — in the vertical of simultaneous interpretation, a nearly unreplicable moat.

A Question No One Can Yet Answer

On year-end outlook, Wu Xiankun spoke a plain truth: "whether AI delivers huge improvement to business — we still haven't answered this question." He raised a specific mismatch: a friend doing AI customer service says he can prove "how many people moved from human agents to Agents," but what customers really care about is "did profit margin rise from 20% to 30%" — the gap between these, over the past half year of token maximization and leaderboards, no one has truly filled.

The other panelists' judgments read more like the industry's "next script":

SenseTime predicts private deployment will rise sharply (data compliance pushing it), generic Agents will enter price wars, and the next entrepreneurial opportunity is packaging industry best practices into "off-the-shelf" industry Agent templates; meanwhile enterprise AI spend will shift from one-off CAPEX to continuous OPEX, and if business owners don't personally participate, Agent projects will likely become "zombie projects."

Google's Ian Li bets on a bigger judgment: just as the e-commerce era eventually produced platforms like Amazon, AI in the second half or next year may also see an "integrated platform" combining multiple Agents and workflows — they're already trying this internally and expect results by September.

Xu Leyang's ambition is simpler: survive first, then slowly migrate users' AI cognition from "clicking search" to "completing tasks" — if Seekee can become the standard for "whether a task was completed by AI" in Latin America, not just a supplement to Google, that's the real moat.

"Putting these together shows a clear main line: this round of AI application commercialization is collectively bypassing the old SaaS road. The zero marginal cost assumed by subscriptions doesn't hold here, and wrapper and traffic dividends have already been disproven in the first half of this year. The players left at the table either go deep in vertical scenarios to build pricing power, or abandon standardized product thinking and go into revenue sharing, outcome-based pricing, or even directly buying companies for equity. As for whether AI can truly hold up enterprise profit margins — even the entrepreneurs inside it can't answer that now."

More Conversation Details

Speakers

Kuse AI co-founder & CEO — Wu Xiankun (吴显昆)

Seekee AI co-founder — Xu Leyang (许乐洋)

SenseTime Group Ecosystem Cooperation Center GM — Cameron Wang (王晓璇)

Google Greater China New Customer Business Manager — Ian Li (李忆)

WeMeet founder & CEO — Gu Xuebin (顾学斌)

Host

Quwan Tech VP & Chief Strategy Officer — Zhuang Minghao (庄明浩)

Zhuang Minghao: First, brief introductions.

Wu Xiankun: Hello everyone, I'm Wu Xiankun, CEO of Kuse; we make digital employees. Our business model differs a bit from what everyone understands as subscription; we can talk later.

Xu Leyang: Hello everyone, I'm Xu Leyang; I did investment and financing for over a decade. Our Seekee is a going-global AI to-C product, mainly for ordinary users and people not yet comfortable with AI. Live a year-plus, over 100 million downloads accumulated. Currently exploring how AI commercialization better fits different groups in ordinary use.

Cameron Wang: Hello everyone, I'm Wang Xiaoxuan, running ecosystem cooperation at SenseTime. SenseTime has its own large model and AI Infra; first time at a Unique event, glad to talk about Agent commercialization.

Ian Li: Hello everyone, I'm Ian Li from Google Greater China New Customer Business. We've watched SaaS going-global since last year or earlier, then AI companies surged starting the year before. Our department largely looks at how to see profitable AI companies; later I can share changes from last year to this year.

Gu Xuebin: Hello everyone, I'm Gu Xuebin, founder of WeMeet, serial entrepreneur. On February 6 this year, WeMeet released the world's first business-meeting Agent in Shanghai, and it's supporting this conference today. We focus on cross-language and cross-platform, from simultaneous interpretation to interaction to lead generation, serving Chinese companies going global.

Money, Cost, Trends

Zhuang Minghao: Today we mainly discuss three questions: taking money, cost, and projecting second-half trends. First, your main pricing model? Who for? Why?

Wu Xiankun: Everyone doing Agents first tries subscription. But you find AI is an open-ended solution, not out-of-the-box SaaS; deployment cost is very high, plus token cost — marginal cost is very high. In the end, whether you grow fast or slow, gross margin is low or near loss. Fast growers might even have negative gross margin. We had a period of very fast growth, but losses expanded just as fast. The only solution is raising investor money, going to capital markets to make up value — not a sustainable road. Second, AI applications in China's funding environment are basically at the bottom of the pecking order; not much meaning.

If you charge along the value chain, subscription is more self-serve; going up, bundled deployment reaches FDE. But no matter how fancy FDE is, it can't escape one category — it's still outsourcing, just more modular and fancy; hard to escape the essence of the business model.

Our chosen path is revenue share with core partners, then roll-ups with some PE. For example, partnering with an overseas education publisher to build an AI workspace for teachers, sold to schools. They bring data — locally equivalent to People's Education Press, with decades of textbooks and teaching materials; the data itself is precious, other large models can't provide it. We bring productized solutions, sell together to teachers, budget depending on location. They bring data, channels, credentials; we bring product, then split. If this were outsourcing or a one-time fee, we might not do it. Including a Shanghai insurance AI case — we did the design proposal, refused implementation, refused coding. Coding is simple; value isn't that great. So at least do revenue sharing.

Going up, mainly doing AI roll-up; very few in Asia, maybe in Singapore some are preparing to try, but we've started one or two months. We'll buy companies, mainly in broad finance. Find ways to forget my own productization, forget which model to use, see if it improves core business, acquisition, or customer service, raise EBITDA, raise enterprise value to profit. Eventually may raise money to buy companies in adjacent industries. Charging software or outsourcing fees is meaningless; deployment is very heavy, and if you're not from the company, interests are hard to align. Very early attempt, also a hard road.

Xu Leyang: Let me add product. We build an overseas AI browser, very C-end, trying to penetrate daily use. From a search entry to sensing needs, converting to simple AI apps. We basically don't do long-horizon or complex tasks — markets Manus and Genspark are proving, easily seen and attacked by big tech.

We chose Latin America. Three years ago internally discussed; previously did to-C products for Southeast Asia and the Middle East. Southeast Asia's feedback was weak local purchasing power. So the second product moved up in purchasing power — Latin America is about 1.5 to 2x Southeast Asia. On product logic, we go down. Chinese or American products essentially try to change user habits, a big behavioral change; we're cautious. Ordinary users' behavioral migration must be gradual. The benefit of an AI browser is that people don't find it jarring; they have a frame, know it's helping find things.

On revenue, first, we have search, several million DAU, embed some ads — cash flow recovery. Second, during search we know what users saw, clicked, wanted to search, why they didn't go next; we introduce different AI tools. Like searching how to edit an image — why introduce an AI tool to solve it, rather than give links. Our understanding of users, accumulated data moat, plus small-language unique advantage, are all reasons for recognition.

Second, slowly increasing is subscription, but it needs time to layer in. Ordinary users have no time, energy, or interest to understand Agents or AI mechanics; today we walk them one step, give solutions, they find and try, feedback if bad — basically short one-or-two tasks, doing very little. Very aligned with overseas ordinary users, a very broad need. Latin America is heavily influenced by North America, willing to pay for vertical products, purchasing power strong. These two revenue streams support us going down, but the company isn't profitable yet. Xiankun said funding started the beginning, but because of fast growth, overseas still has potential.

Gu Xuebin: As an industry old hand, let me be direct. In the AI era, building things means heavy cost reduction, extreme efficiency. Our business-meeting Agent is priced at 99 RMB per hour, capped at 100 people per language, each extra person 1 RMB. Compared to AI simultaneous interpretation, 99% below existing market price. Today this conference has English, Portuguese, French, Malay, Thai; providing this simultaneous interpretation and interaction used to be unimaginably costly, now a few hundred yuan an hour.

We were born global from day one, domestic and overseas simultaneously. Our first overseas real run was Unique Research's June Singapore event; I watched smoothly from Shanghai via link, meaning Chinese compute and cloud can already support it. Domestic 99 yuan/hour; overseas, US dollars.

Changes in Going-Global AI Applications

Zhuang Minghao: For today's AI app companies going global, in traffic operations, promotion, and cloud, Google is the most natural choice. Your department focuses on this; over two years, what changes have you seen in AI?

Ian Li: My department is 0-to-1, helping all AI companies going global become profitable. From last year to this year, at the same time point, companies are very different.

Last year was a bit like muddly fists beating the master. Simple scenarios, not solving Agents or workflows, just simple aggregation platforms wrapping big-vendor Veo 3, Sora, getting overseas traffic and paying users with a single tool. But by June-July this year, we no longer met such companies; instead, most go global as workflows and AI Agents.

For to-C, aggregation-platform companies still exist, but those surviving and doing better are video-tool companies, built into workflows solving overseas top creators' follower-growth needs — critical. This year a single tool isn't suited to going global for better paying users; tools or platforms solving personal or enterprise workflows get better margin and paying users.

Second, sub-verticals. Biomedicine, academia, enterprise professional fields, like WeMeet, very segmented verticals, solving one scenario concept; stronger barriers than general models or general Agents; others can't solve without learning more. Recently helped a medical company with overseas gene sequencing and bio-detection; many tools can't. The more vertical, the higher-paying users this year.

Third, each scenario has different product lines. In March we hit something interesting: overseas and domestic similar companies, strong in AI large models and scanning. Before, upload graphics to make 3D models or avatars; this year felt unprofitable, focused on pet offline track, upload images to generate 3D, but added tangible figurines, pet pendants and physical content; AOV jumped from a few dollars monthly to several hundred. So revenue may have more varied choices.

SenseTime's View: Real Challenges of ToB Agent Deployment

Zhuang Minghao: SenseTime has models and Infra, doing lots of ecosystem cooperation across industries. From a model vendor's view, what changes in commercialization exploration this year?

Cameron Wang: SenseTime has done many private Agent projects. Before talking revenue there's a precondition: the real challenges of ToB Agent deployment. Earlier understanding was that Agent deployment difficulty was mostly at the model layer, but actual challenges come more from engineering and enterprise organization.

Engineering-wise, an Agent running tasks must call existing business software, needs standard APIs, but many enterprises' existing systems lack them; engineers configuring on-site usually take 1-2 weeks, inconsistent with Agent plug-and-play.

Organization-wise, an Agent's precondition is clear documented business processes, but many enterprises rely on employees mouth-to-mouth, so the Agent can't step in; it forces the enterprise to form standard complete business processes. This is no longer an IT project but a management-iteration problem — AI-ify and digitize the whole organization, not just IT but business owners must step in. These are two very realistic problems facing ToB Agent deployment.

Changing Cost Structure and Pain Points

Zhuang Minghao: Second segment, cost. Xiankun, Leyang, Mr. Gu, at today's point, what's the rough cost structure? Most painful module?

Wu Xiankun: After scaling up, especially with our to-C product, 600,000-700,000 users, gross margin positive. But as scale grows, token cost becomes a big problem, very painful — why pay model vendors so much, keep so little. Acquisition cost is relatively low; I'm good at social-media marketing, overall acquisition cost low, but margin still relatively low.

Later shifted more from C to ToB. ToB implementation cost is inherently high; AI engineers or AI-literate engineers are expensive, hourly pay high. To-C acquisition cost is high — lots of funded, burn-willing, profitable companies competing for traffic. Second, it can't be like SaaS — this is core, marginal cost won't be zero but rises linearly; this is everything, other topics revolve around it. There's token cost; turning to B-side hopes to control token cost and raise margin, but deployment cost rises. A triangle loop.

Xu Leyang: As a complete product, we have every cost, plus AI-era costs. Xiankun says token cost rises significantly with users. Besides user use, many who haven't touched AI need education or trial; trial is promotion cost. On one side is acquisition cost. The benefit is Latin America's average acquisition cost is much lower than developed countries, but the other side's cost makes up. We don't have full-time staff in Latin America, but part-timers and local cross-timezone cooperation; cost goes up. This differs from opening offices nationwide; cultural differences and coordination produce hidden costs layered on acquisition.

How to adjust on tokens? The team built orchestration capability. Users are in the browser, daily behavior inside — image editing, documents, conversation, learning. When to let them know a feature might help is behavioral recommendation. As usage data accumulates, under which task what type to recommend, how relevant — understanding users and probing Latin American users.

Third, connect best models, cooperate with all large-model vendors. But across user groups and tasks, whether to give the best or next-tier model ties to user understanding. Image best-model gap is too big, cost very high; we first limit, then charge for best. Some tasks like pure conversation — Gemini Pro and next-tier differ greatly, but simple image modification doesn't need the best. Using previous-gen benefits: when vendors launch next gen, previous-gen cost drops directly 80%. For hundreds of thousands of daily trial users, cost control is critical. Combining AI and going global has big appeal, but cost control is a continuous learning process.

Gu Xuebin: WeMeet sees things a bit differently. First, not selling tokens, not selling credits; from the start priced by hour, based on scenario essence — AI-native service, enterprise AI Service category.

Cost has two big categories. First, vendor compliance cost. As serial entrepreneurs, over the past two years domestically as an AI vertical-model company, we successively obtained large-model and domestic algorithm Cyberspace Administration filing; dual filing done. This cost needed founder team and early investors' support, but after completing it, we can say so in the market.

Second, actual product operating cost. Thanks to large-model vendors, domestic and overseas; over the past year token cost dropped sharply, making 99 yuan/hour domestically still over 60% gross margin. Beyond simultaneous-interpretation text on screen, everyone here speaking Chinese can go to the link to take AI notes, ask the Agent what previous speakers said, minutes — plus AI live streaming. Now there's an opportunity, as AI infrastructure cost drops sharply, to provide users extremely low prices in a highly compliant way — this seems validated.

Balancing Acquisition Cost and Tokens

Zhuang Minghao: To Mr. Ian. Those who find you are very purpose-direct, and you have ad placement, cloud, best model vendors, deep international-market understanding. From a cost view, over the past six-plus months as AI and Agents scaled, what changed?

Ian Li: Tokens are every AI company's potential cost, but I hope everyone focuses more on acquisition cost — the key to whether you can be profitable and develop long-term.

Xiankun controls B-side acquisition cost well. For example, email consultation first charges $50, cutting much invalid traffic at the door; at least labor cost — the $50 is refunded to the inquirer, but it can quickly select highly interested paying users.

From data, B2B vs to-C. On Google, to-C browsing search volume is huge — search "AI tools" or "AI tools for free," monthly actives from tens of thousands to hundreds of thousands. But this traffic may not bring good signups or paying users; maybe free-riders, gratuitously adding huge token consumption. Even with lots of creative user operations, they still won't pay. That's a big challenge facing to-C AI companies. If from the start all marketing or user-operations money goes to attracting these users, you may only keep raising.

B2B terms, like "AI agent for marketing," have very low search volume, just 4,000-plus a month, but per-acquisition click cost can be tens of US dollars. Many paying users cluster in enterprise AI Agent workflow consumption. If you really seek precise paying users, ToB enterprise products may quickly get this increment.

Second, on acquisition cost, assuming going global, each country differs. The US market, many people want to learn enterprise or consumer, talking about AI coding, AI Agent automation. France focuses on compliance; Germany cares about API integration and access standards; UK searches for whether it can quickly help SMB efficiency, job interviews quickly solved. Different market choices or earlier copy prep may mean failure. For example, Leyang's product helps individuals as assistants; if starting in UK Europe, maybe promote "can it quickly help write a resume"; if choosing office automation, maybe not many care. National audiences need deep understanding of what they care about, what stage; then cost control and acquisition cost have more variation.

Finally, cross-trade problem. Same going global, I cover cross-border e-commerce and AI Agents. I strongly encourage going-global AI companies to hire a cross-border e-commerce independent-site operator. Many AI companies watch user growth and operational-cost increases still around technology, but it's a going-global business. Better to directly find someone already successful in cross-border e-commerce to manage and build the growth curve, applying others' experience to AI going global.

How Model Vendors View Token Cost and Commercialization

Zhuang Minghao: SenseTime as a model vendor, model competition fierce; previous guests all raised token cost. You bear heavy model R&D cost and heavy implementation; how do you weigh changes and trends?

Cameron Wang: The whole Agent industry chain is the process of token gradually condensing downstream into business fruit. SenseTime has its own token platform and AI Infra; let me talk more on tokens.

Many ToB users face pain points using tokens. First, cost can't be estimated; open Agent task token use is very probabilistic, possibly spiking from thousands to 100,000+, clients can't budget in advance or understand bills.

Second, token efficiency. When an Agent does a task, different prompts, different model calls, different inference paths cause wildly different token efficiency. Spending 10 vs 1,000 — what's the real difference? No one can say.

Third, effectiveness. Spent so much on tokens, how much business value? If inference fails, all outlay is cost, producing no business value. Clients only want to pay for business value, not uncertainty.

As a token platform, value isn't only producing and distributing tokens but helping clients manage them. A good token platform needs several capabilities: first, full-view token usage observability, providing panoramic granular bills for different Agents and tasks so clients understand. Second, helping clients improve token efficiency in various ways — auto-compress prompts, auto-route unimportant subtasks to cost-effective small models. Third, token governance — platform can set per-Agent or per-department usage caps or auto-approval paths, letting Agents run freely under financial oversight. Finally, the more imaginative: let users pay only for results. Now pay by use, however much you use. Future innovation: token volatility generated during Agent work is absorbed by the platform through technical optimization; users pay only for the final business result, avoiding all usage risk. The token platform is Agent-business infrastructure; only when tokens move from uncertain cost to determined investment can Agent business be reliable.

Year-End Outlook: Milestones and Trends

Zhuang Minghao: Last question: by year-end, what milestones or expectations for your own business, commercialization, or industry trends?

Wu Xiankun: What I want to validate is simple: whether AI delivers huge improvement to company business. We still haven't answered this. The past half year went through many AI movements — token maximization, various leaderboards — but never fully validated whether, in the full process, it can sufficiently justify enterprise value. Earlier I mentioned clients want to pay for results, but facing clients, what is that result? Can it be responsible for profit? A friend doing call center says he can't be responsible for profit, but can be responsible for how many people move from human to Agent — that's his defined result. But the client thinks, using this should raise profit, margin from 20% to 30%. There's a key misalignment; how companies expect to improve and what they define are sometimes different.

We're special in never touching external investment; cooperating with buyouts and PE is because we're neutral, no other external investor gets information access — safer for PE or buyers, don't disclose business to investors. This has nothing to do with orchestrating harness or Agents. But ultimately the acquisition-cost problem remains unsolved; it takes time to see whether a company fully AI-ified significantly improves profit, margin, EBITDA.

Xu Leyang: This can be seen on two levels. As a startup, surviving is most important; posture doesn't matter. Ad revenue or subscription, layering over time is a ratio issue. Longer term, being to-C, we care what users' understanding of AI is today.

In the early era every user click was a billing carrier; later Google became Google IT, users felt actions had value. Today AI's benefit is everyone is educated that using AI costs money, priced in tokens. But users' recognition of tasks is slowly changing. When we did Seekee in Latin America, we migrated users from search behavior to task completion. Search goes original if effective; when recognizing clear intent, convert to a task. "Task" is a bit slave-like; many users naturally dislike it, ordinary life doesn't need such a formal process to show results. But AI can tell them it's efficiency improvement, result optimization, doing what was impossible before. After completion, is there value? Per-use fee or bundled subscription relates to user cognition. Seekee hopes to grow with users, behavior staying in the Seekee browser. Maybe one day in Latin America or emerging markets, Seekee becomes the standard metric for completing tasks on AI, not Google, though Google is already there.

Cameron Wang: Future Agents, especially ToB, may have three directions. First, private Agents will rise sharply. Data-compliance requirements grow; many enterprises will large-scale private deployment, open-source or privatable model use will surge.

Second, industry Agents. Generic Agents start competing on price; the next win point may be industry Agent templates, packaging industry best practices into generic Agent configs, letting clients open and use — possibly the next ToB startup point.

Third, organization. Enterprises using Agents need the mindset that this isn't an IT project but whole-organization management change. Agents grow with model iteration, data accumulation, business change. Enterprise Agent use needs continuous team operation, creating new Agent-related roles; enterprise spend shifts from CAPEX to OPEX. Many enterprises struggle with Agent deployment not due to tech or budget but business owners not personally joining, ending with stunning Agent demos becoming zombie projects no one uses.

Ian Li: I expect fast AI development in the second half or next year. Take e-commerce or the internet era; now cross-border independent sites and platforms both exist, but Amazon has huge traffic. On the traffic side, our internal team has started a workspace model, fusing various Agents or workflows. It reflects that the internet era ended with a few big platforms aggregating all functions; might AI in the second half of this year or first half of next year see a similar integrated platform solving part or most of problems? We've started trying; expect such content to appear.

Second, extend revenue models and methods. Whether combining offline or personal lifestyle or enterprise, besides AI Agents solving problems, the next step is fusing others' offline life and work, or linking to actual business. Expect good results this September.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260803-01html
On-site reading page: https://ffcap.cn/en/research/src-20260803-01html