Yunqi Capital Perspectives | The Second Half of AI Agents: Stop Grinding on Efficiency Tools, the Opportunity Lies in Reconstructing Productivity
If Models Take 80% of the Value, What's Left for AI Applications?
The AI application gold rush is drawing builders from every corner. But here's the brutal truth: roughly 80% of AI software funding and revenue is flowing to the foundation model providers. So how do application-layer players actually unlock value?
Recently, Yunqi Capital and Future Tech — a sub-brand of WAIC (World Artificial Intelligence Conference) — co-hosted the sixth edition of Future Tech Demo Day. Peter Liang, Executive Director at Yunqi Capital, delivered a keynote on "AI Agents Entering the Second Half," cutting straight to the real situation facing application-layer startups and dissecting viable paths for entrepreneurs building Agents. This article shares the highlights.
For more details on this Demo Day, click "Read More" at the end of this article.
*The following is adapted from a live presentation by Peter Liang, Executive Director at Yunqi Capital, with edits for clarity.
This Wave of AI Entrepreneurship Is Harder Than You Think
The popular narrative around AI startups goes like this: barriers to entry have dropped, opportunities have multiplied, and a single person can now build a company. But if you actually look at the track record of this wave of AI application startups, the picture is far less rosy.
In the cloud computing era, application-layer companies captured over 60% of total industry value — B2B SaaS did this, and China's mobile internet companies achieved even greater scale. But from ChatGPT's launch in 2022 to today, despite all the fundraising buzz, roughly 80% of AI software funding and revenue has flowed to the foundation model providers.
The most celebrated application-layer benchmark over the past year has been Cursor, whose success is tightly coupled with both base model capabilities and revenue. Beyond Cursor, most other AI application unicorns are doing under $500 million in ARR.
Looking at user operations data: AI applications currently considered top performers may show explosive user growth, but their paid conversion, retention across various metrics, and gross margins still lag significantly behind both consumer mobile internet and B2B SaaS benchmarks.
Even the most prominent AI companies deserve scrutiny. Referencing research from Shi Xiang: if OpenAI sticks to its current business model — consumer (subscriptions plus internet-style ads and platform take rates) plus enterprise (infrastructure tax) — and considering the global base of individuals and businesses, its revenue ceiling likely sits between $200–300 billion. That's still a considerable gap from today's cloud computing giants. Factor in OpenAI's projected trillion-dollar capital expenditure, and the ROI of its current business model faces serious challenges.
Behind these numbers lies the fundamental question: where exactly is the value and viable business model for the AI application layer?
From Copilot to Autopilot: The Paradigm Hasn't Actually Shifted Yet
Most AI applications today are essentially Copilots — helping people work faster, delivering single-digit multiples of efficiency gains. That sounds substantial, but it's not enough to drive real behavioral change or fundamentally disrupt workflows and organizational structures.
The real paradigm shift is from "assisting humans" to "replacing labor" — AI completing end-to-end tasks directly, not just accelerating individual steps. This is labor substitution, not tool substitution.
The market size difference is an order of magnitude apart. The global B2B software market is roughly $500–600 billion. By public estimates, for every dollar overseas companies spend on software, they spend about six on various outsourced services: HR outsourcing, headhunting, accounting and tax agencies, IT operations... In China, that ratio is likely even higher, perhaps 20x. $500 billion versus $10 trillion — that's a completely different scale. We've also seen growing numbers of overseas VCs adopting "AI Rollup" strategies: acquiring traditional service businesses and reengineering operations with AI to drive efficiency gains.
Of course, this is extremely difficult when the technology isn't mature enough. An Agent task chain requiring dozens of steps, with 85% accuracy per step, compounds into an unacceptably low success rate. But long-horizon Agent engineering — including Harness, context memory, reinforcement learning, and the maturing ecosystem of Agent interfaces — is steadily improving accuracy and generalization on complete tasks.
We're already seeing this happen. Take our portfolio company Creao AI: 80–90% of its R&D and marketing runs on its own Agent platform with multi-domain Agent coordination, replacing repetitive work. The feasibility of AI replacing junior positions is clearly improving.
We've also seen very interesting AI-native hardware form factors in our post-98 investment program Y Tranformers, where founders have achieved genuine organic integration between hardware and software, with each amplifying the other.
What Software Couldn't Do Is What AI Should Do
What scenarios are right for AI entrepreneurship? Here's one criterion: if software could already do it, it's probably not an AI-native opportunity. Software's logic is to find the greatest common denominator of demand and deliver standardized solutions. Competing on the same demand pool tends to end in pixel-for-pixel copying and destructive price wars. AI's bigger opportunity lies in solving personalized, non-standard scenarios that software couldn't address.
The vehicles for this type of demand have existed in commercial society for decades — BPO (Business Process Outsourcing), powered by human labor. The consumer side has numerous industries with the same characteristics: education and training, psychological counseling, and more.
Looking just at the enterprise side: accounting and tax agencies, HR outsourcing, management consulting, insurance brokerage, IT testing, auditing... These industries exist in outsourced form precisely because demand is highly personalized and cyclical — building these capabilities in-house costs far more than outsourcing. Traditional software is "powerless" here: it needs structured inputs and explicit rules, while BPO handles exactly the work in software's "forbidden zone": unstructured information processing, cognitive judgment in complex contexts. This is precisely the dimension where AI excels.
BPO has another key characteristic: it sits naturally outside enterprise processes. AI replacing BPO services doesn't require touching internal organizational structures — it's a frictionless penetration. And these industries are brutally competitive in China — for example, soft labor outsourcing gross margins have been squeezed from over 20% to below 10%. The more cutthroat an industry, the more it needs to be rebuilt. This is definitively an AI-native opportunity. Traditional BPO competitive advantage lies in people management and operational efficiency, in bridging supply-demand peaks and valleys.
Disrupt External First, Then Consume Internal
To understand the rhythm of AI's impact on different types of work, we borrow a four-quadrant framework from overseas funds: the horizontal axis is "does this require proprietary enterprise experience for decisions, or general knowledge decisions," and the vertical axis is "outsourced or internally executed."
The first to be hit is the "outsourced + general knowledge decisions" quadrant — tax and audit, IT testing, compensation and HR, insurance brokerage. These are trillion-dollar markets that don't require changing internal enterprise organization, making them the most directly addressable area for current Agent capabilities.
Management consulting sits at "outsourced + requires some industry internal/external judgment" — and change is already underway: based on my own experience, consulting projects that used to take four- to five-person teams one to six months can now be executed efficiently by one to two people plus AI Agents. We've observed significant compression of consultant headcount over the past two years.
After BPO gets disrupted by AI, the next wave hits internal enterprise processes: supply chain management, wealth management, internal approvals. Once external BPO services are overturned by AI, enterprises will discover that using AI internally works just as well, and internal processes get restructured. The last to be affected are roles highly dependent on proprietary enterprise data and internal experiential judgment — currently the hardest to replace.
One case worth highlighting separately: the investment advisory industry. Yunqi has been deeply rooted in fintech for years. Digitalization and AI have already disrupted multiple industries. The advisory business has been strongly dependent on individual advisor capabilities for client development — user trust in people is hard to replace with products, and even the best advisor has limits on industries and clients they can cover. But Agent changes this ceiling. This is a category that, from capitalization and scaling perspectives leading to industry fragmentation, has historically been very difficult to invest in — now made seriously reconsiderable because of AI.
Don't Become a Single Line in a Foundation Model's Release Notes
Every foundation model update kills another batch of application-layer companies. That's the real situation for AI entrepreneurship today. To build a company with real moats, several dimensions deserve serious attention.
First is the "Thick Agent," not a thin tool. The deeper you embed in customer business, the more complete your ecosystem, the higher the switching cost — that's a genuinely sustainable moat. Companies whose capabilities depend solely on a single model face the constant risk of having their core competence absorbed.
Second is proprietary data. This is the hardest barrier. For example, Harvey in legal and OpenEvidence in healthcare don't just stand on model capabilities alone, but on proprietary data scattered across their industries with extremely high barriers to entry. Continuous data accumulation with continuous model iteration builds ever-higher walls.
Third, traditional business moats still work. Network effects, franchises (data, licenses, qualifications), data flywheels, extremely low marginal costs — these barriers that held in the software era remain valid in the AI era.
Our lead investment last year in Shenzhen Index Technology applies AI to replace China's hundreds of thousands of electronic design solution providers — a classic labor-intensive, highly segmented BPO industry. It combines extremely proprietary data barriers, technical moats, and future hardware-generation entry-point-level imagination space, achieving tenfold-plus improvement in per-capita efficiency. It's the result of stacking these criteria, and one of the most representative cases we see in this direction.
Discussion
The above represents our current observations and judgments. If you're building AI applications, or also thinking through Agent implementation paths, we're genuinely curious about your takes on these questions:
- In your scenario, are you still at "efficiency gains," or have you started "replacing people"?
- With foundation models continuously evolving, where do you think product moats should be built?
- On this category of "what software couldn't do" — BPO — are you seeing new entry opportunities?
Share your judgments in the comments — we'll continue tracking these frontline developments.
More Highlights
This Demo Day also brought together 16 AI startup teams exploring the front lines, covering Agents, AI-native applications, and novel interaction paradigms. Click here for pitch details and project breakdowns — you might spot the contours of the next wave.
