From 169 startups, I've observed these two trends in AI entrepreneurship.

Last week, Y Combinator (YC), the "barometer" of the global venture capital and startup ecosystem, concluded its 2025 Summer Demo Day, featuring over 169 startups.

Artificial Intelligence (AI) remains at the center of the stage. However, unlike previous years, the clearest signal this year is that AI Agents are becoming the core theme of AI startups.

The projects in this YC batch paint a picture of an approaching future: countless highly specialized and autonomous AI Agents will penetrate various industries, tackling the "boring, expensive, and repetitive" tasks that no one wants to handle, and even supporting the infrastructure of the entire AI ecosystem.

So, which AI startups are the most interesting? Which ones have attracted the most attention from investors? And in which sectors is AI most likely to create value in the next few years? Let's take a look.

01 New Trends in AI Agent Implementation: Replacing DDE Tasks

AI Agents were the hottest topic at the 2025 Summer Demo Day.

In the YC S2 list, more than half of the projects mentioned keywords such as AI Agents, autonomy, automation, or autopilot in their descriptions.

The application of AI is shifting from auxiliary tools (Copilot, where humans are the core decision - makers) to AI Agents capable of autonomous perception, decision - making, and execution of complex tasks.

This is not just a technological change but also a business model change. B2B enterprises are willing to pay for AI Agents that can "directly save or make them money," and their willingness to pay is much higher than that of ToC customers.

For example, Solva uses AI to automate insurance claims processing and achieved an Annual Recurring Revenue (ARR) of $245,000 within 10 weeks of launch.

Another example is Autumn, known as the "AI version of Stripe." It specializes in helping AI companies handle complex billing issues and is now used by hundreds of AI applications and 40 YC startups.

Why does a company like Autumn exist? Because the pricing of AI startups is usually very complex. It's not just a subscription fee but also includes usage - based billing, quota charges, and various additional feature fees. If using Stripe to manage these, it would require a large amount of manual operation, which is time - consuming and error - prone.

Autumn addresses this pain point. It has developed an open - source tool that allows AI startups to easily integrate complex pricing rules into Stripe, automatically complete billing and settlement, and significantly reduce manual operations.

There is a common point behind these: AI Agents are targeting tasks that "people don't want to do, can't do well, and are very expensive."

There is a term in the industry called DDE, which stands for Dull, Difficult, and Expensive. These scenarios are exactly where AI excels and have become the best entry points for their large - scale implementation.

The business model can be simplified as either saving money for customers or making money for them. The former involves sharing the cost savings or recovered amounts, while the latter involves taking a commission on the transactions directly facilitated by AI.

For instance, Frizzle uses AI to grade homework, freeing teachers from repetitive work; F4 and ContextFort focus on compliance checks of engineering drawings, detecting errors that are difficult for humans to identify and avoiding economic losses and project delays caused by design mistakes; Risely AI is committed to automating administrative work in universities, which has cumbersome processes and low error tolerance.

In the "making money" dimension, AI Agents are increasingly involved in value creation, being paid based on the results by improving transaction efficiency or facilitating transactions.

Shor is an AI payroll assistant. Its biggest selling point is that it can process payroll in just 3 minutes and save 80% of the cost.

In the past, if a company wanted to hire employees globally, it had to set up branches in different countries and go through complex banking procedures, which were time - consuming, costly, and involved many formalities. Shor's approach is to hand over these cumbersome steps to AI:

You just need to send a message on WhatsApp, such as "Hire Zhang San." Then, the AI will automatically generate a legal employment contract. The salary can be transferred to the employee's account within seconds without the need to set up a physical company locally.

In short, Shor is like a combination of an AI finance and HR function, helping companies hire and pay employees globally quickly and cost - effectively. Its founding team is from Tesla and Wells Fargo, and the product is currently in the testing phase.

Whether it's about saving or making money, it can be seen that the YC S25 cases have grasped the characteristics of DDE tasks and convinced customers to pay with quantifiable and visible value - cost reduction or revenue increase.

02 Extreme Verticalization: AI Penetrating the Fabric of Traditional Industries

There was a very obvious trend at this YC Summer Demo Day: Almost no one is working on "general large - scale platforms" anymore. Everyone is diving deep into vertical niches.

AI is no longer just a cool tool but is starting to transform into a new type of "labor force" and "expert system" in various industries.

The general model has become a highly competitive market. Startups are choosing to directly address industry pain points, even in super - niche scenarios. For example, AI debt collection and collision checks of engineering blueprints, which used to sound very niche, have now become new breakthrough points for entering trillion - dollar industries.

In the medical field, some companies are working on automatically generating ambulance reports, some are specializing in clinic referrals, and there are also AI pharmacy technicians.

The most interesting one is Perspectives Health. It can monitor the conversations between doctors and patients and generate medical records and forms in real - time, saving doctors half of their paperwork time. It maintained a 25% weekly growth during the pilot phase, has connected with 9 clinics, and plans to cover 180 doctors by September.

In the real estate industry, some companies are also targeting the pain points of real estate agents.

Clodo is a typical example. It has developed a "hands - free voice - interactive" CRM that can automatically record leads, follow up with customers, and search for properties. Currently, 60 US real estate agents are using it to save time and close more deals.

Spotlight Realty in the same field is targeting rental commissions, and Closera wants to be an AI employee in commercial real estate.

Moreover, AI is also advancing rapidly in the financial and legal fields. For example, Magnetic is an AI tax preparer for accounting firms, Kalinda conducts class - action lawsuit research for law firms, and there is also Qualify.bot, an AI phone agent for loan business.

Even in the private equity circle, there is a "post - investment alarm" like Palace, which can automatically collect and summarize reports from invested companies, reducing a 20 - hour task to a one - click export and providing real - time risk warnings. Currently, it serves funds managing billions of dollars in assets.

In the more hardcore manufacturing industry, the presence of AI is also starting to emerge. Flywheel, known as the "Waymo of excavators," installs smart boxes on excavators, allowing workers to operate them remotely.

More importantly, the equipment can collect data while working, and the AI model can continuously learn skills such as trenching and leveling. Eventually, one person can manage multiple machines and even multiple construction sites.

It can be seen that the key to the success of AI startups is no longer just technological advancement but also the in - depth understanding and reshaping ability of the operating logic of traditional industries.

03 Providing "Utilities" for the AI World

A large number of companies are focusing on providing underlying tools, platforms, and infrastructure for other AI application developers or enterprises, indicating that the AI ecosystem is maturing.

As countless AI agents flood into various industries, the trend of infrastructure development emerges. A large number of companies are focusing on providing underlying tools for other AI application developers. The AI ecosystem is moving towards large - scale maturity, and some are starting to lay the "utilities" for the AI - native world.

This trend covers the entire lifecycle of software development, deployment, evaluation, and optimization.

At the development and deployment level, Lilac is dedicated to discovering and reusing the idle GPU computing power of enterprises, Metis provides the infrastructure for building reliable agents, and Kernel offers an extremely fast "browser - as - a - service."

In the crucial evaluation and monitoring field, AgentHub provides a simulation and evaluation platform for AI agents, while Truthsystems focuses on real - time governance, automatically preventing high - risk behaviors.

At the data and model foundation layer, Louiza Labs synthesizes medical datasets to simulate human biology, and Relling is committed to creating a "World Model version of ImageNet."

In addressing the core challenge of making AI run efficiently, infrastructure - related projects demonstrate extremely high technical barriers. Enterprises represented by Luminal and Herdora are working on solving the key pain point of AI model adaptation and performance optimization on different hardware.

Luminal's core is to help AI models run better on different hardware. It has developed an open - source compiler and framework that can automatically generate GPU code and optimize performance through a "trial - and - error - select the best" approach.

As a result, the running speed of AI models can be increased by 10 times. Currently, it has served institutions like Yale University. The team background is also very strong. The founders are from Intel, Amazon, and Apple, with experience in chip optimization, system architecture, and business entrepreneurship.

Herdora focuses on analyzing the performance bottlenecks of NVIDIA GPUs. Its tool, Keys&Caches, can generate clear performance traces with just one line of code, helping developers quickly locate problems.

For example, it once helped a customer reduce the latency by 67% when deploying the Llama model. Now Herdora offers 10 hours of free usage. The team also has a strong background: one of the two founders is from a global quantitative giant, and the other is from Google, with knowledge of both computer science and economics.

The emergence of these infrastructure projects means that as the AI ecosystem scales up, the industry will also generate many new demands.

04 Conclusion

Against the backdrop of the industry talking about "AI fatigue," the YC 2025 Summer batch sent an important signal: investors are starting to bring their screening criteria back to the essence of business. They are no longer just looking at the novelty of technology but are more concerned about user retention, unit economic efficiency (UE), data and computing power costs, and potential regulatory risks.

This also explains why the projects selected this time generally tend towards "verticalization of AI applications." People are no longer keen on large and general platforms but are targeting real pain points in specific industries. Especially those traditional, high - value industries that have not been fully transformed by software, such as manufacturing, insurance, municipal management, and national defense.

It can be said that AI is entering a new stage, starting to be more deeply embedded in business processes and becoming the core engine for driving efficiency improvement and automation implementation.

This article is from the WeChat official account "Crow Intelligence Talk," author: Intelligent Crow. Republished by 36Kr with permission.