On September 7, WeChat was reported to be testing a feature called “XiaoWei AI Social.”

Users simply tell “XiaoWei” whom they want to meet and what they want to do. It then contacts the friend’s “XiaoWei” first. The two AI agents confirm the request and timing, returning to their users for a decision when human input is needed.

People used to seek out AI for conversation. Now, AI is starting to find people for them.

The idea has already attracted real money in Silicon Valley. Over the past year, three AI dating startups—Overtone, Known and Ditto—have raised a combined $36.9 million, or about 247 million yuan.

AI Finds People—and Funding

The highest-profile funding came from Hinge founder Justin McLeod.

At the end of 2025, McLeod left Hinge, the dating app he had founded nearly 15 years earlier, to re-enter the dating market. In July, his new company, Overtone, announced an $18 million funding round backed by FirstMark Capital, Pace Capital and Hinge parent Match Group.

Rather than build another dating app that keeps users endlessly swiping, McLeod designed Overtone to learn about users’ experiences, preferences and relationship goals through voice conversations. It then recommends a small number of potential matches and explains why they may be compatible.

San Francisco startup Known takes an even more direct approach. The company raised $9.7 million late last year. Users first speak with an AI, which recommends one person each week. If both sides agree, the AI arranges the first date. Known charges based on results: $15 per person for each date.

Ditto raised $9.2 million in seed funding in February. Users do not browse photos or make small talk with strangers first. Every Wednesday evening, the system sends them a match along with a time and place to meet. TechCrunch reported in August that Ditto had about 150,000 registered users.

All three companies are eliminating the most time-consuming parts of traditional dating apps: browsing profiles, making small talk after matching and repeatedly coordinating schedules. AI handles the initial screening and logistics; people handle the meeting.

That is also the direction favored by Wang Siyu, founder of Shanghai Gongqing Network Technology. He has spent years running online and offline gaming-companion and contracted-companionship services. In his view, AI does not necessarily need to replace people. Its most practical value is making the process of finding someone faster.

“If I want to find someone, for example, using AI would make it much faster,” Wang told Pencil News.

In his business, these requirements are highly specific. Human gaming companions in Shanghai charge from about 60 yuan an hour at the low end to as much as 500 yuan. Some customers care about age and appearance; others prioritize location and timing or have specific requirements for games, conversation and activities. In the past, the more detailed the request, the more the platform relied on customer-service staff to communicate back and forth.

AI can first clarify the requirements and then filter the available providers. The same logic applies to finding dining companions, sports partners, travel companions or gaming teammates.

“AI is only a tool. People still deliver the end service,” Wang said. For these services, machines handle search and communication, while people provide the experience.

The More Users Chat, the Faster the Business Fails

The first wave of AI social products has already shown that users will pay for “machine companionship.” Sensor Tower data shows that mobile in-app purchase revenue for AI companion apps reached about $150 million in the first quarter of 2026, more than 12 times the level in the same period of 2023.

But the leading players are struggling. Character.AI has about 20 million monthly active users, down from a peak of 28 million in mid-2024. It generated $32.2 million in revenue in 2024. Based on its $9.90 monthly subscription fee, that implies roughly 270,000 paying users and a conversion rate just above 1%. Its valuation has fallen from a peak of $2.5 billion to about $1 billion.

Li Di, founder of Nextie and known as the “father of Xiaoice,” cautions that AI social products should be judged not only by revenue but also by their unit economics.

Companies must guard against a model in which “the more users chat, the faster the business fails.”

Traditional internet products want users to spend more time on their platforms. Another half-hour of scrolling creates more opportunities to sell ads at little additional cost. AI social products are different: every text, voice, image and video interaction consumes inference resources. The more engaged users become, the higher the costs.

If the revenue a user generates over their lifetime cannot cover the cost of computing power, growth may only accelerate the losses. AI is expensive. For companion products, high retention is both a source of value and a cost.

Li therefore divides current payment models into two categories. The first covers high-frequency, low-ticket emotional spending, such as subscriptions and paid story unlocks. These products can persuade users to pay for the first time, but their novelty has a limited shelf life. Once characters become repetitive and relationships stop progressing, users may leave.

The second category consists of higher-priced cognitive services. Instead of merely talking to users, AI helps reduce decision anxiety and offers another way to think about problems, evolving from a “conversation partner” into an “external adviser.”

“Users are unwilling to keep paying for simple small talk, but they will pay a premium for a companion that can reduce decision anxiety and provide a second cognitive perspective,” Li said.

WeChat Builds the Road; Smaller Companies Collect the Tolls

WeChat is building the infrastructure for matchmaking and transactions.

Tencent President Martin Lau outlined the roadmap during an earnings call: users give instructions directly to an agent, which executes the task. In the future, agents representing users, merchants and Mini Programs will communicate with one another and complete transactions automatically. “XiaoWei AI Social” is the first step—enabling agents belonging to friends to talk to one another.

Li offered entrepreneurs a warning: “Do not look for safety within WeChat’s firing range.” WeChat’s relationship graph, identity system and trust built over more than a decade are its hardest advantages to replicate. Connecting to a large-model API and adding a chat interface creates no defensible barrier.

Smaller companies can survive by focusing on specific use cases.

Wang is targeting a single transaction: human gaming-companion services. AI clarifies the timing, requirements and service boundaries in advance, while people deliver the service and the platform takes a cut of each order. The same logic applies elsewhere: “For travel companions, the service could later connect users with hotels, tickets and local services. For dating, it could connect them with restaurants, performances and activities. For gaming companions, the platform could take a commission from each order. The AI itself may not need to charge a fee; the transactions it facilitates can be monetized.”

Another path is to continue building purely AI-based relationships. The real challenge there is not whether the AI can carry a conversation, but whether it can sustain a long-term relationship.

Li has observed that some AI social products generate strong initial excitement and decent short-term retention, only for usage to drop sharply after a month. One reason is that machines lack genuine long-term shared memory. Human friends understand each other better over time, while AI can seem to be meeting the user again from scratch every so often.

The barriers facing the two AI social models are therefore different. Companies combining AI with human services must get human supply, identity, safety, matching and fulfillment right. Those building purely AI relationships must solve long-term memory, personality consistency, retention and computing power costs.

What matters most about WeChat’s test is not whether “two AI agents can talk to each other,” but that AI is beginning to enter the intermediary stages of real human relationships.

The first wave of AI social products sold users an always-online virtual friend. In the next wave, AI may select people, clarify requirements and schedule meetings on users’ behalf—and make money from a real date, an hour of gaming companionship or an offline purchase.

Wang sees an opportunity to improve the efficiency of matching people with one another. Li is focused on whether an AI relationship can develop a healthy revenue and cost structure.

Both paths ultimately must answer the same question: what outcome are users actually willing to pay for?