Why Are Companion AIs Being Pulled En Masse? The Triple Dilemma and a Path Forward for Practitioners

A Wave of Removals Exposes the Companion AI Dilemma

Last week, several major players—Doubao, Tongyi Qianwen, Yuanbao, and others—simultaneously pulled a batch of consumer-facing agents. On the surface, this was a response to a regulatory notice from the Cyberspace Administration—compliance management targeting domestic character-based, virtual, and interactive agents. But if you break down the logic behind this wave of removals, you'll find a signal that practitioners should take seriously.

The agents that got pulled were almost uniformly companion-type, role-play, and interactive products; meanwhile, work assistant and office assistant agents were left fully intact. Doubao even went in the opposite direction, heavily promoting a "Work Edition" with monthly subscriptions running as high as several hundred yuan.

This is no coincidence—it's a clear vote by the major players on market value: companion agents, if built as pure software, may have no independent market space at all.

Behind this judgment lie both regulatory pressure and the cold filtering of business logic. From the regulatory perspective, the generative AI governance framework continuously advanced by the Cyberspace Administration of China—including the "Interim Measures for the Management of Generative AI Services" that officially took effect in August 2023—already requires security assessments and algorithm filings for AI services with "public opinion attributes or social mobilization capabilities." These measures are among the world's earliest systematic legislative frameworks specifically targeting generative AI, and their core logic is to manage AI services in tiers according to "social influence": any AI interaction that could affect users' cognition, emotions, or behavior must go through stricter filing and review processes. Role-play and emotional companion agents, because they involve sensitive interaction patterns like identity simulation and emotional inducement, are regarded by regulators as carrying higher risk of psychological impact on users—especially concerns about the potential effect on minors. Research shows that adolescents who engage in long-term interaction with anthropomorphic AI are more prone to avoidance tendencies toward real-life interpersonal relationships. By contrast, work assistant agents have clear interaction goals and verifiable outputs, so their compliance risk is relatively controllable—this is precisely the regulatory logic behind the "keep productivity, remove companionship" pattern in the major players' removal decisions.

The Core Contradiction of Companion AI: High Consumption, Zero Willingness to Pay

Users Are Happy to Chat, but Unwilling to Pay

Companion agents face an almost unsolvable commercial dilemma.

On one hand, users chat with virtual partners and AI avatars at a high frequency, which means the better experience comes with continuously rising compute costs. On the other hand, while users are happy to interact frequently, they are extremely reluctant to pay for it.

The reason is simple: users have no "mental account" for companion products. Whether it's a virtual pet, emotional companionship, or role-play, this kind of need has never carried a clear price tag in users' consumption awareness, and it lacks any accumulated market education.

Mental Accounting is a core concept proposed by behavioral economist Richard Thaler, and one of the key theoretical contributions that earned him the 2017 Nobel Prize in Economics. The theory holds that people mentally categorize money into different "accounts" and set implicit budget ceilings for different categories of spending—even for the same total amount, a consumer's payment mindset for a "work meal" versus a "luxury meal" is completely different. Users are willing to pay monthly fees for "productivity tools" because such spending can be clearly converted into time saved or income gained, offering a quantifiable ROI; but for "emotional companionship" spending, users' mental account is nearly zero—this kind of need has long been satisfied by free social relationships (WeChat friends, offline friends) and free entertainment content (short videos, novels), and has never developed a payment expectation. This phenomenon is especially pronounced in AI companion products: a user might spend two hours a day chatting with a virtual partner, yet feel a subscription fee of 5 yuan/month is "not worth it," because their emotional companionship mental account balance was zero to begin with—it's not a question of how high the amount is.

By contrast, work assistant agents are far more mature. Users are already accustomed to paying for productivity tools like Codex and Claude Code, and even for domestic equivalents, as long as they work well, the payment threshold isn't high. This is precisely the fundamental difference between a mature and an immature mental account.

Even Leading Companies Can't Escape Losses: The Warning from Character.ai's Data

If the plight of a single product isn't enough to prove the point, the financial data of Character.ai—the world's largest independent AI companion company—serves as an even harsher footnote.

Character.ai was founded in 2021 by Noam Shazeer and Daniel De Freitas, core members of Google's former LaMDA team. Noam Shazeer was one of the co-authors of the Transformer architecture paper "Attention Is All You Need"—his technical credentials are beyond doubt. Character.ai was one of the world's earliest platforms to achieve large-scale multi-character AI dialogue, where users can create any custom character and converse with it; the platform already hosts tens of millions of user-created characters. Its business model relies on subscriptions (Character.ai+, $9.99 per month), offering faster response times, queue-skipping, and exclusive features. Yet despite as many as 20 million monthly active users, its annual revenue is only about $30 million. That works out to an average contribution of just $1.5 per user per year—industry estimates put its paying-user penetration rate below 2%. Such per-user value can hardly support a healthy business model.

Compare this with Midjourney (valued at $10 billion, with annual revenue exceeding $200 million and almost no external funding), another generative AI application. The gap in business efficiency between the two reveals a structural chasm in monetization capability between "entertainment/emotional" and "creative/productivity" AI products. Midjourney's users treat it as a professional creative tool, with clear payment logic; Character.ai's users, meanwhile, position it as free entertainment, making the subscription barrier extremely high. This comparison almost perfectly confirms the real-world mapping of mental accounting theory onto AI product monetization.

Even more telling is the valuation trajectory: two years ago, Character.ai was valued at $1 billion; two years later today, it's still out raising funds, its valuation stagnant, with almost no growth.

When even the absolute industry leader struggles this much, one can imagine the challenges facing small teams and independent developers.

"Death on Day 30": The Retention Dilemma of Companion Products

At several recent industry closed-door sessions and KOL discussions, one line left a deep impression: most companion agents die on day 30.

This line cuts to the heart of the problem with companion products. What companion products compete on is not how dazzling the first few minutes of conversation are, but whether they can stably occupy a core, high-frequency scenario in the user's life. Short-term novelty is easy to manufacture; long-term usage habits are extremely hard to establish.

This phenomenon has deep technical roots at the product level. The core experience of companion AI relies on the context memory capability of large language models, yet the long-term memory mechanisms of current mainstream models remain immature—even frontier models like GPT-4 and Claude suffer from information compression and forgetting issues in their effective context windows during actual use. Historical information outside the conversation window is often lost, causing the AI to "forget" a user's personal preferences and past experiences. This runs directly counter to how human memory works: in real interpersonal relationships, the other person remembers what you said three years ago, your favorite foods, your past sorrows—that feeling of "being remembered" is precisely the core source of emotional connection. When AI can't provide this experience, users clearly perceive a sense of "repetition" and "no growth" once the novelty fades. Some products try to introduce emotional state machines and relationship-level mechanisms to simulate deepening relationships, but these mechanisms are essentially still rule-triggered (e.g., unlocking new content after 100 conversations) and struggle to truly replicate the organic evolution of interpersonal relationships. The essence of the retention problem is that companion products have yet to find a reliable technical path that lets users perceive that "the relationship keeps growing."

Once the novelty fades, if the product hasn't embedded itself into a stable part of the user's daily life, churn quietly sets in. No retention means no payment conversion, and no sustainable revenue. This is the deep reason why companion software struggles to form a commercial closed loop.

Where's the Way Out: Enter Productivity, or Go Software-Hardware Integration

Rapidly Enter Specific Productivity Scenarios

For independent developers building AI agents and AI applications, the core advice is: don't be overly obsessed with abstract concepts like "companionship," "growth," or "walking alongside."

The more pragmatic approach is to quickly enter concrete scenarios—productivity gains, work assistance, high-value information and intelligence services. Even providing just one or two high-value features within a vertical scenario, or steadily delivering useful information, may in today's market environment be more reliable and more promising than building a companion agent.

If You Must Do Companionship, Go Software-Hardware Integration

Companionship isn't entirely off the table, but direction is crucial.

If you really want to pursue companionship, the most promising direction is software-hardware integration—giving users a tangible physical object they can see and touch. Only a physical entity that users see every day and can interact with reliably can truly occupy the user's daily scenarios.

Several pioneers have already validated the feasibility of this direction. Japan's Lovot robot, priced at around $3,000, was developed by GROOVE X, a company founded by a former Nintendo engineer. It simulates emotional attachment through temperature-sensitive skin (which senses human touch and responds), eye tracking, and random behavior algorithms—its design philosophy deliberately avoids "usefulness" and focuses on creating the emotional experience of "being needed," achieving relatively stable user retention and repeat purchases in the Japanese market. Embodied's Moxie robot, meanwhile, focuses on children's companionship and education scenarios, monetizing through a subscription model layered on top of hardware sales, and endowing the product with quantifiable educational value through structured social-skills training curricula—thereby breaking through the payment dilemma of pure companion products. Domestically, several companies that have positioned themselves in humanoid robotics in recent years have also listed home companionship as a mid-term application scenario. The core advantage of physical companionship lies in "occupying physical space"—seeing the physical device every day is itself a point of contact, and the product's presence doesn't depend on the user actively opening an app. From product logic to usage habits to mental anchor, the anchoring effect brought about by physical presence is something pure-app products cannot replicate. This is precisely why people's attachment to a pet they've raised for years far exceeds their attachment to a mobile game character—physical presence itself is a trigger for emotional investment.

Final Thoughts

The collective removal of companion agents by the major players is both a result of tightening regulation and a signal of the market's rational return. The triple dilemma of companion AI—high consumption, low payment, and difficult retention—makes it nearly impossible to survive independently in pure software form.

For practitioners, rather than struggling and burning out in the companion track, it's better to turn attention to productivity scenarios where payment willingness is more mature, or to redefine companion products through software-hardware integration. Recognizing this trap is the smarter choice right now.

Key Takeaways

Key Takeaways