中国白领自掏腰包为AI工具付费，企业报销仍是例外

NextFin News — A growing share of China’s white-collar workforce is running two payrolls. One is the salary that hits the bank account. The other is the monthly outlay for model memberships and API credits that make the first job feasible.

Survey data put generative-AI users in China above 700 million by mid-2026, with roughly a third saying they use the tools for work. Lifestyle research this year found more than a third of respondents paying for large-model access and nearly three in ten buying dedicated AI office products. Free tiers run out. Quotas tighten. The invoice often lands on the individual.

Outside a handful of large technology firms, corporate reimbursement remains the exception. Even inside some big platforms, internal free access has shifted toward controlled allowances, and the shortfall is covered from personal pockets. Students and early-career researchers have no corporate buffer at all.

Purchase channels have adjusted accordingly. Telecom operators have sold token packages against phone bills. E-commerce platforms have opened token recharge counters next to mobile top-ups. During major shopping festivals, model credits have appeared in the same promotional halls as ordinary consumer goods. Usage is no longer a curiosity expense. It is a recurring household line.

Interviews and sample studies point to four clusters of tasks: drafting and slide work, meeting notes and translation, data cleaning and light coding, and specialized work such as contract review or visual production. The common thread is volume and repetition. People pay to shrink the part of the day that used to be pure grind.

The returns differ. A journalist described cutting active writing time roughly in half for a few hundred yuan a month and treating the reclaimed hours as rest. A fund manager spending around 3,000 yuan a month cares mainly about first-pass success on research code, arguing that breadth of coverage is otherwise impossible. A high-school teacher funds multiple frontier models to build and mine exam databases, framing the cost as professional duty rather than personal luxury. Doctoral researchers time jobs for off-peak API prices and still watch weekly bills spike when a project runs long; side gigs sometimes finance the tools that substitute for a research assistant.

Global consumer spending on AI has climbed far faster than the user base, according to venture surveys, which implies rising intensity among those already hooked. In China, the share of users on monthly paid plans has edged up while the share of very low monthly spenders has edged down.

Several forces thicken the invoice. Different models still win on different jobs, so serious users stack subscriptions. Prompt and workflow tuning burns credits before useful output appears. Developers widely report that near-miss answers—almost right, not quite—are more expensive than clear failures because they invite endless revision.

Pricing design adds opacity. Caps that once felt unlimited get folded into shared pools. Peak-hour multipliers appear. Products migrate from seat-based or “unlimited” marketing toward pure token metering. Comparable workloads can cost an order of magnitude more under usage billing than under a flat subscription. Official cost guidance from major overseas coding tools has already been revised upward, with the explanation that models improved and people simply used them more.

Ride-hailing shows a fare estimate before the trip starts. Online retail shows a checkout total. AI work still often reveals its true cost only after the tokens are gone. Until employers treat model access as standard productivity infrastructure—or until vendors make spend as legible as a taxi quote—workers will keep subsidizing the efficiency their jobs now quietly assume.

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App