AI in the Factory: Redemption, Rejection, Panic, and an Organizational Efficiency Showdown

The rapid advancement of AI reshaping white-collar work is no longer news. Over the past few years, tech companies have consistently promised society that AI will create unprecedented productivity value.

It was widely believed that this would unfold over a decade-long cycle. However, over the past few months, this promise is being fulfilled at a pace that caught us off guard.

AI is spreading like a contagious virus across various organizations, violently reacting with different job roles and collaborative relationships.

If we view an organization as a living entity, infections of varying intensities are simultaneously occurring in some large enterprises, medium-sized companies, and small teams of just a few people—some are being reshaped, some organizations are rejecting it, and of course, some have not yet realized that the great infection has already begun.

We contacted and interviewed several companies and found that the way they react to AI is rapidly affecting an organization's efficiency. This reaction also inevitably ripples through the workforce.

'It's like the Ice Age is coming, and all the giant animals have to transform themselves to survive.'

01

AI Can Write Code That Technical Teams Can't

'AI has been even crazier than imagined over the past month or so,' said Wei Jiaxing, CEO of CloudBat Intelligence, reflecting on the organizational transformation brought about by their internal use of AI over the past few months and the resulting chain reactions, unable to hide his astonishment.

On the first day back to work after the Chinese New Year this year, this intelligent voice outreach technology service provider with a hundred employees and an eight-year history held an organizational mobilization meeting. In the intelligent voice outreach industry, CloudBat Intelligence has been quick to embrace AI, being one of the first in China to iterate intelligent calling products based on large models and quickly following up with each model upgrade.

However, everything has undergone a qualitative change in recent months.

Starting in September last year, CloudBat Intelligence's R&D department migrated from Copilot and Cursor to a top-tier model. In November, the CTO used AI to develop an automatic noise reduction model for different voices—not only was it developed, but it was also launched, capable of handling high concurrency processing. Recalling this process, Wei Jiaxing's tone carried a shock that has not fully dissipated: 'Our existing technical team couldn't have written this, but it did.'

This narrative is different from the familiar 'AI efficiency boost' story of the past two years. The effectiveness of past AI tools like Copilot, Cursor, and other AI programming tools was generally consistent: helping teams improve efficiency by 20% to 40%. However, since the end of last year, Wei Jiaxing feels that technological progress has reached another level, 'If the effectiveness has improved by two to three times, it can't just be called an efficiency boost.'

CloudBat then overhauled its entire R&D workflow, mandating the use of top-tier AI tools in the technical department and initiating a new round of hiring before the Chinese New Year to address previous capacity shortages. The 30 new employees, regardless of their department, are all vibe coding enthusiasts. 'We haven't laid off anyone; instead, we've added new blood,' Wei Jiaxing said.

CloudBat also designed a reward strategy to promote AI applications—after products collaborating with AI are launched or applied within the organization, the company pays the feature designers. Each requirement is priced according to its difficulty and importance. Wei Jiaxing believes that reimbursing or rewarding based on token consumption does not reflect much real value creation, while tracking how much can be launched using AI is more valuable.

This mechanism operates simultaneously in four roles: R&D, product, UI, and testing. Since frontline employees have generally seen salary increases, the entire team is highly motivated. 'We spent the first two weeks of March exploring and integrating AI, but the last two weeks saw a significant acceleration, resulting in an average per-capita efficiency increase of about 40% compared to the same period last year,' Wei Jiaxing said.

CloudBat Intelligence's experiment is not an isolated case. Remio, an AI startup founded less than two years ago with fewer than 20 employees, has also experienced a sudden shift of AI from an auxiliary tool to a true productivity partner in the past six months.

The head of R&D at Remio told DigitInt that before the end of November last year, AI played more of an auxiliary role in this startup, such as serving as a personal efficiency-boosting tool for document organization and code snippet completion. With a wave of foundational model capability upgrades from model vendors at the end of last year, Vibe Coding began to become popular on a large scale within internal teams from R&D to marketing after November.

A specific case is that when Remio was developing multi-end synchronization functionality, it needed to cover a large number of boundary conditions, for which the team prepared over 500 test cases—a significant portion of which were autonomously completed by AI after the team had left work. During the Chinese New Year period, as Remio transitioned from an AI knowledge base to an Agent version, founder Wang Yuan alone built the entire agent framework from scratch using vibe coding.

Before founding Remio, Wang Yuan was the Vice President of NetEase Group and the Executive Dean of NetEase Hangzhou Research Institute. In 2024, he began to dive into the wave of large-scale AI entrepreneurship. In April last year, Remio, after more than half a year of entrepreneurship, officially launched its AI knowledge base product. Recently, the entire product has completed an upgrade from a knowledge base to an Agent version. The Remio team stated that AI has significantly accelerated the product's R&D and market launch process.

As a startup team, both CloudBat Intelligence and Remio emphasize controlling token costs. At CloudBat, Wei Jiaxing mentioned that they control the monthly AI investment for their 40-person team at 20,000 RMB through account sharing combined with a tiered usage strategy.

Remio's reimbursement policy is 400 RMB per person per month, with the company covering 90% of any excess, and the remaining 10% borne by the employee. This is because if full reimbursement were provided, some employees might 'keep it running in a loop,' burning tens of thousands of dollars a month with output and consumption potentially completely out of proportion.

02

Big Companies Require CEO Approval for Intern and Outsourcing Hires

Actually, it's not just these AI startups that are actively embracing AI; token budgets are becoming a new organizational consensus in more and more organizations. Large tech companies are enthusiastic practitioners of this plan.

Earlier this year, NVIDIA CEO Jensen Huang predicted at the 2026 GTC conference, 'In the future, every engineer will need an annual token budget. Their salary might be hundreds of thousands of dollars, and I would give them an additional token allocation equivalent to half their salary to achieve a tenfold efficiency boost.'

After the Chinese New Year, leading domestic internet giants also began incorporating AI tools and token budgets into their employee benefits systems.

Alibaba, Tencent, and ByteDance have all announced policies or plans related to token budgets. For example, Alibaba's Taotian Group released an AI productivity plan, providing all employees (including interns) with free access to internal paid AI tools like the Wukong and Qoder series, covering technology R&D and general office scenarios. At the same time, employees can apply for full reimbursement of fees for purchasing internal and external development tools.

Tencent has also been reported to equip employees with token packages, with some reports even claiming an exaggerated coverage standard of approximately 220,000 RMB per person per year. DigitInt learned that Tencent employees have a token reimbursement quota of 1,000 USD per person per month and 100,000 points for Tencent's AI enterprise office product, WorkBuddy.

ByteDance announced a new incentive policy earlier this month. In addition to providing a large number of AI tools for work scenarios without specific limits, it also offers employees a reimbursement quota of 50% of the total price for learning and experiencing excellent software and hardware products in their spare time, with an upper limit of 1,000 USD for product/technology teams and 300 USD per year for other departments.

Beyond these generous token budgets, many medium and large tech companies have also accelerated the promotion of various AI tools within their organizations this year.

For example, Xiaohongshu has been vigorously promoting the use of 'Shrimp' (an internal development version of an Openclaw-like tool) among all employees since the beginning of the year. Some employees reported being trained on how to use 'Shrimp' in the morning and finding it already installed on their office computers upon returning. If not opened and used, 'Shrimp' even actively reminds employees to use it.

The workflows involved in these companies cover multiple scenarios and links, including R&D, design, recruitment, data analysis, and marketing. Some large companies have even specifically mentioned the AI contribution rate to code in recent product update announcements. For example, when ByteDance's TRAE SOLO was launched, it was specifically mentioned that the AI code contribution rate in the development of the Solo standalone version reached 93%, with 1 million lines of code and over 9,000 submissions completed by AI.

To overcome organizational inertia and accelerate AI implementation, large companies have taken some measures in their processes and rules, such as lowering the barrier to AI application and increasing the difficulty of processes for expanding new personnel.

An employee from a financial business line of a tech giant told DigitInt that they now need CEO approval to hire an outsourcing worker, forming a stark contrast with the company's strong encouragement for employees to use AI to complete business processes and the generous token budgets.

'When encountering a problem, your first reaction should be to use AI to solve it, rather than asking other departments or seeking other resources,' said a company CEO, expressing his hope to change employees' mindsets and collaboration methods to fully leverage AI's value.

DigitInt also learned that some large companies have set OKR indicators, requiring a certain proportion of work to be completed in collaboration with AI. However, some frontline employees reported that due to relatively vague target settings, it is unclear how the goals will ultimately be implemented. 'How to quantify and assess the specific proportion is difficult to determine with clear data,' they said.

Although some point out that large organizations have multiple reporting relationships, including leadership potentially unfamiliar with the front lines of business, they may lag in perceiving the effectiveness and flexibility of AI use compared to flatter teams, it cannot be denied that large companies also have their inherent advantages. These organizations have abundant resources, including very sufficient token budgets, and they also have the capability to encapsulate tools. Once 'the elephant turns,' they may sometimes adopt cutting-edge tools faster than smaller companies due to their resource advantages.

For example, while many companies had not yet large-scale (massively) adopted AI empowerment last year, large companies like ByteDance and Tencent already had many encapsulated Agent tools for internal use.

03

Redemption, Rejection, and Panic

AI is rapidly spreading like a major infection across various scenarios within organizations, including code, data analysis, marketing strategies, design, and recruitment.

However, this process is not entirely a story of triumphant progress. From the perspectives of market competition, technological advancement, and individual survival, you can see that companies and workers each have their own narrative frameworks.

Some have felt the benefits brought by technological progress. Wei Jiaxing once interviewed a candidate who had previously worked in an office at a civil engineering company. Over the past two years, he had been learning AI programming, using his phone to browse documents and run code on construction sites, with no one around understanding what he was doing.

As a vibe coding enthusiast, he received an opportunity to join an AI agent company. 'He might have been out of place in his previous organization, but this new opportunity feels like redemption,' Wei Jiaxing recalled the interview, believing that people who embrace AI and organizations that embrace AI will attract each other. At the same time, different attitudes toward AI within organizations may also lead to rejection. Those who embrace AI cannot stay in conservative organizations, while those who can use AI will receive higher market rewards.

Immediate changes have already occurred in the employment and recruitment markets. In many companies we contacted, AI proficiency has become a common requirement for candidates.

The Paper previously reviewed 10,221 job postings from five internet companies, including Tencent and ByteDance, for the 2026 spring recruitment season. After analysis, it was found that 47% of the positions had AI requirements. PwC's '2025 Global AI Jobs Barometer,' released in June 2025, clearly states, 'In 2024, the average wage premium for workers with AI skills was 56%, double the 25% of the previous year.'

However, inevitably, attitudes toward AI capabilities vary among different job roles, AI tools used, and even individual positions.

In the brand marketing team of a medium-sized tech company, a senior marketer told DigitInt that, driven by the boss, they are currently frantically seeking scenarios where AI can be implemented. Many scenarios cannot avoid the suspicion of 'AI for the sake of AI,' often failing to deliver content that truly meets requirements, yet the boss frequently questions why the results fall short of expectations.

For example, some copywriting involves complex situational requirements and demand alignment, while AI lacks the relevant knowledge background and material accumulation. Even with complex prompt requirements, the final results often fail to meet the scene's requirements. However, the technically inclined boss often has a catchphrase, 'If it's not used well, it's not CC's (Claude Code) problem; it's a human problem,' which troubles this senior marketer.

Wei Jiaxing from CloudBat Intelligence also lamented, 'If it's not used well, it's not AI's problem; it's a human problem.' He found that in R&D code scenarios, AI generates significant productivity value, and CTOs with a full-stack mindset who can think about product functionality from a holistic perspective achieve far better results with top-tier models than average developers. However, those who cannot precisely position products and requirements and cannot clearly articulate what they want may struggle to produce good results even with AI.

Silicon Valley expert Andrej Karpathy recently explained the widening gap in people's understanding of AI capabilities, attributing it to two things: timeliness and usage hierarchy. Many people form their entire judgment of AI based on the free version of ChatGPT. Even if they subscribe for 200 USD per month, if they do not use it deeply in scenarios like programming, mathematics, or research, the impact remains limited. This is because reinforcement learning works best in areas with verifiable answers, like programming, making it naturally easier to optimize. Meanwhile, search, writing, and advice—the scenarios most commonly used by ordinary people—are not the priority optimization directions for tech companies. 'The gold mine is elsewhere, so attention follows,' Karpathy said.

Wang Yuan from Remio admitted that currently, in many scenarios within enterprises, human intervention remains indispensable in key decision-making and quality control links for Agents. For example, in marketing scenarios, business teams find that AI can complete the workflow of automatic placement, but whether it truly meets real-time business requirements still needs to be verified by professionals.

Regardless of whether AI is effective in all scenarios, a major infection is already underway.

A 'shrimp farming enthusiast' from a large company once mentioned his dilemma to DigitInt. During the craze for 'lobsters,' every time he got a business process running and precipitate ( precipitate here means 'accumulated' or 'documented') a skill, he hesitated whether to report it upward because the skill might represent the work content of his colleagues.

Corresponding to this large company worker's dilemma, two types of skills have been simultaneously popular on GitHub in the past two weeks. 'Colleague.skill' generates AI skill plugins by analyzing employee work records, digitally preserving the work capabilities of departing employees and teaching AI to take over certain colleagues' work. In contrast, someone has introduced 'anti-distillation Skill,' teaching people how to systematically prevent their experience from being distilled by AI.

To a certain extent, this is also a true portrayal of the organizational efficiency anxiety and individual survival anxiety in our time.

04

Disappearance, Integration, and the Competition for Organizational Efficiency

There are also signs of more long-term changes on the horizon.

As AI becomes capable of handling very basic and highly repetitive tasks, the demand for entry-level recruitment is inevitably disappearing. 'AI can be used for administrative invoice reimbursement and personnel screening and hiring processes, so companies may no longer need to recruit interns,' mentioned a business leader.

Of course, some argue that from the perspective of workers, the disappearance of these repetitive tasks is not necessarily a bad thing.

A head of the marketing department at a company told Shuzhi Qianxian that in the past, it was difficult for her to recruit people to post articles on various platforms because this role offered limited growth opportunities, yet it was essential for the company to invest in and maintain a certain standard. The recruited individuals often left due to the lack of career development prospects. Now that AI can handle this part of the work, those who might have languished in these roles now have the opportunity to move into more challenging positions.

The vision of people engaging in more creative work is becoming a more rigid requirement in the job market. But how can an inexperienced entry-level individual progress to a point where they can engage in creative work with professional taste? There is currently no specific solution, as it relies on the collective adjustment and change of the job market, including educational institutions and enterprises.

At the same time, in the job market, roles are also 'integrating,' and the division of labor within various departments is no longer as finely detailed.

Wang Yuan, the founder of Remio, gave an example: the marketing department may no longer have dedicated roles for ad placement, as everyone can collaborate with AI to handle this aspect. In the R&D department, some companies no longer have product managers in the traditional sense; everyone serves as both a product manager and an engineer. Collaboration is already undergoing changes. Workflows that previously required the collaboration of product managers, front-end engineers, back-end engineers, and testing engineers are now being completed by a single individual from requirement to launch. Moreover, Vibe Coding has reduced the cost of execution, while judgment, systems thinking, and product intuition have become even more valuable than before.

In addition to changes in the job market and talent profiles, different organizations are experiencing varying levels of efficiency in this transformation.

The CEO of a startup believes that some top-tier coding tools are sensitive to project complexity. They perform best on tasks with clear boundaries and independent logic—such as an algorithmic function or a localized business closed loop (closed loop). However, their effectiveness drops sharply when tasked with handling a complete integrated business and financial system, especially one with a decade of accumulated legacy code and unclear module coupling.

In his view, this characteristic poses a structural disadvantage for large companies. As large companies' product systems become more extensive and outdated, with more legacy issues, the space for AI to make an impact diminishes. In contrast, small companies, especially those building products from scratch with AI, see more significant efficiency gains.

Furthermore, large companies inherently have higher labor costs. Small companies can leverage individuals earning less than 20,000 RMB per month to accomplish tasks that previously required individuals earning 50,000-60,000 RMB. This makes the cost structure of small companies increasingly competitive for the same output. 'In the future, there may be fewer large factories with tens of thousands of employees, and teams of a few hundred people may grow into giant companies.'

This CEO told Shuzhi Qianxian that what they are now wary of is not just competitors of similar size but OPC-type organizations. These are teams of around ten people, with three or four in R&D and two or three in sales, serving a highly niche local market. They are more agile than startups with a hundred people, and AI provides them with the same tools while imposing a much smaller operational burden.

Of course, there could be another scenario. Large companies, leveraging their systematic resource advantages, could use AI to cover niche long-tail markets that were previously unserviceable due to high labor costs. After both large organizations and small enterprises acquire these 'nuclear weapons,' the pressure will shift between different organizations.

A reshuffling of organizational efficiency has already begun.

Some organizations have transformed amid this 'infection,' while others are still rejecting it, and some companies may not even realize that the 'infection' has already started—but these different stages could determine the fate of various organizations in the next year or two.