Back in 1951, Toyota launched an experiment called the “Creative Idea Suggestion System.” Every assembly line worker was asked to submit improvements they had discovered on the job, and the workers complied.

Over more than half a century, Toyota received roughly 50 million improvement suggestions. At its peak, each employee submitted nearly 50 per year, and in recent years the number has stabilized at around 14 per person annually. This is how the Toyota Production System became a textbook case for global manufacturing.

The reason workers were willing to hand over their know-how was simple: Toyota promised lifetime employment. Your experience made the company stronger, and the company guaranteed your position. Trust came first; knowledge followed.

In 2026, some Chinese companies are demanding something remarkably similar from their employees: hand your work experience over to AI. This includes, but is not limited to, organizing knowledge bases, writing systematic prompts and Skills, inputting decision-making logic, and validating AI outputs.

When Ford introduced the assembly line, workers were moved from scattered workshops to the conveyor belt. Work was broken into finer units, efficiency rose, but no one was asked to write down their craft as a manual and hand it to the machine.

The AI transformation is fundamentally different. It requires every individual to voluntarily add an extra layer of labor, preparing the context AI needs—yet the rewards and the assumption of risk remain uncertain.

Toyota made a promise. Most companies aggressively pushing AI transformation have given nothing at all.

In July this year, the Forbes Technology Council published an article titled The Hidden Context Tax That's Killing Your Enterprise AI Agents, which introduced a concept right in the headline: the context tax.

The research argues that efficiency losses and accuracy declines in enterprise AI systems, caused by insufficient or poorly managed context, constitute a hidden tax burden.

This concept can be pushed one step further: the context tax is, first and foremost, an organizational problem.

To make AI smarter, companies need employees to invest extra effort beyond their core duties to organize, submit, and maintain context. Yet this extra labor is entirely imbalanced across four dimensions: rights, responsibilities, risks, and rewards.

MIT's 2025 report, The GenAI Divide, shows that about 95% of enterprise generative AI pilots fail to produce measurable business returns, and only about 5% reach production. The main obstacles include fragile workflows, insufficient in-context learning, and disconnection from daily operations.

The root cause is not technological; it is that the restructuring of organizational relationships has not kept pace.

Behind that 95% figure lies precisely an uncalculated context tax.

Where the Corporate Anxiety Begins

Companies, as the leviers of the context tax, do have their own anxieties and dilemmas.

For many internet companies in particular, the motivation to push AI transformation is real, multi-layered, and often urgent.

At the most superficial level, there is competitive anxiety. When peers announce that everyone is embracing AI, when shareholders or investors ask during due diligence, “What is your AI strategy?”, and when customers begin comparing prices based on AI product efficiency, not doing AI means losing position in the next round of competition.

By 2026, this anxiety has spread to nearly every industry.

A more tangible motive is solving real problems. Talent attrition in key positions remains persistently high. When core employees leave, both experience and client relationships disappear at once. If that experience can be deposited into the system, the organization's resilience rises by a notch.

For highly repetitive, rule-based work, AI can indeed free up human labor for more judgment-heavy tasks.

There is another motive that is rarely discussed openly: management needs to prove its strategic vision through AI transformation. AI is the dominant narrative of the moment. Pushing early and pushing fast means occupying the cognitive high ground in front of the board and shareholders.

This motive is not necessarily bad, but it makes the pace more aggressive, lowers tolerance for implementation failures, and leaves employees less time to adapt.

These motives stacked together constitute the rationale for aggressively pushing AI transformation—but a rational motive does not automatically resolve the imbalance of rights in how it is implemented.

From the perspective of organizational efficiency, the company sees “I need this experience in the system.” From the perspective of personal interest, the employee sees “What happens to me once the system learns my experience?” Both views are valid on their own terms, but conflict inevitably erupts in the space between them.

Those Who Cannot Hand It Over and Those Who Will Not

Enterprise context comes in two forms. The first is system data: approval records, CRM fields, ERP vouchers. These can be integrated via APIs and connectors—technically, mature solutions exist.

The second is human experience: why a client turned hostile, where in the accounts the problem usually lies, what exactly the third clause of that contract is guarding against. This knowledge exists in no database; only the practitioner knows it.

AI needs the second type of context to actually do the work. If an agent cannot see industry know-how, it can only carry things; it cannot make decisions.

Once enterprise AI transformation reaches deep water, the obstacle always lands in the same place: how do you turn the experience inside a person's head into input AI can use?

Those who cannot hand it over fall into at least three categories.

The first is experience that is inherently fragmented and intuitive. A master with twenty years on the job knows a machine sounds wrong, but cannot say exactly what is wrong—it is a bodily perception, naturally impossible to encode. Organizational theory distinguishes explicit knowledge from tacit knowledge, and much core experience belongs to the latter.

The second is experience that occurs offline. Factory floors, storefronts, face-to-face client negotiations, safety judgments on construction sites—context in these scenarios does not automatically become digital signals. Asking a salesperson to log the negotiation strategy of every client visit into a system is asking them to do the job twice.

The third is that the AI user's capability itself is weak. Some people simply cannot write prompts, cannot translate their experience into structured input AI can understand. When the tool is used poorly, what gets organized is unreadable to AI. Companies can force AI training to address this, but training solves willingness; the capability gap requires a longer period of adaptation.

More troublesome than those who cannot hand it over are those who will not. The reasons for refusal come in two layers, and each points to a failure of institutional design.

The first layer is fear. Once context is handed over, one becomes replaceable. And the speed at which most people expand their abilities to adapt to the new environment is far slower than the speed at which they hand over context.

Wang Qian, a professor at the Shanghai University of Political Science and Law, noted in an analysis this May that companies, by asking employees to “document workflows” and “conduct knowledge management,” are in effect extracting work experience, decision-making logic, and communication tactics to form standardized, reusable skill modules.

In this process, companies gain at least four benefits: knowledge retention, efficiency gains, reduced dependence on individuals, and the creation of proprietary digital assets—but workers' contributions come with no corresponding rights protection.

A FAccT paper co-authored by Imperial College London and Microsoft Research provides an academic framework for this phenomenon. The paper introduces a concept: knowledge extractivism.

The research draws on “accumulation by dispossession” from data colonialism theory to describe the process by which enterprise AI systems extract tacit employee knowledge under conditions of power asymmetry.

The paper even notes that AI can bypass what employees voluntarily submit and infer tacit knowledge directly from behavioral patterns in system interactions—such as who collaborates with whom, and who made which decision at which node. Employees themselves do not know this information is being recorded.

This is an asymmetric transformation characterized by obligation without rights. An employee spends two hours entering a decade of client negotiation strategy into the system; once the system learns it, every colleague can use it. But the problem is that after doing all this, the person who entered it gets nothing—their irreplaceability even declines. The next time, they will submit the bare minimum.

The second layer is the absence of incentives and safeguards. Enterprise AI transformation is almost always top-down: management announces the strategy, middle managers communicate execution requirements, and frontline staff are required to cooperate.

The way to cooperate is by paying the context tax: organizing data, contributing skills, testing and running through workflows, validating outputs, and maintaining knowledge bases. Doing it right earns no extra reward. Doing it wrong—if an AI hallucination causes an incident—means bearing the responsibility.

This April, the enterprise AI company Writer, in collaboration with Workplace Intelligence, released a survey covering 2,400 knowledge workers in the US and Europe. The data is striking: 29% of surveyed employees admitted to sabotaging their company's AI strategy, including using unapproved AI tools privately and refusing to use company-designated AI products. Among those born after 1997, that figure rises to 44%.

Meanwhile, 60% of executives said they plan to lay off employees who cannot or will not use AI.

This contrast itself shows that, amidst the aggressive push of AI adoption, trust is collapsing. Those young employees quietly sabotaging projects are not so much resisting technology as resisting a work arrangement that demands obligation without positive feedback.

ActivTrak, an employee behavior analytics platform, tracked digital work behavior across over a thousand companies and 443 million hours from 2023 to 2025. Its conclusion: after AI implementation, employee workloads did not decrease—they increased. Weekend overtime went up, collaboration time rose 34%, and multitasking time rose 12%.

Feedback from employees at big tech firms is even more direct: companies forcing AI has actually increased work pressure, leaving employees to clean up after AI's outputs. The context tax plus the error-correction tax means total tax burden has not fallen—it has risen.

Paths to Reducing the Tax Are Emerging

Having recognized the problem, some pioneers in the industry have already been exploring several differentiated paths to lower the tax burden.

The first path is to make system data integration as thorough as possible, reducing the burden of manual data transfer on employees. DingTalk has completed CLI-based transformation of over a thousand core capabilities. Feishu supports native MCP protocol access. WorkBuddy uses connectors to bridge structured data scattered across systems.

This path addresses the tax on system context. The effect is direct. As the Feishu product lead put it: no need to manually export PDFs and upload them, no need to organize context. But it cannot eliminate the tax on human brain experience.

The second path is to embed context collection into natural workflows, so employees do not need to do anything extra. Feishu's Doubao Assistant automatically extracts daily work memory and accumulates it as a long-term asset. DingTalk's Wukong inherits enterprise permissions within a secure sandbox, and agents automatically read approval chains and collaboration records.

Even more notable is Anthropic's product design. Its enterprise offering, a digital workspace-type product called Claude Tag, deploys AI directly into Slack channels. When conversation itself becomes context, this path transforms the explicit context tax into implicit background collection.

The efficiency problem is solved, but the ethical problem remains. Painless collection means employees cannot perceive that they are being extracted from. The system continuously accumulates organizational knowledge in the background, and employees do not know what is collected or how it is used.

The Imperial College and Microsoft paper cited earlier calls this mechanism “capture of knowledge-in-use,” noting it creates stickiness and a strategic moat for the platform. As the producers of knowledge, employees have zero bargaining power in this process.

The third path has just begun to be explored by a few companies: building positive incentive mechanisms. If the experience employees contribute can be tracked and quantified, and contribution quality is tied to performance reviews or compensation, the context tax shifts from a hidden obligation to a priced exchange.

The WorkBuddy open platform's Expert Center allows ecosystem partners to package industry experience into AI-invocable expert roles—but how value flows back to the individuals who contributed the experience remains unclear.

WPS Lingxi's model is to let AI help users do new things, rather than doing what users already do. As Zhang Qingyuan emphasizes, work is not documents; AI should move from delivering content to delivering outcomes.

This approach is another way to reduce the context tax. If AI opens up new capability space for employees, they no longer see contributing experience as a threat.

In the solutions section of the Imperial College and Microsoft paper, there is a judgment that serves as a calibration line for all three paths: when developing systems, one must consider a fundamental question—is the system being developed to help users expand the boundaries of their capabilities, or to replace users' current work?

The latter creates strong incentives to dispossess employees of their knowledge. Research shows that the technologies that most improve productivity are those that help users do new things.

The Surging Enclosure of Knowledge

In 18th-century England, the enclosure movement saw landlords fence off land that had been common use. Tenant farmers lost the means of production they depended on and were forced into factories as wage laborers.

Before that, a tenant's land was everything. Losing the land meant losing all bargaining power.

The core means of production for white-collar and knowledge workers are precisely experience, judgment, industry cognition, client relationships, negotiation strategies, and intuition about problems. These things cannot be replicated—and precisely because they cannot be replicated, knowledge workers have bargaining power. A ten-year sales veteran's irreplaceability often lies in knowing when to make the call, whom to call, and what to say.

When companies, in the name of AI transformation, demand that employees organize this experience into knowledge bases, encode it into skill modules, and train it into reusable AI models, what is happening is structurally identical to the enclosure movement: the means of production are transferring from the individual to the organization.

In the classic labor-capital relationship, the enterprise acquires the employee's labor time and pays wages. Exploitation occurs in the gap between the product of labor and the compensation paid. The existence of the context tax brings a new change that even political economy has not yet touched—

What the enterprise covets is not just labor time and labor output, but the very capacity that enables the worker to produce that output in the first place.

Once a salesperson's negotiation strategy is learned by AI, that strategy continues to produce value, but the subject producing that value shifts from the person to the system. The irreplaceability the employee once held through accumulated experience is gradually eroded as the experience becomes systematized.

This is the essential difference between the context tax and ordinary workload. Overtime consumes a person's time. The context tax consumes a person's moat.

More critically, this process unfolds under severe information asymmetry. Employees typically do not know how the experience they submit will be used, how long it will be retained, who will call upon it, or to what extent it will reduce their irreplaceability. The aforementioned “capture of knowledge-in-use” mechanism does not even require active submission—the system infers tacit knowledge from interaction behavior, completely imperceptible to the employee.

No right to know. No bargaining power. No profit-sharing mechanism. No protection of irreplaceability.

The situation of employees facing this enclosure of knowledge is no different in essence from that of tenant farmers losing their land: they are being asked to hand over their most valuable asset, and the terms of return are blank.

Under current labor law, there are clear rules for works made for hire and service inventions. But “training a decade of a person's negotiation experience into an AI-reusable skill library or model” falls outside the range of any existing legal provision.

Professor Wang Qian is already calling for discussion on how to judge the legality of such behavior within the current legal framework. The World Economic Forum's Future of Jobs Report 2025 projects that by 2030, roughly 92 million jobs globally will be displaced by technological change, green transition, and other factors—of which approximately 9 million are attributed specifically to AI and information processing technologies.

But for those currently being asked to hand over their experience, distinguishing among causes of displacement does little to ease the anxiety. Whether legislation can catch up with the pace of enclosure remains uncertain.

In 1951, Toyota workers were willing to submit improvement suggestions because the company answered, with lifetime employment, the question: “if you hand over your experience, what do you get in return?”

In 2026, the vast majority of companies aggressively pushing AI transformation have failed to answer this question. 44% of young employees are quietly sabotaging AI projects. 60% of executives say they will fire those who refuse to use AI. But no executive has said how they will protect those who sincerely teach AI.

The 95% AI pilot failure rate, in the final analysis, is a 95% trust gap.

The fences of the knowledge enclosure movement are already up. But those inside have not yet been told what their land will buy them.