At DingTalk, one of the most revealing management decisions of the year involved no software at all. In early April, according to a resignation letter by product manager Teng Yaxin, the team building Alibaba's workplace AI product was told to stay in the office until midnight. The purpose was to watch when the lights went out at the nearby offices of Feishu, the ByteDance rival known abroad as Lark. Teng's gloss on this is the sharpest line in her 75,000-character account: it was as if lit windows proved the will to fight was still burning, and as if keeping their own lights on could offset the uncertainty in their strategy.
That sentence carries the whole story. The resulting scandal went viral on Chinese social media and ended with founder Chen Hang removed as DingTalk's chief executive. Most commentary has treated it as another episode in China's long-running argument about overwork. It is better read as something more specific. It is a case study in what organisations do when they cannot tell whether their strategy is working, and the AI boom is producing that condition on an industrial scale.
Watching effort instead of results
"Involution" (neijuan) has become the catch-all Chinese term for competition that consumes ever more effort for ever smaller returns. The word is useful, but it can also mislead, because it frames the problem as excess: too many hours, too much rivalry, too much zeal. The DingTalk episode points to a different diagnosis. Firms fall back on inputs when they cannot measure outputs.
A company that knows what its customers want can judge itself on whether customers buy. A company that does not know is left counting what it can see: hours logged, features shipped, meetings held, windows lit. Teng's account of the AI project fits this pattern. The team was driven hard, the product lost users, and the project was eventually scaled back. The effort was real. What was missing was any reliable signal that the effort pointed somewhere.
This is why the AI context matters more than the draft coverage of this story has allowed. Enterprise AI assistants and agents are, as of 2026, a category in which nobody has settled what users will pay for or what "good" looks like. Underlying models are converging in capability. Differentiation is elusive, and switching costs for a workplace AI feature are low compared with the switching costs of the collaboration platform itself. In a market like that, executives cannot easily tell whether they are winning. They can, however, tell whether their rival's lights are on. Fear of falling behind then substitutes for a theory of how to get ahead.
The pattern is not new, and it is not uniquely Chinese. China has run this experiment before: the "thousand group-buying war" of the early 2010s and the bike-sharing battles of 2017 both burned enormous sums on subsidies and speed before most contenders collapsed. Silicon Valley's own history offers the late-1990s rush to capture "eyeballs" before anyone knew how to monetise them. What distinguishes the current round is that the AI race combines genuinely uncertain demand with very high fixed costs and intense state and investor attention. That combination rewards the appearance of momentum.
The tool that measures you
There is a second irony here, and it goes deeper than the one social media enjoyed. The online jokes were about an efficiency product built inefficiently. The more interesting point is that DingTalk is itself an instrument for making effort visible. Its read receipts on task assignments, a feature many Chinese office workers resent, turn attention into data that a manager can inspect. The same logic runs through the KPI and OKR systems that tech firms popularised and that have since spread into state bodies and traditional companies.
Legibility of this kind has a well-known failure mode, usually summarised as Goodhart's law: when a measure becomes a target, it stops being a good measure. Once "responded quickly" and "was visibly online" become the things that are observed, employees rationally optimise for being observed. Teng's observation that organisations are always more agile toward certain people, matters and signals is a precise description of this. Agility flows toward whatever management can see. The midnight vigil was simply the same principle applied to a competitor's building.
This matters for AI specifically. The standard corporate pitch for workplace AI promises to strip away low-value coordination work, the memos, status updates and reporting chains. But if an organisation's incentives reward performed busyness, automating reports does not remove the performance. It lowers the cost of producing it, which tends to produce more of it. An AI agent that drafts ten status updates a day does not make the underlying work more meaningful. Anyone evaluating claims that AI will relieve white-collar drudgery should ask first what the organisation is actually rewarding.
Would an AI do the job better?
Some commentary on the letter has raised a darker question: if the demands placed on workers are this arbitrary, would an AI agent, awake at all hours and indifferent to absurd instructions, simply replace them? The worry is understandable, and "reduce costs and increase efficiency" has become a ubiquitous corporate slogan. But the logic cuts the other way. The midnight vigil produced no output that any machine could produce faster. Its value, to whatever extent it had any, lay in demonstrating human commitment to a human audience. A task whose purpose is to display sacrifice cannot be automated without losing the point.
The real labour risk is more mundane. Firms under cost pressure may use AI as a justification for headcount cuts while leaving the incentive structure unchanged, so that fewer people perform the same rituals with more tools. Whether that is happening at scale in Chinese tech firms is a question the available reporting does not answer. It would require hiring and layoff data rather than anecdote.
What the letter proves, and what it doesn't
Teng's letter deserves to be taken seriously, and it is fair to acknowledge its strengths. It is detailed, specific and reflective rather than merely aggrieved. Its central question, whether a job hones a person's skills, sustains their health and strengthens their dignity, or leaves them numb and adept only at making do, is a better test of workplace health than most corporate engagement surveys.
It is still one account, and a viral one. Virality selects for the vivid and the resonant, not the representative. Several claims deserve more scrutiny than they have received. Teng reports repeated calls from people claiming to be ByteDance recruiters who probed for sensitive information. That may be accurate, but "claiming to be" is doing a lot of work. The calls have not been attributed to ByteDance, and treating them as evidence of ByteDance's own anxiety goes beyond what the letter establishes. The widely repeated figure that DingTalk shapes the working lives of "hundreds of millions" of people needs a current, sourced user number. The story also does not name the specific AI product or quantify its decline in users, both of which would help separate a failed strategy from an ordinary product flop.
The causal link between the letter and Chen Hang's departure needs the same care. Alibaba's senior leadership criticised the team's high-pressure management and called for a more humane culture. Late Post reported that removing a business-unit chief this way is rare at Alibaba. The sequence is suggestive. Still, a public relations crisis may have been the occasion for a change that the product's performance had already made likely. Those two explanations imply very different things about how much a single employee's voice can move a large company.
Treating the symptom as the disease
The DeepSeek comparison often raised in this debate also needs qualifying. The AI start-up reportedly discourages overtime on the grounds that exhausted people make bad decisions, and it is frequently held up as proof that a calmer culture can win. That may be so. But DeepSeek emerged from, and is financed by, the quantitative hedge fund High-Flyer. It has not had to chase market share or answer to venture investors demanding user growth. Its culture may be less a replicable management choice than a luxury of its capital structure. If so, the lesson is not "be more like DeepSeek" but "involution is partly a financing problem." That is a harder message for firms that depend on growth-hungry investors.
The same logic applies at national level. Beijing has made "anti-involution" an explicit economic priority since 2025, targeting destructive price wars among e-commerce platforms, solar manufacturers and electric-vehicle makers, alongside pledges on labour protection. The 2025 e-commerce subsidy war, in which, according to Late Post, platforms raced to break transaction records partly because each feared rivals were quietly doing the same, shows why. Much of the cost eventually landed on merchants and staff, who faced surging order volumes and customers trained to expect discounts.
But price wars and overwork are symptoms of an economy in which many firms chase too little demand. Where consumers are cautious and good jobs are scarce, firms fight harder over a fixed pie and workers have little leverage to refuse. Regulating conduct without addressing that imbalance is the approach China has already tried. The public backlash against the "996" schedule in 2019 and official statements around 2021 that such hours breach labour law did not end the culture that produced them.
There is also an unresolved tension in state policy itself. The government is simultaneously urging companies to deploy AI across the economy at speed and warning them against wasteful, duplicative competition. In a field where every major platform feels obliged to field its own assistant, those two goals conflict. Which one prevails when they collide is the policy question the anti-involution campaign has yet to answer.
The verdict
Teng Yaxin's letter is not primarily a story about long hours, and treating it as one lets its most important lesson slip away. Its real subject is how organisations behave when they have lost confidence in their own direction: they measure what is visible, reward what is performed, and mistake endurance for strategy. The AI race makes that failure more likely, because it pairs enormous pressure to move fast with deep uncertainty about where to go.
A CEO has been removed and a statement about humane culture issued. Neither changes the conditions that made the midnight vigil seem reasonable to the people who ordered it. The question worth watching is not whether Chinese tech firms will work less. It is whether any of them, under pressure from investors and policymakers to win at AI, will be allowed to admit that they don't yet know what winning looks like.
