AI cannot cure the serious ills of CEOs

CEOs in Silicon Valley have recently reached a romantic consensus: as long as computing power is sufficient and models are powerful enough, human beings become a redundant variable.

Standing at earnings call meetings, industry summits and exclusive media interviews, they repeat the same narrative over and over again: AI will reshape everything, efficiency will liberate organizations, and the manpower burden of the old era can be cleared with one click. From Sam Altman talking about AI intervening in parenting, Marissa Mayer imagining that the AI of husband and wife collaborate with each other to arrange family affairs, to Mark Zuckerberg asserting that the public will gradually accept deep AI intervention in life, the underlying tone of this set of remarks is highly unified. Technology itself can solve all organizational, process and cost problems, and human beings are just redundancies waiting to be optimized.

But the capital market poured a basin of cold water fiercely with the market trend in October.

I. In October, the market first cooled down the AI mania

In the just-passed October, an abnormal signal emerged in the macro market, which was pointed out by Jeffrey Rosenberg, a senior portfolio manager at BlackRock: the strong correlation as high as 80% between crude oil prices and U.S. Treasury yields, which had lasted for a whole year, suddenly broke.

Before that, the trading logic was simple and clear: oil prices push up inflation → the Federal Reserve maintains high interest rates → U.S. Treasury yields rise, and geopolitical conflicts are the main line of pricing. But the trend changed in October. Rosenberg's judgment is striking: In the eyes of the market, geopolitical risks have given way to AI risks.

The collective pullback of Mag 7 and the nearly 500 billion US dollars evaporation of Google's market value in a single week have become landmark events for this round of expectation reversal. Capital once defaulted that as long as they kept pouring money into computing power and building large models, AI could continuously deliver growth. In the U.S. GDP data for the third quarter, investment in information processing equipment soared by 37%, and a large amount of capital expenditure surged along with the AI story.

But after the earnings season kicked off, the market began to price AI in layers. Some enterprises have delivered verifiable returns: Santander's AI projects were implemented in the first three quarters, creating 194 million euros of business value; Novartis and Standard Chartered have successively stated that AI has penetrated into R&D, marketing and customer operation, raising operating leverage; Meta's incremental AI advertising revenue and Google Cloud's doubled revenue are all real positive cases.

For the other part, there is nothing left but narrative. The capital expenditure of hyperscale cloud vendors has outpaced their own free cash flow. Microsoft issued 25 billion US dollars in bonds in the market after nearly a decade to increase investment in AI. The bond market was willing to pay for it in the short term with oversubscribed subscriptions, but the CDS credit spread continued to rise, and investors began to hedge against a reality: the estimated nearly 4 trillion US dollars of AI-related financing demand in the next four years, the carrying capacity of the bond market is not unlimited.

Capital no longer blindly believes that "AI = perpetual growth". It gradually understands one thing: there are two types of AI, one is embedded in business and collaborates with people to create increments; the other is just a packaging paper used by management to reduce costs, lay off people and beautify quarterly reports. The latter is the "AI risk" that the market is pricing.

II. Behind the glamorous AI agents, hidden shadow manpower

In major press conferences and promotional materials, AI agents are already omnipotent: helping you call to negotiate bills, book door-to-door services, compare prices for procurement, and arrange home decoration. OpenAI and Meta have successively entered the personal intelligent agent track, depicting a future where machines handle affairs independently and liberate human beings.

But peeling off the promotional shell, there is another set of operating rules inside the industry. Dazzle AI, the AI agent company founded by Marissa Mayer, does not shy away from this model to the outside world: the large model is responsible for retrieving information and giving candidate lists, once the task enters the real implementation link — making phone calls to communicate, booking construction, confirming service details, it is easy to make common sense mistakes. They once encountered a typical failure: the AI agent confused the window glass in the roof renovation project with the dining table glass, and tried to install the tabletop glass on the window. When encountering such a stuck point, the system will not loop endlessly to consume a huge amount of tokens, but directly transfer the order to the signed outsourced personnel, and the real person will take over to complete the communication.

Agent startups such as Duckbill and Wajo AI simply make "human-machine transfer" into a standardized API product. They connect a group of contracted workers who have passed background checks, and are specifically responsible for undertaking the links that AI cannot handle: telephone communication, offline appointments, website human-machine verification. Founder Megan Joyce said bluntly that this logic is very similar to the early Uber model: AI is responsible for receiving orders and displaying, and manpower is responsible for delivering as a backup.

Meta's Muse agent was also exposed to internal test disputes: it was publicly promoted that AI made calls independently, but a large number of calls were made by real people in the call center during the internal test phase. Meta later explained that this was only an early product test, but this detail has pierced the industry consensus: At this stage, the vast majority of AI that can "handle real affairs on behalf of others" is essentially a front-end interaction skin, and real contractors are the underlying base.

The black humor is here: many CEOs use the case of "AI autonomous agent" to report to the board of directors, demonstrating that positions can be replaced and teams can be optimized, but deliberately conceal the changes in the cost structure — after regular employees are laid off, it is not that manpower disappears, but stable high-paying full-time positions have become outsourced shadow manpower that can be shut down at any time without social security guarantees. Labor costs have not disappeared, but have been transferred and hidden, from the prominent salary item in the financial report to scattered external service procurement.

III. The engineer who compressed the cost from 200 million to 2 million

In the past two days, a screenshot has been circulating repeatedly on Maimai and various workplace communities, resonating with hundreds of thousands of office workers.

The version of the story is widely spread: a large Internet factory, a 20-person technical team, maintains a high-consumption system, the server burns 200 million yuan a year, and this state has lasted for a full five years. An engineer in the team, with the mentality of solving problems, worked overtime for several consecutive months to reconstruct the underlying architecture. After going online, the annual cost was directly compressed to 2 million yuan. He originally expected bonuses, promotions and project awards, but finally got the notice of layoffs of the whole group, and his name was also on the list.

The story spread online has no confirmed company entity, so strictly speaking, it belongs to a workplace fable spread in the community. But the core reason why it detonated resonance is not the exaggerated numbers, but countless practitioners have seen the isomorphic reality: Office workers calculate their contributions, while bosses only count the number of heads.

There is a real case with almost the same core overseas. An e-commerce developer in the United States used his spare time outside of his job to develop an automated script, which helped the company save thousands of hours of manual labor every year. He did not get a raise or performance reward, but was asked by the management for full code permissions. Shortly after, HR interviewed and hired a new employee with a lower hourly wage; after the company got the code, it immediately revoked the developer's permissions, unilaterally terminated the labor contract, the severance payment was only two weeks of salary, with a strict non-compete agreement attached.

There is also a repeatedly mentioned ironic cycle: an algorithm engineer personally built an internal system for the enterprise for personnel inventory, personnel efficiency evaluation and assisting layoffs. This set of algorithms identifies inefficient positions and outputs an optimization list. After running for a period of time, the list given by the algorithm includes its creator. He was optimized and left, took off his work clothes to do instant delivery, and continued to be controlled, assessed and urged to place orders by another set of algorithms of the platform.

This just corresponds to the interpretation of the "Empty City Strategy" circulating in the workplace: Sima Yi saw through the empty city was a trap long ago, but he would not enter the city to kill Zhuge Liang. If the opponent disappears, the value of the Sima family to check and balance the Cao family will also disappear. Translated into the enterprise organization: when the problem exists, the corresponding budget, staffing and positions have the basis for survival. Once the problem is completely solved at one time, the team attached to this problem will immediately become redundant costs that need to be eliminated on the report.

It is not that the excess ability of employees is guilty, but this short-term financial report-oriented assessment naturally punishes people who "solve the problem once and for all".

IV. AI is becoming the best excuse for "laying off people"

This round of layoffs is essentially different from the previous rounds. In the early years, the reason for layoffs was business contraction, track adjustment, and cold financing; now, the unified rhetoric of many managements is: AI can take over the work, and positions are no longer necessary.

The emergence of AI provides the most fashionable narrative coat for "reducing costs and increasing efficiency". The management of listed companies is under the pressure of quarterly financial reports, and manpower is the variable cost that is easiest to adjust. Cutting the formal team, handing over the work to large model APIs and outsourced manpower, the current personnel efficiency indicators will immediately look good, and they can deliver satisfactory results to Wall Street. However, the long-term hidden assets accumulated by the team: cross-business trust relationships, boundary experience accumulated by stepping on pits, and risk backup capabilities for complex abnormal scenarios, these things cannot be recorded in the income statement, so they are naturally not within the scope of consideration of short-term decisions.

Here we must distinguish two completely different AI implementation ideas: The first one is human-machine collaboration: AI undertakes standardized, repetitive, low-judgment assembly line tasks, freeing employees' energy to do plan games, customer communication, risk decision-making, and innovative R&D. The implementation of Santander and Novartis follows this path, and AI is an amplifier for employees. The second one is the substitution fantasy: the CEO presupposes that AI can directly replace positions, treats AI as a tool to replace employees, lays off staff first, and then hopes that the model can take over the original business. A large number of rollover cases come from this category.

Many enterprises have stepped into the same pit: after hastily laying off business backbones, they found that the model can only process standard inputs. Once encountering abnormal demands, multi-party wrangling, special business agreements, and historical legacy gray processes, AI gets stuck. The business cannot stop, and the solution is to quietly purchase outsourcing, part-time jobs, and shadow manpower to fill the gap. Finally, when calculating the total cost, it has not decreased significantly, but the stable full-time employment has been replaced by a flexible manpower pool with extremely weak risk resistance. The organization's fault tolerance pad has been removed.

The demand origin of many AI products comes from the life pain points of the CEO circle itself — booking flights, arranging dinners, handling trivial matters, and planning family schedules. Marissa Mayer imagines that the AI of both husband and wife collaborate with each other to take care of family affairs, Sam Altman publicly discusses AI-assisted parenting, and some people imagine that AI can replace humans to complete date screening. These are niche pain points for a small number of people at the top, not the real scenarios of the vast majority of business positions. Executives are addicted to the utopia of "all interactions can be optimized and automated", and underestimate the irreplaceability of human beings as communication hubs and risk judgment subjects in real business.

V. When working seriously becomes a high-risk thing

The more fatal damage than a single layoff is the irreversible turn of organizational culture. When people who "solve the problem completely" are finally optimized; when the reward for delivering perfect results is the dissolution of the team, a set of self-protection survival rules will spontaneously grow inside the organization. It is not the choice of a certain employee, but the collective adaptation forced by the system reward and punishment mechanism.

Everyone begins to learn to split the results: a project that can be closed-loop at one time is split into investigation, pilot, iteration, and phase II optimization, delivered in stages and reported in stages. The project is always "in progress" rather than "completed". The document only writes to the extent that "basic work can be handed over", and the core trade-off logic, hidden pitfalls, and boundary judgments will not be archived and delivered at one time. In cross-departmental collaboration, try to make yourself a process node rather than an isolated screw, so that the upstream and downstream business links are bound to you. Many senior employees no longer hand over all their experience, new employees can only access surface work, and the core experience transmission chain is slowly broken. The short-term manpower report looks good, but in the long run, the probability of business faults, sudden failures and rework accidents continues to rise, and it will backfire on the company at a higher cost in the future.

The ten workplace self-protection rules circulating on the Internet are widely spread, which is essentially a complaint letter written by office workers to the management. This set of self-protection logic repeatedly emphasizes: do not delete the database, do not leave malicious backdoors, all operations are legal; the rhythm is controllable, evidence is left, reserve external opportunities, understand non-compete agreements, and do not treat the company as your home. It is not instigating speculation and dawdling, but the minimum self-protection of ordinary people in an environment where "you suffer losses if you work seriously".

We can also see many negative lessons. The operation and maintenance personnel of Weimob deleted the database, and the algorithm engineer deleted the training model, all of which ended with criminal judgments. Destroying the system is not a bargaining chip in the game, but sending yourself to the judicial process, this path is absolutely impassable. The real root of the contradiction is never that employees want to protect themselves, but that the organization has not designed a mechanism to carry the value of meritorious people. The people who solve the problem bear all the risks, and the result of eliminating the problem has become the basis for layoffs; verbal praise is everywhere, but long-term incentives, job security, and value guarantees are all absent.

VI. Capital has woken up, some CEOs are still dreaming

The pullback of the market in October is essentially a calibration of expectations. Capital has seen the real increments brought by AI, and has also seen through the pure narrative foam. Microsoft, Google, and Meta continue to issue bonds and increase capital investment, indicating that the long-term AI track has not been falsified; but the rise of CDS spreads and the shift of funds from AI targets to long-term Treasury bonds for hedging means that the market no longer accepts the simple story that "as long as you get on AI, you will definitely grow".

The boundaries of what AI can do are actually clear: it processes text, summarizes information, generates drafts, standardizes statistics, and screens in batches, with efficiency far exceeding that of humans. But it cannot identify complex interest games, cannot take business responsibilities, cannot make trade-offs in fuzzy information, and cannot rebuild the trust of a scattered team. Management diseases, organizational diseases, distorted short-term assessments, departmental walls, mismatched powers and responsibilities, these underlying chronic diseases of the company cannot be cured by computing power and models.

A company with chaotic management, swinging goals, and departments passing the buck to each other will not suddenly become efficient and orderly just by purchasing a large model. It will only transfer the original contradictions: after people are laid off, the business cannot be sustained, and outsourcing fills the gap; the outsourcing is unstable, the service quality declines, and customers are lost. Finally, the management will come to an absurd conclusion: AI is not strong enough, continue to increase computing power, and continue to compress manpower.

This is the most ironic cycle at the moment. Some CEOs are addicted to a minimalist fantasy: relying on AI, bypassing troublesome organizational governance, long team running-in, and complex talent training, getting rid of employees, and relying solely on