AI Can't Cure the CEO's Serious Illness: Capital Begins to Price in "AI Risk"

By Shan Zi

Silicon Valley CEOs have recently reached a romantic consensus: as long as there is enough compute and the models are strong enough, humans are a superfluous variable.

Standing at earnings calls, industry summits, and media interviews, they repeat the same narrative over and over: AI reshapes everything, efficiency liberates organizations, and the human baggage of the old era can be cleared out with one click. From Altman talking about AI getting involved in parenting, to Mayer envisioning couples' AIs coordinating household chores with each other, to Zuckerberg being certain that the public will gradually accept AI's deep involvement in life—the undertone of this rhetoric is highly uniform. Technology itself can solve all organizational, process, and cost problems; humans are merely redundancy waiting to be optimized.

But the capital markets, with October's market action, poured a basin of cold water on this.

In the October just past, an anomalous signal appeared in the macro market. As Jeffrey Rosenberg, a senior portfolio manager at BlackRock, pointed out, the strong correlation of up to 80% between crude oil prices and US Treasury yields, which had held for a full year, suddenly broke down.

Before that, the trading logic was simple and clear: oil prices push up inflation → the Fed keeps rates high → Treasury yields rise, with geopolitical conflict as the main thread of pricing. But in October the wind changed. Rosenberg's judgment was striking enough: in the market's eyes, geopolitical risk has given way to AI risk.

The collective pullback of the Mag 7 and Google's market value evaporating by nearly 500 billion in a single week became the landmark events of this round of reversal in expectations. Capital once took it for granted that as long as it kept spending money on compute and stacking large models, AI would continuously deliver growth; in the third-quarter US GDP data, investment in information processing equipment surged 37%, with a large amount of capital expenditure riding the AI story upward.

But once the earnings season began, the market started pricing AI in tiers. Some companies delivered verifiable returns: Santander Bank's AI projects implemented in the first three quarters created 194 million euros in business value; Novartis and Standard Chartered successively said that AI had penetrated R&D, marketing, and customer operations, raising operating leverage; Meta's incremental AI advertising and Google Cloud's doubled revenue are all solid positive cases.

And another group has nothing but narrative. The capital expenditure of hyperscale cloud providers has already outrun their own free cash flow. For the first time in nearly a decade, Microsoft issued 25 billion dollars in bonds in the market to step up AI investment. The bond market was willing to pay in the short term, with oversubscription, but CDS credit spreads kept rising, and investors began hedging against a reality: with an estimated nearly 4 trillion dollars in AI-related financing demand over the next four years, the bond market's capacity to absorb it is not unlimited.

Capital no longer blindly believes that "AI = perpetual growth." It is slowly coming to understand one thing: there are two kinds of AI—one embedded in the business, working with people to create incremental value; the other merely wrapping paper that management uses to cut costs, lay people off, and beautify quarterly reports. The latter is the "AI risk" the market is now pricing.

In major launch events and promotional materials, AI agents are already capable of anything: helping you call to negotiate bills, booking on-site services, comparing prices for purchases, arranging home renovations. OpenAI and Meta have successively entered the personal intelligent agent arena, painting a future in which machines handle affairs autonomously and liberate humanity.

But peel away the promotional shell, and a different set of operating rules exists inside the industry. Dazzle AI, the AI agent company founded by Marissa Mayer, makes no secret of this model: the large model is responsible for retrieving information and producing a candidate list, but once a task enters the real-world execution stage—calling to communicate, booking construction, confirming service details—commonsense errors easily occur. They once encountered a typical failure: the AI agent conflated the window glass and the dining table glass in a roof renovation project, trying to install tabletop glass into a window. When it hits this kind of snag, the system does not loop endlessly consuming massive tokens; instead it directly transfers the order to contracted outsourced staff, with a real person taking over to complete the communication.

Agent startups like Duckbill and Wajo AI simply turn "human-machine handoff" into a standardized API product. They connect with a group of contract workers whose backgrounds have been vetted, specifically taking on the steps AI cannot handle: phone communication, offline appointments, website human verification. Founder Megan Joyce said outright that the logic is much like the early Uber model: AI handles order intake and display, while human labor handles the backstop delivery.

Meta's Muse agent also had an internal testing controversy exposed: it advertised AI autonomous calls externally, but during internal testing a large number of calls were made by real people at call centers. Meta later explained that this was only early product testing, but this detail has already punctured the industry consensus: at this stage, the vast majority of AI that can "handle real-world affairs" is essentially a front-end interaction skin, with real human contractors as the underlying base.

The black humor lies right here: many CEOs take the case of "AI autonomous agents" to report to the board, arguing that positions can be replaced and teams optimized, yet deliberately omit the change in cost structure—laying off formal employees does not mean the labor disappears; it merely turns stable, well-paid full-time positions into outsourced shadow labor that can be shut down at any time and has no social security protection. Labor costs have not disappeared; they have merely been transferred and hidden, turning from the conspicuous salary line in financial reports into scattered purchases of external services.

III. The engineer who cut 200 million in costs to 2 million

These past two days, a screenshot has been circulating repeatedly on Maimai and various workplace communities, striking a chord with hundreds of thousands of workers.

The story has spread widely: at a major internet company, a twenty-person technical team maintained a highly resource-intensive system, with servers burning through 200 million yuan a year—a state that lasted for a full five years. One engineer on the team, hoping to solve the problem, worked overtime for months to refactor the underlying architecture; after launch, the annual cost was directly compressed to 2 million yuan. He had expected a bonus, a promotion, and project recognition, but what finally came was a layoff notice for the entire team, with his own name on the list as well.

The viral story has no confirmed company entity, so strictly speaking, it is a workplace parable circulating in communities. But the reason it struck such a chord is not the exaggerated numbers, but that countless practitioners have seen the same structural reality: workers calculate contributions, while bosses only calculate headcount.

There is an almost identical real case overseas. A US e-commerce developer, outside his regular job and in his spare time, built automation scripts on his own that saved the company thousands of hours of manual labor each year. Instead of a raise or performance reward, management demanded full access to all his code. Soon after, HR interviewed and hired a newcomer at a lower hourly rate; after obtaining the code, the company quickly revoked the developer's access, unilaterally terminated his labor contract, offered only two weeks' salary as severance, and attached a strict non-compete agreement.

There is also a frequently cited ironic loop: an algorithm engineer personally built an internal system for the company to inventory staff, evaluate workforce efficiency, and assist layoffs. The algorithm identified low-efficiency positions and produced optimization lists; after running for a while, the list it generated included its own creator. He was optimized out, took off his work uniform, and went to do on-demand delivery, where he continued to be managed, assessed, and rushed by another platform algorithm.

This maps exactly onto the "Empty Fort Strategy" interpretation circulating in workplaces: Sima Yi had long seen through the empty fort as a trap, but he would not enter the city and kill Zhuge Liang. If the rival disappears, the value of the Sima family in balancing the Cao family disappears as well. Translated into corporate organization: as long as the hard problem exists, the corresponding budget, headcount, and positions have a basis to persist. Once the problem is thoroughly solved once and for all, the team that existed by virtue of that problem immediately becomes redundant cost on the spreadsheet that needs to be cleared.

It is not a crime for employees to have excess capability; it is that this short-term financial-report-oriented assessment naturally punishes the people who "solve the problem once and for all."

The emergence of AI has provided the most fashionable narrative cloak for "cost reduction and efficiency improvement." Management at listed companies faces quarterly earnings pressure, and labor is the most easily adjusted variable cost. Cut the formal team, hand the work to large-model APIs and outsourced labor, and the current-period efficiency metrics immediately look better, allowing them to deliver a satisfactory answer to Wall Street. Meanwhile, the implicit assets a team accumulates over the long term—cross-business trust relationships, boundary experience gained from stepping in pits, the ability to backstop risks in complex and abnormal scenarios—cannot be entered into the income statement, so naturally they are not within the scope of short-term decision-making.

Here we must distinguish two completely different approaches to AI implementation: The first is human-machine collaboration: AI takes on standardized, repetitive, low-judgment pipeline tasks, freeing employees' energy for solution negotiation, client communication, risk decisions, and innovative R&D. The implementations at Santander and Novartis follow this path; AI is an amplifier for employees. The second is the substitution fantasy: the CEO assumes AI can directly replace positions, treats AI as a tool to replace employees, lays people off first, and then hopes the model will take over the original business. A large number of failure cases come from this category.

Many companies fall into the same pit: after hastily laying off key business staff, they discover that the model can only handle standard inputs. Once it encounters abnormal requests, multi-party wrangling, special business agreements, or legacy gray-area processes, AI gets stuck across the board. The business cannot stop, so the solution is to quietly purchase outsourcing, part-time, and shadow labor to fill the gap. When total costs are finally calculated, they have not dropped significantly; stable full-time employment has merely been replaced by a flexible labor pool with extremely weak risk resistance. The organization's buffer for error tolerance has been pulled out.

Many AI products' requirement origins come from the CEO circle's own life pain points—booking flights, arranging dinners, running errands, planning family schedules. Marissa Mayer imagines AI on both sides of a couple coordinating household matters; Altman publicly discusses AI-assisted parenting; some imagine AI replacing humans in screening dates. These are niche pain points of a tiny top-layer minority, not the real scenarios of the vast majority of business positions. Executives are addicted to the utopia that "all interactions can be optimized and automated," underestimating the irreplaceability of humans in real business as communication hubs and subjects of risk judgment.

V. When working seriously becomes a high-risk affair

The more fatal harm than a single layoff is the irreversible turn in organizational culture. When the people who "thoroughly solve problems" are ultimately optimized out; when the reward for delivering perfect results is team dissolution, a set of self-protective survival rules spontaneously grows inside the organization. It is not the choice of any one employee, but a collective adaptation forced by the system's reward and punishment mechanism.

People began to learn to break results into pieces: a project that could be closed out in one go is split into research, pilot, iteration, and phase-two optimization, delivered in stages and reported in stages. The project is forever "in progress" rather than "completed." Documentation only goes as far as "the basic work can be handed over"; the core trade-off logic, hidden pitfalls, and boundary judgments are never archived and handed over all at once. In cross-departmental collaboration, people try to make themselves a node in the process rather than an isolated screw, so that both upstream and downstream business links become bound to them. Many veteran employees no longer share all their experience, newcomers can only touch surface-level work, and the chain of passing on core experience slowly breaks. In the short term the headcount report looks good, but in the long term, the probability of business gaps, sudden failures, and rework accidents keeps rising, and in the future this will backfire on the company at a higher cost.

The ten workplace self-protection rules circulating online are widely spread precisely because they are essentially a complaint letter written by workers to management. This self-protection logic repeatedly emphasizes: don't delete the database, don't leave malicious backdoors, keep all operations legal; keep the pace controllable, leave a trail of evidence, reserve external opportunities, understand non-compete agreements, and don't treat the company as home. It is not teaching people to opportunistically slack off; it is the minimum self-protection for ordinary people in an environment where "taking things seriously means you lose."

We can also see many cautionary tales. The Weimob operations staff member who deleted the database and the algorithm engineer who deleted training models both ended in criminal convictions. Destroying a system is not a bargaining chip; it is putting yourself into the judicial process, and this path absolutely does not work. The real root of the conflict has never been that employees want to protect themselves, but that the organization has not designed a mechanism to carry the value of its contributors. The people who solve problems bear all the risk, while the results of eliminating problems instead become the basis for layoffs; verbal praise is everywhere, while long-term incentives, job security, and value guarantees are all absent.

VI. Capital has sobered up, but some CEOs are still dreaming

The October market pullback was essentially a recalibration of expectations. Capital has seen the real incremental gains brought by AI, and has also seen through the bubble of pure narrative. Microsoft, Google, and Meta continue to issue debt and increase capital investment, showing that the long-term AI track has not been falsified; but the widening of CDS spreads and the shift of funds from AI targets to long-end government bonds for safety represent that the market no longer accepts the simple story that "as long as you adopt AI, you will definitely grow."

What AI can do, its boundaries are actually already clear: it processes text, summarizes information, generates first drafts, performs standardized statistics, and screens in bulk, with efficiency far beyond humans. But it cannot identify complex interest games, cannot bear business responsibility, cannot make trade-offs amid ambiguous information, and cannot rebuild trust in a demoralized team. Management diseases, organizational diseases, distorted short-term assessment, departmental silos, and misaligned authority and responsibility—these underlying corporate ills cannot be cured by computing power and models.

A company with chaotic management, wavering goals, and departments blaming each other will not suddenly become efficient and orderly just because it purchases a large model. It will only shift the original contradictions: people are laid off, the business cannot be sustained, and outsourcing fills the gap; outsourcing is unstable, service quality declines, and customers churn. In the end management will even reach an absurd conclusion: AI is not strong enough, so continue to increase computing power and continue to compress manpower.

English text is being prepared.

AI is a very good efficiency tool, but not a miracle cure for layoffs. Healthy digital transformation is letting AI lift people up; pathological AI fanaticism is using AI as an excuse to abandon people. Capital has already begun to recalculate the accounts, and workers have long seen through this script. The remaining suspense is left to those CEOs still immersed in fantasy: when shadow manpower grows ever larger, everyone in the team protects themselves, and hidden business risks keep piling up, what will they use to catch the backlash propped up by layoffs and AI stories in those pretty financial reports?