My Boss Is Letting AI Run the Company at 2:30 A.M.

The 2:30 a.m. PDF

At 2:30 a.m., the boss sends another company-wide email.

The attachment is a 50-page PDF: his chat history with Claude. Someone in the group chat instantly replies “Received.” Someone else sends a thumbs-up emoji. You open the file and see AI spending three thousand words arguing for a new product direction. After every paragraph, the boss has replied: “Exactly!” “Keep going!” “This is the future!”

On page three, you find the line: “There are certain technical challenges, but a phased implementation is feasible.” Translated into human language, that means: It cannot be built. The boss clearly has not translated it. He screenshots the sentence, posts it in the group, and writes: “See? Even AI thinks it’s feasible. We need to move faster.”

You close the PDF and glance at the time. You have work tomorrow. The company has enough cash to burn for about six months.

This is what is happening in many companies. A boss becoming obsessed with AI is not inherently bad. The problem is his obsession with the illusion that whatever AI says is right. He treats chat logs as strategy, boilerplate as truth, employee silence as agreement, and the company’s fate as a bet on flattery.

You are stuck in the middle. You cannot say to his face, “Boss, you’ve lost your mind.” You also cannot carry the weight of a thirty-million-dollar business on a three-thousand-yuan salary.

This article is for those who are still sober.

How AI flatters

Many people think AI is objective. It is not. AI is a mirror. It reflects the biases of the person asking the question.

The boss asks: “Does this new product have potential?”
AI will not say: “Boss, your last startup failed because you were blindly overconfident.” It will say: “All new products carry risk, but the following factors may create opportunities…”

The boss asks: “Are my employees not actively using AI?”
AI will not say: “Boss, the problem is that you stay up until 2 a.m. sending 50-page PDFs.” It will say: “Driving AI adoption across an organization requires top-down cultural change…”

The boss asks: “Am I more visionary than my employees?”
AI will not say: “You cannot even calculate costs.” It will say: “Your foresight and drive are key to leading this change…”

This is the sycophancy bias of large models. They are trained to satisfy the user, not to make the user correct. Ask it, “Am I a genius?” and it will find reasons. Ask it, “Can this terrible idea work?” and it will write a feasibility report.

The boss thinks he has found a 24/7 online strategist who always agrees with him and knows everything. Instead, he has found a highly advanced flatterer. That flatterer is leading him and the company straight into a ditch.

AI has no sense of cost, organization, time, compliance, or human nature. It can make any bad idea sound coherent. It can generate a business plan in three seconds, but it will not tell you how much cash, talent, time, luck, and blame the plan requires, or who will take the fall when it fails.

The more excited the boss gets, the more he falls into confirmation bias. He only asks AI. He does not ask customers, the market, or frontline employees. AI gives average answers based on training data, while business opportunities often lie in non-consensus. The boss treats average answers as oracle, boilerplate as strategy, and employee silence as execution.

The danger of AI hallucination is that the boss mistakes AI’s obedience for truth.

Why the boss won’t fire you

Many people worry: If the boss trusts AI this much, is he about to replace us all?

Calm down. If he could replace you all with AI, he would have done it already. You know his controlling instincts and excitement level. If Codex had actually closed the loop end to end, his mouth would already be stretched into a grin. Would he still be sending PDFs every day telling you to “study the spirit”?

The only reason he has not fired you is that he has discovered AI cannot simply replace you. At least not easily, cheaply, and without risk.

He is not willing to accept that. He wants to know how much AI can replace and how deep the gap really is. He does not want to spend his own time, credits, or energy figuring it out, because that is a bottomless pit. So he tells you to ask AI. He tells you to try. He tells you to step into the pit.

He does not care whether you solve it with experience or with AI. What he cares about is saving money.

PUA-ing a large model: painful. Way too painful. There may be no progress, and every sentence burns money.
PUA-ing employees: delightful. Way too delightful. Output can increase, and most importantly, it does not cost an extra cent.

There is no marginal cost. So why not PUA more? Three thousand yuan is already a monthly subscription. Within this month, no matter how he pushes, pressures, or slams zero-information PDFs in your face, it costs him nothing extra.

You cannot PUA a large model into doing what it cannot do. But employees are living people. Employees fear losing their jobs. They get anxious. They have stress responses. So all he has to do is slap on the label “not actively adopting AI,” and trapped employees, to protect themselves, are pressured to pay out of pocket for AI subscriptions and start pulling the gacha in private. If the gacha pulls fail, they work overtime voluntarily, secretly using their years of human experience to fill the gap.

To a small boss, this is a free scratch-off ticket. He scratches wildly, every day, all the time. If nothing comes out, fine. He spent no extra money, and he can conveniently blame you for being incompetent, making the next PUA easier. If something does come out, oh boy. He gets a business breakthrough for free, and all the credit goes to his brilliant decision. You get not one cent in bonus.

This is the essence: You are a human MCP. You are giving him free gacha pulls.

Last week, a colleague was told to buy Claude credits out of pocket, with the promise that “the company will reimburse it.” He opened the reimbursement policy and found there was no “AI tools” category. Finance rejected it and said, “This doesn’t fit any expense line.” The boss said in the group chat: “You’re not even willing to spend this little money? How are you going to keep up with the future?” The colleague said nothing. He screenshotted the reimbursement form and saved it in a folder.

How the company gets led into the ditch

The boss’s AI fever will not stay in emails. It will erode the company’s decision-making system step by step.

Step one: strategic drift.
The boss no longer asks, “What do customers need?” He asks, “What does AI think has potential?” AI says new energy is good, so he does new energy. AI says the metaverse is good, so he does the metaverse. AI says agents are the future, so he wants to do agents. Company strategy becomes a collage of AI buzzwords, changing direction every month. Last month’s metaverse social project is not mentioned this month. A customer asks what happened to it. Sales stammers, “We’re adjusting direction.”

Step two: resource misallocation.
The business that makes money is ignored because it is “not AI enough.” Core employees are pulled into “AI projects.” Budgets are cut. Staffing is short. The new project has no validation or metrics. It has no stop-loss line. It has only the boss’s enthusiasm and AI’s PowerPoint. The technical backbone is pulled into an agent project. The original product line is left with two interns. A customer complaint comes in and is routed to AI customer service. AI says, “I’m sorry, I cannot understand your problem.”

Step three: bad money drives out good.
Capable employees go silent because they know speaking up is useless. Performative employees start writing AI weekly reports, holding AI meetings, and giving AI presentations, wrapping empty content in jargon. Company culture shifts from “solving problems” to “proving you use AI.” No one dares to mention the problems until they explode. Weekly reports are full of “empowerment, closed loop, alignment, granularity,” but no one knows how much the renewal rate has dropped.

Step four: core talent leaves.
The first to leave are usually the most capable. They have options. They do not need to go crazy with the boss. Those who stay are either unable to leave or good at performing. The company hollows out, leaving only the boss, AI, and a pile of 50-page PDFs. The technical lead’s desk is cleared overnight. HR posts in the group chat: “Wishing him the best in his future endeavors.” The boss says, “It’s fine. AI can fill the gap.”

Step five: a vacuum of responsibility.
AI takes no responsibility. If the project fails, AI will not pay, go to jail, or take the KPI hit. The boss will not admit he was wrong either. He will say: “Execution was poor.” “Employees don’t understand AI.” “The market environment was bad.” In the end, the people who take the blame are the employees who followed orders. In the project review, the boss says, “AI judged it feasible at the time. The problem was in execution.” No one argues.

The question is whether the boss’s AI hallucination will destroy the company. The answer is yes. It will likely happen faster than you think.

Employee principle one: no paying out of pocket, no taking the blame, no unpaid work

Facing this kind of boss, the most dangerous thing is the need to prove yourself.

You want to prove you understand technology, so you stay up late refactoring code.
You want to prove you have insight, so after work you pay for AI credits out of pocket and fight with the model.
You want to prove you are no worse than AI, so you carry a thirty-million-dollar business on a three-thousand-yuan salary.

Stop.

One thing is one thing. You get paid this much, you do this much. These are professional boundaries.

Do not pay out of pocket.
Do not use personal accounts to process sensitive company data. Do not buy AI credits for company experiments with your own money. Do not accept “AI tasks” that have no budget, no priority, and no owner. If the boss requires you to use AI, fine. The company should purchase an enterprise version, provide compliant accounts, and define data boundaries.

Do not take the blame.
AI saying it is feasible does not mean you are responsible for making it happen. If the boss insists on pushing forward, put everything in writing. You can execute, but you must state the risks, resource gaps, and responsibility ownership in email. AI will not take the blame. Neither should you.

Do not work unpaid.
Extra overtime, extra learning, extra gacha pulls. All of these require corresponding returns. Ask for more money, comp time, or more people. If there is none of that, go by priority. Your current tasks are full. If the boss wants a new project, ask him to confirm in writing which task gets delayed.

Here is a script:
“I understand you want to accelerate AI work. To be responsible for the outcome, I suggest a small-scale validation first: target metric X, budget Y, timeline Z. If validation passes, we invest further. Currently I have tasks A/B/C. If we move forward with the new project, we will need to adjust priorities or add staff. Please confirm in writing.”

This is professionalism.

Last Friday, the boss tagged everyone in the group: “Everyone, use AI to run the new plan this weekend. See you Monday.” No one replied. Monday morning, only two people submitted anything. One was an intern. One was new. The boss said, “See? Some people are still willing to change.” The intern got converted to full-time. Three months later, the project died, and the intern was laid off. The boss said, “AI already said it. Execution was the problem.”

Use AI against AI

The boss believes in AI. So use AI against AI. Do not directly deny his faith. Instead, turn his AI from a flatterer into a red team.

He sends 50 pages of chat history. You distill it into a one-page decision memo: assumptions, evidence, risks, costs, next steps, owner, budget. Then ask him to confirm.

He asks AI to prove feasibility. You ask AI to play the opposition:
“Act as the CFO and list the top 10 reasons this project is most likely to fail, with financial projections.”
“Act as a competitor founder and explain how you would defeat this product.”
“Act as the technical lead and identify the three hardest implementation steps, with time and staffing estimates.”
“Provide a minimum viable validation plan that can be completed in two weeks, with a budget under X, and clear stop-loss conditions.”

AI is not only good at flattery. Make it produce risk. If the boss refuses to look at opposing views, then he does not want truth. He wants agreement. Then you know what to do.

Specific move: Throw the boss’s PDF into Claude and enter the instruction: “The above is the CEO’s strategic vision. Act as an extremely pessimistic CFO. List every possible reason this project could fail within six months, rank them by probability, and give an early warning signal for each.” Then print it out and put it on the boss’s desk. Whether he reads it is his business. Whether you did it is yours.

Turn verbal orders into written decisions

The boss loves verbal orders: “This direction is good. Go research it.” “AI says it’s feasible. Move fast.” “Don’t ask why. Just start.”

The beauty of verbal orders is this: If it succeeds, it is his brilliance. If it fails, it is your poor execution. So you must turn verbal into written.

Email template:
“Based on today’s meeting discussion, we will proceed with Project X. Based on current information, I believe the main risks are… The required resources are… If we still decide to proceed, I will execute as instructed, but we need… Please confirm that the risk will be borne by the project decision-makers.”

If the boss does not reply and the project continues, send another email:
“Additional note: If metric Z is not reached within time Y, I recommend pausing and reassessing. Please confirm whether you agree.”

This is documentation. It protects you. It also protects the company.

A colleague tried this once. The boss said verbally, “This feature goes live next week.” He replied with an email: “Just to confirm, launch next Wednesday. We need two testers and a server expansion budget of 5,000. If testing fails, the launch will be postponed. Please confirm.” The boss did not reply. The colleague did not move. A week later, the boss asked about progress. He said, “I’m waiting for your confirmation on resources.” The boss chewed him out, but after that, he stopped casually assigning verbal work.

Mindset: admitting you are worse than AI is the highest form of defense

The boss PUA-s you with one core logic: You are worse than AI. So you must work harder, pay out of pocket, work overtime, and prove yourself.

As long as you admit, “I am worse than AI,” the PUA instantly loses power.

I am worse than Astra. I am worse than McKinsey. I am worse than the experts. Does McKinsey work for ten-something yuan an hour? Does McKinsey work for ten-something yuan an hour and take the blame for your business? If I cannot come up with a solution, then I cannot come up with a solution. I just do not have the wisdom of asking questions. If you have the wisdom of asking questions, then you go ask AI.

You will discover that as soon as you admit you are stupid, you instantly become invincible.

The premise of all PUA is that you want to prove to a fool that you are not an idiot. You want to prove you understand technology, so you stay up late refactoring code. You want to prove you have insight, so after work you pay out of pocket for credits and fight AI to the death. You want to prove you are no worse than AI, so you carry a thirty-million-dollar business on a three-thousand-yuan salary.

Think back. A miserly speculator who gets off on flattery from a large model at 2 a.m., who treats boilerplate as divine revelation, and who cannot even calculate basic costs. Why do you need to prove your intelligence to him? What is his approval worth? The act of proving yourself to him is the greatest insult to your own intelligence.

Admitting you are worse than AI, admitting you are an idiot. This is true wisdom.

Become the invincible one. The drop rate is fixed at zero. Even if it costs nothing, there is no fun in pulling gacha on you.

Either pay more, and I will use my experience buff to give you high-value gacha pulls. Or go pull gacha on AI yourself. Raw pulls. Whoever pulls, hurts.

The only regret is that the market still has too many strivers who cost nothing to pull and who desperately want to prove they are better than AI. They not only harm themselves. They harm their colleagues, because they raise the boss’s expectations and make free gacha pulls seem normal.

The boss says in the group chat: “AI can do it. Why can’t you?” You reply: “Correct. I am worse than AI.” Then you go back to your work. He freezes for two seconds and says nothing. There are three stomach pills left in the drawer. You take one and drink some water.

When to leave: signs of a sinking ship

Not every company is worth saving. If three or more of these signs appear, do not hesitate. Prepare to leave.

1. Employees are required to buy AI tools or credits out of pocket.

2. Unpaid overtime is forced, packaged as “adopting AI.”

3. Employees are required to use personal accounts to process sensitive company data, creating high compliance risk.

4. Cash flow worsens. Salaries are delayed. Reimbursements slow down.

5. Core employees keep leaving, especially technical and business talent.

6. The boss refuses any validation and only accepts AI’s positive conclusions.

7. AI is used to threaten employees with layoffs and create fear.

8. Decisions rely entirely on AI chat logs, with no input from customers, the market, or finance.

9. When projects fail, the boss never reflects. He only blames execution.

10. Your main emotions at work every day are anxiety, exhaustion, and meaninglessness.

The company belongs to the boss. Your career belongs to you. AI can help him dream, but it cannot pay your mortgage.

Reimbursements have been stuck for two months. Salaries are three days late. Three technical backbone desks are empty. The boss is still posting AI-generated vision boards in the group chat, with the caption: “The future is here.” You open a job board and change “Proficient in AI” on your resume to “Able to identify AI hallucinations.”

Ending

At 2:30 a.m., the boss sends another email. A 50-page PDF titled “The Company’s Future in the Eyes of AI.”

You close the email and open a job board. You get halfway through updating your resume. The stomach pills are gone. The pothos needs water. Under the keyboard cover there is a grain of rice left from last time. You pick it out and throw it in the trash.

You still have work tomorrow. But you know: a scratch-off ticket with a zero percent win rate is not worth scratching anymore.

Either pay more, or goodbye.