The AI Subsidy Is Ending: Why Small Business Owners Are About to Become Renters

I

Lao Zhou deleted the chat history about the PPT revisions.

Not out of anger. There was no need for it anymore. Before, at midnight, he would send one line, “You there?”, and the other side would answer, “Yes, Mr. Zhou, I’ll fix it right away.” Then he would wait half an hour, refresh his feed twice, and wonder whether the guy was cursing him out. Now he dragged the file into the chat box, typed three lines: unify the font to Source Han Sans, change the color of the chart on page three, move the conclusion up one page. Then he went to make tea. By the time he came back, the file was sitting in his downloads folder. It had even renamed itself for him: final_final_rev3_actual_final.

He stared at the screen for ten seconds, then changed the WeChat contact name from “Design – Xiao Li” to “Design – Xiao Li (pending).”

Pending for two weeks. In the third week, Xiao Li’s desk was empty. Lao Zhou paid him one extra month’s salary and no severance, because “I don’t owe him anything.” Xiao Li ran into Lao Chen, the deputy, in the elevator. Lao Chen patted him on the shoulder and said, “Take a rest. It’ll do you good.” Xiao Li said, “You take care too, Brother Chen.” The elevator doors closed. Lao Chen came back and told Lao Zhou, “Before he left, the kid organized all the source files. Pretty careful.” Lao Zhou said, “Mm. GPT can organize files too.”

This scene plays out in small companies every day. No farewell speech. No “thank you for your contributions over the years.” No goodbye dinner. A person leaves, but his work stays. Only now it is done by an account that does not draw a salary, does not pay social insurance, does not take annual leave, and does not send “let’s talk tomorrow” in the group chat at three in the morning. One hundred fifty yuan a month. Lao Zhou had done the math. Xiao Li cost eight thousand a month, plus social insurance, housing fund, desk, electricity, water, over ten thousand. Between one hundred fifty and over ten thousand, there was no need to hesitate.

II

Lao Zhou’s company had nineteen people. He brought in the work himself. Lao Chen managed delivery. One apprentice ran the site. The other sixteen subcontracted the execution. In the past, among those sixteen, there were people who made graphics, wrote proposals, reconciled accounts, haggled with suppliers, and one person whose job was to get yelled at by clients and then relay the message back. Now there were two graphic designers left, one proposal writer, no one doing accounts, and the supplier wrangler was still there, because suppliers recognize people. When AI calls them, they just hang up.

But Lao Zhou found that the change was not the headcount. It was what Lao Chen looked at before signing.

In the past, when Lao Chen reviewed a proposal, he flipped through ten pages: cover, table of contents, executive summary, technical approach, schedule, quote, risk plan, team introduction, case studies, back cover. He looked at every page, because every page could hide a landmine. Now Lao Chen looked at two pages: the first and the last. On the first page, he checked whether the client’s name was misspelled. On the last page, he checked whether the quote was missing a zero. The middle eight pages were written by AI. The formatting was clean, the logic smooth, the cited cases real, the schedule table neater than a human’s. Lao Chen said, “Fine.” Then he signed.

Lao Zhou asked him, “Did you actually look?” Lao Chen said, “I looked. It writes better than I do.” Lao Zhou said, “What if there’s a mistake?” Lao Chen said, “If there’s a mistake, I signed it. I own it.”

Lao Zhou remembered that sentence for a long time. The core value of the twelve thousand a month he paid Lao Chen was not “can write proposals.” It was “dares to sign.” No matter how well AI writes, it does not sign. When the client comes after someone, AI will not say, “I’m responsible.” On the court summons, in the line for legal representative, you cannot write “GPT-4o.”

III

But signing was also something Lao Zhou himself was shrinking away from.

He used to sign everything. Now he let Lao Chen sign first, and then he looked. Not out of trust in Lao Chen. He just did not want to be the last person to read ten pages. He found himself turning into a dispatcher: a client came in, he broke the request into five parts, threw one to GPT for the proposal, one to Gemini for a quote comparison, one to another tool to check supplier qualifications, one to Lao Chen to go look at the site, and kept the last one for himself: dinner with the client.

Dinner with the client could not be replaced by AI. When the client got to his third cup and started cursing his own boss, Lao Zhou had to pick up the conversation. When the client said, “Your quote is higher than XX company’s,” Lao Zhou had to smile. When the client said, “I might get transferred next month,” Lao Zhou had to remember that sentence and go check the client company’s personnel changes. None of this had an API.

So while he cut people, he quietly saved screenshots of his API usage into a folder. The folder was called “Cost Comparison.” Inside was also a table listing: if GPT raised its price to 10x, 20x, or 50x of the current rate, could the company still carry it? He had calculated that at 20x, he would have to let Lao Chen go too. At 50x, he would have to go back to writing proposals himself. At 100x, he said, “Then I won’t use it. I’ll hire people.”

But he knew that was just tough talk. He had already let Xiao Li go, cut two graphic designers, and eliminated the accounts person. The people were gone, the workflow was broken, and the clients were used to “request today, draft tomorrow.” If you made them wait three days again, the client would say, “Are you guys going under?” The money you saved by cutting people was partly profit, partly an advance. An advance on your future options.

IV

Lao Zhou had a friend named Dajun. Dajun ran a small machining shop, about thirty people, half of them old masters. Dajun did not understand programming, and he did not understand AI. His son studied liberal arts in the city and did new media. Every time the son came home, Dajun said, “What can that broken computer of yours even do?” The son said, “Dad, you don’t get it. AI can draw now.” Dajun said, “Draw? Do you know what tolerance fit means? AI knows shit.”

It was not polite, but Lao Zhou thought Dajun was right. In machining, AI could help you write process cards, check standards, convert a client’s PDF drawing into CAD. But on the shop floor, AI had no answer to how much tool compensation, what feed rate, how to clamp the fixture, or how to fix distortion after heat treatment. Not because AI was not smart enough, but because it had never stood in front of that machine tool, never smelled cutting fluid, never heard the sound of a tool chipping. An old master tightened a screw by feel, and the torque was just right. AI calculated for half a day. By the time it finished, the master had done three.

But Dajun had a soft spot too. His youngest master was forty-seven. The oldest was sixty-three. Young people did not want to learn. They thought it was dirty and tiring, three years before they could work independently, and even then the pay was nothing special. Dajun said, “When this batch retires, my shop closes.” Lao Zhou said, “Don’t you have a son?” Dajun said, “Him? He can’t even read a vernier caliper.”

Lao Zhou did not answer. He thought of his own apprentice, who had been with him two years and still could not use Excel properly. When asked to make a budget table, the apprentice asked, “How do I batch-enter all these names?” Lao Zhou wanted to curse at the time. Now he thought: if this person leaves, AI can make the table for him. But AI cannot nod and bow for him in front of the client. It cannot slam the table when the supplier delivers late. It cannot check the fire extinguisher dates one by one before the safety inspection.

No. But Lao Zhou did not want to do those things himself either.

V

So what kept Lao Zhou awake at night was not the cutting. It was the road after the cutting.

He had two roads now. One was to keep cutting, squeeze the execution layer to the minimum, hand everything to AI, keep Lao Chen to sign, the apprentice to run errands, Lao Zhang at the front desk to watch the door. Lowest cost, fastest speed, best-looking profit. But the end of that road was this: he had turned himself into a system with no backup. If AI raised prices, he died. If AI made a mistake, he died. If AI got blocked, he died. If the client found out the proposal was written by AI and demanded a discount, he died. All his eggs were in a basket he could not control, and the owner of the basket was named Sam Altman.

The other road was to keep people. Keep a few who could read AI drafts, a few who could solve problems on site, a few whose faces the clients recognized. Higher cost, slower speed, thinner profit. But the advantage of that road was: when the price went up, he could say, “Then I won’t use it,” and actually mean it. The people were still there, the workflow could still run, the clients still recognized him. AI was a tool, not an organ.

Lao Zhou chose part of the second road. He let Xiao Li go, but raised Lao Chen’s salary by three thousand. He told Lao Chen, “From now on you have one more job. You check the AI output. If you find a mistake, it’s on you. If you don’t find one and something goes wrong, it’s also on you.” Lao Chen said, “Three thousand isn’t enough.” Lao Zhou said, “Five.” Lao Chen said, “Deal.”

Then Lao Zhou made Lao Chen handwrite every process in the company. Not type. Handwrite. From client inquiry to proposal submission to contract signing to delivery acceptance to after-sales follow-up. Who was responsible for each step, what the standard was, what the deadline was, who to find when something went wrong. Written in a kraft-paper notebook, locked in a drawer. Lao Zhou kept the key. The backup was in Lao Chen’s head.

Lao Zhou said, “If one day the internet goes down, or GPT won’t let us use it anymore, you run it for me according to this notebook.” Lao Chen said, “Can’t run it. Not enough people.” Lao Zhou said, “Then call people back.” Lao Chen said, “They won’t come back at the old price.” Lao Zhou said, “I know. So write it carefully. Write it in detail. When we hire new people, they can practice from it and learn fast.”

VI

Last month, Lao Zhou received an email. An AI platform was adjusting enterprise pricing, charging by seat: three hundred ninety a month per seat. Lao Zhou did the math. He was using seven seats, twenty-seven hundred a month. Not a big increase. But he forwarded the email to Lao Chen and said, “Look. Here it comes.” Lao Chen said, “This is nothing.” Lao Zhou said, “Right. This is nothing.”

He filed the email in the “Cost Comparison” folder, next to the price-sensitivity table. One line on the table was highlighted in yellow: “At 50x, keep Lao Chen + apprentice + Lao Zhang, cut the rest, monthly cost about 42,000, 18,000 more than now, acceptable.” Below it was a line in red: “At 100x, switch back to human labor, recall at least 3 experienced hands, estimated recruitment period 2 months, client churn 30%, survivable.”

Lao Zhou did not know whether the table was right. But he knew that having the table and not having the table were two different kinds of people. The person without the table could only curse when the price went up. The person with the table could open it, look, and then decide whether to renew or pick up the phone and call people back.

He sent Dajun a WeChat message: “While your masters are still there, have them train apprentices. One by one. You pay the wages. Don’t skimp.” Dajun replied, “Can’t train them. Young people won’t do it.” Lao Zhou replied, “Then record every sentence the masters say. Put it into a booklet. No matter how good AI gets, someone still has to tell it, ‘This part needs to be tempered twice after quenching.’” Dajun replied, “Mm.”

Lao Zhou put down his phone and looked out the window. Downstairs, Lao Zhang at the front desk was sorting packages. Slow, but not one mistake. Lao Zhou thought, Lao Zhang cannot be let go. If Lao Zhang goes, who sorts the packages? Can AI sort packages? Yes. But the courier does not recognize AI. The courier recognizes Lao Zhang. Lao Zhang will offer a cigarette and say, “Just put it here.”

That was Lao Zhou’s last card: some things AI can do, but some people only people recognize.

He put that card under the kraft-paper notebook, locked it in the drawer, and put the key in his pocket. Then he opened GPT and started writing tomorrow’s proposal.