Lao Gan Ma Cut 600 Workers to 48. The Story Isn't AI.

At the Changming base in Guiyang, a conveyor belt feeds chilies past a camera at three frames per second. After 0.3 seconds, a chili with a black spot gets blown into a waste bin by an air jet. At the next station, 48 people watch screens and make the final experienced call.

The line used to hold more than 600 workers. Now Lao Gan Ma runs it with 48 sorters, and revenue has climbed back to 5.4 billion yuan.

The headlines wrote themselves: AI powers traditional manufacturing. Pull the lens back to the factory floor, though, and the word "AI" does less work than it seems.

1. A chili sorted in 0.3 seconds

The system's main job is sorting raw chilies. Cameras ring the conveyor. They spot discoloration, mold, and spots. They trigger air jets to remove bad fruit. In the past, more than 600 pairs of eyes did this. Eyes tire. Standards drift. Industrial cameras and algorithms do not.

This is artificial intelligence, but in the older sense. It is object detection and image classification in computer vision. Under the hood, it uses convolutional neural networks. It is not a large language model, an autonomous agent, or generative AI. It is not large. It is not linguistic. It will not write a chili procurement report. It finds a bad chili in a frame and triggers a mechanical action.

Hospital radiology has used similar tools for years. Finding lesions in CT, MRI, or pathology slides is also image segmentation and object detection. Food companies beyond Lao Gan Ma use the same logic. Haitian Flavouring has "AI bean face" recognition. Panpan Foods has "AI visual inspection." These are upgrades to industrial visual sorting. They are not suddenly emerged general intelligence.

Calling such a system an "AI management system" stretches the term. Capital and markets like the "AI+" story. Any technology that touches neural networks, machine learning, or automation gets relabeled AI. Lao Gan Ma may not be driving the hype. Media, suppliers, consultants, and investors all translate "automation" into "AI" on their own. Readers still need to separate technical description from marketing.

2. The label becomes a trust substitute

The communication logic is simple. AI is the cheapest authority available.

Imagine two drugs. The first label says AI developed it. AI proposed a theory you cannot follow. It supplied data you cannot follow. It claims FDA Phase III trials. You have no third party verification. The second label says experts developed it. It also gives a theory you cannot follow. It also gives data you cannot follow. It also claims FDA Phase III trials. You still have no third party verification. Which one do you lean toward?

Rationally, the two should carry the same weight. The important question is whether the mechanism can be checked, the data reproduced, and the trials independently reviewed. Who developed it matters less. In practice, many people give the first drug extra trust. "AI" carries a sense of the future, precision, and authority.

Lao Gan Ma's case works the same way. The visual sorting system does not need the AI label to make sense. It saves labor, improves precision, and lowers defects. The math works. The traceability system does not need the AI label either. Every batch of chilies, rapeseed oil, and soybeans gets a digital code. If the finished sauce tastes off, the system can trace the problem back to a farm. That is quality infrastructure, not science fiction. Add "AI," though, and the spending sounds more exciting. It trends. It reads as a traditional company becoming a tech company.

3. What it got right: turning farm variables into factory parameters

Set aside the fight over whether this counts as AI. The part of Lao Gan Ma's system worth watching is not sorting. It is traceability.

Chilies are farm products, not standardized parts. On one farm, pungency, moisture, and aroma precursors change from batch to batch. In the past, that change was an uncontrollable agricultural variable. Now each batch gets a digital code. When the finished sauce changes, the algorithm can trace the shift back to a source farm. It can point to a production area, a batch, or a step.

That matters more than cutting hundreds of sorting jobs. Lao Gan Ma has turned part of agricultural uncertainty into measurable factory parameters. For a condiment company that depends on crops and demands consistent flavor, that is a moat. The defect rate has dropped to 0.02%. The line still keeps 48 experienced checkers. The system does not remove human judgment. It removes the repetitive strain on human eyes so people can make the final sensory call.

The economics are clear. The workforce fell from 600 to 48. Over four years, the number of employees covered by social insurance dropped by more than 2,000. The chili sauce market grows at about 3.3%. Labor costs keep rising. Spending upfront on a visual system buys stable inspection with near zero marginal cost. That is capital replacing labor. The AI label makes the spending sound future-facing. The core is still industrial engineering: automated equipment for repetitive work, data systems for an experiential black box.

4. Where AI fits in food manufacturing

Lao Gan Ma shows visual automation that has already landed. It is narrow and deep. Across food manufacturing, AI shows up in three main paths. They are not equally mature.

Quality inspection and sorting is the most mature. Beyond visual sorting, acoustic signal analysis can check the seal on canned food. Vibration analysis can predict failures in mixing and filling equipment. These cases share features: single data types, clear labeling rules, and manageable demands for interpretability. Here AI is a tool, not a gimmick.

Supply chain and demand forecasting is expanding. Companies such as Kraft Heinz have built monitoring platforms for quality checks, demand forecasts, and process optimization. Food demand still swings with seasons, regional taste, and promotions. Forecasting models have a low accuracy ceiling. Their value shows up in inventory turnover, not in dramatic decisions.

R&D and recipe optimization has the most potential and the least maturity. AI can help screen flavor compounds and design alternative proteins at the molecular level. The bottleneck in food R&D is often not generating candidate molecules. It is that sensory evaluation cannot be fully digitized. A person finding something delicious is hard to encode as a loss function. AI can tell you which recipe is more likely to approach a target flavor. It cannot decide what the target flavor should be.

AI can change the parts of food manufacturing that are measurable, repeatable, and codifiable. It can sort chilies in 0.3 seconds. It can trace a batch from jar to farm. It can warn about equipment failure before a line stops. It cannot replace the final call on taste. It cannot replace a person's experience with flavor.

5. The workers who left, and the taste that stayed

The drop from 600 to 48 is easy to write as a scary headline about machines replacing people. The more accurate description is that jobs shifted. They did not all vanish. The work that left was high-intensity, repetitive, and hard on the eyes. The work that stayed needs experience, judgment, and sensory skill. The company gets efficiency and lower costs. The social shock is still severe. A drop of more than 2,000 people covered by social insurance means more than 2,000 families adjusted. "Industrial upgrading" does not cover that.

AI in food manufacturing follows the pattern of earlier automation. Capital replaces labor. Efficiency replaces repetition. The new jobs it creates, such as equipment maintenance, data labeling, model tuning, and system operations, are often fewer than the old jobs it removes. Society should not try to stop automation. It should provide retraining, reemployment, and social support for displaced workers. Otherwise, the happier the "AI empowerment" story sounds, the quieter the people left behind become.

6. Back to the factory floor

What stands out in the Lao Gan Ma case is that it stood on its own without the AI label. The math of visual sorting and the quality case for traceability fit inside ordinary industrial engineering. Calling it AI is a communication choice, not a technical description.

This does not mean Lao Gan Ma did anything wrong. It did one thing many traditional companies fail to do. It went quiet on social media and got connected in its supply chain and production lines. It did not chase traffic. It put money into quality infrastructure. That is more practical than companies that shout "AI transformation" while their data sits in disconnected systems.

The useful question is what constraint it solved. If AI is only an adjective, it hides technical progress and social costs. If AI is a verb, it should point to specific problems. Which step is slow? Which step has errors? Which batch is hard to trace? Then use the right tool. Academics may call it deep learning object detection. Industry may call it AI visual inspection. On the factory floor, it is the machine that blows bad chilies away.

The conveyor at Changming still runs. The camera scans every 0.3 seconds. The air jet removes bad fruit. 48 people watch the screens. The soul of the sauce still lives on the tongue. The machine only blows away the bad chilies.