A Small Zhejiang Straw-Hat Factory Used AI to Cut Order Costs in Half

A while ago, I traveled to Cixi, in Zhejiang Province, on a research trip.

It's a place known for straw hats. A single county-level city, more than fifty straw-hat factories, an annual output value of 2.5 billion yuan. The gold-wire straw hats (a local specialty — fine woven rush hats) from Changhe Town alone are sold in more than seventy countries worldwide.

Sounds like an industry that couldn't be more traditional.

But after I visited, I was genuinely taken aback.

A local factory called Jusheng wired generative AI straight into its order-processing workflow. The result? A cross-border order that used to require back-and-forth communication, sample revisions, and confirmations — dragging on for weeks — now turns a customer's one-line request into a sample sketch in seconds.

The owner told me: "Just revising the design used to take a week or two. Now the client says 'change the color,' and I've already sent the new image over."

My immediate reaction was—

This is what generative AI in Chinese manufacturing is supposed to look like.

Where Does Straw-Hat Export Get Stuck?

Let me lay out the background first.

The Chinese straw-hat export business is labor-intensive by nature. The Zhejiang coast has plenty of workers and convenient ports, so the industry took root and grew.

But the past two years have been rough.

The first hurdle: labor got expensive. A single cross-border order — from inquiry, prototyping, contract signing, production, all the way to delivery — passes through a stack of hands. The client changes a color today, adds a logo tomorrow, and the designer is working overtime revising the artwork. From my interviews, this "communication plus revision" cost alone can run 30 to 60 percent of the entire order-processing budget.

The second hurdle: nobody knows what foreign buyers actually want.

A straw hat is a fashion accessory — it's about style, about look, about season. But a small factory has no data team and can only stock up based on the boss's gut. The result? Either bestsellers go out of stock, or dead inventory piles up in the warehouse.

I read an empirical study on the Cixi straw-hat industry, and the data was striking.

Three Surprising Findings

The researchers scraped a full year of straw-hat data from a certain B2B platform in 2023 — more than 3,600 records. Then they scraped thousands of buyer reviews from an overseas retail platform and ran LDA topic modeling to mine them.

Finding One: Offering customization doesn't guarantee profit — but skipping it is worse.

Merchants offering customization (logos, made-to-order from buyer sketches, rhinestone work) made up 72 percent. Their median sales, though, were a full 20 percent higher than non-customizing merchants. Customization isn't a selling point — it's the price of admission to stay alive.

Finding Two: Don't slap a "single-scenario" label on a straw hat.

Eighty-seven percent of straw-hat products were tagged with only one use scenario — work, casual, or party. But the study found that the more scenarios a hat covers, the higher its sales. Hats that span five or more scenarios clearly outperformed every other category.

Foreign buyers' logic is dead simple: a hat you can wear to work, to the beach, on a date, and on a trip — that's the one worth buying.

Finding Three: At different price points, users care about completely different things.

Buyers under $25 ask only: "Does it fit my head?" "Can I gift it?" "Is it durable?"

Buyers from $25 to $50 care about: "Does the real thing match the photo?" "Is the service good?" "Is the color accurate?"

Buyers above $50 talk exclusively about: "Design sense." "Material." "Is it handmade?"

Notice — same product, three price tiers, three completely different languages of need. One algorithm trying to serve all of them is basically impossible.

What Did AI Change in the Factory?

What excited me most about Jusheng wasn't the sophistication of the model they used — it was that they had thought one thing through clearly:

AI isn't here to replace humans. It's here to eliminate "communication cost."

Here's what they did.

Step one: pick an open-source model and deploy it locally.

The owner initially approached several AI service providers. The quotes were all steep, and the data would have to be uploaded to the cloud — she wasn't comfortable with that. In the end, she chose an internationally open-source image-generation model — free, capable of running offline, with every byte of data staying on her own servers.

For a small factory, this step matters enormously.

Step two: feed the model with their own historical orders.

Jusheng has been in business for more than twenty years and accumulated over 3,000 pairs of "customer requirement plus finished-product image." The researchers used a low-resource fine-tuning method called LoRA (Low-Rank Adaptation — a technique that compresses training parameters to a fraction of the original), squeezing the training parameters down to one ten-thousandth of the original. Without adding any inference time, the model learned Jusheng's own design style and craft details.

This step is the real moat.

A general-purpose AI model drawing straw hats will produce something that looks right but lacks that "this is unmistakably a Jusheng piece" character. After fine-tuning, the output isn't just faster — the style is right, the craft is right, even the woven texture is right.

Step three: string the whole workflow together.

The old workflow was: client makes a request → salesperson coordinates → designer draws → revisions fly back and forth → sample confirmed → production scheduled.

The new workflow is: client makes a request → AI generates the image → client confirms in seconds → production scheduled.

That endless back-and-forth over revisions in the middle? It's essentially gone.

An Underappreciated Truth

When I talk about AI with friends in traditional manufacturing, they often say: "This stuff has nothing to do with us — it's a matter for internet companies."

Jusheng's story tells us that judgment is wrong.

The value of AI doesn't lie in how smart it is — it lies in whether it can solve the single most expensive, most repetitive, most headache-inducing link in your operation.

For the straw-hat factories of Cixi, that link is "endlessly revising designs with overseas clients." For the small-goods merchants of Yiwu, it might be "cross-language customer service." For the furniture makers of Foshan, it might be "the client sends a photo and asks, can you make this?"

Every industry has its own "communication-cost black hole."

Whoever plugs that hole first will catch the next wave of dividends.

But Don't Deify It Either

I have to say a few words of caution.

Jusheng pulled it off — that doesn't mean every small factory can.

Model fine-tuning needs data, and many small factories haven't even digitized their own historical orders.

The data has to be plentiful enough, clean enough, representative enough. Three thousand records is the family fortune Jusheng accumulated over twenty years. Tell a startup factory to copy this playbook, and there's simply no foundation to build on.

Not to mention — no matter how fast AI generates an image, the final call is still made by a human. Whether the client places the order, whether they reorder, whether they haggle the price — none of those judgments can AI replace.

AI is an amplifier, not a substitute. It amplifies the advantages you already have, but it can't rescue you from a fundamental problem.

In Closing

The day I left Cixi, the owner gave me a straw hat.

She said this one was designed by AI and woven by her own hands.

I wore it the whole way back.

I kept thinking: the next step for Chinese manufacturing probably isn't inside those glittering tech companies — it's in small factories like this one, heads down doing the work, willing to take a chance on something new.

They don't need to understand what a large language model is, or what LoRA is. They only need to know one thing:

This thing can cut my costs in half and double my response speed.

That's enough.