不穿的旧衣服呢？放「AI 版转转」，毛利 80%

不穿的旧衣服呢？放「AI 版转转」，毛利 80%

你的旧衣服怎么处理的？

当然是放在转转上面回收了，转转是一个官方二手平台，比一般二手要正规……

我们今天要介绍的是「AI 版转转」。

它使用 AI 自动识别服装颜色、类别、尺寸，并结合商品质量与市场需求进行定价，降低非标准化商品的处理难度。

每年有约 1700 万件旧衣服，从陌生人的衣柜出发，经过它，走进另一个陌生人的衣柜。

从店主的眼光到 AI 的估价

每件新衣服的背后都有一个确定的库存单位。同款衬衫，同样的颜色，同样的尺码，几千件长得一模一样。仓储、定价、上架，系统处理起来并不复杂。

可 ThredUp 的处理中心每年接收数百万件寄来的衣物，每一件都是孤品。这件格子衬衫洗了二十次，领口有点卷；那件牛仔裤标着 32 码，实际腰围偏大；另一件毛衣摸起来还是新的，标签却掉了。

那平台怎么估价呢，寄售店的店员凭经验收货，摸一摸面料，看一眼牌子，心里估个价，几个月卖不掉就降价，或者退还给寄卖的人。

不少慈善二手店按品类定统一价，一件衬衫几美元，一条牛仔裤十几美元，好货坏货混在同一种价签下面，捡漏的人占了便宜，店里却很难知道一件真正的好东西值多少钱。这在小规模里行得通，一旦量上来，人力就成了天花板。

ThredUp 是这个行业里把流程彻底工业化的一家。据《现代物料搬运》2023 年对其配送中心的报道，进来的衣服要先过 12 项品质检查，合格的才挂上带二维码的衣架，工作人员录入品类、品牌、颜色、尺码、面料和洗涤方式，价格交给专有的机器学习软件来定。高价的衣服拍在人台上，低价的平铺拍摄，照片检查过后才上架。

后来 AI 接手了更多环节，能自动识别颜色、类别和尺寸，再把成色和市场需求放进算法里定价。据 PYMNTS 今年 4 月的报道，平台常年有 400 多万件在售商品，每天新增数万件，一套强化学习模型盯着实时需求，给卖得慢的衣服自动降价。

公司还投入了 4 亿多美元做供应链自动化，覆盖商品识别、尺寸测量和拍摄。

这些投入换来的结果是，ThredUp 每年能卖出约 1700 万件衣服。2025 年第二季度它的毛利率达到 79.5%，而传统服装零售商很少超过 30%。

面向消费者，ThredUp 在 2024 年 8 月，开放了三项 AI 功能，升级后的搜索，以图搜图，还有 Style Chat。

Style Chat 像一位对话式的造型师，输入「秋天婚礼穿什么」，它会给你整套搭配，再输入要求，会跟着调整。以图搜图允许用户上传照片，或者分享一条 Instagram 帖子，去库存里找相似的款式。公司披露的数据显示，使用图像搜索的消费者，最终下单的可能性比普通浏览者高出 85%。

使用这个功能的人，他们带着具体的需求来，AI 帮他们锁定目标。意图越明确，转化越顺畅。因为二手平台的商品实在太散，能帮人在几百万件旧衣里找到那件「对的」，才是门槛所在。

被 AI 带来的人，为什么更愿意下单

Adobe 的最新数据显示，与来自其他流量来源的普通购物者相比，通过 ChatGPT 等 AI 渠道访问零售商网站的购物者更可能延长浏览时间、将商品加入购物车，最重要的是，更可能完成转化。

在今年七月，来自 AI 渠道的购物用户的转化率比来自其他渠道的购物用户高出 60%，每次访问带来的收入高 53%，跳出率低 34%，停留时间多 59%，浏览的页面多 22%。

这种转化率的优势已经连续保持了 11 个月。而一年前的情况则相反，当时非人工智能推荐来的顾客每次访问所花费的金额更多。这意味着人工智能助手在将相关产品展示给具有更高购买意图的顾客方面表现得越来越好。

回到二手衣服的场景，二手平台的货是海量的，ThredUp 光在售的就有 400 多万件，每一件都不一样，没有爆款可以让人闭着眼跟风。逛这样的货架，人很容易累。

心理学里有个很有名的果酱实验，2000 年 Iyengar 和 Lepper 在超市摆摊，摆 24 种果酱时，六成路人会停下来试吃，最后买的只有约 3%，摆 6 种时停下来的只有四成，买的却有约 30%。

选项太多会让人更难做决定，这个说法由此传开。

后来也有人提出不同意见，2010 年一篇汇总了几十项实验的元分析发现，选择过载更容易出现在选项复杂，自己又没有明确偏好的时候。

逛二手的人往往没有具体目标，想要的是「合适的，好看的，便宜的」，正是容易过载的处境。AI 替人把几百万件缩成十几件，用一句话，或者一张截图，换一个小得多的范围。

据 ThredUp 的调查，48% 的二手购物者在购物时用过 AI 工具。被缩小了范围的人来到商品页，已经完成了一半的决定，剩下的只是确认尺码和价格，转化率高也就不奇怪了。

当然，AI 带来的用户并不是铁板一块。他们也有高期待带来的高要求。如果推荐不准，反而会强化失望感——毕竟「AI 推给我的」暗含一种背书，准确性一旦失灵，信任的损耗比随机浏览还要大。

这解释了为什么各家平台都在把推荐模型做得越来越细。知道你喜欢的品牌只是开始，还要知道你上次退货是因为尺码偏大，知道你偏好哑光而非亮面，知道你通常在周五晚上刷手机。

ThredUp 披露，2024 全年 AI 带来的新顾客数量同比增长了 32%。二手服装这个品类本来有门槛，对质地和成色的不确定感让很多人望而却步，AI 的介入正在消解这种犹豫。

从帮你挑到替你买

识别，定价，推荐，「帮你挑」已经成熟，也有了 ThredUp 这样跑通的样本。再往前一步是「替你买」，让 AI 拿着你的授权，自己去下单付款。

2025 年 9 月，OpenAI 和 Stripe 联手在 ChatGPT 内上线了即时结账功能，初始合作商包括 Etsy 和 Shopify 的部分商家。用户在对话里相中了某件东西，可以直接在界面内完成支付，无需跳转。这是第一次有主流 AI 对话平台试图打通「发现」和「购买」之间的最后一段距离。

可到 2026 年 3 月，OpenAI 调整了方向，把重心移回「商品发现」，允许商家使用自己原有的结账流程接单，而不是统一走 ChatGPT 内的支付闭环。

媒体援引分析人士的看法指出，Instant Checkout 没有获得多少用户，原因包括商家接入麻烦，商品数据难以保持准确，多件商品的购物车和会员积分也不好处理。OpenAI 自己认为最初的版本「没有提供我们期望的灵活度」。

Morgan Stanley 在 2025 年 12 月的报告里估计，到 2030 年，AI 购物助手可能占到美国电商的 10% 到 20%，对应 1900 亿到 3850 亿美元的销售额。报告还提到，美国约 23% 的人上个月通过 AI 买过东西，杂货和日用消费品是最先起量的品类。

目前，「替你买」带来的问题已经浮出水面。谁来负责退货纠纷？AI 代购的授权边界在哪里？万一理解偏差，买回来一件根本不想要的东西，损失算谁的？

用户通常给 Agent 的是一份有范围，有预算，有期限的委托，而不是每笔交易都点头。金额上限挡得住贵的错买，挡不住便宜的错买。目前没有公开的规则来处理「Agent 超出了你的本意」，商家往往只能自己扛下损失。

人买衣服前会看图片、尺码表和退货条款，Agent 多半只读数据，这使得退货率上升。ecommerceguide 指出，服装，鞋和家具是风险最高的几类。

如果是二手衣服，一件旧衣服只有一件，卖出即售罄，退回来要重新检验，重新拍照，重新定价，ThredUp 靠规模压低的单件处理成本，或许会在这里被拉高。

Agent 看价格和规格，不看版式和搭配推荐，商家花力气做的陈列会失效，状态描述不够细的二手商品，在 Agent 眼里就排到了后面。

ThredUp 早就把非标准的商品做成了结构化的数据，这些原本为人工作业准备的功夫，到了 Personal Agent 时代可能成了它的优势。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Where's the old clothes? "AI Turn" for "Maori." 80%.

What did you do with your old clothes?

Of course, it's on the spin, and it's an official secondhand platform, more formal than the usual secondhand...

We're going to introduce "Ai Turn" today.

It uses AI to automatically identify clothing colours, categories, sizes and price them in relation to the quality of goods and market demand to reduce the difficulty of handling non-standardized commodities.

Every year, about 17 million old clothes go from the closet of a stranger, through it, into the closet of another stranger.

From owner's perspective to AI's valuation

Each new dress has a defined stock unit behind its back. The same shirt, the same color, the same size, and thousands of it. The system is not complex to process in terms of warehousing, pricing, shelfing.

But ThredUp's processing centre receives millions of pieces of clothing sent every year, each of which is lonely. This plaid shirt was washed 20 times, and the collar was a little curly; the jeans were 32 yards in size and had a large waist; and the other sweater was new and the tag was dropped.

How does the platform be valued? The store clerks, by passing acceptance check, touch the fabric, take a look at the brand, estimate the price in their heart, sell it for months, or return it to the courier.

A number of charitable second-hand stores have uniform prices for goods, several dollars for a shirt, ten dollars for a pair of jeans, and good and bad goods under the same price tags, while the leaky ones take advantage of the price, and it is difficult to find out how much a real good thing is worth. It works in small size, and once measured, manpower becomes the ceiling.

ThredUp is the one in this industry that completely industrialized the process. According to Modern Material Handling, its distribution centre in 2023, the entry of clothing was preceded by 12 quality checks, qualified to carry a two-dimensional-coded hanger, staff entry of items, branding, colour, size, fabric and washing methods, at a price to be determined by proprietary machine learning software. High-priced clothes were taken on the platform, low-priced lay-ups were taken, and photographs were examined and later mounted.

AI subsequently took over more links that automatically identified colours, categories and dimensions, and then placed colours and market demand in algorithms for pricing. According to PYMNTS, in April this year, the platform has more than 4 million items on sale every year, adding tens of thousands of new items per day, with an enhanced learning model focused on real-time demand and an automatic reduction in prices for slow-selling clothing.

The company also invested more than $400 million in supply chain automation, covering commodity identification, dimension measurement and filming.

As a result of these inputs, ThredUp can sell about 17 million clothes a year. In the second quarter of 2025, it reached 79.5 per cent, while traditional garment retailers rarely exceeded 30 per cent.

For consumers, ThredUp opened three AI features in August 2024, an upgraded search to map, and Style Chat.

Style Chat, like a talking plasticist, enters "What to wear at the fall wedding," and it gives you a set of matches, re-enters the requirements, adjusts. A map allows the user to upload a photograph or share an Instagram post to find a similar pattern in the inventory. Company-disclosed data show that consumers who use image searches are 85 per cent more likely than ordinary viewers to end up on a list.

People who use this function, they come with specific needs, AI helps them target. The clearer the intention, the smoother the transformation. Because used platform goods are too loose to help people find that "right" in millions of old clothes, that's the threshold.

Why would I want to put down a list of people brought by AI?

The latest data from Adobe show that, compared to ordinary shoppers from other flow sources, those who visit the retailer ' s website through channel AI, such as ChatGPT, are more likely to spend longer time browsing, add goods to the shopping carts and, most importantly, complete the transformation.

In July this year, shopping users from the AI channel converted 60 per cent more than those from other sources, with a 53 per cent higher income per visit, a 34 per cent lower exit rate, a 59 per cent longer stay and 22 per cent more page views.

The advantage of this conversion rate has been maintained for 11 consecutive months. On the contrary, a year ago, non-analytical recommended customers spent more on each visit. This means that artificial intelligence assistants are increasingly performing well in presenting their products to customers with higher purchasing intentions.

Back to the scene of second-hand clothes, the second-hand platform was big, and ThredUp sold more than 4 million pieces alone, each of which was different, and there was no explosive money to keep people closed. It's easy to get tired of walking on a shelf like this.

There's a famous jam experiment in psychology, and in 2000, Iyengar and Lepper, when they're playing at the supermarket, 24 jams, 60 percent of the people stop to try and eat, and the last buys only about 3 percent, and when they do 6, only 40 percent, and about 30 percent.

Too many options make it more difficult to make decisions, and that is what spreads.

Differing views were subsequently expressed, and a meta-analysis of dozens of experiments in 2010 found that overloading was more likely to occur when options were complex and when there was no clear preference.

Second-hand people often have no specific goals, and what they want is a "right, good-looking, cheap" situation. AI has condensed millions into a dozen pieces in one sentence, or a screenshot, for a much smaller range.

According to ThredUp, 48% of second-hand shoppers used AI tools while shopping. Those who have been narrowed down have come to the commodities page, and half of the decision has been taken, and it is not surprising that the remaining ones are just confirmation of size and price, and that the conversion rate is high.

Of course, AI brought in a user that was not a piece of iron. They also have high expectations. If the recommendation is not made, it strengthens the sense of disappointment -- after all, there's an implicit endorsement of "AI" that, if the accuracy fails, the trust loses more than random browsing.

This explains why the recommended models are being refined by various platforms. Knowing that your favorite brand is just the beginning, and knowing that your last return was due to the size of your size, your preference for dumbness rather than bright face, and your usual use of your phone on Friday night.

ThredUp disclosed a 32 per cent increase in the number of new customers brought by AI throughout 2024. There was a threshold for second-hand clothing, and uncertainty about the mass and coloring was overwhelming, and AI's involvement was defusing that hesitation.

From helping you to buying for you.

Identification, pricing, recommendation, help you pick is mature, and there's a ThredUp run through sample. One more step forward is "buy for you" and let AI take your authorization and pay for it.

In September 2025, Open AI and Stripe joined forces in ChatGPT, with immediate billing functions, with initial collaborators including some of Etsy and Shopify businesses. The user has something in the dialogue that can be paid directly in the interface without jumping. This is the first time that there is a mainstream AI dialogue platform trying to bridge the last distance between "discovery" and "purchase".

By March 2026, Open AI had redirected its focus back to the "Commodity Discovery" and allowed businesses to use their own cut-off process to take orders instead of taking the payment loops in ChatGPT.

The media quoted analysts as pointing out that instant Checkout did not have much access to users, owing to difficulties in commercial access, difficulties in maintaining accurate commodity data, and difficulties in handling the shopping carts and membership credits for many commodities. Open AI itself believes that the original version "does not provide the flexibility we expect".

Morgan Stanley estimated in his December 2025 report that, by 2030, AI shop assistants could account for between 10 and 20 per cent of American electricians, corresponding to $190 billion to $385 billion in sales. The report also mentions that about 23 per cent of people in the United States have purchased through AI last month, and that groceries and daily consumer goods are the first to start with.

So far, the problem of buying for you has surfaced. Who's in charge of the return dispute? AI, where's the authorized boundary for the surrogate? Who's to lose if you buy something you don't want?

Users usually give Agent a scope, budget, time-limited commission, not a nod for every transaction. The monetary ceiling prevents expensive wrongs and cheap wrongs. There are no public rules to deal with "Agent is beyond your will" and businesses often have to pay for their losses.

Before buying clothes, people look at pictures, size sheets and return clauses, and Agent mostly read only data, which increases the return rate. Ecommerceguide states that clothing, shoes and furniture are the most at risk.

In the case of second-hand clothing, only one of the old clothes is sold out, returned for re-testing, re-photos, re-pricing, and ThredUp's single-piece processing costs, which may be raised here.

Agent looks at prices and specifications, doesn't look at layouts and matching recommendations, the displays made by the merchants will fail, and the second-hand goods whose state is not sufficiently detailed will end up in the eyes of Agent.

ThredUp has made non-standards into structured data, which were originally prepared for manual work, and could have been the advantage of the Personal Age.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.