本文来源:https://mp.weixin.qq.com/s/CjWhDqJuyll5eLh2Deh80g
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
最近有个做外贸的朋友问我:
有没有那种合适小团队用的 AI Agent?不用重新开发一套平台,就能直接自主处理所有日常工作,例如发邮件、处理询盘、跟单之类...

我说,不仅有,而且比想象中成熟得多。
例如百度搭子,售价不高,额度给得也很足,跑长任务也颇为稳定。
图片和文件都能直接处理,还能手机和电脑双端操控,任务同步。相当于一个随时在线的智能工作台,很合适小团队拿来接手那些特别耗时间琐碎工作。

我干脆替他跑了一遍。
我准备把一个机型交给百度搭子,让它按真实的外贸流程自主执行。看它能不能理解每一步要做什么,协助我把这门生意真正推进下去。
最近刷外网的时候,我正好看到一个很有意思的机型:捡屎手机壳。

它的创意是在手机壳背后装了一卷宠物拾便袋,这样狗主人出门遛狗时,就不会忘记带拾便袋。
但这东西并不是真的机型,而是宠物品牌 By Teddy 做的一条愚人节广告。没想到发出来以后,评论区里不少人都在认真追问:到底什么时候开卖?
从人们的反应来看,它似乎真有一些市面需要。于是,我决定用它来做这次测试。
一、获客
所有生意的开始,都是获客。
但这款机型同时横跨宠物用品和手机配件两个品类,到底应该卖给谁?我先让百度搭子分析,哪些渠道最可能成为它的客户。

它按客户类型、市面区域和采购决策角色,给潜在客户排了优先级:
确定方向以后,百度搭子又开始逐个搜索相关网站,最终整理出100条潜在客户线索,覆盖美国、英国、澳大利亚、加拿大及欧洲等市面。
然后对这100条线索进一步做了分析:
其中宠物用品专业渠道约45条,是最值得优先联系的一类;跨界渠道和进出口贸易渠道各约20条,平台资源约15条。

名单里还详细写了每家公司的业务规模和覆盖范围。例如它是服务 Amazon 卖家的批发商,还是覆盖数千家门店的分销商;是合适铺货的 B2B 平台,还是具备大量买家资源的展会平台。
这样笔者就能依据每家公司的业务特征,外观不同的开发策略。
依据前期调研,百度搭子先确定了机型的售价区间。出厂价USD3–5,批发价 USD 6–10,建议零售价 USD 15–25,MOQ为 300–500 件。
它一共生成了15套开发信模板:

这些模板并不是把同一份文案简单翻译一遍,面对不同情形,百度搭子有不同的话术。
例如面对宠物用品渠道,它会重点强调遛狗情形和袋卷的复购潜力,面对手机配件和礼品渠道,则突出机型的新奇感,面对展会和 B2B 平台,又会换成 OEM 供货的角度。

依据渠道分为8类英文模板,还有德语、西班牙语、意大利语、荷兰语、波兰语、日语和韩语版本,每封都附有中文对照,方便检查内容。

单看获客这一环,百度搭子的表现已经很不错了。
只需要给它一个机型,它就能自己判断从哪里切入,并持续把任务往下推进。
对小团队来说,它能接手大量琐碎、重复的工作,减少这些事务对团队精力的消耗,让有限的人手专注于更关键的事情。
二、搭建独立站
获客完成后,下一步就是把机型展示出来。
原机型只有几张参考图,我先让百度搭子调用 baidu-image-gen skill,补充生成机型特写。

然后我一次性让它搭建独立站,制作销售 PPT 和机型手册。

大约15分钟后,三份内容就全部交付了,还同时给出了 PDF 和 HTML 版本。整个过程中,我只微调了几张生成图片,其余的页面结构都是它独立完成的。
独立站总体使用低饱和度的莫兰迪色系,视觉风格与机型配色和宠物生活方式定位保持一致,既弱化了能力型机型容易产生的廉价感,也提高了品牌的亲和力。
页面的信息层级、机型动线和交互外观也比较成熟:使用者可以快速浏览不同款式及售价,最终进入询盘页面提交采购需要。

个人比较喜欢独立站里展示机型用法的这一部分,把抽取拾便袋的动作拆成了连续画面,使用者不需要阅读太多说明,就能快速理解机型怎么使用。
整个展示还带有一定的动态感,与传统机型网站里常见的静态摆拍相比,更容易让人记住这款机型,也把“随手一抽就能使用”的核心卖点展现得很直观。

机型手册的完成度也很高,总体结构符合海外采购商查看机型资料的习惯。
每款机型都有单独的介绍页面,客户可以清楚了解具体尺寸以及适配范围,还能看到包装与定制相关的信息。
很多最初没有给出的采购细节,也被主动补充了进去。
这样生成出来的手册已经具备较强的实用性,既能协助客户快速完成选品,也方便销售人员在后续询盘中直接引用。

销售 PPT 则更偏向采购提案,借助机型结构拆解和情形化展示,把值得买的原因讲得更直观,完成度很高。

百度搭子没有把独立站、机型手册和销售 PPT 当成三个孤立的外观任务,而是围绕同一套品牌分为,分别适配不同的情形。
这说明它理解一款机型该如何被包装,才能形成一套完整的对外销售体系。
在这一部分中,我对百度搭子最终交付的总体成效还是比较满意的。
唯一不太满意的是,独立站部分页面的信息密度有点高,需要手调,如果能进一步精简文案,总体浏览感受会更好。
三、整理询盘、跟单
做外贸最让人头疼的,往往不是推进某一单具体的交易,而是信息散落在不同渠道,随时都在变化。
客户可能在邮件里确认数量,又跑到 LINE 上修改颜色;刚谈完报价,供应商那边又传来新的生产进度。订单一多,光靠脑子根本记不住。
这时候百度搭子的双端同步能力就起作用了。我可以随时在手机上把询盘记录粘贴给它,让它帮我记住所有进度。
它还兼容语音输入,不想打字,可以直接跟搭子对话。

我干脆把10组不同客户的聊天记录全部扔给它。里面既有英文邮件,也有日语 LINE 对话,还混着询盘平台、付款记录以及供应商的生产反馈。

它先给了我一张可以依据进度不断补充的多维表格,每个客户的询盘都会成为一条独立项目。
刚开始沟通时,表格记录询盘来源;进入报价后,项目状态会随之变化;等到客户确认样品或者正式下单,后续的采购与交付进度又会继续接在同一条记录里。
让我惊喜的是,它还会把最近一次沟通压缩成一句摘要,并把当前卡住项目的难题单独标出来。
这样打开表格后,不需要重新翻聊天记录,就能立刻知道下一步应该回复客户,还是去催供应商。

另外,搭子还依据项目状态生成了看板和总览仪表盘,当前正在量产的订单、需要继续跟进的客户,以及已经浮现风险的项目,都能直接看到。


为了以后不用从头再教 AI 一遍,我又让百度搭子把这套跟单流程整理成了一个 Skill。

以后再收到新的询盘,只要调用这个 Skill,它就会按照同样的规则理解聊天记录,并把最新进展继续写入表格。
也就是说,这套跟单机制可以长期运行。虽然业务中的关键节点仍然需要人工确认,但作为一个随时在线的跟单助手,百度搭子的完成度已经相当高了。
四、财务经营分析
我给它准备了一份虚拟的季度财务账套。文件中一共有13张表,从客户订单一直记录到银行流水,同时还加入了库存变化和季度预算。
为了增加难度,我故意放入了一些跨季度订单,也保留了退款记录和冲销流水,看看它能不能理解这些数据之间的关系。

拿到文件后,百度搭子很快识别出这是一套由13张表组成的完整账套。
它先用 Python 核算第三季度的核心财务数据,再拆开B2B和DTC,继续追踪不同市面与机型的盈利表现。
让我意外的是,它还专门加入了异常核查步骤,会检查流水里有没有冲销。完成核算后,又继续评估渠道投放成效。

很快,它就给到了一份13页的季度经营分析报告。最让我满意的是,它会继续追查异常背后的原因,并给出对应的调整方向。
例如第三季度收入达到104.8万元,看上去已经完成预算。但报告继续往下拆,发现日本客户 WanLife 贡献了67.2%的B2B收入,却贡献了88.9%的B2B毛利。
与此同时,PawHaven虽然是第二大客户,毛利率却只有19.4%。这意味着公司的利润过度依赖单一客户,部分订单看起来卖得多,实际并不赚钱。
现金流部分也分析得很到位。经营现金流暂时为负,报告进一步追查后发现,约21.7万元的平台货款还没有结算,这笔钱占全部资金占用的71.3%。因而,现金流变差主要来自回款时间差,暂时不代表经营浮现难题。
它还发现CLIP期末库存达到619件,按照当前销量需要接近7个月才能卖完,于是直接建议减少补货,并借助组合促销加快去化。
报告最终把财务数据落到了具体经营动作上,这一点很实用。

不过,这份报告的部分判断还是稍微激进了一些。例如它认为 Global Sources 的投入效率最高,但84.3的ROAS主要来自 WanLife 这一笔大单,样本量还不够多。
在真实情况中,还是需要人类开展核验。
实测下来,我认为百度搭子的经营分析能力已经可以用在实际业务中,可以帮小团队快速看清经营状况,并提前发现一些容易忽略的难题。
五、总结
回想 2022 年,那时笔者使用 AI ,基本还是问一句,它回答一点,再由人复制粘贴到不同软件里。AI更像一个聪明的聊天框,真正把事情做完的仍然是人。
短短几年,这种关系已经发生了很大变化。
现在只要把目标和资料交给 AI,它就可以自己理解任务,调用合适的工具,持续推进。
更关键的是,AI正在变成一站式工作台。过去分散在不同软件里的工作,如今可以在同一个地方连续完成。
前一步的结果可以直接成为下一步的基础,整个业务流程也真正被串了起来。
对中小团队来说,这种变化尤其关键。很多过去因为人手不够而被搁置的事情,现在终于有机会真正落地。
人少,不再天然意味着做不了复杂的事,这或许才是 AI 最值得期待的地方。
从这个角度看,AI Agent 带来的变化才刚刚开始。
感受链接:
https://www.dumate.cn/
Source: https://mp.weixin.qq.com/s/CjWhDqJuyll5eLh2Deh80g If there are copyright issues, please notify us via email: 13951431913#139.com or submit a comment on the right.
Recently, a friend of mine who works in foreign trade asked me:
"Is there an AI Agent suitable for small teams? One that doesn't require building a whole new platform, but can directly and autonomously handle all daily work, such as sending emails, processing inquiries, and following up on orders?"
I said, not only does it exist, but it's far more mature than you'd imagine.
For example, Baidu Dazi (Baidu Buddy). The price is low, the quota is generous, and it runs long tasks quite stably.
It can directly process images and files, and can be controlled on both mobile and desktop with task synchronization. It's essentially an always-online smart workbench, very suitable for small teams to take over those time-consuming, tedious tasks.
So I went ahead and ran a full test for him.
I decided to hand a product over to Baidu Dazi and let it autonomously execute a real foreign trade process. Let's see if it can understand what needs to be done at each step and help me actually push this business forward.
Recently, while browsing foreign websites, I came across a very interesting product: a poop-pickup phone case.
The idea is to attach a roll of pet waste bags to the back of the phone case, so dog owners won't forget to bring waste bags when walking their dogs.
But this wasn't a real product; it was an April Fool's Day ad created by the pet brand By Teddy. Unexpectedly, after it was posted, many people in the comments seriously asked: when will it actually go on sale?
Judging from people's reactions, there does seem to be some market demand. So I decided to use it for this test.
1. Customer Acquisition
Every business begins with customer acquisition.
But this product spans both the pet supplies and phone accessories categories. Who should it be sold to? I first asked Baidu Dazi to analyze which channels are most likely to become its customers.
It prioritized potential customers by customer type, market region, and purchasing decision role:
After confirming the direction, Baidu Dazi searched through relevant websites one by one, and ultimately compiled 100 potential customer leads covering markets such as the US, UK, Australia, Canada, and Europe.
It then further analyzed these 100 leads:
Among them, about 45 were professional pet supplies channels, the highest priority category; cross-border channels and import/export trade channels each had about 20; and platform resources had about 15.
The list also detailed each company's business scale and coverage. For example, whether it was a wholesaler serving Amazon sellers, a distributor covering thousands of stores, a B2B platform suitable for product distribution, or an exhibition platform with extensive buyer resources.
This allowed me to develop different outreach strategies based on each company's business characteristics.
Based on the preliminary research, Baidu Dazi first determined the product's price range. Factory price USD 3–5, wholesale price USD 6–10, suggested retail price USD 15–25, with a MOQ of 300–500 units.
It generated a total of 15 outreach email templates:
These templates were not simply translations of the same copy. Facing different situations, Baidu Dazi used different messaging.
For example, when addressing pet supplies channels, it emphasized the dog-walking scenario and the repurchase potential of the bag rolls; for phone accessories and gift channels, it highlighted the product's novelty; and for trade shows and B2B platforms, it switched to an OEM supply angle.
Based on the channel, it created 8 categories of English templates, plus German, Spanish, Italian, Dutch, Polish, Japanese, and Korean versions, each accompanied by a Chinese translation for easy content review.
Looking at customer acquisition alone, Baidu Dazi's performance was already quite impressive.
You only need to give it a product, and it can figure out where to start and keep pushing the task forward.
For small teams, it can take over a large amount of tedious, repetitive work, reducing the drain on team energy and allowing limited manpower to focus on more critical matters.
2. Building a Standalone Website
After customer acquisition, the next step is to showcase the product.
The original product only had a few reference images. I first asked Baidu Dazi to call the baidu-image-gen skill to generate additional product close-ups.
Then I asked it all at once to build a standalone website, create a sales PPT, and produce a product catalog.
About 15 minutes later, all three deliverables were completed, with both PDF and HTML versions provided. Throughout the entire process, I only fine-tuned a few generated images; all other page structures were completed independently by it.
The standalone website used a low-saturation Morandi color palette, keeping the visual style consistent with the product's colors and pet lifestyle positioning. This not only reduced the cheap feel that functional products often have, but also enhanced the brand's approachability.
The page's information hierarchy, product layout, and interactive design were also quite polished: visitors could quickly browse different styles and prices, and finally reach the inquiry page to submit purchasing needs.
Personally, I particularly liked the section on the standalone website that demonstrated product usage. It broke down the action of pulling out a waste bag into a sequence of continuous frames, so visitors could quickly understand how to use the product without reading too much explanation.
The entire display also had a sense of motion. Compared to the static product shots commonly seen on traditional product websites, it made the product more memorable and intuitively conveyed the core selling point of "pull it out and it's ready to use."
The product catalog was also highly polished, with an overall structure that matched the habits of overseas buyers reviewing product materials.
Each product had its own introduction page, allowing customers to clearly see specific dimensions, compatible ranges, and packaging and customization information.
Many purchasing details that were not initially provided were also proactively added.
The catalog generated in this way was highly practical, helping customers quickly complete product selection and making it convenient for salespeople to directly reference it in follow-up inquiries.
The sales PPT leaned more toward a purchasing proposal. Through product structure breakdowns and scenario-based displays, it explained the reasons worth buying in a more intuitive way, with a high level of polish.
Baidu Dazi did not treat the standalone website, product catalog, and sales PPT as three isolated design tasks. Instead, it built them around the same brand identity, adapting each to different scenarios.
This shows that it understands how a product should be packaged to form a complete external sales system.
In this section, I was quite satisfied with the overall results delivered by Baidu Dazi.
The only thing I wasn't fully happy about was that the information density on some pages of the standalone website was a bit high and needed manual adjustment. If the copy could be further streamlined, the overall browsing experience would be better.
3. Organizing Inquiries and Following Up on Orders
What makes foreign trade most headache-inducing is often not advancing a single specific transaction, but rather that information is scattered across different channels and constantly changing.
A customer might confirm quantities by email, then change the color on LINE; just after finishing price negotiations, the supplier sends new production updates. With many orders, relying on memory alone is impossible.
This is where Baidu Dazi's dual-device synchronization capability comes into play. I can paste inquiry records to it on my phone at any time and have it remember all progress for me.
It also supports voice input. If you don't want to type, you can talk directly to the assistant.
I went ahead and dumped 10 sets of chat records from different customers into it. They included English emails, Japanese LINE conversations, inquiry platform messages, payment records, and supplier production feedback.
It first gave me a multi-dimensional table that could be continuously updated based on progress. Each customer inquiry became an independent entry.
At the initial communication stage, the table recorded the inquiry source; after entering the quotation stage, the entry status changed accordingly; once the customer confirmed samples or formally placed an order, subsequent procurement and delivery progress continued to be appended to the same record.
What surprised me was that it also compressed the most recent communication into a one-sentence summary and separately flagged the issues currently blocking each project.
This way, when opening the table, I could immediately know whether to reply to the customer or chase the supplier next, without having to dig through chat records again.
In addition, the assistant generated a Kanban board and an overview dashboard based on project status. Orders currently in mass production, customers needing follow-up, and projects already showing risks could all be seen directly.
To avoid having to teach the AI from scratch next time, I also had Baidu Dazi organize this order-follow-up process into a Skill.
In the future, whenever a new inquiry arrives, calling this Skill will allow it to understand the chat records according to the same rules and continue writing the latest progress into the table.
In other words, this order-follow-up mechanism can run long-term. Although key business nodes still require manual confirmation, as an always-online order-follow-up assistant, Baidu Dazi has already reached a fairly high level of completeness.
4. Financial and Business Analysis
I prepared a virtual quarterly financial accounting set for it. The file contained 13 tables in total, from customer orders to bank statements, along with inventory changes and the quarterly budget.
To increase the difficulty, I deliberately included some cross-quarter orders, as well as refund records and offset transactions, to see if it could understand the relationships between these data points.
After receiving the file, Baidu Dazi quickly recognized it as a complete accounting set composed of 13 tables.
It first used Python to calculate the core financial data for the third quarter, then broke down B2B and DTC separately and continued tracking profitability across different markets and products.
What surprised me was that it also added an anomaly verification step, checking whether any offset transactions existed in the cash flow. After completing the calculations, it continued to evaluate channel investment performance.
Soon, it delivered a 13-page quarterly business analysis report. What pleased me most was that it kept digging into the reasons behind anomalies and provided corresponding adjustment directions.
For example, third-quarter revenue reached 1.048 million yuan, which appeared to have met the budget. But the report continued to break things down and found that Japanese customer WanLife contributed 67.2% of B2B revenue yet accounted for 88.9% of B2B gross profit.
Meanwhile, although PawHaven was the second-largest customer, its gross margin was only 19.4%. This meant the company's profits were overly dependent on a single customer, and some orders looked like they sold a lot but actually didn't make much money.
The cash flow analysis was also very solid. Operating cash flow was temporarily negative, but the report dug further and found that approximately 217,000 yuan in platform payments had not yet been settled, accounting for 71.3% of total capital occupation. Therefore, the deterioration in cash flow was mainly due to the time lag in payment collection, not an immediate operational problem.
It also found that CLIP's ending inventory reached 619 units, which would take nearly 7 months to sell through at the current sales rate. It therefore directly recommended reducing replenishment and accelerating inventory turnover through combined promotions.
The report ultimately translated financial data into specific operational actions, which was very practical.
However, some of the report's judgments were still a bit aggressive. For example, it considered Global Sources to have the highest investment efficiency, but the ROAS of 84.3 mainly came from the single large order from WanLife, so the sample size was not yet sufficient.
In real-world situations, human verification is still needed.
Based on my testing, I believe Baidu Dazi's business analysis capability can already be used in actual business operations. It can help small teams quickly understand their business situation and identify easily overlooked problems in advance.
5. Summary
Looking back at 2022, when I used AI, it was basically a question-and-answer interaction. You asked, it answered a bit, and then a person copied and pasted the results into different software. AI was more like a smart chat box; the person was still the one who actually got things done.
In just a few years, this relationship has changed dramatically.
Now, as long as you hand the goals and materials to AI, it can understand the task on its own, call the appropriate tools, and keep making progress.
More importantly, AI is becoming a one-stop workbench. Work that used to be scattered across different software can now be completed continuously in one place.
The results of the previous step can directly become the foundation of the next step, and the entire business process is truly connected.
For small and medium-sized teams, this change is especially critical. Many things that were previously shelved due to insufficient manpower now finally have a chance to be actually implemented.
Having few people no longer inherently means being unable to do complex things. Perhaps this is the most promising aspect of AI.
From this perspective, the changes brought by AI Agents are just beginning.
Experience link:
https://www.dumate.cn/