I Looked at Xinggu Cloud's AI Sales Agent: WorkBuddy Is Actually Doing Its Upstream Work

I open my inbox at 7:30 in the morning and there are over forty new emails. Fewer than ten are actual customers asking for quotes — the rest are all platform notifications, auto-replies, and receipts that are obviously mass-sent. I've been in foreign trade for ten years. Inquiries are never in short supply; the hassle is that they're always mixed in with junk, and I have to spend the first half hour as a sorting worker.

A couple of days ago I came across that article from Xinggu Cloud about their AI sales agent DeepSales. It says it can handle overseas inquiries 24/7 in multiple languages, read the company knowledge base to answer spec questions, and while chatting, dig out the customer's purchase quantity, target market, and project timeline. It keeps nurturing low-to-mid intent leads and hands high-intent ones over to a human. It covers channels like the official website, WhatsApp, Facebook, and LinkedIn, and the communication records can be written back to the CRM.

My first reaction after reading it was: the direction is right. My second reaction was: this solves the downstream work. Who's going to deal with that pile of mess upstream?

The leads that come in on my end are probably a bit more typical. On WhatsApp there's half a sentence saying "how much," the phone number on a trade show business card is printed blurry, the same company shows up three times in an Excel export from Tendata, and there are auto-replies mixed into the emails. If you feed all of that straight into even the smartest sales agent, all it catches is noise. So for the past two months I've been using WorkBuddy to do the front-end work — washing the material clean first.

WorkBuddy and DeepSales aren't even the same species. One is a salesperson that can talk to customers for you; the other is an assistant sitting next to you helping you lay out your materials flat. But a lot of people insist on comparing them, so I'll just go through my judgment point by point.

First, receiving customers. This is something WorkBuddy can't do, and shouldn't do. It won't reply to WhatsApp for you, won't negotiate MOQ with a customer in Spanish at two in the morning. That's DeepSales's turf. The 24/7 multilingual reception it advertises — I buy it. I've only been on WhatsApp for three weeks, and the time difference with customers is just there. A message at midnight, replied to the next morning — too late, the dish is cold. This really is a hard need.

Now, organizing materials. Here it's the reverse — DeepSales can't help much, because it assumes what you feed it is clean data. What I actually do is create a folder, throw in photos of trade show business cards, exported WhatsApp chat logs, and customer tables pulled out as CSV, and have WorkBuddy dedupe by company name, extract the columns for company, contact, country, inquired product, and quote status, and output a table that can be imported straight into the CRM. I used to do this by hand and an afternoon would be gone. Now it's about twenty minutes, and I use the time saved to check the fields it isn't sure about.

One detail is worth mentioning. That Xinggu Cloud article says DeepSales reads historical information before replying, bases model and price on the product library, only uses customer cases if it can retrieve the relevant material, and records the opportunity info already collected before handing off to a human. I quite agree with this set of constraints.

What I fear most with AI is it just making things up. I've stepped in the same pit with WorkBuddy. Once I had it organize a quote sheet and it swapped the unit prices of two models across rows — I almost sent it out. After that I learned my lesson: anything involving models and prices, I always go through it myself. It's only responsible for building the table, not for making the final call.

Then there's intent judgment. DeepSales says it can identify needs and judge intent during the conversation, as a reference for sales. I don't doubt the direction, but I noticed one detail: on its own page, that "21x improvement in initial screening success rate" has a footnote saying it's a design goal, illustrative, with data sources listed as Salesforce and MIT. That kind of thing I discount. Intent is tricky — the same customer asks about price this week and lead time next week; whether they're in a hurry or stalling often comes down to human intuition. What AI gives you is a label, not a conclusion.

WorkBuddy can do rough screening too. My current usage is to have it do a preliminary classification by inquiry content — quote type, sample type, complaint type, pure tire-kicking — and then I scan through it myself. That's good enough; the rest of the judgment I do myself.

Let me also mention two pitfalls I've hit. One is email. I tried having WorkBuddy process customer inquiry emails and then classify them, but it couldn't tell auto-replies and bounce notifications apart — they all got mixed in as real inquiries. I still haven't found a good filter configuration for this; I can only manually scan the subject lines first. Two, don't be greedy. At first I always wanted to state the requirements all at once, stuffing seven or eight demands into one instruction, and everything that came out was wrong. Later I changed to doing one thing at a time — clean first, then classify — and it was actually faster and more accurate. I've walked this detour; I'm writing it out to save the next person some time.

So my judgment is clear: these two things sit upstream and downstream of each other. WorkBuddy stands at the very front, flattening messy material into a usable table. Sales agents like DeepSales stand behind, taking the clean table to receive customers, follow up, and close deals. The CRM layer in the middle — I've used it for two weeks — manages the deal flow. Only when the three segments connect is it a complete line. No matter how pretty anyone's assembly line is, if the material is dirty, what comes out is still dirty.

If you force me to rank priorities: for most small and medium foreign trade teams, getting the upstream cleaning done first is more worth it than rushing to bring on a bot that can auto-reply. Because if what the bot catches is duplicate, wrong, or vague leads, it only amplifies the chaos — and amplifies it fast.

Looking ahead, I'd guess these two types of tools will each move a step toward the other's side. Sales agents will sooner or later add an entry point for data cleaning, and office tools like WorkBuddy may grow lightweight lead scoring. But in the short term I won't switch. A shovel is a shovel, an assembly line is an assembly line. Getting that pile of dirty material in your own hands figured out is more real than anything.