All the 'AI-native CRM' hype doesn't really matter for a small trading company like mine

A few days ago I came across a review published by ShareCRM, a head-to-head comparison of six AI-native CRMs, with Salesforce, SAP, Oracle, and HubSpot all included. It was broken down in detail, explaining the gap between bolt-on AI and true native AI across four layers: data, semantics, permissions, and execution. There's a line in the article I agree with: where AI actually lives in a CRM isn't about whether there's a chat box in the UI.

That's true. But my first reaction after reading it was that this framework has limited use for a small company like mine with a few million dollars a year in volume.

That stuff is prepared for customers like Digital China, Mengniu, and TBEA—seven thousand large and medium enterprises, multi-organization, multi-format, multi-language, with permissions so complex you need a dedicated governance layer. My foreign trade team is just about ten people, with fewer than four hundred customers on the list. The most troublesome permission issue is just making sure new salespeople can't see the boss's bottom-line quote. In that scenario, talking to me about the semantic layer of AI-native feels like using a sledgehammer to crack a nut.

But there's something here I think that article led people astray on. Everyone's fixated on "is AI the core of the system," as if only AI grown inside the CRM counts. After using things for two months, my thinking is a bit different. For small and medium foreign trade companies, most people haven't even gotten to the stage where they can use a CRM—arguing now about whether the AI in the CRM is native enough is premature.

Let me give an example from last week.

Last Wednesday morning, an old German customer sent an email saying they wanted to swap three products in their procurement list for new models. The attachment was an Excel file, twenty-plus rows, mixed Chinese and English, with a column of their internal material numbers. Two years ago, this would have taken me about a morning: first check what our three old models correspond to in our product library, then see whether the new models are in stock, at what price, with what lead time, then redo a quote, match the customer's material number column, and finally reply with an email.

Can a CRM do this for me? Honestly, no. What's stored in the CRM is what this customer bought last year, how payment collection is going, what stage the order tracking is at. It doesn't know how much stock I have left in the warehouse right now, nor whether the supplier's quote for the new model has changed. This work lives outside the CRM—in folders on my computer, in PDFs sent by suppliers, in WeChat and WhatsApp chat histories.

My current approach is to throw this whole chunk at WorkBuddy. It's a general-purpose office agent that can read Excel, PDF, email, and images, and can also handle spreadsheets and documents. I drag the customer's attachment in, along with our own product cross-reference table and the supplier's latest quote, and tell it clearly what to do: find the corresponding new models for the three old ones, use our latest prices, take lead times from the supplier's PDF, and finally produce a quote with the customer's material number column preserved so they can check it.

About forty minutes later, I went through what came out. Of the twenty-plus rows, it got the lead time wrong on two rows, because the supplier's PDF gave two price tiers for the same product—one for orders starting at five hundred units and one for a thousand units—and it defaulted to the lower one. I had to fix that manually. But it finished the rest: the quote format was correct, the customer's number column wasn't lost, and after fixing it I could send it directly.

This is the first thing I want to say. That CRM review talked about "whether AI can turn suggestions into the next action in the system," which is very true, but the premise is that you first have to have a system. For a company like mine, a lot of things just never entered the system—they're scattered in folders. WorkBuddy solves that part of the work; it doesn't care where the data is, you just hand it over.

There's another judgment in that article I don't quite agree with: it says bolt-on AI's experience stays in personal prompts, while native CRM can solidify effective practices into reusable Skills. That's right for large enterprises. But for me, the biggest takeaway from two months of using WorkBuddy is actually the opposite: for the tasks I have to do repeatedly, I've turned them into a fixed workflow.

For example, organizing materials after a trade show. Since coming back from the show in late September, I've been using a fixed approach: first take photos of business cards and throw them in, have it extract all the recognizable text, with company name, person name, email, and phone in separate columns; then cross-check this batch of leads against existing customers in our CRM, marking which are old customers and which are new; finally group by country and by product interest, and export a table ready for follow-up. I've run this workflow for three trade shows, each time with the same wording, the same grouping rules, and the same output table structure—I don't have to teach it again.

Whether this counts as a Skill, I don't know. But this stuff really did stick around—in how I talk to it, and in the grouping logic it remembers. For a foreign trade company where one person does ten jobs, this is more practical than a native CRM that costs hundreds of thousands to implement.

Of course, WorkBuddy isn't without pitfalls.

The biggest problem is that it doesn't understand business judgment. That quote last week—getting the lead time wrong was just one thing. More troublesome was one time it merged two customers with very similar names into one lead: one a trader in Shenzhen, one an end customer in Hong Kong, with email prefixes differing by only one letter. It can't judge this on its own, because literally they do look like the same person. If you send this out without checking, the customer will think you're unprofessional.

Another one: when it processes poorly scanned quotes, recognition accuracy drops. I have a batch of materials sent by Southeast Asian suppliers—some taken with phones, some fax scans, with text blurred into a mess—and half of what it recognized was wrong. In this situation I don't force it now; I go through it myself first, and re-shoot or re-request whatever can be made clearer.

So my judgment is this. No matter how fiercely those six AI-native CRMs argue, they're talking about the question of "which layer AI lives in," and that's prepared for companies that already have a CRM and where the CRM is the business hub. For a company like mine that hasn't reached that stage and where a lot of daily work gets done outside the system, a general-purpose foundation like WorkBuddy is what's actually usable right now. No matter how native the AI in a CRM is, it can't manage the work in your folders that hasn't entered the system yet.

I don't think these two things conflict. Someday when my company grows to a dozen-plus salespeople, thousands of customers, and processes so complex that it must be managed by a system, I'll probably have to go back and look at those native CRMs. But before that, what I need more is to first kill off these scattered, dirty, repetitive tasks on hand. That's what WorkBuddy does.

That German customer's quote last week—after I fixed the two lead-time rows, I sent it out that same afternoon. In the past, this order would have dragged into the next day.