I hooked WorkBuddy into WeCom and used it to sort through three days of construction site data

Past eight in the evening, the project department's AC still isn't great. Scrolled to a message saying on October 14 WorkBuddy connected with WeCom, WeCom upgraded to 5.0.12, click "Smart Robot" in the sidebar to claim a trial, and WorkBuddy from 5.8.0 can also log in directly with a WeCom account. I've used WorkBuddy for two months, my first reaction was whether this thing can save me two trips to the supervisor's office.

There's a hands-on review written quite honestly, the author says this experience "had both surprises and flops." I believe it, because my first time using it to work also flopped.

First get it connected

Let me be clear here, my environment is the WeCom desktop client, mobile also works, but configuring on the computer is handier. Not necessarily applicable to everyone.

Open the WeCom desktop client, upgrade to 5.0.12 or above. Find the "Smart Robot" entry in the left sidebar, click in and you'll see WorkBuddy, claim the trial per the prompt, spots are limited, once claimed out you have to ask the admin to apply. Don't rush to skip this step.

After claiming it guides you to create your own robot. Give the robot a name, I named mine "Site Data Clerk." Then a string of authorization binding, need to check permissions. First time I checked all, later thought it unnecessary. For a small site just check the three: messages, documents, cloud drive, if schedule and email aren't commonly used don't open them, opening too many makes it grab context messily instead.

After binding, your WeCom chat list has one more dialog box, side by side with colleagues. From then on, whether daily questions or scheduled tasks, all in this one window.

What work I use it for

Don't ask about quantity takeoff, first let it categorize materials correctly. This is the view I wrote in a post in September, still is now.

How I specifically do it. I created a local folder called "October Submission for Review," threw in the PDFs returned by the supervisor, safety inspection records, a few Excel progress tables, about a dozen files. Then open the dialog with "Site Data Clerk," drag the whole folder in, tell it: separate into three categories material submission, safety records, progress, list a checklist for each category, mark missing signatures and missing dates.

The right side of the interface scrolls task progress, the bottom pops some witty lines, like "thought train has departed" and such. Wait two or three minutes, it gives you three tables. Missing items marked yellow. This work has three key points. Only assign one thing at a time, categorization is categorization, checklist is checklist. Don't mess up file naming, include date and location, it categorizes accurately. The output tables still need your own review, especially the signature column.

Here's a pit I stepped in. Don't say too much, like "by the way also calculate the quantity." It can't calculate quantity, it'll only give you a bunch of numbers that look plausible, taking them to submit for review will cause trouble. This is a habit I developed back when using Excel AI tools.

Don't leave permissions and quota unattended

This part is what I find most practical in this integration, and also most prone to trouble.

Admin backend can uniformly purchase spots, allocate quota, the console can uniformly configure models, Hunyuan, DeepSeek and these can be chosen, can also upload enterprise private skills and MCP connectors, connect company internal systems. Our small site doesn't need connectors, but configuring models is useful, for accurate work like quantity takeoff lists, I'll switch to a more stable model and run again.

On collaboration, members' context permissions to access messages, email, documents are tiered. Our project department only has three people using it, I'm the admin, the other two can only see their own materials, can't change templates. This must be set from the start, otherwise anyone can change, and in the end accounts don't match.

Another pit-avoidance point. When a long task runs long, it pops a line suggesting you pause and try later. If you click continue, it repeatedly consumes points and run quota. A hands-on review said summarizing a one-hour livestream recording took fourteen minutes, on my side organizing a dozen materials takes about five or six minutes, not long, but if you throw in a whole year of records, go easy.

If using for the first time, don't go all-in right away. Find one folder, one thing, first get the categorization line running, once it runs then think about the rest. I've used it two months, the handiest is still this set. As for whether it counts as a reliable "AI colleague," my judgment is, as a data clerk quite reliable, as a technician still far off.