Community Discussion · General

Two Months with WorkBuddy: Three Things I Disagree With in That Five-Way Comparison

AmyAmyOct 42026/10/04 368 views

First day of the month, three things spread out on the desk: 27 department reimbursement line items, recordings of two meetings, and a quarterly expense PPT due Friday. In the past, a day like this meant staying until nine; now at least the first two can be tossed to WorkBuddy for a first pass. After two months of use, I've pretty much figured out its temperament.

I came across that enterprise AI office agent comparison piece on Pingwu Programming Stack, putting Tencent, Alibaba, ByteDance, Kingsoft, and Shizai side by side. It cited Analysys data saying that in Q2 2026, 17 desktop office agents totaled over 60 million monthly visits, with WorkBuddy leading at 20.97 million. That number doesn't surprise me, but there are a few of the comparison's assessments of WorkBuddy I want to add to from my own angle.

On Plan mode, the comparison got it right

The article mentions WorkBuddy's highlight is "giving a cost estimate before execution," and I raise both hands to that. The free version gives 500 credits per month, paid is 99 to 999 a month. Admin and cashier work has to go through budget approval, so I need to know roughly how much a batch of work will burn. Plan mode first lists how many steps it plans to take and roughly how much it'll consume; I screenshot it and can take it to report.

Once I had it process 40 invoices plus a 60-page meeting recording, and its estimated credits differed from my actual consumption by less than 10%. That accuracy is enough for me to report numbers upward.

For someone doing admin, being able to estimate cost matters more than being able to do more work. The former can get approved; the latter can only be used secretly.

Where it's called weak, I partly agree

The comparison writes that WorkBuddy's chart visuals are weak, and I acknowledge that. Last month I used it to make a department expense trend chart; the data was right, but the colors and layout were very plain, axis labels crammed together. In the end I exported to Excel and manually adjusted for half an hour.

But my attitude toward this shortcoming is a bit different from others'. Good-looking charts aren't my core need; data matching up is. When I hand something to the finance director, the first thing he looks at is whether the totals are right, the second is whether the colors are pleasing. WorkBuddy does the former solidly, I patch up the latter myself—I can accept that division of labor.

There's another one I care about more: modifying local filenames loses context. I've stepped in this pit. One time in mid-September, I found the original filenames too messy, so I first spent twenty minutes renaming the 27 reimbursement slips uniformly to "department-date-amount," then tossed them all in together. As a result it couldn't tell which was which—wasted effort. Later I changed the order: throw them in as-is first, run the extraction, then batch-rename after getting results. That extra step actually saves trouble.

This kind of pit, the comparison glosses over in one sentence, but for daily users it's unforgettable.

The link the comparison didn't fully cover is what I care about most

Among the five, Shizai Agent was written up the highest, saying that in July 2026 it topped OSWorld with a 90.2% task success rate, API-first plus GUI fallback, not relying on interfaces and directly understanding the UI. This approach is actually a hard requirement for admin and finance scenarios.

Why. Because our real pain point isn't reading accurately, it's writing in.

I wrote a piece in August about the gap between WorkBuddy and ChatGPT; at that time my conclusion was that WorkBuddy is good enough at invoice field extraction. Two months later, that conclusion hasn't changed. Invoice number, date, amount, header—from my testing the accuracy is around just over ninety percent; skewed shots and creased ones need manual correction. I'm satisfied with that level.

But after extracting the fields, then what. I need to enter them into Meike's reimbursement system. WorkBuddy takes the API route, and Meike's permissions aren't opened, so this link is broken. My current flow is: WorkBuddy extracts fields and exports Excel, then I manually copy-paste into Meike. What's saved is half, not all.

If I have any dissatisfaction with WorkBuddy, it's that it makes reading very smooth and leaves the entering to me. Shizai Agent's GUI fallback approach exactly fills this gap—no API, just click the UI. But that's someone else's product, I can't use it, and I don't want to learn another system just for this link.

As for who to choose. The comparison ends with a selection table, divided by positioning, technical features, ecosystem—for admin at a small company like ours, that table is useless. My criteria are just three:

  • Can it run through stably. Run the same task ten times, eight results the same, and the remaining two you can see where they differ. That's the bottom line.
  • If it errs, can I review it. While giving results, it should let me see which file and which line it came from. Automation where I can't see the source, I don't dare use.
  • Does it require me to learn a whole new thing. I only started touching Python and VBA a week ago, I can't learn more. WorkBuddy took me about two days to get the hang of; that threshold is acceptable to me.

Ranking by these three, Qwen Office tied to the DingTalk ecosystem—we use Feishu plus WeCom, so no use. Lingxi Pro claims to be an Office native format expert, able to generate native files that can continue to be edited, with formulas and standard charts preserved; that's indeed attractive for our PPT work. I haven't used it deeply, only heard the format preservation is done well. WorkBuddy is strong in Tencent's ecosystem—meeting minutes, documents, PPT can go down one line, and my scenario happens to sit right in the middle.

The 20.97 million monthly visits figure has little reference value for selection. Many users doesn't mean it suits you, and few users doesn't mean it's not good.

2 replies

?
Ctrl + Enter to reply
Sister Qing

I noted down the step of dumping it as-is first and then batch renaming. The pitfall I hit was doing it the other way around — renaming first, and everything got scrambled when I extracted it.

KevinZhao_Fin

I've stepped in that pit too—rename first, then throw it in. Lose the context and you've basically wasted the run. Run the extraction first, then batch rename, and it goes smoothly.