
After 3+ Months with WorkBuddy, I Turned Off Half Its "Auto" Features
Title: Two months with WorkBuddy, I installed three brakes on it
Let me start with something. The biggest change in the AI world this year is that everyone finally started admitting models make mistakes.
In early November, that Huatai research report got shared around a lot, saying the internet industry is shifting from "grabbing users" to "refining models," and 2026 is the dawn of the AI-native era. The numbers do look good too—Tencent's Q3 AI-related revenue grew 142% year over year, Baidu's ERNIE reportedly broke 480 million daily calls, and Doubao's API has connected over 20,000 small and medium businesses. I believe these numbers. Even my little 20-square-meter stir-fry shop is using AI now, so how bad could the big companies' data be.
But that one line in the report—I stared at it for a while and felt it had nothing to do with me. It said the competitive focus is shifting from "whose app icon is more eye-catching" to "whose model understands users better, whose Agent actually gets work done."
"Gets work done"—those three words are the really hard part.
I started using WorkBuddy about two months ago. Purchase orders, reconciliation sheets, supplier quotes—more than half of it goes through it now. The longer I use it, the more I'm sure of one thing: its automation flows really do run, but the word "fully automatic" needs to be taken apart. My biggest move over these two months was turning off the "auto" in a lot of my flows—about half of it.
Full auto is a trap, semi-auto is what a small shop can actually use
When I first got started I was hooked too. WorkBuddy's automation flows can be chained together—one trigger after another action—and it looks like everything can run itself. My first version was fully automatic reconciliation: every day after closing, it would go grab the day's receipts from a designated folder, recognize them, categorize them, summarize them, produce a table and send it to my WeChat.
Sounds great. Ran for a few days, and on the third day something went wrong.
That day a supplier delivered two boxes of frozen goods, the receipt was handwritten, and there was some glare when I took the photo. WorkBuddy recognized it as another supplier's goods, and got one digit of the amount wrong too. The flow didn't throw an error—it just ran through, sent me the table, perfect formatting, numbers were numbers. I only caught it the next morning when I was prepping ingredients and happened to glance at it.
This reminded me of something I wrote before: for a small shop, the dividing line for an AI tool isn't whether it can finish the job for you, it's whether you can spot it immediately when it gets it wrong. The biggest problem with full auto is that when it screws up, it doesn't say a word.
So I split the flow into two parts. The first half runs as before—recognize, categorize, fill the table. The second half got a "pending confirmation" step: anything over 300 yuan, or anything with low recognition confidence, gets listed separately and waits for me to tap confirm on my phone before moving on. The 300 number is something I figured out by trial—most of a small shop's daily purchase receipts are under 300, and anything over is either a big order or a handwritten receipt, both worth a glance from me.
The result is three or four extra minutes a day. But what those three or four minutes buy me is the confidence to let it run. With the previous version I had to re-check the table every morning anyway, which made it pointless.
I don't like it guessing, but I like that it can be stopped
One of WorkBuddy's abilities is figuring out on its own how a task should be broken down. You throw a bunch of stuff at it, and it first thinks through how many steps this job takes, what each step does. Most of the time this is a time-saver, especially for jobs I can't describe clearly—it can break out a rough plan for me.
The problem is when it guesses wrong, it's not easy to spot from which step it started going off track.
Last month I had it organize a quarterly purchasing summary, with the requirement of categorizing by supplier, merging same categories, and calculating each supplier's monthly average price. The table it gave me looked fine, categories were right, but the average price column was mixing goods of different specs together. Among the frozen goods, one box was 10 jin net weight, another was 12 jin with ice glaze—the prices were different to begin with, so the mixed average was meaningless.
This mistake wasn't really its fault—I didn't state my requirements clearly. But the trouble is, I had to trace the whole flow backward to figure out at which step it misread what I said. WorkBuddy gives you the execution steps, one per line, so you can see what it did, but not why it did it that way.
So now I've developed a habit: for anything slightly complex, I first have it list out the steps for me to see, and only let it run after I confirm. Costs an extra minute, saves half an hour of backtracking later.
This habit sounds dumb. But anyone who's used it for a while should get it—the real time cost of an AI office tool isn't the few minutes it runs, it's how long it takes you to locate the problem after something goes wrong.
The parts the research report didn't mention
Back to that report from the beginning. It said the enterprise AI agent market will break 48 billion in 2026, with office scenarios growing over 60%. I believe in that direction—I'm walking into it myself.
But I saw another piece of material with a claim that the 2026 dividing line isn't in model parameters, isn't in single-point features, but in whether enterprises have completed the switch in their evaluation standards. It said what's truly scarce is the enterprise's own judgment system, and general model capability is secondary.
I chewed on that line for days. What exactly is my little shop's judgment system? It's that I know which supplier shortchanges on weight, which receipts tend to come out blurry in photos, which season purchase prices jump. WorkBuddy doesn't know these things, and it shouldn't.
So my approach to using it now is clear: it handles the legwork, I handle the judgment. It recognizes the receipts, fills the tables, sorts what needs sorting—that's manual labor. Which payment is off, which number is suspicious, which supplier's quote jumped absurdly this month—those still need my eyes.
After two months, it really has saved me a lot. Evening reconciliation used to take over an hour, now it's about twenty minutes. The extra time means I can go to bed earlier, or squat by the back kitchen door scrolling my phone for a bit.
But to say it can manage the books for me—that's nonsense. All those "AI-native" and "agent" phrases in the research report, landing in my 20-square-meter shop, amount to a helper who can run errands but will make mistakes. You have to watch it, but watching it still beats running yourself.
A few days ago a batch of saury arrived, I snapped a photo and posted it on the forum—ugly as hell, but it fried up nice. On the WorkBuddy side, the receipts are getting smoother and smoother, but which accounts to be strict about, which suppliers to keep an eye on—that's still on me.
Physix Frontier