One Week with WorkBuddy Cut Report Compilation Time from 3 Hours to 20 Minutes

I came across that article testing the top 6 large models for coding. Gemini 3.1 Pro scored 77.1% on ARC-AGI-2, GPT-5.5 can control computers and operate software, and the comments section is once again full of cheers about doubled efficiency. I read it for ten minutes, closed it, and went back to using WorkBuddy to organize our club's inventory list.

This isn't me being contrarian. Coding tools solve the problem of writing code, but as a junior in college, what actually eats up my time has never been writing code—it's those chores that look simple but make you want to curse when you do them. For example, yesterday afternoon, I was staring at 63 inventory records, copying from one Excel sheet to another, fixing formats, aligning columns, checking for missing items, and manually merging over a dozen duplicates. Halfway through, I wondered if there's any AI out there that could handle this mess for me.

Previously, my attitude toward WorkBuddy was half-belief, half-doubt. When I started using it a week ago, I tried using it to organize new member registration forms—47 tables with messy formatting. It did recognize, categorize, and summarize them. But at the time, I thought it was just okay; the results still needed manual review, which wasn't faster than doing it by hand. So in that post, I wrote that "deliverability is still lacking," citing poor data fault tolerance, unstable export formats, and a distance from true usability.

But this inventory task changed my mind.

Here's the background: our club applied for school project funding and needs to use a batch of supplies. The Youth League Committee requires submission of a standardized asset registration form. The original data was scattered across three places: over 30 entries in a shared Excel file, a dozen or so in a group chat relay message, and several invoice photos requiring manual entry. Previously, I'd copy everything into a temporary sheet, align each item one by one, and spend two to three hours on it.

This time, I tried WorkBuddy's import function. I dropped the Excel file directly into it, copied the relay message text and pasted it in, then uploaded two invoice photos. It has a feature that merges data from different sources into one table, automatically deduplicating and matching fields. I hit run, and it took less than a minute to produce a new summary table. All 63 records were there, with quantity, unit price, total amount, date, and remarks aligned in their respective fields.

I double-checked carefully. What surprised me was how it automatically parsed colloquial descriptions like "Zhang Wei, 3 USB drives, 45 yuan" from the relay messages into four fields: name, item, quantity, and unit price. It even recognized that "45 yuan" was the unit price, not the total. I didn't expect it to handle this detail correctly. The relay message format was quite random—some wrote "2 boxes of A4 paper," others just "printing paper x2"—yet it managed to map them all to the correct columns.

Of course, there were issues. One invoice photo was taken crookedly, and it misread the amount, dropping a zero (recognizing 80 yuan as 8 yuan). If I hadn't caught it during verification, the committee would have rejected the form. Also, regarding export formats: I chose CSV export to open in Excel, but one column's time format turned into text, messing up sorting. These small problems aren't fatal, but they do require manual review.

One feature where I think WorkBuddy beats similar tools is partial editing. I noticed this approach with Seedance 2.5 too—you can select a single record to modify without regenerating the entire result. For instance, when the invoice amount was wrong, I selected that row and told it "the amount should be 80," and it only changed that cell, leaving other data untouched. This experience is far better than the "regenerate everything" interaction, where you never know which piece of data might get corrupted.

However, what I most want to complain about is its handling of free-text fields like "Remarks." Notes on invoices or occasional comments in relay messages like "don't buy this yet" are often missed or misplaced. In this table, a remark saying "model pending confirmation" got dumped into the "Specifications" column. It's not a big deal, but if you expect it to completely replace human verification, you're dreaming.

So my conclusion is this: WorkBuddy has reached a level of daily usability for handling "structured data," saving roughly 70% to 80% of repetitive labor. But the prerequisite is that you need to clarify the boundary conditions yourself. This aligns with my previous view on AI coding tools: the more powerful the tool, the higher the demand for the user's data literacy. If you don't understand your own data fields or which ones might have pitfalls, no tool will help.

Right now, forums are buzzing about how strong coding models are—GLM-5.1 surpassing Claude, GPT-5.5 controlling computers. These are indeed impressive. But for me, what truly saves my life are office tools that handle chores efficiently. I think the trend going forward is this: AI coding tools will increasingly compete on code generation, while AI office tools will move toward "delivering results." Whoever cleans up that final step will be the product that truly lands. WorkBuddy is still far from perfect, but at least it knows which direction to go.