After two months with WorkBuddy, I'm convinced office AI is won on handling, not generation

What annoys me most is that nobody can pick up what AI writes.

I've been holding this in for a while. At the end of last month the club handed in annual materials, I threw a bunch of stuff at AI to process, and the summaries, lists, and summary tables it produced all looked decent, but when I handed them to the club president and the advisor, they scanned for two seconds and sent them back. Format was a mess, no sources found, and no way to tell which parts could be edited and which couldn't be touched. I thought right then — no matter how strong the generation capability is, when it comes down to actual collaboration, the bottleneck isn't generation at all.

Later I saw an e-commerce report and the thinking clicked all at once. Shero Commerce's September Commerce Jam report, covering 73 independent sites, about 96.4 million sessions. There's a counterintuitive phenomenon in it — AI-referred traffic basically didn't grow this year, but orders from AI referrals went from 5.9 per day to 16.1 per day, and conversion rate rose from 0.66% to 1.75%. Traffic scale didn't change, monetization capability did.

The report breaks it down one more layer — AI visitors landing on product detail pages convert at 1.77%, landing on the homepage only 0.57%, a difference of more than three times. The report is about e-commerce, but reading this, what popped into my head was office work. Same visitor, same intent, landing in different receiving structures, completely different results.

This analogy maps almost one-to-one onto office AI. Same instruction, output landing in a folder nobody can edit, with no permissions, no sources — it's waste paper. Landing in a workspace with boundaries, ownership, and versions — it's something usable. The value of office AI isn't in generation, it's in receiving. I wrote a piece on this a while back, and at the time it was still a bit vague; now after two months of running it, I'm confirmed.

I've used WorkBuddy for about two months, at first for club weekly reports, then gradually moved the whole material flow into it. This time for the annual materials I rebuilt it from scratch, let me note the process, good and bad.

First I created a workspace in WorkBuddy and split materials into three layers by source. Feishu-exported PDFs in one layer, WeChat screenshots and event photos in one layer, Excel sign-up sheets and Word in one layer. This step sounds dumb, but it's the most critical step in the whole thing. I used to dump all files into one folder and let AI find them itself, and it often confused the advisor-annotated Word with the un-annotated version, twice piecing together a summary missing the annotations.

After splitting, I set permissions for each layer. PDFs and photos set to read-only, Excel and Word set to editable. This way it won't mess with the originals while organizing, and new output files go into a separate output directory. From my testing, this setting blocks most mis-edits, much less hassle than checking one by one afterward.

I split the task into four steps, letting it do only one thing at a time. Step one, categorize — sort thirty-plus screenshots into different subfolders by event batch, took about ten minutes. Step two, extract — pull names, student IDs, and contact info from the Excel sign-up sheet into one master table, this step it did cleanly, format didn't break. Step three, summarize — piece together the activity records from the Feishu PDFs into a first draft of the annual summary by timeline. Step four, check — have it list a manifest against the original files, clearly marking which data came from which file and which page.

Four steps done, about forty minutes total. I did this manually once last semester, from 8pm to almost midnight, and even missed photos from two events. Forty minutes versus four hours — that's the direct reason I'm willing to keep using it.

But there are a few places it really can't do. Advisor-annotated Word — handwritten annotations aren't fully recognized, three of them I had to fill in myself. Another time I cut corners and stuffed all requirements into one sentence, the output was all messed up, categorized two batches wrong, actually took more time than manual. Later I learned — one thing at a time, and it's both faster and more accurate. This lesson is on me, not its fault, but beginners easily step in it.

Also, once there are many files in the workspace, retrieval slows down — from my testing, you start feeling it after about two hundred files, opening a subfolder takes two or three seconds. Not fatal, but if your materials are in the thousands, better to split into multiple workspaces by year or project first, don't pile it all in one.

So my judgment is, WorkBuddy suits scenarios with many files, mixed sources, and collaboration boundary requirements. Club materials, course archives, research group data, small team reimbursement attachments — it can receive these, and more steadily than general chat models. Its value isn't in helping you write a pretty sentence, it's in helping you put a sentence into a container others can also understand and continue editing.

Who's it not for? If you just want it to write a WeChat public account copy, make a poster, or revise a resume, then WorkBuddy's whole workspace logic is pure burden for you. You'll probably find it ten times more troublesome than ChatGPT, needing several extra clicks to get one sentence out. That's the truth, I won't cover for it.

That report mentioned at the end that the iteration of AI assistants' own capabilities is also changing who users will see and who they'll trust. Office work is probably similar. After I rebuilt the workspace, the club president didn't send it back this time.