After Reviewing 46 AI Office Tools, I Kept WorkBuddy Pre-Reimbursement

I just saw that article from aitoollab recommending 46 AI office tools. The section on Feishu (Lark) 8.0 sounded exciting—you can search names in group chats to pull in Agents, multiple Agents can collaborate, write docs, operate multidimensional tables, check cloud drives, and initiate approvals. There was also ChatGPT Work Finance Edition, claiming it can generate valuation models and research reports with permission and compliance exports. Help! These lists scare me because every time I read one, it reminds me that other companies are already using AI to manage processes while I'm still staring blankly at invoice headers in Excel.

But after reading it, I wasn't tempted. Instead, I became more certain of one thing: Roles like administrative cashiers lack an execution layer that can finish repetitive actions and stop when things go wrong. Using WorkBuddy for a month, I placed it exactly in this role, and it fit perfectly.

Last Wednesday, I cleared a batch of invoices as usual. Taxi receipts, dining bills, conference venue fees—96 PDFs and screenshots mixed together. Previously, this task was grueling: open each invoice, log into Meike (expense system) to export reimbursed records, check for duplicates in Excel, and verify amount, date, header, tax ID, reimbursement city, and policy limits item by item. Smoothly, it took about two hours. But with tilted photos, stamps covering numbers, or split notes in remarks, time spiraled out of control. Amazing how fixed actions felt like daily repetition.

This time, I used WorkBuddy's inbox folder as the entry point. I had it read the files, then output fields like invoice number, issue date, amount, tax, buyer name, seller name, reimbursement type, and remarks. It didn't directly query Meike because our API permissions weren't set up, and I didn't want to casually send sensitive data out. I asked it to generate an exception list first, exported it as CSV, and threw it into a Feishu multidimensional table for review. In my tests, most fields for the 96 invoices were usable, with accuracy around 90%. The rest needed manual fixing. Tilted shots, compressed screenshots, and long remarks still caused misses. But the key is that it changed "opening each one" to "reviewing the exceptions it filtered." For a worker like me, this step is more practical than another Agent that writes meeting minutes.

Feishu 8.0 is indeed strong, especially in organizational collaboration. Its ability to create colleague-like Agents fits the imagination of team work. Pulling them into groups, @-mentioning them to check docs, build tables, or send approvals makes the process look smooth. I've used Feishu for three weeks and integrated Tongyi Tingwu and Feishu Meeting Minutes into daily meetings. The problem is, meetings and reimbursements are different beasts. Wrong meeting summaries mean rework; wrong reimbursement actions mean minor ticket corrections or major audit visits. So I don't want AI bypassing human confirmation to submit approvals directly. WorkBuddy seems restrained here, more like a screwdriver requiring my manual confirmation.

I tested this workflow for two more days with clear division of labor. WorkBuddy handles the dirty work on the file side: identification, classification, field extraction, and anomaly marking. Feishu multidimensional tables handle status and auditing trails—who changed what, why, and if reviewed. Feishu 8.0 Agents notify relevant claimants of anomalies and initiate flows after confirmation. This combo isn't flashy, but it feels like a functioning automated admin-cashier workflow. It places each AI within its competent boundaries.

This made me rethink those 46 tools. Airtable AI turns business data into interfaces and automations; Beautiful.ai makes presentations look good; OfficeCLI lets AI process Word, Excel, and PPT via command line; Octarine suits local Markdown notes. They all have their place. But the daily life of an admin cashier involves massive amounts of low-creativity, high-repetition, low-tolerance tickets and spreadsheets. Capabilities like PitchBook, LSEG, and Daloopa in ChatGPT Work Finance Edition are useful for investment research but useless for me checking if taxi receipts cross months. Valuation models don't help me; I care more about it not entering the same invoice twice.

What satisfies me about WorkBuddy is that it bridges the gap between file organization and business actions. I've tried many AI tools before; they're great at summarizing and generating professional-looking judgments. But summaries don't help deduplicate, and judgments don't fill out exception lists. If it outputs a structured CSV, I can continue with subsequent tables, approvals, and audits. This capability looks unsophisticated, even a bit rustic, but WorkBuddy's position is completing verifiable actions for me.

It has issues too. Financial domain knowledge is lacking. Rules like "reimburse intra-city transport based on actual occurrence" or "note reasons for cross-month claims" aren't natively understood; I have to break rules into fields, thresholds, and anomaly types. APIs aren't fully connected yet. WorkBuddy can extract fields and generate lists, but writing stably into the reimbursement system and waiting for approval status callbacks is difficult. This is more about permissions, configuration, and system boundaries. Undo and rollback features are missing. Admin scenarios fear auto-submission most; if wrong, we must retract and explain. Currently, I block this step with manual confirmation. The process slows down, but being slower is acceptable.

If I were to define the target audience for WorkBuddy, I'd say: people facing massive files, tables, and tickets daily who are willing to break down processes. Such people should use WorkBuddy to standardize inbox handling, exports, comparisons, and anomaly marking. I wouldn't recommend it to those hoping for one-click full automation, dumping sensitive financial data into external models, or lacking anyone responsible for review.

After a month, my view on AI office tools has changed. Before, seeing lists made me feel left behind if I missed a tool. Now, I think instead of scrolling through rankings, I should ask: Can this tool assume a clear role in my company? For me, WorkBuddy is more like an execution layer before reimbursement. Feishu 8.0 Agents are more like a collaboration layer, better suited for meeting minutes and approval flows. I occasionally use OpenAI and DeepSeek for text processing but won't let them touch ticket actions directly. Meike remains the system of record; WorkBuddy doesn't touch its books.

Going forward, I'll keep WorkBuddy at that pre-reimbursement step.