After a month with WorkBuddy: I see it as a file pipeline, not an office assistant

I have a somewhat unpopular judgment lately: tools like WorkBuddy, which are AI office assistants, should primarily redefine themselves as auditable data preprocessing. They shouldn't rush to generate PPTs, write weekly reports, or take over external collaboration portals, nor do they need to package results prettily. Their better fit is compressing the messy pile of files on my desktop into a data table that can be queried, calculated, and reviewed.

I recently saw an image claiming Google shoved AI image generation and fine-grained editing directly into Workspace. The competitive focus isn't just models, but Workflow.

That statement holds up, but my first reaction was exhaustion. I'd just finished overtime, eyes blurry, and seeing 'integrated into office workflows' made me think of buttons, dashboards, templates, and auto-generated images. The bottlenecks in office workflows are often invoice amounts, duplicate supplier quotes, or contract payment milestones buried on page X. The more AI fills the gap with generation capabilities, the more I worry it conflates 'looks done' with 'data is usable.'

I've used WorkBuddy for about a month, and the biggest change is re-understanding where it fits in the stack. I actually care less about whether it can one-click takeover chat portals. I've used WeChat for about a month, and just started touching Enterprise WeChat, Feishu, and DingTalk in the last few days. To me, they are notification channels at best, not direct sources of data conclusions. I prefer placing it in a traceable processing pipeline, letting it do one thing: eat files, spit out structured data.

Last week, a client dumped another zip file on me. Inside were three scanned PDFs, two invoice screenshots, a chaotic Excel sheet, and a quote. This kind of task used to be the most annoying, specifically because of the dirtiness. You have to open each file, check dates, suppliers, amounts, taxes, notes, and merge them into a CSV. In my tests, doing this manually takes about forty minutes. Along the way, you deal with file naming, duplicate rows, date formats, and currency units—it's easy to burn an entire evening.

This time, I didn't throw it straight into the master table. Instead, I created an independent batch ID for these files, called batch-001, containing only unprocessed files. I ran WorkBuddy on this batch. I didn't ask it to generate a report or beautify tables; I only asked it to extract four fields: date, supplier, amount, and tax. Output: CSV. The first run wasn't great; one of the scanned PDFs recognized "1,200.00" as "12000," and the notes column was dropped entirely. It loves throwing summaries at me, as if knowing I'm working overtime and trying to offer some humanistic care. But what I need is auditable data.

Later, I learned my lesson: make it do one thing at a time. Extract first, don't aggregate. CSV first, not Markdown. Fields first, not reports. I defined a very narrow field constraint using Pydantic and JSON Schema: dates must be YYYY-MM-DD, amounts must be numbers, suppliers cannot be empty. I've used Pydantic and JSON Schema for a month—not expert level, but enough to catch obvious errors. Then I forced WorkBuddy to output only structures conforming to this schema. Results became much more stable. In my tests, processing eighteen files in a batch with default aggregation takes about six minutes and often merges amounts incorrectly. Switching to extracting CSV first, then aggregating via SQL and Pandas, produces the table in about two minutes, followed by eight minutes of manual review. Saving thirty minutes net isn't huge, but if it happens three times a week, you save an afternoon a month.

This is my understanding of WorkBuddy's architecture. It looks like an all-in-one office assistant, but it's really more like an extraction pipeline. Input layer: PDFs, CSVs, Excels, images, email attachments. Processing layer: recognition, extraction, cleaning, classification. Output layer: tables, JSON, reports, to-dos. Many users skip the first two layers immediately, demanding the final pretty result. The outcome is complete reports with suspicious data. The easiest way office AI fails is by packaging unauditable guesses into deliverable conclusions.

WorkBuddy's greatest value to me is turning dirty files into computable objects.

So I set some crude but effective rules for it. Every batch gets an independent batch ID and timestamp; output filenames must include the batch ID and timestamp. Failed recognitions go to a separate 'pending review' directory for manual inspection. All amounts, dates, and supplier names must retain source filename and page number for traceback. It cannot generate final reports directly, only auditable intermediate tables. The last rule came from a mistake I made. Previously, WorkBuddy helped me merge quotes; it looked neat, but two supplier names were similar and amounts close. I almost treated a duplicate quote as normal fluctuation. Later, I required it to include source_file on every row before I dared continue using it.

I also compared it with Excel Copilot. I've used Excel for a month and tried Copilot during this period. It's strong within the spreadsheet: formulas, pivots, and cleaning hints are smooth. But if the data isn't in Excel but scattered across PDFs, screenshots, and email attachments, it's not fast enough. WorkBuddy has a wider entry point, suitable for compressing miscellaneous files into tables first. Tools like Formula Bot and CHATEXCEL—I've seen comparisons—one leans toward data prep and formula generation, the other toward conversational table editing. Both have value, but my pain point is that the data before the formula hasn't been organized yet. So I let WorkBuddy handle the entry and the dirty work, then pass clean CSVs to Excel Copilot or Pandas.

I also tried combining it with n8n and Webhooks. I've used n8n and Webhooks for a month. The idea is simple: WorkBuddy parses files, n8n triggers and distributes. For example, when a file enters the batch directory, WorkBuddy outputs a CSV, n8n checks for abnormal amounts, and sends me a notification. I treat this flow strictly as a review signal, sending notifications only to my own computer. Because pushing unreviewed extraction results to departmental portals turns saved ten minutes into an hour of rework. Recently, I saw a WorkBuddy deployment case study mentioning it connects to Enterprise WeChat, QQ, Feishu, and DingTalk, and is compatible with the OpenClaw skill system. I just touched OpenClaw for the first time in the last few days, and I'm new to QQ, Feishu, Enterprise WeChat, and DingTalk too, so I'm not qualified to judge. But my judgment is clear: stabilize the extraction pipeline first, then talk about other entry points. If a tool can turn messy files into auditable tables, that's already great. Expanding entry points is an organizational process issue; I'll secure the data quality layer first.

This has changed my understanding of automation. Previously, I thought automation meant end-to-end: file in, report out, no humans involved. Now I think end-to-end is dangerous. More stable is a minimal process: one input, one output, one review point. WorkBuddy handles the dirtiest middle-mile transport; humans make boundary judgments. After it extracts, I scan amounts and suppliers. After it merges, I query duplicate keys with SQL. After it generates, I adjust formatting with Excel Copilot. Don't let it finish everything in one breath. It's fast, but its speed doesn't equal my correctness.

Looking ahead, I think what WorkBuddy-type tools should compete on next is auditability. Big tech shoving AI image gen, editing, and computer operation into office flows looks lively, but enterprise offices lack trust capabilities. If it's unclear which page a field came from, or whether an OCR altered an amount, the smarter the tool, the messier it gets. I hope WorkBuddy adds field-level provenance in the future, where every extraction result links back to the original file location. I also want diffs, showing differences between this extraction and the last. And sandbox mode, where write operations default to preview-only, not landing on disk. As for natural language PPT generation, auto-writing weekly reports, or one-click delivery, I don't expect that from it yet.

I dislike framing AI office tools as liberating humanity. It doesn't have that grand meaning. If it saves me from copy-pasting twenty times when I receive a client's zip file at 9:30 PM, it's already worth the price. Over this month, I haven't worshipped it as an 'office brain.' I placed it in a very narrow role: an extraction pipeline. Dirty files come in, clean data goes out, judgment stays with me.