Using WorkBuddy for Expense Audits: Moving from Chat Interface to Execution Layer

Yesterday I came across an article about WorkBuddy acting as an ops assistant. It said it can automatically perform inspections, troubleshoot faults, and renew certificates, with monthly active users reportedly reaching the 20 million mark. My first reaction was: help! The pile of reimbursement anomalies every Monday morning is back again. For admin and cashiers, what really saves time is having someone handle the repetitive actions. After using WorkBuddy for a month, my biggest takeaway is that it shouldn't be used to write essays for me; its real value lies in executing verifiable actions on my behalf. If AI only generates summaries, it's just secondary labor for admin and cashiers. I adapted the ops inspection mindset to reimbursements, and it took about a week to get things running smoothly.

Recently, when reconciling accounts, I'd open the CSV exported from Youcheng Reimbursement (used for two weeks) and cross-reference it with project definitions in Feishu meeting minutes, checking invoice images one by one. Up to 20 invoices were tolerable, but beyond that, it became a bottleneck. With WorkBuddy, I created a Inbox. This term is new here—it's essentially a local folder authorized for automatic reading; drop new files in, and tasks will scan them. To set it up: open the WorkBuddy desktop app, click Inbox on the left sidebar (in some versions, it might be under Automation), then click New, select Local Folder, name it Reimbursement Inspection Inbox, and authorize read access. Put exported details, Markdown minutes, and invoice screenshots inside. Next, click Automation on the right, create a scheduled task named Monday Reimbursement Anomaly Inspection, and set the trigger time to 9 AM on Mondays. Here's a pitfall: initially, I didn't disable cloud summarization. If the task uses a cloud model, it requires an API Key (interface call credential). Later, feeling that invoice amounts were too sensitive, I changed the task to use only local File Reading and Table Generation. My setup involves a company intranet plus local WorkBuddy, which may not apply to everyone, but don't dump raw attachments to external models if data is sensitive. I only include field tables after export and screenshot filenames.

When configuring instructions, beginners often write "summarize expenses for me," which doesn't work. I wrote very strict instructions: read all files in the inbox, match by Reimbursement ID, generate an anomaly table with fixed headers: ID / Amount / Invoice Status / Project / Anomaly Type / Suggested Action. Only five anomaly types are allowed: missing water bill, amount mismatch, project not filled, missing remarks, invoice date expired. Only three suggested actions are allowed: supplement materials, return to employee, enter approval draft. After saving, the task list shows Enabled. Click Run Now, and logs will scroll line by line on the right, changing status from Running to Completed in about 3-5 minutes. Under the Output tab, an Anomalous Reimbursements.xlsx file appears. This step is similar to ops inspections; the difference is ops look at CPU, certificates, and interface health, while I look at invoices, projects, and approval statuses. Ops get fault reports; I get anomaly lists, but I only want results that lead directly to the next action.

I split permissions finely. WorkBuddy cannot directly submit to Youcheng Reimbursement; it only generates a Pending Confirmation column. In Sharing, I set the output folder to Read-only. The finance manager views the anomaly table, employees see only their own rows, and I retain edit rights. Each anomaly must be clicked Confirm or Reject. Only after confirmation does WorkBuddy generate a Feishu notification draft. Previously, I accidentally enabled auto-notifications, and it sent "Suggest supplementing materials" to the employee group. Wow, Admin almost died of embarrassment. Later, I changed the rule to send notifications only to me, and I forward them manually.

Daily maintenance is also routine. Every Friday, I check Run Logs to see if any fields are misaligned. When I used DeepSeek for summaries for a month, it would also squeeze remarks behind amounts. Early versions of WorkBuddy did the same. I later enforced One Row Per Item in the template, filling empty values with NA. Before modifying templates, I copy a _Backup_0905. Those five seconds have saved me. When it breaks files, rolling back is much more dignified than re-reconciling.

WorkBuddy acts like an intern who organizes words in the chat box, but behaves more like process ops in the inbox, scheduled tasks, and confirmation actions. It turns "take a look" into "verifiable." Now I only let it generate anomaly lists and notification drafts; submission actions still require human confirmation.