Extract purchasing-card receipt data into a report cover sheet

By: Stephen Toback. Task: Stephen Toback tested Claude Cowork, a desktop research preview, to automate manual receipt processing. He provided a PDF purchasing card receipt and requested data extraction and a TPS report cover sheet. Cowork executed a Python script in a local Linux VM, debugging errors autonomously to produce final files. The author monitored the chat interface but did not manually run code or install dependencies.; tools: Claude Cowork. Human role: The original post does not say how a person stayed involved.; limitations: Cowork is a research preview with known instability, particularly during network interruptions. It is tied to a single computer, lacks collaborative features, and does not sync session context across devices. Users must manage local VM resources and accept that output files sync via cloud storage, but workflow state does not. Human review is essential due to potential crashes and inconsistent states..

Tools

Claude Cowork

Published on Agent Work Atlas:

Overview

By: Stephen Toback

Workflow

Stephen Toback tested Claude Cowork, a desktop research preview, to automate manual receipt processing. He provided a PDF purchasing card receipt and requested data extraction and a TPS report cover sheet. Cowork executed a Python script in a local Linux VM, debugging errors autonomously to produce final files. The author monitored the chat interface but did not manually run code or install dependencies.

Reported result

The initial task succeeded, but the tool crashed during a connectivity loss, entering an inconsistent state where the local VM ran but cloud reasoning failed. Recovery required clearing app data and using external diagnostics. No quantitative metrics on time saved or accuracy were reported. The author concluded Cowork adds value by executing code for non-developers but is unstable for time-sensitive tasks.

Human role

The original post does not say how a person stayed involved.

Limitations

Cowork is a research preview with known instability, particularly during network interruptions. It is tied to a single computer, lacks collaborative features, and does not sync session context across devices. Users must manage local VM resources and accept that output files sync via cloud storage, but workflow state does not. Human review is essential due to potential crashes and inconsistent states.

Try it

Act as a workflow automation specialist. I need to process [replaceable: number] PDF receipts from [replaceable: folder path] into a summary spreadsheet and individual cover sheets. Use Claude Cowork to execute this. First, perform a dry run on [replaceable: 1 sample receipt] to verify data extraction accuracy and formatting. Show me the output files and the script logic used. Do not proceed to batch processing until I confirm the sample is correct. If the tool crashes or loses connectivity, stop immediately and report the error state; do not attempt to restart the VM without my approval. Ensure all output files are saved to [replaceable: destination folder]. Verify that no sensitive data is leaked in logs. If Cowork is unavailable, fallback to generating a standalone Python script for manual execution, but flag this as a deviation from the requested automated workflow.

Original source

Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.

More cases · Office & operations