Case Study
Agentic Web & Desktop Reporting Workflow
A fully autonomous agentic workflow powered by Anthropic's Computer Use API. It eliminates the error-prone, manual process of pulling monthly advertising performance data across heterogeneous platforms, mapping it directly into enterprise SharePoint reports.

Updating complex Excel reports requires analysts to navigate multiple ad platforms (Amazon, Meta, Google), adjust attribution windows, and carefully inject data into live formulas. Utilizing the official Anthropic Computer Use architecture (often referred to in the desktop environment as Claude Cowork), this agent automates the entire sequence—visually navigating UIs, fetching data, updating the workbook in-place, and drafting executive commentary.
Our implementation follows Anthropic's official reference architecture for the Computer Use tool. It utilizes an iterative "Agent Loop" where the model receives visual context, decides on an action, and our system executes that mouse or keyboard command in a secure, sandboxed environment.
Analyzes screenshots and requests tool actions (e.g., left_click, type).
Translates tool requests into physical OS-level mouse and keyboard actions.
Virtual display running Chrome, Excel, and local applications.
Automatically pulls performance data across Amazon Ads, Meta Marketing, Google Ads, and Shopify Admin APIs.
Handles differing attribution models and standardizes calendar boundaries without manual intervention.
Leverages the official Anthropic Computer Use tool (computer_20251124) to visually navigate interfaces and type directly into spreadsheets.
Updates existing Excel workbooks in-place, adding new month columns without overwriting historical records.
Spot-checks totals against source data and validates cell formulas to ensure 100% reporting accuracy.
Drafts automated Outlook notifications summarizing Month-over-Month (MoM) and Year-over-Year (YoY) performance.
Removes the risk of delayed reports by running entirely autonomously on a scheduled trigger.
Manual reporting processes suffer from single-person dependencies, attribution drift, and silent formula errors. Adding a new month column in complex Excel files often leads to broken references, misplaced pastes, and inaccurate Year-over-Year (YoY) metrics.
We designed a robust agentic workflow that connects directly to the underlying APIs. The system intelligently handles rate limits, respects locked attribution windows, and safely modifies SharePoint documents. It includes a built-in validation engine to ensure totals match before drafting a review notification.
Reporting time was reduced to zero. Formula errors and attribution drift were completely eliminated, providing leadership with 100% accurate, timely data on the first business day of every month.