RMEP – Portfolio Reporting Automated with AI Agents

Starting point

RMEP reports monthly to its investment committee and shareholders. The figures of several portfolio companies were spread across Excel files and a complex cash flow model.

Every reporting round meant a lot of manual work: collecting, checking, preparing, and formatting data. The process depended on individual people and was hard to scale, yet errors were not an option.

Solution

An agentic end-to-end pipeline built on Claude Code and Agent Skills that starts automatically every month:

• Pulling the figures from the data sources via API
• Four agent skills: completeness, consistency, valuation, and PDF generation
• All calculations in deterministic Python tools, not in the language model
• A hard QA gate with sum and consistency checks: if a check fails, nothing goes out
• Human approval before the report is sent
• Controlled self-repair via automatic pull requests

GitHub Actions serves as a serverless runner for scheduling and versioning. The human-in-the-loop approval stays in place on purpose: the AI prepares, the human decides.

Tech stack

• Claude Code + Agent Skills (agentic execution layer and report generation)
• GitHub Actions (serverless workflow runner, scheduling, versioning)
• Python (deterministic calculation and QA tools)
• Google Workspace and other data sources via API
• GitHub pull requests (controlled self-repair mechanism)

Results & impact

• First fully automated monthly report in August 2026; a human reviews and approves
• Roughly a tenth of the cost (estimate): instead of several hundred euros in working time per report, running costs are now €20 to €50
• Bottleneck removed: the report no longer depends on individual people
• Versioned and traceable, with QA gates before every send-out
• Blueprint for further AI automation at the client, such as ad-hoc analyses via chat directly on the portfolio data