Compare agentic invoice routing with n8n

Felix rebuilt an invoice-classification workflow with an n8n AI Agent to test whether agentic routing improved the task. The earlier version classified uploaded PDF and image invoices, moved each file to the appropriate Google Drive folder and flagged uncertain documents in Slack.

The new version used Gemini and seven tools: six folder-move actions and one Slack alert. The agent received a file identifier, decided which category applied and called the corresponding tool. Felix says it worked, but he chose to retain the deterministic version for this process.

His reasons are concrete. The agent could route the same invoice differently across runs, made incorrect decisions harder to explain and added tool loops, missing parameters, schema mismatches and binary-transfer problems. He also rejected treating the agent's own confidence score as equivalent to a clean extraction failure signal. Both workflows were linked for comparison. The account records a completed agent experiment with an unfavorable deployment choice, showing where added reasoning created extra maintenance without a reported gain in capability.

Prompt
Original prompt

Source: dev.to · Felix