Why a Personal Expense Bot Moved From a General AI Agent to a Deterministic Backend

The author first built an expense tracker on a general-purpose agent harness backed by a large model, which spent over 1,000 tokens just to log a single coffee purchase. The broad harness reloaded instructions and tool catalogs on every message, and its open-ended behavior caused context buildup, slower responses, and inconsistent categorization of repeated entries. The fix was a deterministic backend application, with a model used only at the entry point to convert voice or text into structured data. This cut model costs to nearly zero, removed latency, and made the data consistent. The experience led the author to a 4-level intelligence pyramid for balancing autonomy against strict rules in production AI systems.