AI Agent Data Errors Linked to Pasting Method, File Paths Improve Accuracy

An MCP server developer tested methods for AI agents passing financial return data to an audit tool. When numbers were pasted directly into prompts, accuracy dropped significantly and token usage surged. Using file paths instead of pasted data nearly doubled accuracy and reduced processing tokens by 75%. The errors occurred because models sometimes dropped numbers when copying large datasets into tool calls. The findings suggest tools requiring substantial data should use references rather than requiring models to copy values.