Specialized AI Agents Can't Do Spreadsheets. General Agents Can.

A new benchmark tested AI agents on real financial modeling tasks — DCF valuations, 3-statement models, debt schedules.

The most surprising finding: Claude's general web interface scored 69.1. Claude's specialized Excel add-in scored 60.4. ChatGPT's Excel add-in scored 59.0. The tools built specifically for spreadsheets performed worse than the general-purpose interface.

The deeper finding: the same Claude model accessed via API with an agentic framework scored 18.2. The web interface scored 69.1. A 3.8× gap — same model, different harness.

The failure mode: specialized Excel agents compute correct numbers but hardcode them into cells instead of building formulas. The workbooks look complete but are professionally useless — no one can audit, edit, or extend a model full of magic numbers.

On hard tasks, golden solutions contain 20,925 formulas. The best agent produced 3,138 — covering 15% of the reference.

For enterprise teams evaluating AI tools: a product labeled "AI for finance" may underperform a general-purpose model with a better-engineered harness. The label is marketing. The harness is performance.

Same model, different wrapper, 3.8× performance gap. Your tool selection process should be testing the harness, not the model.