My underwriter asked the AI that built his rent roll model if the numbers were right. It said yes. It had checked everything. It had extreme confidence. But should he trust it? A June 2026 benchmark tested the top AI models on real business spreadsheets. The best one got 34.9% of tasks fully right. On finding and fixing errors, it got 12%. The five failures I see most: • Extraction: wrong number pulled. • Categorization: right number, wrong bucket. • Formulas: broken ranges or typed-in numbers. • Units: monthly vs. annual, dollars vs. thousands. • Reconciliation: totals don't tie back to the source. The sixth one keeps the other five alive. You ask the AI that built the model to check its own work. In one study, a model counted 19 rows in a 20-row table. It rechecked three times. It gave the same wrong answer all three times. My rule: run a separate checker in a fresh session, on a different model when you can. I have my subagents do this work. Then maybe trust the output. The same as with any big decision: you still have to spot-check it. I do it with my employees; why would I not do it with my AI agents? The video walks through each error and what to look for. What other errors are you seeing when working with AI and spreadsheets?