Suppose an AI assistant builds a rolling forecast. The actuals tie out, the formulas calculate and the summary looks ready for the board pack. Then the team closes another month and adds a cost center. The new expenses appear on the department tab, while the consolidated summary still stops at the old range. The file still opens, and the check cells stay green. Before that workbook becomes part of the monthly routine, finance needs to put it through the changes the team will actually make and check what comes out.

In MBABench-2.0, submitted on October 22, researchers asked AI systems to build 101 financial spreadsheets from scratch. The leading tested system passed 16% of tasks under a workbook-quality rubric that included accuracy, flexibility and documentation. The study used an AI judge validated against experts and didn't include a human-performance baseline. Its results came from one recorded attempt per task and depended on the model and tools used. That pass rate describes acceptance of the whole workbook under the study's rules.

ICAEW's spreadsheet principles, updated in 2024, recommend adaptable designs and testing that reflects the workbook's risk. For a rolling forecast, start with a separate test copy and move it forward one month. Load the next actual period, advance the forecast cutoff and extend the horizon. Check that the closed month contains actuals, the remaining forecast still covers twelve months, and the opening cash balance agrees with the prior month's closing balance. Follow the new month through supporting schedules, summary formulas and charts, including the ranges that feed the board summary.

Then add a cost center with a small, known set of expenses. Its payroll, occupancy and allocated costs should appear once in the department view and once in the consolidated total. Check the allocation denominator too: adding a department can change how shared costs are distributed across every other department. Remove the test cost center and confirm that the original totals return. This catches formulas with fixed ranges, missing lookup entries and allocations that leave part of the expense unassigned. If the workbook needs a structural edit for this routine addition, record the edit and test the revised version before the team starts repeating it every month.

Change an assumption whose effect can be calculated independently. Suppose the test forecast has 100 employees working the full month, with salary expense spread evenly across twelve months. Increasing each annual salary by $1,200 should add $10,000 to monthly base salary expense before benefits and payroll taxes. Calculate that expectation outside the generated workbook, then compare the relevant schedules and consolidated result. Change the assumption back and confirm that the baseline returns. A formula that references a hidden constant or a pasted value can produce a convincing original answer and fail this simple exercise. Test a few drivers separately before combining them into a scenario.

The error checks need their own tests. In the test copy, remove a required assumption, enter text in a numeric input and deliberately break a reconciliation. Decide beforehand whether each condition should stop calculation or produce a visible exception, then confirm that it does. A green check cell is reassuring until somebody tests it. Inspect whether the summary flag captures formula errors as well as failed reconciliations, and whether a missing input becomes an unexplained zero. Finance should also check selected outputs against an independent calculation or an agreed business expectation. Two schedules can agree perfectly because both reference the same incorrect assumption.

Give another analyst the saved workbook, its instructions and the test inputs, and ask them to reproduce the update without the original chat. They should be able to identify editable assumptions, refresh the calculations, trace an output and explain any exceptions. Keep the expected results and review notes with the approved version so a later change can be checked against the same cases. This also gives teams using AI agents and their operating tools something concrete to rerun after a model, prompt or tool update. The analyst's work belongs in the delivery estimate alongside the initial build.

By next quarter, a team may have several AI-built workbooks moving through the forecast cycle. Each needs a defined purpose, an owner and evidence that ordinary changes produce the expected results. When evaluating AI tools for FP&A, start with one recurring model and preserve the tests that expose its weak points. When the next version arrives, run it through the same month, cost center and assumption changes before accepting it. Over time, finance can compare tools on how much usable, maintainable work they deliver, including the effort needed to keep that work running when the calendar changes.