How to Check an AI-Generated Spreadsheet Before It Reaches the Board
Before using an AI-generated spreadsheet for a board decision, check where its inputs came from, whether its formulas represent the business correctly and how its outputs respond when assumptions change. Reconcile the important totals, inspect the saved workbook and record what remains uncertain. A spreadsheet is useful when a reviewer can understand what drives its conclusion.
Imagine a new branch proposal. The workbook is neatly formatted. Its dashboard shows revenue, operating profit and a promising return. The supporting commentary recommends proceeding.
One assumption deserves immediate attention: the model treats every sale as cash received in the same month. Customers are expected to pay later, while rent, wages and some suppliers must be paid earlier.
The profit calculation might be internally consistent. The funding requirement could still be understated. The question is whether the model represents the decision well enough to support it.
Start with the decision, then inspect the model
A model should make clear what decision it supports and which outputs matter to that decision.
For a new branch, those might include the investment required, the point of maximum cash need and the sales level required to cover operating costs. For a project, they might include the cost of a delay and the effect of different delivery options.
Ask what would change the recommendation. That identifies where review effort belongs.
If the proposal depends on reaching a particular sales volume, inspect the volume assumption and its supporting evidence. If it depends on funding availability, examine cash timing and the constraints on that funding.
The review should follow the business logic from the assumption to the decision.
1. Establish where the inputs came from
Separate historical actuals, management estimates, external assumptions and values inserted to make a demonstration work.
A well-presented model can make these categories look equally authoritative. They are not equally evidenced.
For material inputs, identify the source, period and units. Confirm whether a figure is monthly or annual, whether money is expressed in rands or thousands of rands, and whether amounts are inclusive or exclusive of relevant taxes.
Pay particular attention to missing information. A blank field, a zero and “not yet known” mean different things. Replacing an unknown cost with zero can make a proposal look attractive without improving the evidence behind it.
These are established spreadsheet disciplines. ICAEW’s guidance recommends a clear flow from inputs through calculations to outputs, with appropriate checks and review. AI-generated work benefits from those same disciplines. Source: ICAEW’s Twenty Principles for Good Spreadsheet Practice.
2. Recalculate one material relationship independently
Choose a calculation that matters to the recommendation and work through it outside the workbook. This gives you a known result against which to compare the model.
Here is a fictional, deliberately simplified operating example:
Monthly sales volume: 1,000 units.
Selling price: R500 per unit.
Variable cost: R300 per unit.
Monthly fixed operating costs: R150,000.
Revenue is 1,000 multiplied by R500, giving R500,000. Variable costs are R300,000, leaving R200,000 before fixed operating costs. After those fixed costs, the operating surplus in this simplified example is R50,000.
The contribution per unit is R200. Dividing fixed costs of R150,000 by R200 gives a break-even volume of 750 units.
Now reduce sales to 700 units while holding the other assumptions constant. Revenue becomes R350,000, variable costs become R210,000 and the result becomes a R10,000 operating deficit.
The model should reproduce those outcomes. If the dashboard still shows a surplus, investigate which input, formula or display failed to update.
This example excludes tax, financing, depreciation, investment and working capital. Its purpose is to test one relationship. A full business model must address the additional mechanics relevant to its decision.
3. Change an assumption and follow the effect
A useful model should respond coherently when an input changes.
Reduce volume. Increase a material cost. Delay the start date. Change a collection assumption. Then follow the effect through the calculations and into the decision summary.
Look for results that remain fixed when they should move, charts that use an outdated range and formulas that refer to a different scenario from the one selected.
Test a difficult but plausible case as well as the preferred case. Ask what happens if demand arrives later, if a key cost rises or if implementation takes longer. The purpose is to understand the conditions under which the recommendation changes.
Record the assumptions behind each scenario. Three scenarios are only informative if the reader can explain why their results differ.
4. Distinguish profit from the cash required
A business can show an operating profit while needing additional cash to fund its activity.
Return to the fictional example. Suppose customers pay the month’s sales in the following month, while all the stated variable and fixed costs are paid in the current month. Assume no opening customer receipts or other cash movements.
The first month then requires R450,000 to cover R300,000 of variable costs and R150,000 of fixed costs, before collecting those sales. The operating surplus of R50,000 does not remove that timing requirement.
Real businesses have more complicated terms, opening balances and payment schedules. The review question is whether the model represents them explicitly enough to reveal the funding need.
Check the opening cash position, the timing of receipts and payments, and the lowest projected cash balance. Where the workbook contains linked financial statements, examine whether changes flow through them consistently.
5. Check the file that will be used
An AI assistant may describe its intended model accurately while the delivered workbook contains a broken reference, an incomplete formula range or a summary that no longer matches the calculation.
Open the saved file in the spreadsheet application the team will use. Confirm that it recalculates, that the scenario controls work and that the important outputs match the underlying sheets.
Review formulas as well as displayed values. A manually entered result can look correct for the initial assumptions and fail as soon as the user changes them.
Distinguish different kinds of verification. Checking that formulas and references were written correctly is a structural check. Recalculating the workbook and comparing its results is a different check. Neither establishes that a forecast will occur.
How ExecutiveNavigants builds this discipline into the workflow
ExecutiveNavigants’ advanced Excel tool separates the proposed model structure from the software that builds the workbook.
The AI develops the structure and assumptions. A dedicated engine validates that structure, evaluates the model and builds the Excel file. The workflow includes a review of the model before the workbook is produced.
The builder also reopens the saved workbook and checks defined structural properties, including formula placement and references. Where relevant, model validation includes balance checks. These are specific engineering checks; they should be described according to what they actually test.
The resulting workbook still needs a business owner who understands the inputs, the intended use and the limitations. Software can faithfully calculate an assumption that management has not yet substantiated.
Put the assumptions beside the recommendation
Before a model supports a board paper, give the reviewer a short account of its purpose, the assumptions that drive the outcome, the checks performed and the unresolved questions.
Name the person responsible for the model and identify the version used for the recommendation. If a subsequent change affects the result, the reader should be able to establish what changed and why.
AI can reduce the effort of building and explaining a spreadsheet. The opportunity is to use some of that time to examine the business more carefully: what must be true, how the result changes and whether the organisation can live with the downside.
Explore the advanced Excel modelling workflow at ExecutiveNavigants.