Model checks and responsible AI assistance

Overview

A manager at established distributor Rowan Supplies receives an AI-suggested profit formula. It returns a number without an error message. Before using it to approve a campaign, she checks whether it represents the business correctly.

Both businesses, figures, and proposed AI outputs are fictional. No AI account or spreadsheet is required.

Recall Lesson 031: Why could category totals reconcile even when a duplicated sale overstated revenue?

Decide what should be true

A model check compares a calculated result with a stated expectation. The expectation might be an independently calculated answer, a total from source records, or an operating limit. Decide the expectation before looking at the model's answer.

A reconciliation check compares amounts that should agree. For example, total revenue should equal the sum of revenue across all included months. Subtract one total from the other: the expected difference is zero.

ICAEW recommends visible checks and systematic testing, including extreme or impossible inputs. ICAEW: Twenty principles for good spreadsheet practice

Use more than one check. Matching totals do not establish correct prices, and correct arithmetic does not establish demand.

Worked example

Rowan is considering a one-month campaign selling 30 identical kits at $50 each. Variable cost is $30 per kit; the campaign also incurs a fixed $300 expense. Assume all kits sell, sufficient capacity and funding, no displaced work, and no other additional costs or income taxes. Existing business results remain unchanged.

First calculate by hand. Revenue is 30 × $50 = $1,500. Variable costs are 30 × $30 = $900. Additional profit is $1,500 − $900 − $300 = $300.

Suppose an AI proposes =B2*B3-B4-B5 for profit in B6. This gives 30 × $50 − $30 − $300 = $1,170. It subtracts the per-kit cost only once instead of multiplying it by 30 kits.

The correct formula is =B2*(B3-B4)-B5. Parentheses calculate contribution per kit first: $50 − $30 = $20. Then 30 × $20 − $300 = $300.

Put the independently calculated $300 in C6, labeled “Hand-check expectation for 30 kits.” In D6, enter =B6-C6. The incorrect formula gives an $870 difference; the corrected formula gives zero. Restore the stated inputs before rerunning this test.

Test the rule, not just one answer

One correct result can be accidental. If quantity were one, both Rowan formulas would return −$280. Testing only that case would miss the error.

Try zero sales while retaining the committed $300 campaign expense. Correct profit is −$300. The incorrect formula returns −$330 because it charges one kit's variable cost even when no kits sell.

For a rule with a cutoff, test values below, at, and above it. In Lesson 032, equipment rental started above 100 bags. Testing 99, 100, and 101 would expose a formula that mistakenly charged rental at exactly 100.

Checks need independent reasoning. Copying the same faulty formula into a second cell and comparing the cells gives no useful evidence. Likewise, if every month's profit omits a cost, adding those profits correctly still produces an incorrect business total.

Use AI with a reviewable request

Give an assistant the cell map, units, period, business rule, and limits. Ask for a formula, a plain-language explanation, and test cases. For Rowan, specify that B4 is cost per kit, B5 is charged once per month, and zero sales still incurs B5.

Treat the response as a draft. NIST identifies the risk that generative AI can confidently produce incorrect content. Confidence in the wording does not validate the calculation. NIST: Generative Artificial Intelligence Profile

Work one example yourself and inspect references, parentheses, and boundaries. Asking the same assistant to reassure you is not an independent check. If you cannot explain a formula, simplify it or obtain qualified review.

Use synthetic examples or the minimum authorized data needed. Keep the original version, record the correction and its reason, and rerun checks after edits.