An AI Spreadsheet Formula Needs a Unit, Not Just a Result

A generated formula should explain the quantities it connects, their units and the period measured before a plausible number becomes a business decision.

Define the quantity before requesting the formula

A request for an AI tool to produce a spreadsheet formula should describe what the result means. Naming a column is a beginning, not a complete definition. Revenue per customer, cost per unit and hours per project each depend on choices about what is counted and the period covered. A number can be arithmetically tidy while answering a different question from the one management intended.

This is a proposed review discipline, not an evaluation of a particular AI model or a measured claim about formula errors. Its argument is that the person accepting a generated formula should own an intelligible account of the quantities. The tool can help express a calculation; it should not silently decide the business meaning of the result.

Put the unit beside the input

Consider a hypothetical planner estimating work required for an order. One input gives minutes per item, another gives items per batch and a third describes hours available each day. Before using a generated calculation, the planner should explain how those quantities relate and which conversion belongs between them.

The purpose is not to make the spreadsheet elaborate. Clear labels and a short definition may be sufficient. The important distinction is between a value and the kind of value it represents. A result labelled simply capacity can conceal several possible questions: items per hour, batches per day or orders before a deadline. The formula should answer the one the planning decision actually needs.

Match the period on both sides

A calculation comparing quantities should make their time periods explicit. A monthly payment and an annual budget do not become comparable because both are expressed in the same currency. The reviewer should identify which period each input represents and what assumption permits the comparison.

For a hypothetical service report, one team might count customers active at the end of a month while another counts every customer served during it. Neither definition is automatically wrong. They describe different populations. If the formula uses one as though it were the other, the issue is conceptual before it is computational. The review should settle that definition before asking the AI to make the syntax more elegant.

Use an example you can explain without the spreadsheet

A small hypothetical example can expose the intended relationship. Choose values simple enough that the reviewer can describe the expected result in ordinary language and check the reasoning independently. The point is not to prove the entire workbook through one convenient case. It is to establish that the accepted formula expresses the relationship the reviewer thinks it does.

Then examine a different kind of case. What if there is no activity during the period? What if the denominator is absent? What if a value is a correction rather than a new transaction? The actual workbook needs an appropriate treatment for these situations. The treatment should be deliberate rather than whatever outcome happens to appear first.

Where currencies differ, the reviewer should identify the conversion assumption and the date or period to which it applies. A currency symbol added after calculation does not explain that choice. The workbook should show whether a comparison uses documented transaction values or an explicitly assumed rate for a hypothetical scenario.

Separate missing information from a measured zero

A blank cell can mean that information has not arrived, that a question does not apply or that nothing was recorded. A zero can be a legitimate observation. The workbook’s design should explain the difference where it matters to the result. A generated formula should not be accepted merely because it avoids displaying an inconvenient message.

Imagine a hypothetical utilisation report in which one team has not submitted its available hours. Treating that absence as a measured zero could answer a different question from the one the report is intended to show. The appropriate response might be an incomplete-result label or a separate review state. The choice belongs to the responsible designer of the measurement, not to a cosmetic preference for every cell containing a number.

A copied formula needs the same meaning in its new place

A formula that fits one row may be reused where the surrounding definitions differ. Before extending it, check whether the inputs still represent the same quantities. A convenience in construction should not become a reason to assume that every sheet follows the same business arrangement.

The review should also identify who can change the definitions later. A new reporting period, revised product unit or different customer classification may require more than editing a heading. Keeping a short explanation of the calculation helps a later reviewer notice that the underlying question changed. The workbook should be maintainable as an account of work, not merely preserved as a collection of expressions that once produced acceptable-looking totals.

Accept responsibility for the measurement

The final question is whether someone can explain what the result establishes and what it leaves out. The explanation should identify the quantity, unit, period, population and treatment of exceptions relevant to its use. It need not expose every spreadsheet detail to every reader. It should give the responsible reviewer enough information to challenge the calculation intelligently.

An AI-generated formula can be a useful contribution to that process. Its usefulness grows when the request is precise and the acceptance decision remains inspectable. A result deserves attention because its relationship to the business question has been examined, not because the tool returned a fluent explanation. Name the unit, settle the definitions and test the intended relationship before allowing the number to govern a consequential choice.

Gabriel Nicolaev — CyGuru

Editorial photograph for illustration.

FOLLOW THE CONNECTION

Another angle on this subject

Continue through related ideas, decisions and consequences.

Technology & AI

An AI Supplier Update Is an Operating Change, Not Just a New Feature

Procurement needs a plan for changed model behaviour, revised integrations and rollback before an AI service becomes embedded in essential work.

Technology & AI

An AI Workflow Needs an Owner for What It Keeps

An operating perspective on retained prompts, uploaded files and generated records: decide what deserves to remain before storage becomes the default.

Money & Business

A Business Partnership Needs a Rule for the Decision Nobody Can Win

A commercial essay on partnership deadlock: define authority, reversible experiments and a credible route through disagreements before the business depends on personal victory.