A survey workbook contains 100 submissions. The AI summary says, “Thirty-eight percent of respondents never used the feature.” It may have performed flawless arithmetic on a bad premise: 38 cells were blank, not recorded zeroes.
Some people did not answer. Others were correctly routed past a question that did not apply to them. A third group may have lost data during import. Zero is an observed value. Blank is a data condition whose meaning has not yet been established. Asking the model to “double-check the math” will not recover that distinction.
The risk is not unique to generative AI. The official pandas guide documents NaN, NaT, and NA, notes that reductions generally skip missing values, and shows a missing-only series summing to 0 by default. Code can behave exactly as designed while the denominator answers the wrong question.
Give each column a missingness rule
Do not begin with “fill with zero” or “drop the row.” First identify one of three states. An observed zero means the field was measured and the result really was zero, such as no returns this week; it can usually remain in the valid denominator. An unknown is an unanswered survey, failed sync, or missing source field. Keep and count it, but do not turn it into zero. Not applicable means the row was never eligible for the field, such as asking a non-user for monthly usage. It is a boundary on the question, not a smaller value.
The UK Government Analysis Function warns that a blank does not explain why data is absent, so readers may assume zero or a mistake. Its machine-readable guidance permits blanks when metadata explains the reason. AI needs that explanation before analysis, not under the finished chart.
Decide meaning before calculation
A blank, N/A, a dash, and an explicit 0 may all appear in one workbook. Their meaning can change by column and source system, so a global find-and-replace is not a data policy. Apply this branch separately to every field used in the answer:
- What did this empty cell mean in the source process?
- The measured result was confirmed as zeroKeep 0, retain the evidence, and include it in the valid denominator
- The value was not collected or failed to arriveKeep it unknown, count it separately, and do not impute or conclude
- This row was not eligible for the questionExclude it from this denominator; document why
If the workbook documentation, survey routing, source-system definition, or data owner cannot resolve the branch, stop at unknown. Similar-looking rows are not evidence for imputation.
The W3C CSV on the Web recommendations likewise use metadata to annotate tables, columns, cells, and null markers. A bare CSV carries cell strings, not enough business meaning to infer why a value is absent. Metadata makes the interpretation reviewable.
Put the denominator next to the answer
Consider those 100 survey records again. Sixty-two people supplied a valid usage count, including 24 explicit zeroes. Twenty-eight were routed past the question as not applicable, and 10 did not answer.
“Twenty-four percent of received records show zero use” divides by all 100. “Among 62 valid responses, 38.7% report zero use” divides by valid observations. They answer different questions: adoption may need the second, while collection quality makes the 10 unknowns the result. AI must not choose that policy for the reader.
A deliverable summary should put the source row count, filter scope, valid observations, explicit zeroes, unknowns, not-applicable records, exact denominator, and classification rule together. Record the rule version when comparing periods: a trend may move because this month’s import began treating dashes as null. This is the column-level companion to fixing the source and definition before AI answers company numbers.
NIST’s AI RMF Playbook MAP 1.1 calls for purpose, context, assumptions, and limitations to be documented. Here that means: Whose behavior is being counted? Who is eligible for the denominator? Which values remain unknown? Until those answers exist, AI may inventory gaps, but should not publish a rate or performance judgment.
Preserve the raw sheet
Even after some blanks are confirmed as zero, preserve the source field and add a derived field with rule, version, date, and approver. For health, credit, insurance, hiring, performance, or safety data, do not let a general assistant invent values or receive confidential files through an unapproved tool. Test a de-identified sample and require domain review.
For handoff, include the source, scope, owner, and limitations described in a transferable AI summary. Do not make the next analyst reverse-engineer the denominator from a chart.
AI handoff card
Read only the workbook I authorized. Do not fill blanks, delete rows, change formulas, overwrite the file, or issue final rates. Identify sheets, ranges, columns, formulas, blanks, explicit 0, N/A, dashes, and source codes. Per relevant column, report data type; row, nonblank, zero, code, and blank counts; plus any definition, survey skip rule, or source document. Classify each marker only as confirmed zero, unknown, not applicable, or unresolved, and show numerator and denominator effects. Without evidence, keep it unresolved and never infer from neighboring rows. End with at most five owner questions and a derived-field plan that preserves raw values. Stop until I confirm the rules.
The mistake in four panels
- The analyst hands the AI a bounded stack of survey sheets containing both deliberate circular zero markers and blanks that have not been interpreted.
- The AI sweeps every blank into the zero tally, making the blue counters and chart balloon far beyond the original task.
- She raises a clear stop hand, blocks the conveyor, and uses a coral divider to separate observed zeroes, blanks, and not-applicable records.
- She completes one modest result from verified zeroes only; every unknown card remains intact in a side tray for the data owner.
An empty shelf slot can mean sold out, waiting for restock, or never assigned to a product. A careful stocktake does not record all three as “zero sales.” Give spreadsheet blanks the same respect before an AI summary turns them into a decision.
References
- pandas documentation: Working with missing data — https://pandas.pydata.org/docs/user_guide/missing_data.html [accessed: 2026-07-28]
- UK Government Analysis Function: Releasing statistics in spreadsheets — https://analysisfunction.civilservice.gov.uk/policy-store/releasing-statistics-in-spreadsheets [published: 2021-06-30]
- W3C Recommendation: Metadata Vocabulary for Tabular Data — https://www.w3.org/TR/tabular-metadata/ [published: 2015-12-17]
- NIST AI Resource Center: AI RMF Playbook, MAP 1.1 — https://airc.nist.gov/airmf-resources/playbook/map [accessed: 2026-07-28]
