How to verify inconsistent numbers in AI-generated work

Freeze the claims before changing them

Copy every conflicting numerical claim into a grid exactly as presented. Assign a claim ID and preserve its location, including sheet, cell, chart label, paragraph, or AI response. Do not edit the source artifact during collection. The NIST Information Quality Standards distinguish utility, objectivity, and integrity; preserving the original claim protects the integrity of the review.

For each claim, record:

This inventory often reveals that the values are not competing answers to the same question. One may be year-to-date revenue in thousands while another is quarterly revenue in whole currency units.

• quantity and stated unit
• reporting period and cutoff time
• population, geography, or other filter
• source version and retrieval date
• formula or transformation, if known
• rounding rule and displayed precision

Sources: [1]

Normalize definitions before recalculating

Choose a comparison basis only after the decision owner confirms the intended definition. Convert units, align periods, and reproduce filters in separate columns. Keep the original values alongside the normalized ones. The NIST AI 800-3 Cybersecurity Profile for Artificial Intelligence addresses the need to manage AI-related data and system risks; in numerical work, traceable data provenance is a practical control against confident but unsupported output.

Do not ask an AI system merely to “fix the numbers.” Give it the definitions and source excerpts, request a proposed reconciliation, and independently recompute decisive values. AI can help enumerate hypotheses, but the spreadsheet formula, source system, or signed record remains the authority chosen by the task owner.

Sources: [2]

Work a reconciliation grid

Consider a synthetic quarterly plan with three statements: a table says projected subscriptions are 12,480; a chart label says 12.8k; and the executive summary says 13,200. The grid shows that the table uses active accounts at the August 31 cutoff, the chart rounds a September 7 export, and the summary adds 400 trial conversions twice.

A useful resolution record would read:

The resolution is not “13,200 was wrong” in isolation. It explains which question each value answered and why one definition fits the decision.

• N-01: 12,480 active accounts; August 31; account-status filter excludes trials; source is planning workbook version 6.
• N-02: 12,800 displayed as 12.8k; September 7; same filter; source is later dashboard extract.
• N-03: 13,200; September 7; trial conversion adjustment duplicated; source is AI-assisted summary draft.
• Decision: use 12,800 for the September outlook, update the cutoff label, remove the duplicate adjustment, and retain the workbook formula used to reproduce the figure.

Test every downstream surface

After correcting the authoritative calculation, trace the number into every dependent table, chart, narrative sentence, percentage, and recommendation. Recalculate derived values from unrounded inputs. Search for both the old value and its formatted variants. If a chart is image-based, inspect its label manually.

Use four final checks:

The NIST AI Risk Management Framework Core supports documenting risks, measurement, and responses across the lifecycle. A reconciliation record is one concrete way to connect a detected inconsistency to a managed correction.

• Reperformance: another reviewer can reproduce the result from the cited source and formula.
• Dimensional check: units cancel or combine as expected.
• Boundary check: totals match the stated population, filters, and period.
• Cross-artifact check: all published surfaces use the resolved definition and disclose material differences.

Sources: [3]

State what remains uncertain

Some mismatches cannot be resolved before delivery. Mark the disputed value, list the candidate sources, explain the decision impact, and identify who must resolve it. Use a range only when the range itself is supported; do not manufacture a midpoint to make the document look complete.

In a Two Prune assessment, inconsistent-number verification can be observed through source review, calculations, AI interactions, revisions, and the final business artifact. The evidence supports an assessment of what the participant demonstrated in that scenario. It does not establish that every future numerical task will be handled the same way or substitute for a domain owner's approval of the underlying definition.