When an AI assistant combines spreadsheets, require it to preserve the intended calendar dates rather than blindly copying their underlying numbers. Excel’s two date systems can assign different numbers to the same date. A transfer that preserves every digit can therefore change a deadline by about four years.
For an analyst or operations team consolidating renewal dates, delivery schedules or training records, the useful check is specific: establish each workbook’s date system, identify how the transfer interprets dates, then reopen the saved result and compare known records. Changing the visible date format does not establish that the underlying date is right.
The worked example below uses synthetic workbooks checked locally with Python and openpyxl 3.1.5. It demonstrates a file-conversion hazard, not an observed failure of an AI assistant. No Excel desktop application, Microsoft 365 Copilot or other spreadsheet assistant was independently tested.
The workbook supplies part of a date’s meaning
Microsoft’s date-system documentation describes the 1900 and 1904 systems and their 1,462-day difference. For the same ordinary modern date, the 1900-system serial number is larger. The documentation also describes automatic conversion during supported copying operations, with settings and exceptions affecting that behavior. A mismatch is a possibility to investigate, not an inevitable result of copying between workbooks.
The openpyxl date documentation explains that XLSX dates can use a numeric representation or ISO 8601 strings. For numeric dates, the workbook’s date system determines interpretation. The library exposes that system through wb.epoch and translates supported date values to and from Python date-time objects.
An assistant may work through a spreadsheet application, a Python script, an export file or a connector. Those routes need not carry the same information. Ask which representation crosses the boundary: a recognized date, an untyped number or a text string. The conversion should be reviewed at that boundary.
Do not infer the date system from the sender’s computer. A workbook can retain settings from its history. The current file and its actual interpretation matter more than an assumption that “Mac means 1904” or “Windows means 1900.”
One deadline, two correct numbers and one wrong transfer
Our fictional record X has a deadline of 5 November 2026 at 12:00 and a separate quantity of 45,000. Two small workbooks store the same deadline using different date systems. Their date cells use the same display format.
The local fixture saved and reopened both files, inspected their worksheet values and workbook properties, then tested two transfer paths into the 1904-system workbook:
| Case | Date system | Numeric date value | Date returned after reopening |
|---|---|---|---|
| Original source | 1900 | 46331.5 | 2026-11-05 12:00 |
| Correct destination representation | 1904 | 44869.5 | 2026-11-05 12:00 |
| Source number written directly into destination date cell | 1904 | 46331.5 | 2030-11-06 12:00 |
| Decoded date-time object written into destination | 1904 | 44869.5 | 2026-11-05 12:00 |
The first two numbers differ by 1,462. The raw-copy path preserves the source number but changes its meaning. The typed-copy path preserves the intended date while allowing the writer to encode it for the destination system.
In the inspected XML, the 1904 workbook declared date1904="1"; the default 1900 fixture did not include that attribute. Both saved transfer files retained the separate quantity of 45,000. The noon component also survived. These are observations from this small openpyxl fixture, not a certification of other libraries or transfer methods.
The quantity is deliberately similar in size to a date serial. Subtracting 1,462 from every large number would corrupt it. A cell’s business role must determine whether date conversion belongs there.
Diagnose before adding an offset
A four-year discrepancy is a useful clue, but it is not a diagnosis on its own. Compare a known source date with the saved destination, inspect both workbook settings and trace the transfer operation. A wrong reporting year, a text-parsing error or a formula referring to the wrong column needs a different correction.
J. Carlton Collins’s February 2012 practitioner column describes a pasted-date problem caused by a workbook using the 1904 system. It is a historical troubleshooting account, not evidence about present AI failure rates. Its workbook-specific remedy should not become an instruction to change the date-system setting of every file an assistant encounters.
For our proposed workflow, review the conversion before altering the whole workbook. A workbook-wide setting affects the context in which other numeric dates are interpreted. Changing it to make one imported column look right may leave existing dates wrong.
Likewise, do not add an offset when the library or application has already converted the date. That would apply the correction twice. Ask the assistant to show the input type, source interpretation and destination write operation. Inspect the saved result rather than accepting an explanation that merely repeats the desired date.
Keep an untouched source copy while investigating. If several files were already consolidated, retain their origins so a correction can target the affected records instead of shifting every date in the combined report.
Test meaning across the actual transfer path
The original SpreadsheetBench paper, published at NeurIPS 2024, evaluates generated solutions using multiple workbook cases with varied values. That evaluation design motivates testing beyond one plausible output. It does not establish date-system error prevalence, validate this fixture or predict the performance of current assistants.
Our recommendation is to vary the representation as well as the date. Use a copy of the actual workflow and prepare expected results independently of the generated conversion:
- Transfer a known modern date between workbooks using the same system. It should retain its meaning without an offset.
- Transfer the same date in each direction between different systems. Verify the calendar date after saving and reopening.
- Include a date-time with a nonzero time component so integer-only conversion cannot pass unnoticed.
- Include an ordinary quantity or identifier in a neighboring numeric field. It must remain unchanged.
- Include an empty date and a text date with an explicitly defined interpretation. Require visible handling of missing or unresolved values.
Only the fixture paths described above were executed for this article. The other cases are proposed acceptance checks for the reader’s implementation. In particular, text-date parsing and a full operational workbook were not tested here.
For a real delivery, open the result in the spreadsheet application the team uses. Check the known records and important dependent calculations. A library round trip provides useful evidence about that library’s file handling; it does not establish how every application displays, recalculates or imports the file.
Preserve a date contract with the handoff
Give an AI-assisted consolidation task a short date contract. Name the source field, its business meaning, the source workbook’s date system and the destination’s intended representation. Include one expected calendar date and, where relevant, its time component. State that unrelated numbers and workbook settings must remain unchanged unless the task specifically requires otherwise.
If a connector exposes only numbers and omits the workbook context, mark that interpretation unresolved. Do not ask the assistant to guess an epoch from a plausible-looking range. Obtain the missing context or export dates through a documented, explicit representation before continuing.
Keep this separate from the row-inclusion and formula checks used to validate totals. A formula can select the intended rows while every imported deadline is shifted. Conversely, correct date conversion does not establish that a report includes the right population.
For the next spreadsheet merge, choose one known date before the assistant starts, inspect the source and destination systems, and compare that record after the file is saved. Record the method and any unverified application checks. The acceptance question is whether the business date survived the transfer, even when the number representing it needed to change.