Can AI read the spreadsheets our team built?

A modern workbook is a bundle of smaller files. The format behind .xlsx is ECMA-376, Office Open XML, whose Part 1 is in its fifth edition of December 2016 and which is standardized internationally as ISO/IEC 29500. Rename a workbook to .zip and the parts are there: one XML file per worksheet, a styles part, and a shared string table holding each piece of text once. The cell holds only an index into that table, so a 0 is a number or a pointer depending on the type the cell declares.

Formulas sit as text beside a cached value, which Microsoft's documentation describes as the result from the last time the formula was calculated. Conditional formatting is saved as rules: a shading or font colour the application applies when a condition is true. So a program opens the values as last saved and the rules behind them, not the picture on your screen.

A CSV holds less. RFC 4180, the IETF's informational note from October 2005, says no formal specification existed and that implementations differ considerably, then describes one table of comma-separated text. There are no formulas, no colour, no merged cells and no second tab.

None of these is a mistake. Each made the sheet quicker for the person who built it, and each hides meaning from anything reading the file cell by cell:

• Merged cells. Microsoft's guidance says merging keeps only the upper-left cell's contents and deletes the rest. A brand name merged down a block of rows belongs to one cell, and every row under it reads as blank.
• Several tables on one tab. A section per manufacturer, each under its own header row, a freight table off to the right, terms typed underneath. The format can mark a range as a table, but only when someone has defined one.
• Colour as meaning. Yellow for a price awaiting approval, red for discontinued. A hand fill is stored as a style and a conditional colour as a rule, and neither says what it means. The legend, if there is one, is on another tab or in somebody's head.
• Notes typed into cells. "Call for price" or "2 wk lead" in a column of numbers. Those cells become text, and the column stops being a column of prices.
• Long numbers and codes. Excel keeps 15 significant digits, rounds anything past the fifteenth down to zero and removes leading zeros from codes such as 00123. A part number typed as a number can be wrong in the file before anyone reads it.
• Dates. A date is a serial number counted from the start of 1900 or of 1904, depending on the workbook's date system, and the two differ by 1,462, four years and a day. A reader that assumes the wrong system shifts every date on the sheet.
• Hidden rows and tabs. Hiding deletes nothing. Last year's prices are still in the file, and a tab can be set as very hidden, which a user can't bring back from the menu. A program sees all of it; only your team knows which parts still count.

A formula can point at another workbook, and the format calls that a link. A takeoff sheet that prices from the master list on the shared drive keeps the last value it calculated. Microsoft's page on workbook links warns that a source file may have been moved or deleted, and that when links don't refresh, the people using the workbook won't know the data is out of date.

Versions are the second trap. Nothing inside "Price list FINAL v3 (2).xlsx" says which copy was in force when a given quote went out. Public auditors treat this as a control question. In a 2018 audit of key Sarbanes-Oxley spreadsheets (Audit 2017-15451), the Tennessee Valley Authority's Office of the Inspector General found that an unenforced naming convention limited the agency's ability to see whether critical sheets were stored properly, and it recommended baselining each one so later changes could be understood. For a test, send the version your team used for each example, not today's, and say where its links pointed.

The sheet as input. A contractor's takeoff, a buyer's list of part numbers, an employer's census. The draft reads it and produces a quote, an order or a clean census. Reference sheets belong here too, the price list or rate table a draft prices from, each as it stood on the date of every example.

The sheet as answer key. The tracker your coordinator keeps, the WIP schedule your controller closes, the bordereau that went to a carrier. Here the sheet is what your team produced, corrections included, and the draft has to fill the same columns from the same source documents.

Many jobs have both: a customer's sheet comes in and your priced sheet goes out. Say which role each sheet plays before the test starts, because the success criteria differ. For an input, the question is whether every value was read as your team read it. For an answer key, it is whether each cell matches what your team entered.

Our free AI proof of concept runs on one job and copies of 20 to 30 finished examples. For a spreadsheet job, each example is a set: the incoming sheet or document, the reference sheets as they stood that day, and the finished output. Your reviewer, the person on your team who knows the job best, picks the sets before any draft exists. An NDA is signed before a file moves, and nothing logs in to a shared drive or any system of yours. If sheets can't leave the building, we build look-alikes with you that keep their merged headings, colour codes and odd columns. Bring each set like this:

• The working copy, not a tidied one. Leave the merges, colours, notes and hidden tabs in, unless a hidden tab holds something that must not leave.
• The linked files, or a note saying the link was broken and which values your team typed over it.
• The colour legend in words. If nobody can write it down, that is a finding before the first draft.

Each draft names the tab, row and cell every value came from, and your reviewer compares it with the finished output cell by cell, against success criteria agreed before the test starts along with a price range for the live job. Your reviewer's call decides it. After a yes, the job runs on new sheets for about 30 days, with a person approving each item before it goes anywhere, and nothing writes into the sheets your team keeps.

Which version is current, what a note in a price cell means, whether a hidden row is retired or just out of the way: those are decisions, and the person who knows the sheet makes them. A draft can point at each one. It shouldn't settle any of them.

Bank supervisors drew the same line. The Basel Committee's BCBS 239 (January 2013) says a bank relying on manual processes and desktop applications such as spreadsheets should have effective mitigants, end-user computing policies among them, asks for a single authoritative source for each type of risk data, and allows that where professional judgements are required, human intervention may be appropriate. It was written for banks, not price lists, and none of this is legal or audit advice, but the reasoning travels: name the one sheet that counts, and leave the judgment calls with a person.