A spreadsheet in an AI workflow: where data types, formulas and responsibility can be lost
A spreadsheet connected to AI needs checks on data types, formulas and versions. Learn how to find silent errors and establish safe rules for writing results.
Article language: English

In this article
Product code 00417 comes back from an automation as 417. The spreadsheet looks tidy, the summary sounds professional, and the warehouse system knows no such product. In this hypothetical example, AI did not even need to invent anything. A data type changed along the way. Before allowing results to be written to a company workbook, check the entire path: from the source cell through export and the model to the receiving program's reading of the data.
The most important test is not merely whether the answer makes sense. Did the right result reach the right record, retain its meaning, and avoid overwriting calculations or a later human correction? This applies to manually pasting a chat response and to an integration that runs without opening Excel. A spreadsheet can be a full participant in a process, however inconspicuous its file costume may be.
Establish what work the workbook actually does
Start with one workflow, such as assigning categories to sales enquiries. Record where rows originate, who corrects them, what AI reads and who uses the result. Do not describe the entire company. One properly reconstructed workflow will show more than a list of every file with an XLSX extension.
Distinguish three roles a spreadsheet can play. It may store source material, calculate values or present a result. One workbook may do all three, but each calls for a different check. A summary of the visible range does not verify the source. Correct classification of a description does not confirm the formula in the adjacent column.
Include that description in an audit of business process automation. The scope of such an audit is discussed in the article on audit of business process automation (in Polish). It reveals work employees still do out of sight: removing spaces from codes, supplying units, or recognizing that “urgent” is a comment rather than a category. Automation cannot take this over sensibly while the rules remain in the file author's head. Add the spreadsheet connection and the owners of the file and integration to an inventory of AI tools. For how to prepare that inventory, consult the article on inventory of AI tools (in Polish). This helps locate dependent processes when someone changes the data layout or moves the workbook.
Digits can be text, and a blank cell can carry information
A case number does not become a numerical quantity just because it consists of digits. You do not add two customer numbers to get a premium customer. For each column, define the permitted type, meaning, unit and way to represent missing data. If the result passes through JSON, preserve those distinctions when checking the structure and meaning of AI-generated data. Valid syntax does not determine an identifier's meaning or unit, as explained in the article on JSON structure and meaning (in Polish).
Microsoft explains that Excel can remove leading zeros and transform long digit strings, and that numbers are handled with precision up to 15 significant digits. Identifiers may need to be stored as text during import. Changing formatting afterward will not recover lost data. These are specific limits described in its instructions for keeping leading zeros and long numbers.
Build your own test with a few artificial records: a code with a leading zero, a long identifier, a name with a comma, a negative value, a blank field and a number written with a decimal comma. Send them along exactly the same path as real data. Check not only their appearance but also the value read by the next system.
Separately agree what a blank means. A missing deadline might mean “not agreed,” “not applicable” or “not provided by the source.” Automatically inserting zero or today's date erases the difference. If the model cannot decide, the result should preserve uncertainty rather than force order onto it.
A formula and its last result are two different objects
You see 2400 in a cell, but a process may retrieve the formula text, a stored value or the result of the last recalculation. Check the behavior of the specific exporter and integration. Do not assume that every way of reading a workbook produces the same material.
Choose a formula with a simple result calculated by hand. Change one input, run the normal workflow and compare the data seen by the model with the spreadsheet. Then check a row added at the end of the range and a cell replaced with a constant value. The latter may look like a member of the family even though it stopped taking part in the calculations long ago.
Excel provides rules for detecting, among other things, inconsistent formulas and omitted cells. Its documentation also explains that ignored warnings do not reappear in later checks until reset. Detecting formula errors is a technical aid; it does not decide whether the company uses the right rate or definition of the range.
Therefore supplement the meaning check with review by the person responsible for the calculation. When a spreadsheet feeds AI-generated commercial offers, check units, rounding and the separation of one-off and recurring items as well. The use of source values in proposals is examined further in the article on auditing AI-generated sales proposals (in Polish). A neat total does not explain what was added together.
A row needs an identity independent of its position
“Write the result to row 38” works only until someone sorts, deletes an item or edits in parallel. Every record needs a stable identifier. When writing, the integration must locate that identifier and check that it is still working on the expected data version.
Test a situation in which AI reads a description while an employee corrects it. A result for the old description must not silently replace the result for the new one. Plan a clear concurrent-edit conflict message, so the spreadsheet user knows what to compare and what to do next. Guidance on that message appears in the article on warnings for concurrent editing conflicts (in Polish). A sensible response may be to stop the write and request reprocessing of the particular record. The rule needs to be visible to the operator, not merely buried in a script.
If a spreadsheet contains model-assigned labels, a separate message-classification test helps. The message-classification testing guide (in Polish) sets out checks for the assigned category. Assess category accuracy and correct assignment to a record separately. A correct label on someone else's case is still an error.
Restrict writes to specified fields
An integration adding a category does not need permission to rebuild the entire workbook. Identify the destination file, sheet, range and permitted columns. Agree who may change the schema. Renaming a “status” header to “stage” should not trigger creative guessing about which column to write to now.
When reviewing AI integration permissions, check the account performing the write and how access can be revoked. The scope of integration access is addressed in the AI integration permissions guide (in Polish). Separate permission to read material for analysis from permission to approve a result. Cell protection and file organization can help, but do not replace access control for the whole workflow.
Send the model only what the task needs. Data minimization before AI use also covers extra tabs, comments and columns a user normally overlooks. That broader scope is the focus of the guide to minimising data before using AI (in Polish). Inspect the scope sent by the integration in the actual request instead of inferring it from what happens to be visible on screen.

Accept the workflow on a copy, with a recovery plan
Prepare a controlled test file with expected results. Include sorting, a filter, a blank row, new columns and an error in one record. Decide whether one faulty item stops the whole batch or is sent for clarification while valid items continue.
Keep a before-and-after comparison. “Updated 40 rows” does not show the scope of changes. The reviewer should check new values, intact formulas and skipped records. Add these checks to AI quality regression tests so they can run after a change to the model, template or integration. See the article on AI quality regression tests (in Polish).
Agree how to recover after an error. Restoring an entire old copy can remove valid work other people performed later. Safer correction often requires a list of the specific changed records and a check of their current state. Reversal needs a scope as precise as the write.
See whether the model received every required row
After a filter, eleven items are visible on screen, but the integration may read the entire sheet or only a specified range. Check in practice. Prepare three marked records: one visible, one hidden by the filter and one on another tab. Ask about each and inspect what was sent to the model. An answer's failure to mention a record does not prove that the record was not read.
Also establish the table boundary. An item added below may fall outside the export range even though a person sees it as part of the same list. Moving a header, merging a cell or adding a title row can change the interpretation of data. The process should recognize an incompatible layout and stop the import instead of treating the first text it finds as a header.
In an acceptance exercise compare three numbers: records intended for analysis, records actually transmitted and records with a result. Explain every difference. Not all differences need be errors; some may follow an agreed exclusion rule. The rule must nevertheless be explicit and the omitted items findable.
For large datasets, check the limits of the specific integration. Do not assume an attached file was read in full simply because the tool accepted it. Divide the test into recognizable portions and verify their presence in both input and results. A technical limit should produce a clear message, not a shortened analysis pretending to be complete.
Change the structure through a controlled data contract
The workbook owner can improve the file, but the automation needs to know about changes on which its operation depends. Briefly describe the data contract: required field names, types, permitted values and a layout-version identifier. It need not be a large document. What matters is being able to distinguish a cosmetic change from one requiring a new acceptance check.
Suppose the “deadline” column is split into expected and confirmed deadlines. An integration using the old field should not choose one of the new ones because its name looks similar. A decision about meaning, an updated mapping and repeated tests are needed. Otherwise the team improves the spreadsheet while the model worsens the way it is used.
Prepare a small reference copy containing only artificial data. After a layout change, run it before the real workflow. The person changing the file should know where to check the result and who confirms that the integration is ready. Agree how to respond to an incompatible version: stop that import, notify the owner and retain the file for clarification.
A well-named version later helps reconstruct an error. “Monday's spreadsheet” may be insufficient if the file changed several times that day. Retain identification of the input and process configuration according to company needs, without making unnecessary copies of the entire contents.
Try handing the workbook to another person
Ask an authorized employee who did not create the automation to perform a task on a copy. They should find the right file, run the analysis, inspect the result and resolve one planned error. Observe where they ask about a column's meaning or look for the author. These are clues for the instructions, not reasons to judge the employee.
Introduce an artificial conflict: change one record's description during processing. The reviewer should see why the result was not written and what to do next. If one item simply disappears from the summary, the process needs a clearer message. Similarly, check missing write permission and an unavailable file.
Finally ask the person to explain the result without opening the author's chat. Where did the category come from, who approved it and what does a blank mean? If the answer requires guessing, improve the description or execution trail. A spreadsheet does not have to become a complex system to have understandable working rules.
Finish acceptance with a short decision: what scope runs automatically, what needs checking and after which change the test must be repeated. This helps prevent gradual expansion of use merely because the file still opens. Opening reliably is a pleasant feature of a workbook, but a modest measure of process trustworthiness.
In the instructions, also name the boundary between a proposal and a value accepted for further work. You can reserve one field for the model result and another for the reviewer's decision. A later user then need not infer from color whether a category was approved. Color is a useful signal, but a poor archivist of decisions.
For manual copying of responses, run the same test as for an integration. Pasting a whole column into a filtered view or over an existing formula still changes data. The person doing the step should know the permitted area and check the list of changed records. A hand on the mouse does not remove the risk of shifting a row.
Keep one correct example and one typical conflict with the process response in the documentation. This helps a new person reconstruct the work without learning every possible failure. A short instruction that someone has actually followed in a test is more valuable than an extensive description nobody has checked yet.
If the spreadsheet is an important part of company automation, the Aeneia Audit offering includes a discussion of how to check such a process.







