AI in hospitality

Asking an AI assistant to check a spreadsheet: formulas, totals and the questions to put to it

An AI assistant can find formula differences and explain sheet structure, but it cannot recalculate live data. Learn which questions to ask, how to verify findings by hand, and how to protect sensitive information before uploading.

  • 10 min read
  • AI Otel Editorial

AI draft · awaiting editorial review

Contents

Key takeaways

  1. An assistant is good at finding differences and explaining structure, not at producing a number you can put in a report.
  2. Ask for cell addresses and the formula text in every answer, never for a summary that hides them.
  3. Recompute three numbers after every session: the total that matters most, the largest error it claims, and one picked at random.
  4. Take guest names, identity numbers, salaries and contract rates out of the file before it leaves your building.
  5. Log every accepted change with the cell, the old formula, the new formula and who approved it.

An AI assistant can read a workbook's structure and formulas, list the ones that differ from their neighbors, explain a sheet in plain language and reconcile two totals. It cannot recalculate like a spreadsheet engine, so treat each finding as a lead and recompute the important numbers by hand.

The budget is off by a small amount and nobody can find it. Forty tabs, years of edits, and one cell in March that no longer matches the row beneath it. This is the work an AI assistant is genuinely useful for, and also the work it quietly gets wrong. After reading this, you will know how to run an AI check on hotel spreadsheet formulas, which questions produce findings you can act on, and how to verify the results before anyone relies on them.

What can an AI assistant actually see in a hotel workbook?

It sees the file as text and structure: cell addresses, the formula written in each cell, and the values your spreadsheet last saved. It does not open the file the way your spreadsheet program does.

Separate three layers before you start. The formula as written is the expression typed in the cell. The stored value is the number the file saved at the last calculation. The live recalculation is what the spreadsheet would compute right now, and that is the spreadsheet's job, not the assistant's. An assistant usually gets the first two layers and never the third.

LayerWhat it isWhat the assistant gets
Formula textThe expression typed in the cell, for example a sum over a rangeRead exactly
Stored valueThe number the file saved at the last calculationRead, but may be out of date
Live recalculationWhat the spreadsheet would compute nowNot available
Source systemReservations, invoices, payrollNot available, only whatever was exported into the file

Within those limits, it handles several jobs well. It can list every formula in a column. It can compare one formula with its neighbors and flag the ones that differ, which is the single most useful check on a workbook built by copying formulas down. It can translate a formula into plain words. It can find hard-coded numbers sitting inside a block of formulas, and spot a range that stops one row short of the data.

It cannot recalculate, run macros, refresh external links or query the system the numbers came from. If the workbook was built from a reservation or billing export, the assistant sees that export, not today's records. It is checking a snapshot, and you should know how old the snapshot is.

Some structure trips it. Merged cells, hidden rows and columns, pivot tables, several sheets with the same layout, and totals placed above the data instead of below all cause misreadings. Tell it the layout in the prompt: sheet name, header row, first and last data row, and where the totals sit. Give it one sheet at a time. A whole workbook in one go produces a description of the workbook, not a check of it.

If the terms around this work are new, the vocabulary of reconciliation, extraction and the review queue in plain terms is worth a minute before you start.

How do you take the names and the private data out first?

Replace anything that identifies a person or a contract with a code, keep the mapping outside the file, and send only the copy.

In a hotel workbook, treat these columns as sensitive: guest names, identity and passport numbers, staff names and salaries, agency and corporate contract rates, and supplier terms. The first three are personal data; the last two are commercial terms you have promised to keep to yourself.

Replace values with stable codes in a working copy. Keep the code-to-name mapping in a separate file that never leaves your building. Delete rather than hide: hidden rows and columns travel with the file, and so do comments, tracked changes, the author name in the file properties and last year's leftover sheets.

Decide who may send a workbook out at all, and write it as one line that everyone can quote back to you. Work on a copy named with the date, and leave the original untouched so you can always return to it. Which columns may leave the building is a decision for your data protection lead, not for the person running the check.

ChecklistPre-upload checklist

  • Working copy saved with today's date.
  • Guest names, identity and passport numbers removed or coded.
  • Staff names and salaries removed or coded.
  • Agency and corporate contract rates coded.
  • Hidden rows, hidden columns and unused sheets deleted.
  • Comments and tracked changes cleared.
  • File properties checked for an author name.
  • Mapping file stored outside the shared folder.
  • One named person has approved the send.

Which questions to put to it?

Narrow questions about structure and difference get usable answers. "Check my budget" does not.

Ask for a list, not a verdict. The strongest single prompt is: "list every formula in column F that differs from the formula in F5, with the cell address and the formula text". Demand the address and the formula in every answer. An answer without a cell reference cannot be checked and should be discarded.

Start with "explain this sheet in plain language". You will often spot your own layout error in its description before it finds anything. Then move to hard-coded numbers inside a block of formulas, ranges that stop short of the last data row, and reconciliations between two totals.

One question per turn. A prompt with five questions gets one careful answer and four guesses. And never ask a leading question. "Is this total wrong?" invites agreement; "compare this total with the sum of column D and show both figures" does not.

Question to askWhat a good answer looks likeHow you check it
List formulas in column F that differ from F5Cell addresses with both formulas side by sideOpen two of them and read the cells
Find numbers typed directly into a block of formulasAddresses and valuesConfirm each is a deliberate input, not a leftover
Which ranges end before the last data rowAddresses plus the last row of dataScroll to the last row and compare
Reconcile the departmental total with the summary totalBoth figures, the difference, and the cells that explain itRecompute the difference yourself
Explain what this sheet calculatesA short plain-language descriptionRead it against the sheet and correct it where it is wrong

If the workbook feeds your monthly variance review, pair this with reading a monthly budget-to-actual variance, because a formula error and a real variance look the same until you separate them.

Where does it fail with numbers?

It predicts plausible text; it does not run a calculation. Long sums, unit conversions and number formats are where it drifts.

It may sum a column it inferred rather than the one you named, and present the result with full confidence. It may read a stored value that is out of date and report it as the current figure. It may drop a row without saying so, and the total still looks reasonable.

Number formats cause real errors. A Turkish workbook writes 63.917.057, with a dot for thousands, and 65,2, with a comma for decimals. State the convention in the prompt and still expect mistakes.

The most dangerous failure is different: it may invent a formula that would produce the answer you seem to want. This reads like help, which is exactly why it gets past people. It also carries a conversation forward, so a wrong figure accepted in one turn becomes the basis of the next answer.

Watch outA lead, not a result

Any number the assistant produces is a lead, not a result. Never paste its total into a report, a board pack or a payment file. Recompute it in the spreadsheet first, and if the two disagree, the spreadsheet wins.

How do you check its findings by hand?

Recompute three numbers yourself and open every cell it names.

The three numbers: the total that matters most, usually revenue or payroll; the largest error it claims; and one picked at random. Recompute by a second route, for example a pivot table against a conditional sum, or the workbook against the report the figures originally came from.

Open the flagged cell and read the formula. Check that the range starts below the header and ends on the last data row. Check the sign convention: a refund or a discount entered as a positive number passes a formula comparison and still distorts the total. Check the unit: a night count compared with a room count, or a percentage compared with a currency amount, will look like a discrepancy and is not one.

Write down which figure you used and why. The next person to open the file needs that sentence more than the correction itself.

ChecklistHand-check routine

  • The most important total recomputed by a second method.
  • The largest claimed error verified at the cell.
  • One random finding verified at the cell.
  • Range start and end confirmed against the data.
  • Sign convention confirmed for refunds and discounts.
  • Units confirmed.
  • Agreement or disagreement with the assistant recorded in one line.

For the surrounding vocabulary, zero-based, rolling and flexed budgets and what a variance means explains the terms your findings will land in.

How do you document a change and keep the problem from coming back?

Version the file, log each accepted change in one line, and run the same questions again next month.

Keep a before copy and an after copy. Never edit the file you audited, and never overwrite the before copy. One log line per accepted change: date, sheet, cell, old formula, new formula, reason, who found it, who approved it. Paste the assistant's answer that triggered the change into the log, so the reasoning survives after the conversation is gone.

DateSheetCellOld formulaNew formulaReasonFound byApproved by
12.03.2026PayrollF41Sum ending at row 40Sum ending at row 41Last employee row was outside the rangeAssistant checkFinancial controller

The row above is an example, not a real entry. Keep the questions in a saved list. Running the same list next month is a regression check on your own edits.

Where the workbook is only a copy of numbers that live in the hotel system, fix the export or the report as well as the cell, or the same error returns with the next export. The records themselves belong in the hotel operating system that keeps reservations, billing and reports in one place, and a fuller picture of that layer is on the platform overview. A change nobody can explain six months later is a change that will be undone by accident.

The same discipline applies when AI writes prose rather than checks numbers, as in checking an AI-written menu description against what the kitchen must confirm: the machine drafts, a person with the facts confirms.

What to do next

Pick one workbook you already distrust, take one sheet, and run three questions on it: list the formulas that differ from their neighbors, find numbers typed inside formula blocks, and reconcile two totals. Recompute three numbers by hand, log what you changed, and keep the question list for next month. The assistant shortens the search; the arithmetic and the approval stay with you.

Frequently asked questions

Can an AI assistant open my Excel file and change it?

You normally upload a copy, and the assistant works on that copy. Nothing it says edits your workbook, and any change to the real workbook is made by a person, in the workbook, after the check. Check your provider's retention settings before you upload anything sensitive.

Should I send the whole workbook or one sheet at a time?

One sheet at a time, with the header row and the totals row included. Say where the data starts and ends. A whole workbook produces a general description; a single sheet produces findings you can act on.

Can it tell me which formula is wrong?

It can point at a cell whose formula differs from the ones around it, which is a strong signal. Whether that difference is an error depends on the business rule behind the cell, and only the hotel knows that rule. Treat the finding as a question to answer, not an answer.

Is it safe to upload a hotel budget?

It depends on what is in the file. Remove or code guest names, identity numbers, salaries and contract rates, delete hidden rows and comments, and work on a dated copy. Who may send a workbook out is a decision for your data protection lead.

What about Turkish number and date formats?

Say the convention in the prompt: dot for thousands, comma for decimals, dates as day.month.year. Expect confusion anyway, and check any figure that crosses a thousand or sits near a decimal point by hand.

AI Otel EditorialEditorial team

Drafted with AI assistance from AI Otel’s published fact base. Not yet reviewed by an editor.

We can brief you on this topic.

If this topic touches your work, we can walk you through what is built and what is designed.

Request a briefing