- Sort by task, not by product. Writing a formula and explaining an inherited one are reliable. Reading numbers off a large sheet and summarising them is where wrong answers hide.
- The safety rule that makes the whole thing usable: ask for a formula you can inspect, not for a value you cannot audit.
- A benchmark of spreadsheet tasks found models competent on simple work and failing on multi step operations, producing plausible but incorrect output.
- In a study of maths over tables, accuracy fell by about 20 percent on average once the lookup was embedded inside reasoning rather than asked for directly.
- Spreadsheet errors were already the norm before any of this. The research on human spreadsheets says we are bad at spotting our own mistakes and confident anyway.
The panel appears one Tuesday on the right of the window and offers to help. You have a sheet with three years of orders in it, a pivot you built in 2024 and never quite trusted, and a monthly figure you send to your accountant. The honest question is not whether the panel is impressive. It is whether it can touch this file without quietly breaking the one number that matters.
The answer depends entirely on what you ask it to do, and the difference between the safe tasks and the dangerous ones is sharper than any product page will tell you.
Which spreadsheet tasks is it actually reliable at?
The reliable ones share a property: the output is something you can read and check in a few seconds. The unreliable ones share the opposite property, which is that checking the output means redoing the work.
| Task | How reliable | Why | Your verification step |
|---|---|---|---|
| Write a formula from a description | Strong | The answer is short and self describing | Read it, then test on three rows you know |
| Explain a formula you inherited | Strong | Nothing is computed, only described | Change one input and see if the explanation holds |
| Clean and normalise messy text columns | Good | Pattern work with a visible result | Sort the output column and scan the extremes |
| Build a pivot or a summary layout | Good | Structure rather than arithmetic | Check one total against a manual filter |
| Analyse numbers across a large sheet | Weak | The model reads and calculates in its head | Impossible to verify without redoing it |
| Write the narrative around your figures | Weak in a specific way | The prose is fluent whether the numbers are right or not | Check every figure quoted in the paragraph |
| Generate a macro or script | Mixed | Readable, but it runs before you fully read it | Run it on a copy of the file, never the original |
The weak rows are not weak because the technology is immature. They are weak because of what the task asks: produce a number, with the working done somewhere you cannot see.
What does the research actually say?
Two findings matter here, and they point in the same direction from different angles.
The first comes from a spreadsheet specific benchmark. Simon Thorne's FLARE benchmark, presented at the EuSpRIG 2025 conference, evaluated leading models on formula creation, auditing and reasoning over real spreadsheet scenarios. The headline finding is that models handle straightforward tasks well and stumble on complex multi step operations, where they tend to return output that looks right and is not. That last property is the dangerous one. A wrong answer that announces itself gets caught.
The second is broader and more diagnostic. The TabularMath study took maths problems that models solve well in plain prose and re-expressed the same information as tables. Accuracy fell across every model tested, and fell further as the tables got more complex. The researchers then separated the two things the model has to do: find the right cell, and reason with it. Asked to retrieve a value directly, models did fine. Asked to retrieve it as one step inside a longer chain, performance dropped by roughly 20 percent on average.
Their error taxonomy is worth keeping in your head, because it names what goes wrong. Retrieval omission is not realising a lookup was needed. Retrieval mismatch is pulling the wrong cell. Expression errors are building the wrong equation. Numerical calculation errors are getting the arithmetic wrong from a correct equation. Retrieval mismatch was the biggest single source of failure, and that is the one that leaves no trace at all: the shape of the answer is perfect and the input was the wrong row.
What the built in assistant does out of the box
Before deciding what to trust, it helps to know what is actually on offer. Microsoft's getting started guide for Copilot in Excel describes a panel that works against data formatted as a table and offers to highlight, sort and filter, to add formula columns, and to suggest charts and PivotTables from what it sees.
Notice the shape of that list. Most of it is formula generation and layout, which lands in the reliable half of the table above. The panel produces an artefact in the sheet rather than a claim about the sheet, and an artefact is something you can inspect. That is a better default than a general chat window offers, and it is worth knowing that the reliability advice here matches how the product was designed rather than fighting it.
The requirement to format your range as a table is not a nuisance either. It is the same thing that makes any ai spreadsheet feature work better: named columns give the model something to reason about besides cell coordinates, which is precisely the weakness the table reasoning research identifies. Ten minutes spent naming your columns properly improves every answer you will get for the next year.
The three jobs people actually use it for
Talk to anyone running a small business and the tasks converge on the same short list, which is worth stating because it is much narrower than the marketing.
Data cleaning comes first. Supplier exports, marketplace order downloads and bank statements all arrive in a shape nobody would choose, and turning them into something joinable is an hour a week that nobody enjoys. This is where excel formulas written to a description save real time, and the output is inspectable by construction.
Understanding an inherited workbook comes second, and it is the one people are quietest about. Someone built the pricing model, the stock reorder sheet or the VAT calculation, and that someone has left. The same problem turns up in a ledger, where an automated categorisation nobody can explain later is harder to defend than a formula you can at least trace. A spreadsheet audit used to mean tracing precedents by hand through six tabs. Being able to ask what a formula depends on, in plain language, changes that from a day to an afternoon.
Third is the recurring report, and this is where ai for excel goes wrong most often, because the request naturally comes out as a question about numbers rather than a request for a mechanism. The fix is the same one as everywhere else in this piece, and it is worth building the habit on this task specifically because it repeats every month.
The rule that makes it safe
Prefer output you can audit. In a spreadsheet that has a precise meaning: ask for a formula, not for a number.
A formula shows its work. You can read =SUMIFS(D:D,B:B,"Retail",A:A,">="&DATE(2026,1,1)) and see immediately which column is being summed and which conditions filter it. If the model picked the wrong column, the mistake is visible on the face of it. And Excel, not the model, does the arithmetic, so the failure mode of the fourth error type disappears entirely.
A pasted number shows nothing. It is correct or it is not, it does not recompute when a row is added, and six months later nobody can tell where it came from.
This rule also tells you what to do when a task genuinely needs computation over the whole sheet. Ask for the formula, the pivot layout or the script that produces the answer, and let the spreadsheet execute it. You get the same answer with an audit trail attached.
Three prompts, before and after
The pattern in each of these is the same: replace a request for an answer with a request for a mechanism.
Working out which products are slipping
Before: which of my products sold worse this quarter than last?
After: write a formula for column H that returns this quarter's units minus last quarter's units for the product on each row, using the columns Q1 units in E and Q2 units in F. Then tell me how to sort by it.
The first version asks the model to read hundreds of rows and compare them silently. The second gives you a column you can sort, filter and check, and it stays correct next quarter.
Cleaning a supplier export
Before: clean up this list of product names.
After: the names in column A have inconsistent capitalisation, trailing spaces and the supplier code appended in brackets. Write a formula for column B that strips the bracketed code, trims whitespace and applies proper case, and show me what row 4 becomes.
Asking what row 4 becomes is the cheap trick that makes this reviewable. You get a worked example next to the rule, so you can see the rule is right before applying it to nine thousand rows.
Understanding a formula somebody else wrote
Before: what does this spreadsheet do?
After: explain this formula piece by piece, name every cell reference it depends on, and tell me what it returns if column C is empty.
Explanation is the one task where the model is working from something fully visible to it, which is why this is the most reliable thing in the list. The edge case question at the end is what turns an explanation into an audit.
Which sheets should never leave the building?
A spreadsheet is usually the most sensitive file a small business owns, because it is where things end up: customer lists, supplier prices, payroll, the real margin numbers. Before any of it goes into a hosted assistant, the question is what happens to it on the other side.
Read the vendor's own statement rather than a summary of it. Microsoft's data, privacy and security documentation for Microsoft 365 Copilot states that prompts, responses and data accessed through Microsoft Graph are not used to train the foundation models, and describes the service boundary the processing stays inside. Whether an equivalent commitment exists for a browser extension somebody in your business installed last month is a different question, and usually the answer is that nobody checked.
A workable rule for a small team is three tiers rather than a policy document. Sheets that contain only your own operational data, like stock counts and order volumes, go to an approved tool freely. Sheets containing named customers, staff pay or supplier contract terms go only to a tool with a written commitment you have read. Sheets containing payment details, identity documents or anything covered by a confidentiality clause do not go anywhere, and the answer to a request to make an exception is a formula you write yourself. If you want the longer version of that, our AI policy template built around a data classification table puts the same idea into text you can adapt.
The reconciliation habit
One habit catches most of what the rest of this misses, and it takes under a minute.
Whenever an assistant touches a sheet, check one number you already know. The total row against last month's filed figure. The count of orders against your order tool. A single customer's total against their invoice. You are not auditing the change, you are testing whether the file still agrees with reality at one point you can independently confirm.
This works because of how spreadsheet errors behave in general, not just AI assisted ones. Raymond Panko's review of the research, published as What We Don't Know About Spreadsheet Errors Today, reaches three conclusions that have held up for decades: errors are rare per cell but almost inevitable somewhere in a large workbook, they are extremely hard to detect and correct once made, and the people who build spreadsheets are heavily overconfident about their accuracy. He argues that the biggest need is massively more testing than anyone currently does.
That is the honest frame for all of this. You did not have a reliable spreadsheet before the AI panel arrived. What changes is the speed at which new formulas enter the file and the confidence with which they arrive, and neither of those helps with detection.
Questions people ask
Can it work on a sheet with 200,000 rows?
It can write a formula for one, and that scales perfectly because the spreadsheet does the work. What does not scale is asking it to read the rows and tell you what it sees, because the whole sheet has to fit into the model's context and the reasoning failures above get worse as the table grows. The formula route sidesteps the size question completely.
Should I use the built in panel or a chat window?
The panel has your data and the chat window does not, which cuts both ways. The panel gives better answers because it sees the columns. The chat window is safer for a sensitive sheet because you paste in only the twenty rows you are working with. For formula writing, the chat window is usually enough, since a description of your columns is all the context the task needs.
What about the macros it offers to write?
Treat a generated macro like a script from a stranger, because that is what it is. Run it on a duplicate of the file. Read what it deletes before you read what it creates. The failure that hurts is not a macro that does nothing, it is one that does something to the wrong range and gets saved.
Does this replace learning formulas?
It replaces memorising the argument order, which was never the valuable part. It does not replace knowing that a lookup can silently match the wrong row, or that a sum over a filtered range is not a sum over a visible range. You still need enough grasp to read the answer, and that grasp is exactly what the reliable tasks in the table above assume.
Where this leaves you
The useful version of AI in a spreadsheet is narrower than the demo and more durable than it looks. It writes the formula you would have found on a forum in ten minutes, it explains the workbook your predecessor left behind, and it turns a messy export into a clean column. Those three are real time savings, repeated weekly.
The version that costs you money is the one that answers a question about your business by reading the sheet and telling you a number. Ask for the mechanism instead and the risk mostly disappears, because now the spreadsheet is doing the arithmetic and you are checking the logic, which is the division of labour that worked before any of this existed.
Two adjacent pieces go deeper on the underlying habits. Our guide to forcing a model to hold a fixed output shape covers why a structured answer is easier to check than a prose one, and five prompt patterns with a before and after for each shows the same rewrite move applied outside spreadsheets. If you are weighing what a habit like this costs in practice, our plans and credit pricing page sets out what a working month of assisted drafting actually comes to.