AI in Excel: three kinds of helpers, and why a wrong number does not come from a formula

How an assistant inside Excel differs from a chat with an uploaded file and from an agent that reads the accounting system and marketplace accounts itself; three causes of a wrong number (ambiguity, staleness, a similar column); a three-sheet file that survives the next export; a scheduled task with blanks instead of zeros; the check before the Monday meeting and three scenarios people start with.

The template is out of date the day you download it

Search for "sales table in Excel" or "Excel dashboard" and the first pages are templates: unit economics for a marketplace, a payment calendar, a budget of income and expenses, warehouse in-and-out. Accounting systems and CRMs give them away for free, and they are honest, carefully made files. They have one problem, and it is not in the formulas.

A template assumes you will be the one entering the data. By hand, from an export that first has to be made in the accounting system or in the Ozon seller account, then cleaned of merged cells and mixed date formats, then pasted into the right sheet without shifting the ranges. The first time it takes an evening; by the third month it takes an hour, and during that hour the seller account already holds new numbers. The template honestly shows the state at export time, but it is read as the state today.

That is why demand has shifted from "Excel dashboard" to "AI in Excel": people do not need a file, they need someone to keep it up to date. Our Yandex Wordstat measurement for August — September 2026 shows queries about Excel dashboards flat, while "AI in Excel" and "AI for Excel" together bring about three thousand impressions a month, and the first of them grows month over month. The question "which AI" is the wrong one, though: there are three, and they do not replace each other.

Three kinds of "AI in Excel"

What it is	What it does	Where it stops
Assistant inside Excel (Copilot, agent mode, add-ins)	Writes and explains formulas, cleans a range, builds a pivot from a description, finds an error in a model	Sees only what is already in the file. The export from your accounting system and seller accounts is still on you
Chat with an uploaded file	Answers questions about the export, calculates, charts, hands back a new file	One-off: next week you upload the next file and explain everything again
Agent with access to the sources	Reads your accounting system, Ozon, Wildberries, Bitrix24 and Google Sheets itself, merges them into one file and refreshes it on a schedule	Needs the accounts connected and a task written in five or six lines. The first run takes 25–40 minutes of your attention

The first two are about speed with a file you already have. The third is about not having the file "with you" at all: it lives at a link and updates itself. The rest of this article is about the third, because that is where the main mistake hides — one the first two never make, for a simple reason: they do not go and fetch the data.

A wrong number is not a formula error

An observation from those who have run AI analytics on their own data warehouse: writing the query or the formula is the trivial part of the job for a model. Errors come almost entirely from three causes, and all three are about context, not computation. For a business owner with exports from an accounting system and two marketplaces they look like this.

Ambiguity. Is "revenue for the week" shipments or payments? With VAT or without? After returns are deducted or before? By order date, ship date or the date the marketplace booked it? Your accounting system, Ozon and Wildberries answer these four questions differently, and each has one column with the right answer and three columns with similar names. An agent that was not told which one will pick a plausible one. The number will match the seller account's total to within a few percent — which is exactly why nobody will notice.

Staleness. An export made on Monday no longer matches the seller account by Wednesday: the marketplace books returns and corrects commissions retroactively. A dashboard that takes the data once honestly shows the state at export time, but it is read as the state today. A separate case of the same illness: a formula that was correct for an 800-row export and silently truncated the data once there were 1,200 rows.

Missed it, though it was there. The right column exists in the export and is named correctly, but the agent matched a product from the accounting system to a product on Ozon by a similar name rather than by SKU, because the SKU sat in the next column in a different case. Or it took "Loft armchair, grey" and "Loft armchair grey" as two items. The formulas are flawless throughout.

The uncomfortable conclusion: what needs checking is not formulas but decisions. Which column was taken, which date bounded the period, which key was used for matching. The good news is that all three questions can be put to the agent directly, and it will answer them — more on that below.

A file that survives the next export

A dashboard built by an agent differs from a template not in looks but in structure. There is one rule: data goes into exactly one place, and everything else is computed from it. In practice that is three sheets.

Sheet	What is on it	What a person does with it
Source	A flat export: no merged cells, dates in one format, one row per transaction	The only place a new export is pasted
Summary	Pivots by channel, month, category, manager	Nothing
Dashboard	Charts and slicers tied to the pivots, not to "Source" directly	Looks at it

Three details that separate a living file from a pretty one:

Ranges are defined as tables or named ranges, not as A2:A800. Then an export with more rows is neither truncated nor shifted.
Not a single number on the dashboard is typed in as a value. If a cell holds 4,920,000 rather than a formula, it stays the same at the next refresh, and it will be taken for a fact.
A separate "How to update" sheet: where to paste, what not to touch by hand, which slicers will not survive opening in another program.

The last point deserves a closer look. Google Sheets and Excel treat slicers and pivots differently, and a file built for both at once will lose some of its charts and slicers in one of them. So the first question to the agent is not "which charts" but "where will the file be opened more often". If accounting has no access to the cloud, it is XLSX. If five people need the report at a link and it has to update itself, it is Google Sheets — and that is a different scenario: not a file but a schedule.

Then on a schedule, with blanks instead of zeros

Once the shape of the dashboard has settled, you stop building it and start refreshing it. The task fits in a few lines, and one of them is the line everybody skips. The source sheet is called "Data" here rather than "Source": the task picks the name; what matters is that there is exactly one such sheet.

Build a revenue-by-channel dashboard in Google Sheets for Ozon, Wildberries and retail from the accounting system, and refresh it every Monday at 8:00. Overwrite the "Data" sheet entirely; do not touch the charts sheet — it references "Data". Do not erase past weeks: each week is its own row. If a source did not respond or returned a clearly incomplete period, do not put in a zero and do not carry last week forward. Leave the cell blank, put the week on a "Gaps" sheet with the reason, and say so in the message. After each run send a short summary: what was updated, which channels did not come through.

The line about zero is the most important one. A zero in the cell "Wildberries, week of 9–15 June" looks like a sales collapse on the chart, while a blank cell marked "no data" looks like a gap. The first takes half an hour of a Monday meeting; the second is settled by one line in the summary: "WB did not return data, the row is not in the total." An agent that was not told this will most often act cautiously, but caution is better named explicitly than relied on.

The second trap is data the marketplace changes retroactively. Revenue for the first days of a week is being corrected by Ozon and Wildberries for several days, and a dashboard that reads the seller account every Monday will see different numbers for the same week next Monday. A good task says what to do about it: overwrite last week or keep both versions.

What to ask the agent before the Monday meeting

In the scenarios below every task comes with a ready check request, and they are all built the same way: not "check that everything is right" but a list of specific things the agent must enumerate. For a dashboard it looks like this:

Check the structure of the file. List: formulas and ranges that will shift if the new export has more or fewer rows; cells where the value is typed in rather than computed from "Source"; pivots and slicers that will lose their binding if the column order or names change; elements Excel does not support that will break on opening; what exactly a person has to do when updating and what they must not touch by hand. Whatever you can rebuild yourself, show as a "proposed changes" list; leave the rest to me.

For merging sales from four sources the list is different, and it hits exactly the three causes from the section above: items matched by name rather than SKU; rows that may have remained duplicates between the accounting system and the CRM; channels where revenue was taken with VAT while the others are without; orders that fell into the period by different dates in different systems; the share of unmatched items and the amount they carry.

This check works because the agent really does remember which decisions it made: a run has a record of every call to a source and every file it created.

The same record carries a second line of defence, and it works without you. When the agent closes a task with "file built, data taken from the accounting system", the words are checked against the run's record. A file was promised and there is no file; a source is named that the run never went to; "sent" is claimed and nothing was sent — in each such case the agent goes back to finish, and the answer does not go out. For a scheduled task this is joined by acceptance criteria you set in words. If some of them are impossible for an external reason — the seller account did not respond, access was not granted — the run closes with a named boundary rather than an invented number.

This does not make the number correct. It makes a wrong number visible — which, in reports read by people without access to the raw data, matters more.

Three scenarios people start with

Sales dashboard on accounting-system data — the "Excel dashboard" scenario. One run, 40 minutes, output is an XLSX or a Google Sheet with three sheets. Right when you need a file you can hand to a colleague: they replace the export and everything recalculates.

Build a sales dashboard for last quarter in XLSX on accounting-system data. Structure the file like this: a "Source" sheet — a flat table with no merged cells and a date column in one format; a "Summary" sheet — pivots by channel, month and manager; a "Dashboard" sheet — charts and slicers tied to the pivots, not to the source directly. Define pivot ranges as whole columns or named ranges. Do not hard-code computed values into dashboard cells. On a separate "How to update" sheet describe where to paste the new export and what not to touch.

Sales from the accounting system, Bitrix24, Ozon and Wildberries in one table — the "Sales in one table" scenario. Monthly, 30 minutes. This is not ready conclusions but a raw base for any reports that follow. The key part of the task is the matching rule: by SKU or barcode, not by name, and what to do with what did not match (put it on a separate sheet with candidates rather than glue it together by eye).

Profit by business unit from the accounting system — the "P&L by unit" scenario. Monthly, 25 minutes. Here the number most often breaks on shared costs — rent, bookkeeping, management. If there is no rule for splitting them, the honest result is an "unallocated" row with an amount, not costs spread across units by someone's feel.

In all three scenarios the agent reads the sources itself: nothing to export or upload. If the data exists only as a file — a supplier's export, say — the file can be attached in the chat: CSV, TSV, XLS, XLSX and XLSB are accepted, and the agent works with the original file in its environment, not with a preview of the first rows.

Frequently asked questions

Do I need Copilot or a paid Excel? No. The agent works with the file outside Excel and hands back a finished XLSX or writes to a Google Sheet. The assistant inside Excel is a separate tool for a different job: refining a formula in a file that is already open.

Will this replace an analyst? No. It replaces the part of an analyst's work done by copy-paste: export, clean, merge, refresh. Decisions about what counts as revenue and how to split shared costs stay with a person, and the agent asks about them rather than guessing.

What if products have no common SKU? Then there is nothing to match on, and the agent will say so: the rows go to an "Unmatched" sheet with the amount they carry. That beats a table where half the pairs are glued together by similar names.

What happens if the accounting system or the seller account does not respond at refresh time? Whatever you wrote in the task: a blank cell, a separate "Gaps" sheet, a line in the summary. If you wrote nothing, the behaviour is left to the agent's judgement, and that is a poor thing to rest a report on.

Can I start with Google Sheets and move to XLSX later? You can, but these are two different builds, not an export: slicers and some pivots are not supported in one of the programs. Better to say up front where the file will be opened more often.

How much time does it cost? The first build — 25–40 minutes of your attention: name the breakdowns, answer the questions about VAT and dates, look at the check. After that — zero if the task is on a schedule, and five minutes per new export if the file lives as XLSX.

Last updated: 8 September 2026.

Sources: Anthropic — How Anthropic enables self-service data analytics with Claude; Anthropic — Advancing Claude for Excel and PowerPoint; Zapier — Copilot in Excel.