AI Tools for Business Analysis: 6 Questions to Ask Your Sales Spreadsheet

Six questions are worth asking your sales spreadsheet: who are my top customers, which products make the most margin, what is slow-moving, how seasonal am I, who has stopped buying, and what happened after my last price change. Ask each in ChatGPT or Claude, make it show its working, then check one figure against your own filter before acting.

The checks are the point. A chat assistant will answer any question about a sheet, in confident prose, whether or not the sheet supports it. Blank cells, duplicates, refunds and text-formatted numbers all move the answer, and the tool rarely mentions them unless you ask. Each question below comes with a prompt you can paste, a sample answer, and a short check that takes under five minutes. For a wider look at what to upload, see what to upload and check when AI analyses a sales spreadsheet.

Follow me on Instagram@sagnikteaches

Get the sheet ready: ten minutes that decide everything

Before any question, prepare one clean table. The tool can only be as right as the data it reads.

Connect on LinkedInSagnik Bhattacharya

1. One row per sale line. Date, customer, product, quantity, revenue, cost. No merged cells, no blank header rows, no totals row at the bottom.
2. Real dates and numbers. A date that is text ("3rd Nov") or a price with a currency symbol stored as text will not add up. Select the column and check that it right-aligns, which numbers do by default.
3. Customer codes. Replace names with codes (C001, C002) in a copy if you prefer, and keep the key to yourself. Remove emails, addresses and phone numbers entirely.
4. Refunds and credit notes. Keep them as negative rows or in their own column, and tell the tool which it is.
5. A control total. Sum the revenue column yourself and write the figure down. You will use it to check that the tool read the whole file.

The control total is the cheapest check there is. Ask the tool first, "How many rows did you read and what is the total of the Revenue column?" If either number differs from yours, stop and fix the file before asking anything else.

Subscribe on YouTube@sagnikteaches

The example used throughout: a catering company

Illustration only, with made-up figures. Say a catering company has a twelve-month sales sheet of 3,100 rows with these columns: Date, CustomerCode, ProductLine, Item, Units, Revenue, Cost. Four product lines: corporate lunches, weddings, canapés, hampers. The owner's control totals are:

These totals are what the owner knows in advance, from their accounting software. Everything the assistant says afterwards is tested against them. The first prompt in every session is the same:

I have attached my sales sheet. Before analysing anything, tell me: how many data rows you read,
the date range, the sum of Revenue and Cost, and any rows you excluded or could not read.
Do not interpret the data yet.

Here is an illustrative first reply that should stop the session:

Rows read: 3,101. Date range: 1 Nov to 31 Oct.
Revenue total: $824,000. Cost total: $573,200. Rows excluded: none.

The owner's sheet has 3,100 rows and $412,000 of revenue. Both totals have exactly doubled and there is one extra row, which is the signature of a totals row left at the bottom of the export: the tool counted it as one more sale worth $412,000. Delete the row, upload again and ask the same question. Only move on when the row count and both totals match yours, because every later answer, from top customers to seasonality, would otherwise be built on figures twice their real size.

Question 1: Who are my top customers, and how dependent am I on them?

Why ask. A business where one customer pays 15 per cent of revenue behaves differently from one where the top customer pays 3. Concentration decides how worried you should be when someone leaves.

Rank customers by total Revenue over the whole sheet. Show the top 10 with revenue, share of total revenue,
and cumulative share. Then tell me how many customers make up 50% and 80% of revenue.
Show the formula or code you used.

Illustrative answer. The top customer, a corporate client, accounts for $61,800, or 15.0 per cent of $412,000. The top five total $156,560, 38 per cent. It takes 31 customers to reach 80 per cent of revenue, out of 214.

How to check. Use a pivot table in your own sheet: customers in rows, sum of revenue as values, sorted largest first. Compare the top customer's figure to the tool's, and check that the cumulative share reaches 100 per cent. Watch two traps. The same customer entered under two codes splits the total, so check the next few names for near-duplicates. And a customer who only appears via refunds may show a negative total that distorts the ranking.

What to do with it. A 15 per cent customer deserves a named contact and a standing check-in. A long tail of one-off customers is a marketing question. The related tutorial on spotting at-risk retainers shows what to do when concentration is high.

Question 2: Which products really make the margin?

Why ask. Revenue rankings and margin rankings disagree more often than owners expect. A line that dominates sales can be your thinnest.

For each ProductLine, calculate total Revenue, total Cost, gross margin in dollars and as a percentage of revenue.
Then do the same for each Item within the two lines with the lowest margin percentage.
Show the table and say which lines are below the overall margin. Show your working.

Illustrative answer. The tool reproduces the table above, then notes that hampers earn 20 per cent against an overall 30.4 per cent, and that within hampers the "luxury" hamper earns 12 per cent because its cost includes an outsourced gift box.

How to check. Sum the Cost column for one product line with a filter and =SUBTOTAL(109, G2:G3101) on the filtered range. If your cost column is missing for some rows, the margin is overstated, so count blank cells in Cost first: =COUNTBLANK(G2:G3101). A margin that looks too good is most often a cost column that was never filled in.

Limits. Gross margin here means revenue minus direct cost. It leaves out labour and overheads unless you put them into Cost. If you want a fuller picture, the tutorial on checking margins product by product goes further.

Question 3: What is slow-moving, and is it dead or just seasonal?

Why ask. Items that do not sell tie up stock, menu space or attention. But an item that is quiet in March might be your best in December, so "slow" needs a time window.

List every Item that sold fewer than 5 units in the last 90 days of the sheet. For each, show total units sold
in the whole year, date of last sale, and the month with its highest sales. Separate items that sold well earlier
in the year (possibly seasonal) from items that never sold much.

Illustrative answer. Seven of 48 items sold fewer than five units in the last 90 days. Four of them, such as a summer salad box and an iced dessert platter, peaked in June and July, so they are seasonal. Three, including a gluten-free tart box, never exceeded 12 units in any month.

How to check. Filter the sheet to one of the flagged items and sort by date. Confirm the last sale date and count the recent units by hand. Check also that the item name is spelled consistently, because "Tart box GF" and "GF tart box" would be treated as two items and one of them would look dead. A quick pivot with items as rows exposes these.

What to do with it. For the seasonal four, nothing, except note when to bring them back. For the three genuinely slow items, decide: drop, reprice, bundle, or push once. The decision is yours, and the tool does not know that the gluten-free tart box is there for one regular customer.

Question 4: How seasonal am I, and when do I need extra capacity?

Why ask. A monthly profile shows when cash comes in and when staff and stock are needed. It also stops you judging a quiet January against a busy December.

Total Revenue by calendar month, with each month's share of the year and the average of the 12 months.
Mark months more than 25% above or below average. Repeat for each ProductLine as a small table.
If there is only one year of data, say clearly that the seasonal pattern is based on a single year.

Illustrative answer. Average monthly revenue is $34,333. December is $58,000 (14.1 per cent of the year), 69 per cent above average. January is $21,000 (5.1 per cent), 39 per cent below. Weddings peak in June and September, corporate lunches in March and November, hampers almost entirely in December.

How to check. Make a pivot with months in rows and revenue as values, and confirm December and January. Check the twelve months add up to your control total. If the sheet starts mid-year, partial months at either end will look falsely low, so look at the first and last dates.

A caution on single years. One year of data cannot separate a seasonal pattern from a one-off event, such as one large wedding or a client who hosted an annual conference. The tool will usually mention this only if you tell it to, which is why the prompt asks. With two years, ask for the same month in each year side by side.

Question 5: Who has stopped buying?

Why ask. Lost customers are quiet. Nothing happens; an order just does not arrive. A list of who has gone silent, and how much they used to spend, tells you who to call first.

Today is [date]. Find customers who placed at least 2 orders in the first 6 months of the sheet and none in the last
90 days. For each, show total revenue, number of orders, average gap between orders and date of last order.
Sort by total revenue. Separately, count customers who ordered only once.

Illustrative answer. Of 52 customers who placed at least two orders in the first six months, 14 (27 per cent) have not ordered in the last 90 days. Their combined revenue was $38,900. Five of the 14 had an average gap of about two weeks and have now been silent for over 90 days, which is unusual for them.

How to check. Pick the three highest-value names and filter the sheet by their codes. Confirm the last order date. Then ask whether silence is real: a customer with an average gap of four months is not lost after 90 days. The prompt asks for the average gap so you can tell a quiet buyer from a gone one. The tutorial on running a churn analysis goes further into reasons.

What to do with it. Send a personal message to the top five, not a mass offer. A sample: "Hi [first name], I noticed it has been a while since your last order. Is there anything we could have done better, or is it simply a quiet spell?" Record the answers, since a reason is worth more than a number.

Question 6: What happened after my last price change?

Why ask. Owners often raise prices without knowing the effect. A before-and-after comparison shows whether the volume dropped by more than the price rose.

On 1 July I raised the price of the "Corporate lunch box" from $14.50 to $15.50. Compare the three months before
with the three months after: units per month, revenue per month, number of distinct customers, and average units
per order. Tell me if the sample is too small to draw a conclusion. Show your working.

Illustrative answer. Before: 1,150 units a month and $16,675 revenue. After: 1,090 units and $16,895 revenue. Units fell 5.2 per cent, the price rose 6.9 per cent, and monthly revenue rose $220, or 1.3 per cent. Distinct customers fell from 63 to 59.

How to check. Multiply the numbers yourself: 1,150 x 14.50 = 16,675 and 1,090 x 15.50 = 16,895. Then check that nothing else changed in the same window, because a price rise coinciding with a lost big client or a seasonal dip proves nothing. Look at the four lost customers by name and ask whether price was their reason.

Be careful with the conclusion. Revenue rose slightly, but if cost per unit also rose in the same months, margin may have fallen. Ask for margin before and after, too. If you want a model to guess at causes, ask it for hypotheses to test, not for a verdict. A tool that says "the price rise caused the churn" has gone beyond what the sheet can show.

The same six questions in a coworking space

The questions travel to any business with a sales sheet. Say a coworking space exports its membership and day-pass invoices: Date, MemberCode, Product (hot desk, fixed desk, private office, meeting room, day pass), Revenue, and a Cost column that holds only direct costs such as cleaning supplies and refreshments. Here is how each question changes shape (all figures illustrative):

• Top customers. A private office tenant paying $2,400 a month is 20 per cent of a $144,000 year. The check is a pivot by member code, and the usual trap is a company that holds three offices under three invoices.
• Margin by product. Meeting rooms often look cheap to run and earn well per hour, while hot desks carry the shared costs. Ask for margin per product and also revenue per square metre if you hold that column.
• Slow movers. Here the slow item is a product, such as a fixed desk at the far end of the floor that has been empty for four months. The tool lists it, and the decision is yours: reprice, convert to a meeting room, or leave it.
• Seasonality. Day passes may dip in August and spike in September. Ask for a monthly profile of each product, because the total hides which line moves.
• Churn. Members leaving is the central question. Ask for members who paid in each of the first six months and none in the last two, plus how long each stayed. A pattern such as "most leavers go in months four to six" points to onboarding rather than price.
• Price changes. After a hot desk rise from $180 to $195 a month, compare new sign-ups per month before and after, not only revenue from existing members, who may not have noticed yet.

The checks stay the same: a control total, one figure rebuilt by hand, and a sceptical eye on any "because".

Follow-up prompts that make each answer more useful

The first answer is rarely the last. These follow-ups turn a number into something you can act on, and each has an in-built check:

• "Show me the five rows that most influence this answer." Opens up the actual lines behind a total, so you can see whether one oversized order is driving it.
• "What would change this answer if I excluded the single largest customer?" Tests how fragile a conclusion is.
• "List anything in the data that looks wrong or unusual." Surfaces negative quantities, dates in the future, or a price of zero. Treat the list as a set of leads and verify each against your records.
• "Write the answer as three sentences for someone who has not seen the sheet." Good for a team update. Check the sentences against the table before you send them.
• "What question should I ask next, and what data would I need?" Often points out that you need a column you do not have, such as acquisition source or cost of delivery.

An illustrative before and after for the last prompt: the catering owner asks it after question 5 and gets "you would need the reason each customer stopped, which the sheet does not record". The owner adds a "ReasonLost" column to the customer list, fills it in from the five calls, and next month has something the data alone could never give.

One more habit worth building is a "what I decided" line under each answer in your notes. Write the figure, the check you ran, and the action you took, for example: "Top five customers are 38 per cent of revenue, checked with a pivot, booked quarterly catch-ups with the top three." Three months later you can see which answers led to action and which were interesting but unused, and you can drop the second kind from the routine without guilt. A review that produces no decisions is worth shortening.

A short checklist for any answer the tool gives you

1. Totals match. The row count and revenue total in the answer match your control total.
2. One figure reproduced. Pick one number and rebuild it with a pivot table or filter.
3. Window is stated. The date range of the answer is the one you meant.
4. Exclusions are listed. Ask what it left out and why.
5. Method is shown. Formula, code or step-by-step working is visible, not only the result.
6. Causes are hypotheses. Any "because" is treated as a question to test.
7. Sensitive data is out. No personal details were uploaded that the question did not need.

What goes wrong: three realistic mistakes

The text-number trap. A seller exports sales from an invoicing tool, and the Revenue column arrives as text with a currency symbol. The assistant reads some rows and skips others, then reports a total 9 per cent too low. Nobody notices until the control total is compared. The fix is to convert the column to numbers, and the habit is to compare totals first.

The duplicate customer. One events client appears under two codes, once with "Ltd" after its name and once without. The top customer ranking understates a $40,000 client by splitting it into $22,000 and $18,000, and the churn list shows one of them as lost when it is actively ordering under the other. A pivot of customer names sorted alphabetically finds this in a minute.

The over-reading. The tool is asked "why did sales drop in March?" and gives three reasons, including "a possible increase in competition". There is nothing in the sheet about competitors. When you ask why, it invents a plausible story. Ask instead "what in this data changed in March?", and take causes from your own knowledge of the business.

Which tool, and how to set the session up

Either ChatGPT or Claude can do all six jobs on a file of a few thousand rows, provided it runs the numbers in code rather than estimating them. ChatGPT writes and runs Python for data analysis and lets you review the code; Claude needs its "Code execution and file creation" setting (labelled "Create and edit files" in the newer settings layout), which is on by default on individual plans and can be switched off by an admin on Team and Enterprise. If you prefer to work inside a spreadsheet, the guides to ChatGPT with Excel and the comparison in Julius AI versus Claude for sales data are good next reads. A few setup habits that save time:

• Start a new chat for each month's review, and attach the new export. Long chats drift.
• Keep your six prompts in a document, with your column names filled in, so a monthly review is copy and paste.
• Use a project or saved instructions to state your column meanings once: "Revenue excludes tax; Cost is direct cost; negative rows are refunds."
• Save each answer with the date next to your own check. Over time this becomes a record of how the business changed.
• For privacy, use the plan's data controls and strip personal fields. Check the vendor's current data settings page before you upload anything about real customers.

A monthly rhythm that takes under an hour

Once the prompts exist, a review is short. Illustrative timing for a small business with a clean export:

1. 10 minutes: export and clean the sheet, note the control totals.
2. 5 minutes: run the control-total prompt.
3. 20 minutes: run the six questions, skimming for anything unexpected.
4. 15 minutes: check the two or three answers that matter, with a pivot or filter.
5. 10 minutes: write three actions and one question for next month.

An action list might look like this: call the five silent customers this week; reprice or drop the three slow tart-box items by month end; check the hamper cost for the outsourced gift box and ask for a second quote. Each action came from one question and each can be tested next month. If you want a second person to look at your setup, the 1:1 call on this page covers it, but nothing here needs it.

The six questions are not the only ones, but they cover the choices small owners keep making without evidence: who to protect, what to sell, what to drop, when to hire, who to call and whether a price held. A tool that answers them in minutes is useful. A habit of checking each answer is what makes it safe.

Questions owners ask before pasting sales data into AI

Is it safe to upload customer names to ChatGPT or Claude?

Check the plan's data settings first. On individual plans you can switch off training on your chats ('Improve the model for everyone' in ChatGPT, the Model Improvement toggle in Claude), and both vendors say business plans are not used for training by default. Still, replace customer names with codes and drop emails, addresses and phone numbers before you upload.

How big a spreadsheet can I upload?

Limits change by plan and tool, so check the current help page. As a rule, send a clean export with only the columns you need. A summary table of a few hundred rows is easier for the tool, and easier for you to verify, than a raw file with fifty columns.

Which is better for this, ChatGPT or Claude?

Both can read a spreadsheet and run calculations when their analysis tools are on. The better choice is the one that shows its working. Whichever you use, ask it to show the calculation or the code it ran, so you can reproduce a number in your own sheet.

What if the AI's answer differs from my own sums?

Trust your sheet until you find the reason. The usual causes are blank rows, text stored as numbers, duplicate lines, refunds counted as sales, and different date ranges. Ask the tool to list the rows it included and excluded, then compare with a filter in your own file.

Further reads

• AI Food Insights Tool: Toast IQ, Square AI or a Spreadsheet? — Built-in analytics versus a spreadsheet in hospitality.
• How to Price Catering: A Cost-per-Person Sheet With AI — Build the cost-per-person sheet this example relies on.
• How Small Shops Can Use AI to Forecast Stock and Reorder — Use your sales history to forecast reorders.
• Can AI Help a Small Shop Set Prices and Promotions? — Test price and promotion ideas on your own data.
• What Is AI Churn Prediction and Can a Small Gym Use It? — What churn prediction means for a membership business.
• How to Run a Stocktake Faster With AI and a Phone Scanner — Reconcile sales with a faster stocktake.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help centre (Data analysis with ChatGPT, Data controls, Projects) and business data page; Claude help and privacy centres (Create and edit files, model training for consumer and commercial plans); figures in the worked example are illustrative, not real client data.