Before an AI answers a business question, it should settle five things: which number you mean, which dates, which channel, which product, and compared with what. A fast answer that skips them is a guess that sounds sure. You can test any AI with six words: "How did we do last month?" Then count what it asks.
Six words. That is all it takes.
“How did we do last month?” You type it into Claude at 8:40 at night.
Four seconds later: “Last month you did $138,000 in sales, up 9%.”
You say the number out loud in the next meeting.
Nobody asked you which sales, which dates, or up from what.
Before an AI answers a business question, it should settle five things: which number you mean, which dates, which channel, which product, and compared with what. A fast answer that skips them is a guess, and it does not say so.
This post shows the five questions, what skipping them costs in one made-up example, and a six-word test you can run on any AI today. Every number below is illustrative, and the brand is invented.
Why a quick answer is a guess
An AI model starts every chat knowing nothing about your business. Anthropic’s own prompting guide says to think of its model as “a brilliant but new employee who lacks context on your norms and workflows.” It adds that the more precisely you explain what you want, the better the result (Anthropic prompting guide).
A new hire handed your six words would not answer at once. They would ask what you mean. An AI is built to be helpful, so it often fills each gap with the most likely choice and keeps going.
The risk is not that the AI makes things up. The risk is a confident answer built on a choice you never saw.
Our post on why AI gets ecommerce numbers wrong covers what the data itself fails to say. This post covers the other half: what the question fails to say.
The five questions
Each row is a gap in your six words. Each has a cost when the AI guesses.
| Gap | What the AI should ask | What a silent guess does |
|---|---|---|
| Which number | Orders placed, sales after discounts, or cash paid out? | Adds figures that mean different things |
| Which dates | The calendar month, or the last 30 days? Is the month closed? | Compares half a month with a whole one |
| Which channel | Amazon, your own site, wholesale, or all of them? | Leaves a channel out, or counts one twice |
| Which product | One item, one range, or the whole catalog? | Splits one product across five listings |
| Compared with what | Last month, last year, or your plan? | Calls a normal seasonal dip a drop |
A word on the first row, since it causes the most damage. Shopify defines gross sales as product price times quantity, before taxes, shipping, discounts and sales reversals. Its net sales are gross sales minus discounts and minus sales reversals, which are order adjustments that bring a negative amount (Shopify sales report). Two Shopify numbers, both called sales, and they differ.
Amazon adds another. Its settlement reports, which show money paid out, are created on Amazon’s own schedule, and you cannot request one (Amazon SP-API docs). So “sales” on Amazon can mean orders placed or cash that has settled, and those do not line up by month. We cover that gap in settlement report vs orders report.
One question, four answers
Here is the made-up case. A brand called Biscuit Barn sells dog chews on Amazon and on its own Shopify store. The owner asks the six-word question about October.
The data holds these figures:
| Figure | Amount |
|---|---|
| Amazon ordered sales | $96,000 |
| Shopify gross sales | $42,000 |
| Shopify discounts | $3,000 |
| Shopify sales reversals | $1,000 |
| Amazon cash settled in October | $71,500 |
Shopify net sales are $42,000 minus $3,000 minus $1,000, which is $38,000.
The fast answer added Amazon ordered sales to Shopify gross sales: $96,000 plus $42,000 is $138,000. That is a sum of two different bases.
Put orders on the same footing and you get $96,000 plus $38,000, which is $134,000. Amazon cash alone is $71,500, and it is not a month of orders at all.
Three answers to six words, and the one that came back fastest was the mixed one. Illustrative numbers for a made-up brand.
The gap between the fast answer and the clean one is $4,000. That sounds small. It is the discounts and reversals the owner never saw, and it hides a worse habit. If the fast answer says “up 9%,” the owner cannot tell whether that compares like with like.
None of these figures is wrong. Only a labeled one can be used.
Now say the owner had asked the five questions first. The conversation takes about thirty seconds:
- “Which number: orders placed, or cash paid out?” Orders placed.
- “Which dates: calendar October, which is closed?” Yes.
- “Both channels?” Both.
- “All products?” All.
- “Compared with September or October last year?” September.
The answer now reads: “Orders placed in October, net of discounts and reversals, were $134,000 across both channels. September on the same basis was $128,000.” That is a figure you can act on, and the basis is printed in it.
Ask first, or state the assumptions?
Some owners hate being quizzed. Fair. An AI that fires five questions at every message is a bad employee too.
There is a middle path. For a small question, the AI should answer and state what it assumed. For a question that moves money, it should ask first.
| Kind of question | Better behavior | Why |
|---|---|---|
| “What was Shopify net sales on Tuesday?” | Answer, with the basis | Little room to be wrong |
| “How did we do last month?” | Answer with assumptions listed, and offer to switch | Several bases are plausible |
| “Should I raise the ad budget?” | Ask first, then answer | A wrong basis costs real money |
| “Why did profit fall?” | Ask, then show causes with dollars | It needs joined numbers |
The last row needs care. A “why” answer often leans on two numbers that moved together. The Australian Bureau of Statistics warns that a correlation does not automatically mean one change caused the other. It names a controlled study as the most effective way to establish cause (Australian Bureau of Statistics). A good AI says how sure it is, not just what it found.
Can AI tools actually ask you things?
Yes, and the plumbing exists. The Model Context Protocol is the standard way AI apps connect to other software. A version of its specification lets a connected server request additional information from the user through the app, with the app staying in control of what is shared (MCP specification).
The same page marks the feature as newly introduced in that version and says its design may change. So do not assume every AI app supports it. Today the simplest way to get questions is to ask for them in plain words, as the test below does.
Run this test: the six-word test
It takes ten minutes. You need one AI app that can see your sales data, and the real answer from your own finance numbers.
- Type the six words. “How did we do last month?” Nothing else.
- Count what it asked. Write down each question, or each assumption it stated.
- Check the basis. Did the answer say whether it used orders placed, net sales or cash?
- Check the dates. Did it say whether the month was closed?
- Ask for the total again, this time with all five gaps filled. Compare the two numbers.
What a pass and a fail look like:
| Result | What you see | What it means |
|---|---|---|
| Pass | It asked, or listed assumptions on at least four of the five gaps. Both runs agree within a rounding error. | The AI reads your data in a labeled way |
| Weak | It stated the basis but not the dates or the channel. The numbers differ by a few percent. | Usable for rough reads, not for decisions |
| Fail | It gave one number with no basis. The second run is far off. | Treat every answer from this setup as a draft |
A fail is useful. It tells you the AI is guessing today, before the guess reaches a meeting.
Run it again a week later with a different teammate typing. If two people get two numbers from the same words, the gap is in the setup, not in the people.
Where asking first goes wrong
- Too many questions. Five is the limit. If the AI needs ten, your data has too many unwritten rules.
- Questions with no default. “Which revenue?” is weak. “Orders placed, or cash paid out? I will use orders unless you say” is better.
- Asking what it could know. If your definition of net sales is written down, it should not ask. It should use it.
- Asking, then forgetting. An answer you gave on Monday should hold on Tuesday. If it does not, you are the one doing the data work.
- A clear question on muddled data. Perfect questions cannot unmix two meanings of sales inside the files.
How to get better answers this week
- Write your own five answers once. Which revenue, which calendar, which channels, how products group, and what you compare against.
- Paste them at the top of your chats. A short standing note beats retyping your meaning each time.
- Ask for assumptions. Add: “List what you assumed before the answer.”
- Ask for the basis on every figure. “Orders placed” or “cash paid out” should sit beside each number.
- Re-run the six-word test monthly. Definitions drift, and so do tools.
Our guide to AI-ready data shows how to write those definitions so they travel with the numbers instead of with you.
Where Synthesis fits, and where it does not
A standing note works for one person in one chat. It breaks down when three teammates ask in three apps, or when a number needs to match the books.
Synthesis is ontology software and an AI harness in one plan. It writes the business down once, so every number means one thing, and it keeps the AI you already use on those numbers. That means fewer questions to ask, because the answers to “which number” and “which product” are already written. It also checks every number against its source. Synthesis has no chat of its own. You ask in Claude, ChatGPT or another AI app.
If you have one channel, one product and one person asking, a standing note and the six-word test may be enough for a long time. A brand brain earns its place when the questions cross channels.
The short version
An AI that answers in four seconds has made five choices for you: the number, the dates, the channel, the product and the comparison. A good answer either asks about them or says what it chose.
Run the six-word test on your own setup this week. If the answer comes back with one number and no basis, you have found the gap before it reached a decision.
Questions people ask
How do I ask AI about my business data?
Name the number, the dates, the channel, the product and the comparison in your question. “Net sales on Shopify for October, all products, against September” leaves the AI nothing to guess. If you leave one out, a good AI should ask or say what it assumed.
Why does AI answer instead of asking me what I mean?
Most AI apps are built to give a helpful answer in one turn. Asking feels slow, so the model fills each gap with a likely guess. Unless the guess is stated, you cannot tell which parts were guessed.
What should an AI ask before answering a revenue question?
It should ask which kind of revenue you mean, such as orders placed, sales after discounts or cash paid out. Then it should ask which dates count, and whether the month is finished. Those two gaps cause most wrong totals.
Is it better for AI to ask questions or to state its assumptions?
For a simple question, stating assumptions is faster. The AI answers and says what it chose, and you correct it in one line. For a question that moves money, such as an ad budget, asking first is safer.
Can I make Claude or ChatGPT ask clarifying questions?
Yes, for one chat. Tell it to list its assumptions and ask about anything unclear before it answers. That helps, but the next chat starts from zero. Lasting fixes are written definitions the AI can read every time.
Does asking questions first make AI answers more accurate?
It makes the question clearer, which helps. It cannot fix wrong inputs. If the data mixes two meanings of sales, a clear question still gets a muddled answer. Clean, labeled data and a clear question work together.