Ask an AI for revenue and you get someone's definition. Make sure it's yours

One question, three answers

Here's a scene that's playing out in a lot of marketing teams this fall. Someone asks an AI assistant, "What did email bring in last month?" It answers with a confident number. A colleague asks the same thing in a different chat and gets a different number. The finance lead pulls the figure from the store's reports and gets a third.

Nobody made a mistake, exactly. Each answer used a different definition of revenue. One counted gross order value. One subtracted refunds. One counted only orders a marketing tool attributed to an email. All three are reasonable. Only one is the number your company reports, and the AI didn't know which.

This used to be an analyst problem, solved by an analyst who knew the house rules. Now it's everyone's problem, because anyone can ask an AI tool a business question in plain language and get an answer in seconds. A number without its definition is an opinion with a decimal point.

What Klaviyo just changed

Klaviyo, the email and SMS platform many online stores use, shipped something on this front in its October release notes. It's called SQL in Klaviyo Data Platform. You ask a plain-language question, and Klaviyo turns it into a database query using its own definitions of terms like revenue, runs it against your account's data, and returns the answer along with the definition it used and the query that produced it.

You can ask in Composer, Klaviyo's built-in assistant, or through Klaviyo's MCP server, the connection that lets outside AI tools such as Claude, ChatGPT or an agent you've built talk to your Klaviyo data. Klaviyo's own write-up names the exact problem above: an AI guessing at what revenue means, gross or net, refunds in or out, can give you different answers to the same question in two chats.

That's a real improvement, and worth turning on if your account has Klaviyo Data Platform. But notice what it solves and what it doesn't. It makes the definition visible. It doesn't make the definition yours. Klaviyo's definition of revenue is a sensible one for a marketing platform. Your finance team may use another, and your board deck almost certainly does.

Answer one: the tool's definition

The first kind of answer uses whatever the tool decides. With a well-built product like the one Klaviyo describes, you at least get told what that was. With a general chatbot connected to a spreadsheet, you often don't; it picks something plausible and moves on.

The tool's definition is fine for exploring. "Which flows seem to be doing the most work?" doesn't need finance-grade precision. It isn't fine for anything that ends up in a forecast, a budget request or a bonus calculation. The test is simple: if the number will be quoted to someone who wasn't in the chat, the definition has to be quoted with it.

Answer two: the analyst's definition

The second kind of answer comes from a person who knows the house rules, in their head. They know that wholesale orders are excluded, that the spring promotion's refunds were booked late, and that "last month" means the fiscal calendar. Their numbers match finance because they've been corrected by finance before.

The trouble is that none of that knowledge is in the AI tools. When your team switches from asking the analyst to asking an assistant, you lose those rules unless someone writes them down. We've suggested before that every key operating number deserve a short written definition: what's counted, what's excluded, where it comes from and who owns it. The AI era makes that note a working document instead of a nice idea.

Answer three: your definition, written where the tools can see it

The third kind of answer is the one you want: an AI tool working from your company's written definition and showing that it did.

Getting there takes a little setup. Write down the three or four metrics people ask about most, usually revenue, orders, new customers and repeat rate, each in a sentence or two. Decide which system is the source of truth for each. Then put those definitions where your AI tools will read them: in the instructions for any custom agent, in a pinned note for your team's assistant, in the prompt templates people reuse. When a tool's built-in definition differs, as Klaviyo's may, write down the difference so people know which number they're looking at.

Keep the definitions short and concrete. "Revenue: order value after discounts, before tax and shipping, minus refunds issued in the same month, excluding wholesale and test orders, from the store's order export" is more useful than a paragraph about principles. If two teams genuinely need different versions, give them different names, such as "booked revenue" and "attributed revenue," rather than letting one word mean two things. A lot of arguments about AI numbers turn out to be two teams using one word.

Then make showing the working a habit. Any number that leaves the chat travels with its definition and the query or report it came from. Klaviyo's tool does this for you. For tools that don't, ask for it.

The spot-check, and when to stop asking AI

Klaviyo's release notes give the right advice for anything new: check early answers against a number you already trust before you rely on them. Pick a month you've closed, ask the question, and compare. If the answers match, good. If they don't, the difference is almost always a definition, and finding it teaches you something about both systems.

Some questions shouldn't go to a chat at all. Numbers that pay people, numbers you file and numbers in a board pack should come from a report with an owner, not a fresh conversation. AI is a fast way to explore and a slow way to settle arguments. We've reviewed Klaviyo's Composer separately; the same rule applies there.

If you're connecting AI tools to your store, email and finance data and want them to agree with each other, that's the kind of work we do in AI applications and agents. Get in touch and bring the three numbers that never match.

Related work: Sutton Place.