Can you trust AI with your business numbers?

The short answer

Yes, you can trust an AI answer about your business numbers when three things are true. Each number has one written meaning that someone owns. The AI may answer only from what is written. And you can open the calculation behind every answer. If any of the three is missing, treat the answer as a draft, however confident it sounds.

The common setup today fails all three. Someone pastes a spreadsheet export into an AI chat, or connects the chat straight to the company’s database, and asks for revenue. The AI decides for itself what “revenue” means, and it does not say which meaning it chose. It fails quietly: the number looks right. If your team already pastes exports into a chat, there is no need to ban it; the section on pasted exports says what to do instead.

You should know who is telling you this. We are OWOX, and we make OWOX Data Marts, a product where an analyst defines each number once and people then get that number in their spreadsheets, their dashboards and their AI chat. So we have a view. We have tried to be fair to the other ways of doing this, and we say below when one of them is the better choice.

Two ways an AI answer can be wrong

People usually worry about the first. The second does more damage.

A different answer each time. Two people ask about last quarter’s revenue and get two figures. It is embarrassing, and it is also visible: someone notices, and someone checks.

The same wrong answer each time. The AI settles on one meaning of revenue, for example one that still counts cancelled orders, and repeats it faithfully. Nobody checks, because the number never changes. “It gave me the same figure twice” tells you the AI is consistent. It does not tell you the figure is the one your finance team would sign.

One illustration from our own product, not a measurement: on 9 October 2026 we asked a demo shop’s data for revenue three ways, in three separate chats. Two answers led with net revenue of $849,177. The third led with gross revenue of $998,141 without looking at the data: the AI reused an earlier conversation, and said so. Both figures were correct for what they measured; which came first was the AI’s choice. It shows the failure can happen, not how often. The full run is in our guide to AI analytics on BigQuery.

Why the wording of a question changes the answer is covered in why AI gives different answers to the same data question. And when a supplier says its AI is “90% accurate”, that figure was measured on someone else’s questions and someone else’s data, not on yours.

Why it happens: the AI has to guess what you mean

It is rarely the arithmetic. The AI adds up correctly; it adds up the wrong things.

“Revenue” has several honest meanings inside one company: before or after refunds, with or without cancelled orders, with or without tax, counted on the order date or the payment date. A new colleague would ask which one you want. An AI picks one and carries on.

How much it has to guess depends on what it was given:

Three things follow from that table:

• When the AI writes its own database query, the guesses include table and column names that do not exist. That is covered in AI analytics hallucination: why it happens.
• Loading your documentation into the AI helps it find a definition, not the right one. If three documents describe revenue three ways, it still has to choose.
• When a colleague builds a dashboard by describing it to an AI, the same guessing happens out of sight. Vibe analytics is a hallucination machine lists what goes wrong.

If people already paste exports into a chat

This is the most common setup in smaller teams, and banning it rarely works. Set three rules instead:

• Treat the answers as drafts, not decisions, until the number is written down and its calculation can be opened.
• Keep customer-level and personal data out of pasted files: names, emails, addresses, and any figure that must not leave the company. Everything in a pasted file is sent to the AI company.
• Find out what has already been pasted, by whom, and where those answers ended up.

Three things that have to be true before you rely on an answer

They hold whichever tool you use, ours included.

1. The number has one written meaning, and a named person owns it. Written means it says what the number counts and what it leaves out. Owning it means three things: you decide when the meaning changes, you agree the change with the people who use the number, and you write the change where the AI reads it. It does not mean vouching for every answer a chat gives. You may already have this without buying anything: a model in your reporting tool, metric definitions your team keeps in code, or a documented table can all count.
2. The AI may answer only from what is written, and says so when it cannot. We tested this on our demo project on 9 October 2026 by asking for headcount, which the project holds no data on. The reply was: “I can’t give you a headcount from this connector: it has no employee or HR data.” A refusal is the correct answer there.
3. Every answer has a record you can open. The record shows the question, who asked it and when, which definition the answer used, and the calculation that ran. If an answer has no record behind it, you cannot tell whether it came from your data.

In OWOX Data Marts that record is called Run History. Each entry shows when the question came, who asked, the dataset used (here, Purchases) and the calculation that ran, in full.

If your team already keeps its definitions in a tool that meets these three, that is the tool to use.

Where your figures go when you ask

To write a reply, the AI has to read the figures behind it. So those figures are sent to the company that runs the AI: Anthropic if you use Claude, OpenAI if you use ChatGPT. Not only your question. The figures.

You choose what the AI may read. It can only send on what it is allowed to see. Decide that list before anything is connected, and leave out whatever must not leave the company, such as figures that identify a client, or payroll.

Put four questions to any supplier, and ask for the answers in writing:

1. Which AI company receives our figures?
2. What is kept, and for how long?
3. Is anything used to train the AI?
4. How is the list of what the AI may read enforced, and who can change it?

Our own answers, in one sentence: the AI gets only the figures an analyst has chosen to share, we keep a note of each question but not the figures, and our documentation does not say how long that note is kept or what the AI company does with the figures, so ask us in writing.

The detail is for your analyst, from our documentation as it reads on 10 October 2026. The connection between OWOX Data Marts and an AI chat, called the MCP server, is part of OWOX Data Marts Cloud.

1. Who receives the figures: the company behind your AI chat. Our documentation says that when the AI asks a Data Mart (a dataset an analyst has defined) for figures, “the resulting data rows and totals are sent to that provider”.
2. What is kept: on our side, the record of each question keeps “the query definition and executed SQL only”, never the figures themselves, according to the same documentation. How long that record is kept, and what the AI company keeps, our documentation does not state. Ask us, and the AI company, in writing.
3. Training: our documentation does not state this. It is the AI company’s policy; ask it, and us, in writing.
4. Enforcement: the AI can ask only for Data Marts an analyst has published, and only for “data you are permitted to query”. What each person may see is set by their role and by business areas called Contexts, and only a project admin can create or change a Context.

Inside your own company, the people to ask are whoever handles security, or your auditors. In a small company with neither, agree the list between you, the founder and whoever runs finance, and write it down.

What it takes, and what it does not fix

The long step is agreement, not connection. Connecting an AI chat to your data is short. Getting the people who use a number to agree what it means is long, and no tool shortens it. We will not give you hours; it depends on how many meanings are in use today.

If nothing is written down, start with one number, the one your leadership argues about most:

1. Find every place it is calculated today: reports, spreadsheet formulas, saved queries.
2. Write what it counts and what it leaves out.
3. Put that definition in one place the AI will read.
4. Get the person who uses the number to sign it. In a small company that is usually the founder or whoever runs finance.

That can be done by hand, this week, with no new tool. Step 4 depends on other people and is usually the slow one.

It is not a one-off. The business changes its mind about what counts, and the written meaning has to follow.

Three things stay unfixed, in every tool:

• The AI can pick the wrong number from a correct list.
• It can do its own sums on correct figures and get the sums wrong.
• It cannot combine numbers that are not stored in one place. If yours sit in several tools or several warehouses, start with why cross-tool joins need a warehouse.

And sometimes none of this is needed. If you want help writing a query and will check the result yourself, a general AI chat is the right tool. Writing SQL with ChatGPT covers that.

Find your situation

Start from the line that sounds like you.

Pürblack’s chief marketing officer, Mark Simmons, puts it the way we would: “we know the data underneath is solid, so the risk of hallucination is much, much smaller.” Smaller, not gone.

Coming next in this series

• A comparison of AI data analyst tools, including Snowflake’s and Databricks’ own assistants and Copilot in Power BI.
• Whether you need a semantic layer for AI agents, and how to write a definition the AI can use.
• What a text-to-SQL accuracy score says about your own data.
• Why loading documents into the AI does not settle which definition is right.
• How colleagues can build dashboards with AI without each choosing their own revenue.

A check to run this week

You can test whatever AI tool your team already uses, in one sitting, once it is connected to your data.

1. Pick the number your leadership argues about most.
2. Ask for it three ways, each in a new chat.
3. Write down the three numbers, and which one each answer led with.
4. Open the calculation behind each answer. If there is none, you cannot tell where the number came from.
5. Ask one thing your data cannot answer, and see whether the tool refuses.

If your team only pastes exports into a chat, run the same check on one export: ask the same question three ways, ask the AI to state which definition of revenue it used, and compare each answer with the number you would sign.

The full steps, with the prompts we used and what we found, are in our BigQuery guide.

What to send back to the person who asked

Short enough to forward as it is:

• Can we trust AI with our numbers? Yes for the numbers we have written down. Not yet for the rest.
• The risk is not that the AI invents figures out of nothing. It is that it chooses its own meaning of “revenue”. It can do that consistently, so an answer can look stable and still be wrong.
• What I need from you is agreement on what our most-argued number means, and your sign-off on it. That is the slow part, and it is not technical.
• For any figure that matters, I will be able to show you the calculation behind it. If a tool cannot show that, we do not use it for decisions.
• If AI figures have already gone into decisions, including anything in board or investor material, I will re-check the important ones against their calculation.
• People can keep using AI chats on pasted files to explore. Until a number is written down, those answers are drafts, and customer names, emails and other personal details stay out of what is pasted.
• What the AI reads to answer us is sent to the company that makes the AI. We choose what it may read. Let us agree that list, with security or our auditors if we have them, before anything is connected.
• This will not make the AI right every time. It makes a wrong answer something we can catch.

If you want to see the three conditions working on your own data, our product is where we built them, and this page shows how an AI chat connects to it.

AI in AnalyticsAI analytics agents replace your backlog, not your analystJuly 2, 2026→AI in AnalyticsAI analytics hallucination: why it happens and the SQL fixJuly 1, 2026→AI in AnalyticsAI analytics hallucinations: trace every number to SQLJuly 14, 2026→

See all articles →