Picture a leadership team in a mid-sized trading company at an AI product demo. The vendor connects an assistant to the company's data and asks for the first question. The CEO asks for last month's margin. The assistant answers instantly: a clean sentence, a chart, a confident tone. The finance director looks at the screen and says her number is different. The head of sales has a third one.
(Illustrative scenario. It does not describe a specific company or project.)
The tool did not fail. It did exactly what it was built to do: it took the data the company has and served it faster and more confidently than anyone had before. The problem is that this data was already speaking in three voices.
An AI project makes sense when five things about your data are true before you sign the contract: revenue and margin are one number, every number has an owner, the knowledge of how it is calculated does not live in one person's head, manual workarounds are counted rather than hidden, and the project is meant to improve one specific decision. Below, each condition in turn, followed by a one-page exercise that shows how many your company meets.
Why AI does not fix a mess, it speeds it up
AI answers based on what it is given. It does not know that two departments calculate revenue differently. It does not know there is a manual correction in a spreadsheet that only one person remembers. It will not ask which number is the right one. It will pick one and state it with the same certainty as any other.
That is why AI on inconsistent data produces inconsistent results, only faster, better worded and harder to challenge. A person bringing the board a report that does not add up usually adds "but let's double-check." An AI assistant does not.
Data Debt is the backlog of disorder in your data that never appears on the balance sheet, grows every quarter, and is always paid by someone in the end. An AI project is one of the most expensive moments for that bill to arrive. Data Debt stops being a reporting nuisance and becomes a problem with an investment you have just approved.
Condition 1: revenue and margin are one number, not three
The first test is simple. Ask three people (sales, finance, and whoever builds the board reports) what last month's revenue was. Ask about margin too. If the answers differ by more than rounding, you do not have one source. You have several versions.
That does not mean anyone is wrong. Usually every version is correct under its own definition: gross or net, with adjustments or without, by order date or by payment date. The trouble starts when nobody has decided which definition the company uses. Then, as the article on Single Source of Truth describes, the board picks the number it likes.
An AI model cannot settle that. It will receive all three versions and blend them, or it will receive one, chosen by chance by whoever prepared the data for the project.
What "true" looks like: every figure you plan to ask AI about has one definition, written down in one place, and everyone who reports it reports the same one.
Condition 2: every number has a name next to it
The second condition sounds organizational, but it is about data. Ask: who is responsible for this number being correct? Not who prepares it. Who answers when it turns out to be wrong?
When the answer is "it depends" or "sales and finance together," there is no owner. In practice that means nobody will notice when the number stops being true: a system changes, a new product gets a different code, someone edits a formula. For a person this is annoying. For an AI project it is a delayed-action problem, because the model will keep answering on new data using old rules.
What "true" looks like: each key figure (revenue, margin, customer acquisition cost, inventory or occupancy, depending on the industry) has a named person responsible, and that person knows it.
Condition 3: the knowledge of how to calculate it is not in one person's head
The third test is the most uncomfortable. Imagine that tomorrow the person who "always knows how to calculate it" is gone: the analyst, the accountant, someone in controlling. Can anyone else rebuild the management report from the same data and arrive at the same number?
In many companies the honest answer is: not right away, because the knowledge lives in one person's head, in their spreadsheet, and in a few adjustments written down nowhere. I wrote about this in why one analyst leaving can bring a company to a standstill.
In an AI project this problem does not disappear, it moves. Someone has to explain to the system how to calculate. If only one person knows, the whole project depends on whether that person has the time, the will and the memory to explain everything. And then on whether they stay.
What "true" looks like: the way key figures are calculated can be reproduced from a document or a system, not from a conversation with a particular person.
Condition 4: manual workarounds are counted, not hidden
The fourth condition is about what happens between the system and the report. In most companies there is a layer of manual work: export to a spreadsheet, cleaning, filling gaps, a correction the system does not make. Each step has a reason. Together they form an area where data changes without a trace.
For an AI project this is two separate problems. First, the model will not see these corrections, so it will work on the version of the data from before them. Second, if the project's goal is to replace manual work, someone first has to know what that work actually includes. Without that you will automate a step that is already wrong, or skip the one that makes all the difference.
This is also the moment for the question from the article on who should actually be doing the repetitive work in your company. A Digital Worker, meaning automation that does repetitive work instead of an expensive specialist, makes sense when it is clear what that specialist does and why.
What "true" looks like: you know which manual steps sit between the source data and the report, who performs them, and how much time they take in a typical week.
Condition 5: you know which decision the project is supposed to improve
The last condition is about the goal, not the data, but without it the other four are not enough. Before you ask whether your data is ready for AI, ask: ready for what?
"We want to use AI" is not a goal. A goal is a decision someone in the company makes today worse, slower or at higher cost than they could. For example: how much to spend on a campaign, which customers to extend terms to, when to reorder stock, who to move to a different task. If you cannot name such a decision and say in which direction cost or revenue should move, the project is an experiment, not an investment. That can be fine, but it is worth calling it that and budgeting for it that way.
AI makes sense where you can count its effect on cost or revenue. Outside that area it is usually a nice extra that someone pays for month after month.
What "true" looks like: you can name one decision and one measure (cost, revenue, margin, hours of work) that should change if the project succeeds.
The one-page exercise: check your company before you meet the vendor
Take one sheet of paper and five lines, one per condition. Next to each, write one of three answers.
- Yes, I can show it. There is a document, a system or a person who confirms it, and you can verify it.
- Yes, but only according to one person. The condition is met, but there is no trace of it beyond someone's memory.
- I don't know. That is also an answer, and often the most valuable one.
Count the result. It is not a grade. It is a sequence for your decisions.
What this means for your AI decision
If most lines say "yes, I can show it," the conversation with the vendor can begin. You will know what to ask, and you will quickly see whether the proposed solution understands your data.
If several lines say "yes, but only according to one person," that knowledge needs to be written down first. It is not a large job, but it has to be done before the project, not during it. Otherwise the cost of documenting it is folded into the implementation cost, and you pay the vendor's rate for it.
If most lines say "I don't know," the first step is not choosing a tool. It is a diagnosis: what is actually underneath. Only with that map can you see where automation or AI makes sense on your data and what has to be done before it. That is an assessment, not a promise of implementation. It lets you spend the budget on a project that has a chance of working, instead of a demo that looks good until the first question about margin.
In every one of these cases, the same EBITDA Leak, the place where margin quietly escapes through manual work, inconsistent numbers and dependence on one person, becomes visible earlier than after the budget is approved.
An AI vendor can show you how fast the system answers. Only you can check whether it answers on the right data.
What to do this week
You do not have to decide on an AI project. It is enough to do the one-page exercise before your next demo. Ask the vendor one question: where will the system get its definition of margin, and who confirms it? The answer will tell you more than the whole presentation.
If you want to run this check together on your own company, book a 15-min call about your situation.

