The five questions to ask before you start with AI

Before you start with AI, ask five questions at the level of the organisation: which problem demonstrably costs us money or time today, who owns the outcome and who will work with it every day, which data do we already have for it and does it mean the same thing everywhere, what is the smallest reversible step and how will we know after a few weeks whether it works, and what has to stay in house once the external party has left. All five are about the company around an application; choosing a tool comes afterwards. If you can answer them in a few sentences per question, that choice has become a small decision.

This piece is written for the director or owner of a company with 50 to 350 people who is about to free up time and money for AI. You can forward it to your management team and answer the five questions together in a one-hour meeting. At the bottom there is a one-page overview you can share as it is.

We have worked out a few related subjects elsewhere, and we will not repeat them here. Whether your organisation is ready, and which of the four kinds of ready is the bottleneck for you, is in How do you know whether your organisation is ready for AI?. Examples of good first applications per sector are in What can AI actually do for my company?. This piece sits before both: it is about the decisions you take before there is a project, a vendor or a budget.

Why these questions are about the organisation

AI enters your company even if nobody takes a decision about it. Statistics Netherlands (CBS) reports that the share of companies with 50 to 250 employees using AI rose from 20 percent in 2023 to 45 percent in 2025. Among companies with 250 or more people it is 66 percent. These are provisional figures. Companies that use AI deploy it most often for marketing or sales (35 percent) and for business administration or management tasks (32 percent), and least for logistics (4 percent).

Much of that use starts with a feature in software you already had, or with an employee who tries out a tool on their own. That is fine in itself. It does mean that in many companies the question of whether to start with AI has already been answered. The question that remains is whether you give it direction as an organisation: which work, who is responsible, with which data, how you measure and what you keep in your own hands.

There is also a less cheerful reason. A widely discussed 2025 report from MIT Project NANDA stated that about 95 percent of the organisations studied saw no measurable effect of generative AI on their profit and loss. Read that figure with care. The report calls itself preliminary findings, is not peer-reviewed and relies partly on interviews and self-reported outcomes. The authors themselves write that their sample does not represent all sectors and regions and that organisations define success differently. What is useful is mainly the explanation the report gives: the outcome seemed to depend less on the quality of the model and more on the approach. That is the territory of the five questions below.

How to use the five questions

Three agreements make the difference between a good conversation and one you have again next month.

Answer them in writing. A few sentences per question, with names and numbers where possible. An answer you cannot write down is an answer you do not have yet.

Answer them together. The director, the person who knows the process and someone who knows the systems. Each question has a different owner, and whoever answers alone fills the gaps with assumptions.

Leave a weak answer standing. An honest weak answer tells you where you have work to do first. That is more useful than a fine answer that is not true.

For each question below you will find why it matters, what a strong and a weak answer look like, and a hypothetical example. The example is always about the same invented company: a technical wholesaler with 180 employees. That company does not exist. It only helps to make the questions concrete.

Question 1: Which problem demonstrably costs us money or time today?

If you start with the technology, you then go looking for a problem that fits it. If you start with the problem, you can decide later whether AI is the right means at all. Sometimes the answer is a better agreement, a fixed form or a piece of ordinary automation. Then you have saved money before spending any.

The word demonstrably does the work in this question. You do not need a twenty-page business case. You do need to be able to say how often the work occurs, how much time it takes and what happens when it goes wrong. An estimate from the people who do the work is good enough for that, as long as you write down where the estimate comes from. Without that baseline you cannot establish at question four whether something has worked.

Hypothetical example. At the wholesaler, orders come in as PDFs or as e-mails. Four people in the sales office retype the order lines into the ERP. By their own estimate this takes about sixty hours a week between them, and about one in twenty orders contains an error that later costs money in returns. That is a strong answer: named work, an estimate with an origin and a consequence you can count. The answer that was first on the table was "we want to do something with AI in customer service". There is nothing to measure in that, and so nothing to decide on either.

A colleague writes on sticky notes on a wall. Mapping the work first shows which problem actually costs time.

Question 2: Who owns the outcome, and who will work with it every day?

An application without an owner belongs to nobody after the first failure. The owner is the person who decides what the application may do, who judges whether the outcome is good enough and who is given time for that. Usually this is someone from the line, such as the head of the department where the work sits. IT can think along and help build, but knows the work less well than the people who do it every day.

The second half of the question is often forgotten. The people who will work with it every day decide whether it gets used. If they are at the table from the start, they tell you which exceptions exist and where it will chafe. If they only hear about it at delivery, there is a good chance they will work around it. They know cases nobody else knows, and those cases are then missing from the design.

Hypothetical example. At the wholesaler, the head of the sales office becomes the owner. She is given four hours a week while the application is being built, and she may decide herself which customers take part first. Two of the four people who now retype orders take part from the first week. One of them knows the customers who always send deviating article numbers. That knowledge is not written down anywhere, and without her the first version would get stuck on exactly those customers.

Question 3: Which data do we already have for it, and does it mean the same thing everywhere?

Most companies of this size have more data than they think, spread across an ERP, a CRM, mailboxes and spreadsheets. The question is whether you can get to it, and whether the words in those systems mean the same thing everywhere.

That second part is the hardest. An order can mean something different to sales than to finance, and a customer in the CRM is not always the same customer as in the ERP. People carry that translation in their heads. An AI application does not have that head and gives a confident answer based on two definitions that run into each other. A semantic layer records that meaning once, on top of the systems you already have, so you do not have to replace those systems first. For this question you do not need to build such a layer yet. You only need to know whether the problem exists in your company.

A quick test: ask three people from three departments what an order is in your company, and when it counts. If three different answers come back, you know where the first work is.

Hypothetical example. At the wholesaler, the article data sits in the ERP and the customer agreements partly in the CRM and partly in the heads of the sales office. The test shows that sales counts an order when it is received and the finance team only when it is invoiced. For retyping order lines that is not a blocker, because the application only needs to recognise the article number and the quantity. For a later application that forecasts revenue per customer, it would become a problem. That is useful to know now, even if you do not solve it yet.

Question 4: What is the smallest reversible step, and how will we know after a few weeks whether it works?

A large first project is a large risk. It takes a long time before you know whether it works, and if it disappoints, a lot of money and goodwill is gone. A small step you can undo lets you learn quickly at low cost.

Reversible means: if it does not work after a few weeks, the work goes back to how it was, without damage. The application prepares something, for example, and a person approves it. Or it runs first on some of the customers or some of the orders. The old work continues until the new way has proven itself.

The second half of the question belongs with it. Agree beforehand what you will measure after a few weeks and which result is good enough to continue. Also agree at which outcome you stop. That feels uncomfortable, because nobody likes to start with a plan for stopping. If you do not agree it beforehand, you will argue afterwards about whether it succeeded, and then the person with the most enthusiasm usually wins.

Hypothetical example. The wholesaler chooses as its first step: the application reads the PDF orders of ten regular customers and prepares the order lines in the ERP, after which an employee checks and confirms them. After six weeks they look at two numbers: the time per order and the number of errors found during the check. Beforehand they agreed to continue if the time per order falls by at least a third without more errors, and to stop if it does not. If it goes wrong, the sales office goes back to retyping those ten customers themselves. Nothing has been thrown away except six weeks of a small team's time, and you know more than before.

Two people go through figures on paper next to a laptop. Agree beforehand which numbers decide whether you continue.

Question 5: What has to stay in house once the external party has left?

Most companies of this size bring in outside help for their first application. That is sensible. It becomes a problem if at the end you have a working system and nobody inside who can change it. Then you stay dependent, or you let the system quietly decay as soon as something changes in your ERP, your customers or your way of working.

So ask this question at the start, before you choose a party. What does your team need to be able to do itself in six months? Who learns that, and how? Where does the knowledge sit: in documentation you own, or in a consultant's head? And who owns the instructions, the integrations and the data when the contract ends? How that works out in practice and which agreements to make beforehand is in What happens when the external party leaves. How to stop the knowledge leaving with one person is in How to anchor AI knowledge.

Hypothetical example. The wholesaler agrees with the external party that the two sales office employees and one IT colleague can change the application themselves after three months: add a new customer and teach it to recognise a deviating article number. That is the test at handover. The instructions and the integration with the ERP sit in an environment owned by the wholesaler. If that first change succeeds without a phone call to the outside, the knowledge is in house.

The five questions on one page

You can share this overview with your management team or take it to a meeting. The column with red flags is the most important part. If you see one, you know where to go first.

What to do when an answer stays weak

Hardly any company answers all five questions strongly at the first attempt. That is normal, and it is what the questions are for. Below are three rules of thumb. They are our own way of working, not a measured standard.

Weak on question 1. Do not start yet. First spend a week mapping where the time goes in one or two teams. Without a named problem, every application is a gamble.

Weak on question 2 or 5. This calls for a decision by the board. Appoint an owner and free up time before you look for a party. If you cannot free anyone up, that is useful information too: it is not the moment yet.

Weak on question 3 or 4. You usually solve this during the work. Choose a first step that asks less of the data, or make it smaller until you can measure it.

If these questions mainly remind you of what is already going wrong in projects that are running, read Five signals that your AI ambition is stalling on your organisation. That piece describes what the same gaps look like once you have already started.

From five answers to a next step

Once you have gone through the five questions with your management team, you have a first picture. If you want to hold that picture against a fixed yardstick, the AI-Foto is a free self-assessment: ten minutes, twelve questions, and you see where your organisation stands.

If you then want a well-founded decision, there is the AI Readiness Scan. It is a two-week diagnosis that maps where you stand, ranks the opportunities, calculates the return and delivers a roadmap the board can decide on. The scan returns to the same kind of organisational questions as this piece: who decides, who owns the process, and whether the same term means the same thing in two departments.

Whichever route you take, the five answers on paper are the most useful thing you can bring to any conversation with an external party, including a conversation with us.

About this page

This piece was written on 5 October 2026 by the Radical AI team. The figures on AI use by company size and by purpose come from CBS and are provisional figures for 2025. The 95 percent figure comes from the report The GenAI Divide: State of AI in Business 2025 by MIT Project NANDA. That report is not peer-reviewed, calls itself preliminary findings and relies partly on interviews and self-reported outcomes. The five questions, the tables and the rules of thumb are our own way of working, not a measured standard. The wholesaler with 180 employees is invented and no client data was used. The descriptions of the AI-Foto, the semantic layer and the AI Readiness Scan come from our own pages.