Four questions for every AI project

Without concrete answers to four questions there is no AI transformation, only a new tool.

I now ask four questions on every AI project. They are short, and most teams find them too simple. Then they try to answer them. And the room goes quiet.

The questions are not theory. They are the result of 76 transformation projects in which the technology almost always worked. The demo ran. Employees saved time. And surprisingly little of it reached the P&L. The four questions are the filter that closes this gap.

1. Which work really goes away afterwards?

Not: which work gets faster. But: which work does nobody do any more afterwards. That is the difference between an assistant and a transformation. AI must not just make people faster. AI must make work disappear.

Most projects answer this question at the level of positions. Which position goes away? The answer is almost always: none. Then the project counts as socially acceptable and stays without effect. The right level is tasks. AI does not see jobs. It sees tasks.

The test: Break the affected role down into its individual tasks, as concretely as possible. Accounts payable processing, for example, breaks down into eight tasks. Behind each task write: human or machine. If “machine” stands behind none of them, no work goes away. Then you have bought a tool, not started a project.

Second test: name the minutes. Which task costs how much time per week today, and how much of that stands at zero after the project? Whoever cannot name minutes has no answer.

2. What happens at the exception?

Every process has a standard case and exceptions. The demo shows the standard case. Daily work consists of exceptions. The supplier without a purchase order number. The customer with the old discount. The approval given verbally. At the first exception, the answer is: “Ask Sabine. She knows how it works.”

If the answer to this question is “then a human takes over”, that is not yet an answer. It is a description of the problem. Because then the machine runs the easy part, and the human stays for the hard part. The work does not disappear. It only gets more unpleasant.

The test: Take ten real cases from last week. Not the clean ones. The ones where someone had to ask. Run each of them through the new process. Count how often the machine stops and who then takes over. If it stops on most of them, you have not rebuilt the process. You have put a tool on top of the old workflow.

Also define who decides the exception and where the decision goes. Not into Sabine's head. Into a rule that applies next time.

3. Does the system learn from it?

This is the question skipped most often. A tool that handles the standard case today will still handle only the standard case in a year if nothing flows back. The share of work that stays with humans then does not shrink. It grows, because exceptions grow with the business.

Learning here does not mean the model gets trained. Learning means every exception a human decides ends up as a rule in the process. The second time, the exception is no longer an exception. Only then does the work left with humans fall over time instead of rising.

The test: Ask the project team: if Sabine decides an exception tomorrow, where is that decision the day after? If the answer is “in the email” or “in her head”, the system does not learn. If the answer is “in the rule list the process reads”, it learns.

Second test: is there a person who maintains the rule list, and is that in their job description? Without a name there is no maintenance. Without maintenance there is no learning.

4. What do we do with the time gained?

The most uncomfortable question. Many projects save time, and then nothing happens with that time. Employees are relieved, costs stay the same, and so does revenue. Nothing shows in the P&L. The project still counts as a success because usage is high.

I do not measure whether an AI tool is used. I measure which work really goes away afterwards and what happens with the time gained. There are exactly three options. The time turns into more revenue with the same people. The time turns into lower costs. Or the time turns into better work that the customer pays for. Everything else is relief, not transformation.

The test: Before the project starts, write one line into the P&L plan: which line item changes, by how much, from when. If you cannot write that line, you do not yet know what you are doing the project for. After six months, check exactly that line. Not the number of users. Not satisfaction. The line.

In 2023 we tried AI out. In 2024 we built copilots. In 2025 the agents arrived. In 2026, AI finally has to reach the P&L. The four questions are not a framework and not a strategy. They are a filter you can apply to every running project tomorrow morning. If you have no concrete answer to one of the questions, you probably do not have an AI transformation. You have a new tool.