Niko Föhr explains how project profitability can be predicted already before projects begin.

Who are you and what do you do?

I am Niko Föhr, Senior AI and Data Scientist. I work with artificial intelligence and data and help customers solve business problems using analytics and modeling. In practice, I build predictive models and carry out analyses that enable customers to make better decisions. My work combines two areas: understanding how models work and the ability to translate results into the language of business.You have long experience in implementing AI and machine learning projects.

‍

Is there a project that has remained particularly memorable?

One particularly memorable project was one where project profitability was predicted already before the projects started.

The model does not make decisions on behalf of the customer. Instead, it provides a prediction of the margin percentage and indicates whether the project appears risky. This supports decision making: the model can flag a project as higher risk. After that, the customer can decide how to proceed. There were several options:

• the project can be declined
• contract terms can be renegotiated
• the project structure can be changed
• or the situation can be examined more carefully before making a final decision

‍

What was the project about?

The customer wanted to avoid situations where a project turns unprofitable. From a solution perspective, it was essential to identify factors from historical data that had influenced the success or failure of past projects. Based on these factors, it was possible to assess what the risk profile of new projects looks like.

‍

Why did this project stand out for you?

This project stood out because we did not settle for the first and simplest solution. Models are easy to build, but the real challenge is understanding how a point estimate can be used most effectively in decision making.

For example, a single number such as margin 5 percent does not say much on its own. But when you can say how likely it is that a project will become unprofitable, decision making becomes significantly more confident and better justified. This made the solution clearly more useful than a traditional point estimate. It shifted decision making from forecasting to risk management.

‍

What kind of solution was chosen?

We ended up with a solution where the model was not evaluated solely based on a single prediction. Instead, its effects were examined using Monte Carlo simulation. The model was trained multiple times with different data samples, resulting in a probability distribution of possible outcomes. This provided decision makers with information not only about likely results, but also about their variability and risks.

From a business perspective, this meant that decision makers could see risks immediately. If the margin prediction could fluctuate into negative territory, they could renegotiate better contract terms, increase project monitoring, or even decide not to take on the project.

Simulation also made it possible to assess the financial value of artificial intelligence. The results showed that supporting decisions with the model improved the average margin by approximately 60 percent, as poor projects were identified in advance. Finally, the customer received clear instructions on how to interpret and use the model, ensuring that the solution would support decision making consistently in the future as well.

‍

What was the reception of the solution like?

The reception was very positive. The customer was positively surprised by the predictive capability of the model, as they initially did not have high expectations regarding the quality or predictability of their data. Now they have a tool that helps them avoid unprofitable projects and identify situations where contract terms should be renegotiated.

‍

What challenges emerged during the project?

The biggest challenge in projects like this is often not the modeling itself, but communication. How do you condense a complex machine learning model and its probability distribution in a way that business management can understand immediately and use as a basis for decisions?

In this project, this was achieved well. This was also supported by the fact that the customer’s data was exceptionally high quality and easy to use.

‍

Any other interesting observations?

Customers are often interested in which factors carry the most weight in the model and how they affect the outcome. This touches on one of the key choices in machine learning: accuracy or explainability.

A simple linear model is easy to explain: when the monetary value of subcontracting increases, the risk increases proportionally. However, reality is rarely that simple. A more complex model is more flexible and produces more accurate predictions because it can account for interactions between multiple variables, but its internal logic is harder to explain in a simple way.

Significant variables can be examined on a project by project basis, making it possible to see how different factors have influenced an individual prediction. Ultimately, the question is what matters most to the customer: maximum accuracy or the ability to explain the model’s behavior clearly.

‍

Could the same solution concept be applied elsewhere?

Yes, the concept is highly generalizable. It is suitable for any problem where a numerical value needs to be predicted and where there are many variables. It is especially useful in situations where an average alone is not sufficient and decision making requires an understanding of variation and probabilities. Examples include estimating customer purchasing potential or forecasting production volume.

‍

What was the best part?

The best part was that the modeling performed beyond expectations and that the customer’s data was in an unusually good and ready to use state. This allowed us to focus directly on solving the problem instead of spending time on cleaning the data.

‍