Hi Stefan, your timesheet approval AI is making waves in the company. How did it come about?
It’s best to start with some background info: at Arineo, around 30,000 timesheet entries accrue every month. At the end of the month, the project managers review the timesheets posted to their projects. A tedious job that no one really enjoys. So I thought about how to speed up this process. I quickly concluded that hard-coding criteria for approval or rejection would be too costly and error-prone. After all, the criteria are not the same in every project: one customer wants the timesheets in English, and another prefers specific wording. Programming all of this rigidly into software would not lead to the desired results.
So a dynamic system was needed.
Exactly. I needed a system that could recognize patterns in past corrections and constantly update itself. AI is ideal for this. So I got to work with Ansgar and Jonas from the AI team and programmed the algorithm.
How does the algorithm work specifically?
Our AI observes the project management in its work. It remembers when a timesheet entry is approved, rejected, or corrected – and how. Depending on the project leader, up to 100 factors come into play. The AI always learns from the past 2-3 months, so its knowledge dynamically adapts over time.
And what does the workload relief for the project management now look like?
The AI pre-screens the timesheet entries and assigns a probability of the need for action. Now, the project managers can easily sort the entries from the most problematic to the least problematic. This greatly speeds up the approval process and contributes to the timesheets’ quality and transparency.
Show me the numbers! What is the recognition rate of your AI?
The recognition rate is remarkable: in a list that was pre-sorted by the AI, a majority of records requiring action are in the top 15 percent. Of the entries at the top of the list, 95 percent actually get corrected. Before using AI, the entries that needed action were hidden throughout the list because it was sorted by date. The fact that the error rate was higher back then is only human.
But there is one thing I have to mention: the algorithm delivers the best results for long-running projects because its project-specific knowledge increases over time. For new projects, it transfers knowledge from previous projects. Based on the corrections, it then learns what factors are used to judge the timesheet entries for the new project. The right amount of time stability of the underlying rules is crucial: they must be stable enough that updating from the past promises good results, and at the same time, flexible enough that static programming of rigid rules makes no sense. Then the task is suitable for a self-learning algorithm.
Wouldn’t it make sense for the algorithm to give warnings when project staff are entering timesheets if the entry seems problematic?
Yes, it’s a paradox. In fact, this would probably reduce the error rate, and, thus, the learning potential for the AI. It could become a self-fulfilling prophecy, thereby reducing its significance – and that would be counterproductive. The corrections based on human judgment are how the flexibility of the previously mentioned rules finds its way into the algorithm.
Now we are talking about a very common, but still very specific use case – timesheet approval. Could your algorithm be applied to other business cases?
Yes, this artificial intelligence can basically be used anywhere where a mass of abstract structured data is processed by humans. Structured data is usually found in ERP processes because they are designed for repetition – and, thus, we find structured data.
Another field of application that spontaneously comes to mind is the checking of travel expense reports. Or inventory. Here, the algorithm could increase efficiency by analyzing which items were particularly problematic from past inventories. With this information, people could focus on the stock items that are likely to be problematic. It would also be conceivable to extend the checking intervals in line with actual needs: while screws would be inventoried once a quarter, leaf blowers would only need to be inspected every two years.
These are very pragmatic applications of artificial intelligence. Very far from the representation of AI in public discourse or culture.
That’s right. But, in my opinion, that’s exactly the strength of this kind of AI. Here, it brings mainly advantages. It learns unnoticed in the background, does not compete with people’s jobs, and creates free space for value-adding, human tasks. With a use case like this, the barrier to entry into artificial intelligence for companies is very low because it is also so specific.
What would the procedure be if a company was interested in your algorithm?
That would make me very happy! We would first create a proof of concept and determine if the data and features are suitable for artificial intelligence. We would also interview the people to be supported by the AI. Their needs are essential. After all, AI is supposed to serve people, not the other way around. And if the proof of concept is positive, we would set up the algorithm and train it.
Stefan, thank you for the interview!