Artificial Intelligence, or simply AI, is an area you have probably heard an increasing amount of buzz and hype about over the last couple of years. And rightly so; the potential within this technological area is no less than mind-blowing.
Since we at Azets always strive to utilize the best technology available to serve our customers in the best possible way, we have – of course – also looked into different ways we could utilize AI to improve our customers’ business.
In this blog post I will tell you more about one of our first attempts to accomplish this. An obvious place for us to start our AI initiatives, was to see how it could be helpful for the users in our customer portal Azets Work.
Worth pointing out before we get down to business, is that we would never want to use AI, or any new cool technology for that matter, just for the sake of using it. It is crucial that the technologies we decide to utilize provide a good value for our customers. If our customers like what we do, Azets as a company will also prosper.
Looking for errors
In Azets Work, we have about 6300 managers, approving roughly 55,000 time reports every month across the Nordic countries. As you probably know, managers are often very busy, and to be honest, approving time reports is not the most exciting part of their work. Quite understandably, then, there are time reports with errors being approved incorrectly every now and then.
These errors can be a result of mistakes made by the employee, or actual intended fraud attempts. It can be difficult for a manager to notice if an employee have added 11 hours of overtime instead of the actual or intended 1 hour (regardless if it’s a mistake or by intention).
When you outsource your payroll to Azets, our payroll consultants will of course notice these errors most of the time. However, it is hard for a human to detect everything. An error might not be noticed until after the incorrect salary has been paid, or sometimes not at all. We can of course only quantify the cost of mistakes that are found. But we can certainly say that mistakes are easier and cheaper to fix the earlier in the process they are found.
Identifying faulty time reports
So, the idea was that we should try to utilize AI to support the managers in identifying these faulty time reports before they get approved. Then the manager can send the report back to their employees for correction before it is approved. This would potentially reduce the amount of incorrect reports being approved and save a lot of time on fixing the errors later on.
In our time and absence application, Azets Employee (available in Azets Work), we have over 50 million (!) reported events (overtime, parental leave, vacation, and so on), so we did not lack data to analyze and test on.
And if anyone is wondering – yes, we are of course compliant with GDPR regulation when it comes to storing and using personal information. We anonymize all data we no longer need to keep records of.
What we ended up with
After testing different AI algorithms and methods, analyzing the results, and thinking about this for a couple of weeks we ended up implementing an “Anomaly detection” algorithm. What it does is to look for reported events that is out of the ordinary, something that does not appear to be normal for that specific employee, and then highlight this for the approving manager.
This was a bit trickier than we first thought – it turned out that a “normal” reported event can differ a lot between our users. Time reporting is done very differently in different countries, industries and companies.
We are now rolling out this functionality for our customers. The first couple of hundred companies will have AI helping them out when approving February’s time reports. We will closely monitor our AI’s behavior in “real life” to see how much it will end up supporting our customers.
Going forward, we plan to see how we can utilize Machine learning to further improve our AI’s efficiency.
I hope that I can come back and tell you more about that later on!
Want some more?
How about a bit more in-depth description, featuring technical terminology and graphs? Get into the details on our AI thought process. I strongly recommend that you check out our detailed description on how we implemented AI in Azets Work.