Use AI in Human Resources? Think again!
Most recently I noticed that a company was proudly telling their social media audience that their shift planning is supported by AI already today. And isn’t that a great idea?
Shift planning is a tedious and repetitive task that should be automated. Initially this repeated work might be perceived to be as easy as filling blanks in an Excel sheet. But in practice, there are a lot of dimensions to be taken into account: vacations, shift time limits, the proper composition of teams (seniority and qualifications), regulations for working at night and other aspects of work, adopting to customer demand and the amount of work to be done, individual needs, contractual obligations etc.
Clearly it’s a great idea to automate this. A machine isn’t easily bribed into giving away the best shifts to its buddies. An algorithm will not get confused by complicated multi-dimensional problems. And I’d guess that with a long backlog of shifts from the past an organisation probably has enough data to train a machine learning model to solve this multidimensional optimization task. If some employee complains about allegedly always getting the bad shifts, Human Resources would just shrug and blame the model. This approach is called “Automated Decision-Making” (ADM).
Let’s take a step back and reflect what AI can deliver. Artificial Intelligence is a game of stochastics, it’s not about “the perfect fit” or about being absolutely accurate. If you’d analyze a picture of a cat, the model won’t say “That’s definitively a cat!”. Instead, it will respond along the lines of “I think it could be a cat, with around 89.1% probability.”
Imagine receiving a shift plan for the next month, and you’re not on it, because the model determined it would be a better fit going on without you this time. Lucky you, taking a paid month off.
Or, the AI has planned a particular shift where none of the workers holds some required certification. Nobody notices, they can start working, no problem. Now it just so happens that in that very shift some work is executed leading to one person getting injured. Ouch. Later it is discovered the team formally wasn’t qualified to do that kind of work because the shift planning left the required qualification out of the equation this time. Blame the model.
Shift planning is a strictly rule-based problem, where some important rules must not be ignored. Throwing AI to the task may not be working that well at all.
This is why such AI systems probably are regarded as being attached to some risk. And in fact, the upcoming EU AI Act regulates such a system (the new law is currently in draft only). Human Resources is a major domain where AI-based systems are considered high or low risk, but not no-risk.
Now let’s turn back to the company already using AI for a HR task. Surely they are the front-runner technologically. Still, my recommendation would also be that at some point they take a second look at their AI-based shift planning system and determine its risks, look into biases and error conditions, and in general invest some additional work into assessing the system in order to being able to continue using it. Finally, all this needs to be documented.
There’s no need to wait until the regulation laws are in effect. The work can already be started with low risk of creating waste. You’re only wasting precious time when waiting until the final regulation is signed into law. Start today to assess your AI based systems! Some simple initial steps will be universal enough: Define the system, assess its risk, keep the data and document everything.
Further Reading: There’s a great paper about ADM in HR (PDF) by algorithwath.org.