Don’t model individuals but predict group behaviour

We are all unique. That’s why it does not make sense to deploy AI tools at a personal level. And even if it did, you would never want to create an ML model of every single identifiable person in the company as that would be privacy-intrusive. What you should be focusing on is the group level: for example, predicting the availability of your workforce.

That might sound obvious, but the importance of this basic concept was recently affirmed in a situation at one of our clients. In this specific case for every incoming task, people had to decide to accept the task or not. Of course, that depends on the task: content, circumstances, estimated duration of the task, timing,… But it also depends on the people and their personal situation and preferences: one person doesn’t mind accepting long-duration tasks requiring overtime or even stay-over or might even prefer this kind of task due to the interesting financial compensation, whilst somebody else might prefer not to accept this kind of task as he or she has another family situation resulting in other preferences. But how do you manage to have a view on the availability of your staff, on the capacity to accept incoming tasks and make sure you can guarantee your service to your users?

Lots of insights were brought by analyzing historical data. By feature-engineering the executed jobs and the decisions of the staff members, and applying unsupervised learning, we were able to “group” these members (I hate this word, as I don’t have the intention to label people). Using the groups, it was easy to predict the chances of acceptance of a certain parameterized job. While the individual choices are not accurate, at the group level they are. And that’s all you need to know to accurately predict the capacity of your workforce and the SLA you can offer your clients.

Towards an AI-based Chief Happiness Officer?

In another case we recently identified, a company is dealing with a rather high rate of churn of newly hired staff. The environment is quite challenging, with sometimes some travel time and an irregular workload including peak moments with lots of overtime and jobs during nights and weekends. For every job, team managers have to find the right balance between experienced staff who get a certain kind of job done fast and efficient, and junior profiles who need some more time but for whom training on the job leads to a more durable workforce longtime. Certainly, in these kinds of situations, people have their own personal situation and preferences on the times they want to do which kind of jobs. Not even mentioning the team and colleagues they feel themselves the most comfortable with. Especially in a context where it is hard to find people suited and willing for these kinds of jobs, you want to avoid the high cost of every person leaving the company after seeing a mismatch between the expectations and the actual situation, being either too many times scheduled for certain jobs or just too few. And of course, you want to avoid having some teams where people quit the company more often than in other teams.

In small-sized companies, this kind of situation can be handled with a good HR manager knowing all the staff members, asking the right person at the right time how they are doing in the company and their team, especially to those people who don’t easily raise their hand themselves but bottle up their frustrations until it’s too much. Then it’s simply too late.

That’s where a good HR team can get support from data-driven insights. Think of historical data of jobs assigned to people, being both people still active and people who left the company, combined with data derived from interviews with newly hired staff asking for their preferences and expectations. Learning methods reduce the complexity and enhance generalization in order to create an AI-based HR support tool. A system that gives an indication of the chances a person is close to leaving the company. Name it a staff retention predictor, or even a staff happiness indicator if you want.

It’s a system that gives warns you when trying to schedule a person for another weekend job when that person has a family and has been doing weekend jobs for the past five weeks already, whilst another colleague would be happy to do an extra weekend job as he is in another personal situation preferring the extra pay. It’s a system that optimizes team composition and scheduling for the long run. It’s a system that gives the HR manager or team leader an indication of which colleagues have been in non-preferred situations too much lately and could need a conversation, asking them whether they’re still fine. Not with a complex rule-based system, but based on experience, on historical data, on ever-growing intelligence, incorporating kind of a gut feeling. A fairer system for everyone. Especially in large teams or companies, this kind of tool makes it possible to scale by supporting an HR team to prioritize and not losing connection with certain people. It’s a system that combines the knowledge and skills of data science and machine learning experts, the ERP system integrator and internal and/or external HR specialists.

As machine learning engineers, these are applications we didn’t have in mind when founding ML2Grow. But they’ve been challenging us, delivering insights to us and driving us to go further along this journey.

Do you recognize these situations? We would be glad to read your comments on this.