In the last edition of On Target, we discussed the power of cloud computing, which makes our workforce management algorithms more accurate, truly optimal, and much faster.
I expect that all workforce management vendors are currently investing in cloud-based algorithms to enable AI-based forecasters and true mathematical optimization. In that article, I teased the concept of a type of what-if analysis that is made real by cloud-computing-based scheduling and forecasting algorithms.
I’d like to introduce the mathematical concept of multi-objective programming.
In our contact centers we care about a lot of stuff. We would like to hit our service level or ASA goals. We would like to hit these service levels more consistently, and primarily throughout our core contact hours. We want our agents to be happy with their schedules. We want to offer our employees flexibility. We would like lower attrition. We would like to be able to flex our workforce for system issues and outages, or volume spikes. We want to hit our NPS scores and First Call Resolution goals. We want to minimize cost and maximize sales. That’s just a sample, but still a lot.
So, isn’t it strange that we only schedule to a service level target?
Not Just Service Level
This is going to change. As our models get more sophisticated and faster, we will be able to do what-ifs around our schedules and staff plans that draw out the repercussions of the different decisions we make.
Here is a fun example. A new call center reality is the ability to route calls, not to just specific work groups, but to the actual individual call center agent most likely to be successful with this specific customer. This AI-based technology is called predictive routing, and it turns much of the old school queue and skill logic on its head. We let AI determine who is the agent most likely to provide a happy outcome for our agents.
Results from these technologies are impressive. But they also change the concept of workforce management.
When should an agent work? Maybe it makes sense to schedule that agent at the time that more customers, needing the talent of that specific agent, reach out to the contact center? Think about this—how would our processes change if we could match calls to agents? I expect we would want to schedule our agents so they are available when their gifts are most useful. It’s a different way of thinking about scheduling.
Say we have a great sales agent, and one of our important goals is to maximize sales. We should make that agent available during the hours that sales calls typically come to the center.
But we also want to make sure that the agent is happy, and schedule that agent when they want to work. Hence, the conundrum. How do we handle trade-offs like these?
Multi-Objective Programming (MOP)
One of my favorite classes in college was Multi-Objective Programming (MOP). The purpose of MOP is to help decision-makers understand the repercussions of their decisions on all aspects of their business. An emphasis on employee shift preference affects service delivery. Minimizing costs affects the number of abandons and employee occupancy and satisfaction. When we make a staffing decision, we should understand the trade-offs.
MOP can help, but it has three very important prerequisites. First, in our example, whatever method used to develop agent schedules must produce them optimally. In other words, it is critical that our scheduling algorithms produce the best set of schedules, given our objectives, not just a good one (see “The Side Effects of True Optimization,” in On Target Winter 2018). Second, it needs to solve quickly to enable timely what-if analyses. Third, it should be developed on a cloud-based system with multi-tenancy, so it can solve multiple scenarios in parallel.
Let’s walk through a simple example. Let’s say that our contact center only has two main objectives. We would like to maintain our service levels, and we would also like to give as many agents their favorite shift as possible. In this example, we could schedule our agents one of two ways: we could maximize our service levels, or we could maximize the percentage of agents who get their favorite shifts.
Think about a graph that plots the service level attained on the X-axis, and percentage of agents who get their preferred shifts on the Y-axis. If we developed schedules maximizing service levels, we would land on point “B,” a high service level but a low agent preference. Similarly, if we were to give all the agents their favorite shifts, we would land on point “A,” terrible service delivery but happy employees.
Neither of these points is truly business-optimal, although point “B” is the likely outcome of old scheduling algorithms. What we would like to do is to run our scheduling algorithms with both objectives in mind, and we would like to understand the trade-off associated with them. If our scheduling algorithms had both objectives coded, we could run the scheduler with different weighting. Meaning, we could weight service 70% and preference 30%, or vice-versa. By playing with the weighting of the objectives you get points, “C,” “D,” and “E” and a nice curve starts to emerge.
The Next Frontier
Ready to impress your friends? This curve is called the “Pareto Frontier” and each point belongs to the “non-inferior set.” Each point along “A” to “B” on this curve are all equally optimal points, meaning they are derived optimally and an improvement in one goal means the degradation of the other goal. What is left is a value judgment: how do our decision-makers feel about each of these points and where on the trade-off curve do they want to be?
In ancient times, in 2018, for call center scheduling problems, these curves were very hard to derive. But today, in our modern world (ha!), it has become simple. Parallel computing means we can solve this entire curve in a few minutes, optimally, and show the trade-offs of our competing objectives.
Our ability to provide these super smart business analyses is near. And juggling business objectives is getting easier and easier.
Ric Kosiba, Ph.D. is a charter member of SWPP and vice president of Genesys’ Workforce Systems. He can be reached at Ric.Kosiba@Genesys.com or (410) 224-9883.