Artificial Intelligence and Workforce Planning
By Ric Kosiba, Ph.D. Vice President, Genesys
Shaping a personality
Funny and true: When I was a boy, my mother always lovingly called me “Cynic.” Like, “Good night, Cynic” or “Time for dinner, Cynic.” I was always mystified by the name, but it became part of my view of myself: I assumed from an early age that I must be weirdly cynical. Heck, my mom thought so. It probably became the basis for my being more than a little distrustful – especially of boastful people or new fads. Vicious cycle and all that.
When I was around 40 years old, my mom was visiting, and called me “cynic” again. And it occurred to me to ask her why. Was it because I was distrusting as a child? Skeptical of new people? Mean? She howled, almost fell off of her chair laughing. No! It wasn’t “cynic,” it was “synek,” which was Polish for “baby boy!” She always was calling me her “baby boy!” But the damage was done, and here I sit, cynical.
You can’t take three steps without stepping in some AI
Artificial Intelligence (AI) is everywhere nowadays. We hear it in TV ads and there are news stories everywhere (yesterday in the news an AI program gave birth to a baby AI program). IBM is trying to sell us a boatload of AI, and news outlets tell us that AI will murder us all (you know, all the killer robots). We’ve been told we are interacting with AI all the time, but we just don’t know it. It is the new craze, the new buzzword.
But it is also very cool and powerful. The cynical part of me wants to let you all in on a little secret: much of new development is now being called “AI” even though it sort of isn’t. And many of the things we have been doing for years is AI, it’s just that we have never called it that. So I thought I would try and clear it up a bit in this article.
AI is very good at a few things
Here is a fun sentence: AI is very good at solving specific problems, generally. What I mean is that certain tasks can be given to it without a lot of attention given to the particulars. For example, AI is great at pattern matching. Data scientists and operations researchers can give a well-designed AI program tons of data and ask the program to find patterns, and the program will likely find real patterns if they exist. Humans are also great at pattern matching, but a great AI program can find patterns among datasets too big for humans to get their heads around.
AI programs are great at evaluating many, many options. Because computers are now so quick, you can present it a seemingly infinite number of possible solutions to a problem, and AI will find the best solution, or combination of solutions. You can train it to search only those solutions that its experience has taught it would be likely to be a good one (called machine learning).
AI is good at classification. This is important in customer service or sales centers where it is very useful to know which unknown customers are most like known profitable customers. AI can look at tons of data to classify people into likely behaviors, such as likely to buy our products or most likely to need special customer retention efforts.
AI can also mimic humans and their thought processes, allowing computers to follow the decision-making steps of known human experts. Early AI programs were used to help diagnose car problems and human health. These systems were called expert systems (I built a prototype expert system for evaluating satellite state-of-health when I was first out of college. Sounds fancy, but it really wasn’t).
So AI programs can work tirelessly chugging through gazillions of emails searching for contextual data (a pattern) to find those that might be a security concern. Or it can look at a picture super quick and understand the differences between a balloon and a human head. And because it can do that quickly, and it can follow rules of human experts, it can be trained to drive a car.
AI is not good at everything
One of the first things that they teach you in AI school is that domain knowledge will most often trump AI for a significant number of problems. In our world, we have operations that have known correlations between the decision inputs and the center performance outputs. If they are known, we don’t need AI to find them for us.
Within the workforce management and planning function, we have several examples of analytic processes that are decidedly not best solved using an AI system. The most obvious is the calculation of staffing requirements. Discrete-event simulation modeling has been proved to be accurate, because it is “easy” to understand the relationships between inputs (staffing, volumes, handle times, efficiency, customer patience) and performance metrics (service levels, ASA, abandons, costs, revenues). AI would not solve this problem well or fast, and certainly not better.
Schedule optimization or hiring/staffing optimization is of a very different class of problem that also does not lend itself to AI. When we optimize schedules or hiring plans, we are solving something called a combinatorics problem. Combinatoric problems are well-defined, the set of solutions are just huge. Specialized algorithms can be used to rapidly search and find the combination of schedules that most improve our objective (usually reduce costs). Algorithms like integer programming are great at solving these sorts of problems, but it is not AI.
Where will AI help us?
Forecasting has been researched for years. Many a Ph.D student has made her bones developing a new statistical method for forecasting. The art of data science has therefore been to find that one method that most reduces forecast error among dozens of statistical methods and thousands of combinations of method parameters. It is an art, because no human being could work through that many combinations to build a best forecast.
AI changes this for us. Standard computing, machine learning, and pattern recognition opens the door to something I have written about a few times over the years. It would be exciting to build a system that can evaluate not five or ten, but thousands of possible forecasting models. And then have the computer choose not only that method which is best, but which subset of the thousands can be best combined in an ensemble of models, to reduce forecast error. This is fantastic stuff, and it is here.
But wait, there’s more. AI can also be used to understand the patterns associated with known, but rare events, like holidays and marketing calendars. Or the relationships between weather, website hits, and our sales call volumes. Other exciting possible uses includes using AI to derive the relationships between standard contact center metrics, like ASA, occupancy, attrition, and abandons, and qualitative metrics like net promotor score or employee satisfaction. These relationships have been hard to find, but AI could take a crack at it.
Validation is still key
A common nag from this column is that any method can be used, so long as it has been proved to be correct, and that our industry has historically used models that are known to be incorrect (remember the Erlang C calculation?).
No matter what method we use to forecast, to build schedules, to build staff plans, or to develop staffing requirements, we should develop the discipline to always prove to ourselves and our senior management that these methods are accurate. This is true if we are forecasting using an old regression technique or the latest, greatest AI model. Validation graphs are a necessary minimum anytime we use any math model.
AI has tremendous potential for our industry, but, cynically, it is still just another math method. For us humans, we still need to use I (intelligence) to prove that these technologies are accurate.
Ric Kosiba, Ph.D. is a charter member of SWPP and vice president of Genesys’ Decisions Group. He can be reached at Ric.Kosiba@Genesys.com or (410) 224-9883.