How Erlang Formulas Killed Skill-Based Routing (and what's bringing it back)


Posted by Antony Brydon

One of the most exciting technologies I've worked with is skill-based routing.

Skill-based routing is the ability to connect a consumer with the best possible agent to handle their problem. In 1998, I ran a project for a call center startup that had developed advanced skill-based routing and wanted to gauge the demand from contact centers. I was excited about it from a technology perspective (it was really cool) but also from a consumer perspective - I'd felt the frustration first hand when the person on the other end of the phone seemed to recite answers rather than use their expertise to solve my particular problem. This seemed like a great solution.

Over two months, I reached out to over 150 contact center leaders and interviewed close to 50. I was surprised that there was relatively little enthusiasm among managers for skill-based routing, and what I learned opened my eyes to Erlang formulas and the complex math used in staffing.

Erlang formulas help contact centers figure out how many people to staff in the contact center on a given shift. They start with some basic data (the average number of calls during the shift, the average length of these calls, the number of agents working) the use statistical models to forecast what will happen if when there is natural and inevitable randomness in the actual calls and call lengths. See the chart below:

Erlang 1

This chart shows 300 calls that average 5 minutes each coming in over a 15 minute interval. With 120 agents available, each agent is busy 83% of the time and only 1.22% of customers have to wait on hold for more than the 15 second target wait time.

Now to understand why these contact center managers were so unenthusiastic about skill based routing, look at what happens when we break that same group into three groups of agents (you can imagine this is a cell phone company, with one group of agents for iPhone, one for Android, and one for everything else).

Erlang 2

This chart shows that with only a small variation in the distribution of calls among these three groups, performance plummets, with 30.23% of customers waiting on hold and the agents in the third skill group idle for 33.33% of the time. These contact center managers would have to hire and staff an additional 30 agents to bring performance in line with what they get from single groups with a universal skill set.

The math for the managers was clear. First, they told me the skill groups would require the higher staffing described above (+20%), higher compensation (+20%) and more training (+10%). In the simple example, moving to skill based routing would immediately increase contact center expenses by +44%. The call center managers agreed that this would likely be offset by shorter call durations and higher first call resolution that comes with more skilled agents, but skeptical that this could offset an immediate 44% increase.

In my interviews, the biggest challenge for skill-based routing that the managers cited was the big swings in performance that can come from even small variations in contact flow. While a larger and scripted group of agents can absorb this variation, the impact on distributed group of skilled agents is more significant. In the example above, event a slight variation that distributes 115 calls to the iPhone group instead of the forecast 100 results in 65.66% of customers waiting longer than the target time.

I was disappointed by the results. The skill-based routing technology was powerful and promised customers a superior experience, but it was destined to sit on the shelf because the fixed business models and economics of the contact centers couldn't afford the consequential costs.

15 years later, we started Directly, and I started thinking about fluid models that allow more skilled experts into the equation and improve efficiency and economics at the same time. The models revolve around on-demand workers, latent expertise and skill-based routing. The fact that they're all coming together in 2015 means we can finally deliver on the promise of matching customers with the best possible agents without increasing costs. In my next post, I'll delve into how.

You might also like: The On-Demand Enterprise