On Target

A quarterly publication of Society of Workforce Planning Professionals

Digging a Little Deeper to Choose
the Best Metrics for WFM

by Tiffany LaReau, Human Numbers

Most call center managers are familiar with the term “EIGHTY-THIRTY” because it represents a common service level goal. This breaks down into the literal definition of 80% of calls that we answered within a 30-second limit. There are no industry standards to benchmark service levels. 80/30 became a commonly used goal in our industry (as well as 80/20) because it feels like a number that would represent adequate productivity levels without leaving a lot of unused, inefficient overstaffing. It’s also the default setting in many ACD systems.

The problem with any service level goal is that these are not one-size-fits-all. When we staff to meet a service level, this produces a variable phone occupancy rate, and that is an output based on the workload (volume × handle time).

In the following example, each interval is perfectly staffed for the minimum requirement to meet 80% in 30 seconds. If your idea of an optimum occupancy level is 80-85%, this service level goal is a good fit, as long as your call center receives ~330 calls per hour with a 5-minute handle time.

But if you’re only getting ~20 calls/hour, your agent’s occupancy is less than 50%. That means they are spending more than half their time in the idle state, waiting on the next call, unless you’re also occupying them with in-between call work. If you can’t afford that much unused time, this service goal is a poor fit. You might look at the 7:30 interval and see a higher service level at 88%, well above our 80% goal, and ask, why not take another person away in that case? It’s true that would indeed raise phone occupancy up to 58%, but it would also create a negative net staff with a failing 68% service level. In a forecast group that size, one person equals a -20% drop in service level. Another way to say that is:  Even when they’re performing at a 68% service level, occupancy can only reach as high as 58%. That’s how the rules of Erlang Math work, and it’s better to set expectations on occupancy potential early on.

On the other hand, if you’re building this plan for a large center, an 80/30 service goal with perfect staffing means you’re running your agents at a 95% occupancy level, and that WILL lead to burnout. Using a simple Erlang calculator in Excel allows you to look inward at your own contact center’s metrics and determine the best-fitting service goal that also fits your budget. (I’m happy to send you one if you don’t already have it – hit me up on LinkedIn.)

You can also use your own reports and raw switch data to find out what your customers are telling you about how long they will tolerate for wait times. A scatter diagram with a slice using each interval’s abandonment rate and service level will instantly show you where that sweet spot lies between the customers’ happiness and a working budget for headcount.

Making good choices on what service goal to use is a great place to start, but you can generate even more improvements by turning a critical eye to other metrics.  Here are a few examples of more things to consider as you work with your reports:

Forecast Accuracy

The deviation of a forecast from the actual represents how our customers’ behavior is changing compared to the original model. When measuring this, the first thing I consider is the frequency of the forecast/reforecast process. I like to see interval data reforecasted weekly. Anything less frequent produces less accurate results, which becomes very visible compared to the forecast’s baseline.

The ideal forecasting accuracy percentage will be relative to both the length of time being reviewed and the size of the group.  Bigger chunks of time have more averages to smooth things out, so they should have stricter goals.  Monthly results should be tighter than weekly or daily numbers, and interval results should have the most flexibility of all.  Larger volume groups, especially older ones with more stable call drivers, should produce tighter results than small groups with unknown external factors.

I never reforecast for the sake of making my forecast results look better. I’d always prefer to see my original forecast anyway, so I can compare it to the actual data and keep all the red flags and alerts clearly visible because that helps with the follow-up conversation when it’s time to analyze the history with the rest of the team.

Handle Time

Handle time is the other half of the workload metric, but it’s also a metric that tells us a story about current affairs in the contact center.  There are four discretionary components to handle time:

Ring time is essential because if the call went through an IVR and now it’s ringing at someone’s desk, it’s still being counted as part of the cost, and the center is paying for that time.  If one agent answers their call on the 2nd ring, but their neighbor answers on the 6th ring, that difference in service should be captured (especially if they ring out.) This won’t apply to centers using auto-answer but is helpful to isolate problems if the abuse is intentional.  Hold time also tells its own story and may need to be isolated if agents hide out in there.

You can pull extra value out of forecasting handle time by applying seasonality. This can happen on a big scale by comparing winter to summer and on a smaller scale by comparing first shift workers to third shift workers.  Even the day of the week can impact handle time, especially with weekends vs. weekdays.

Handle time combines the callers’ behavior and the agents’ behavior. In some cases, both behaviors can be altered through education, self-service tools, or service level results.  When a customer is not spending    on the front end of the call complaining about how long it took to get to a human, the talk time may be lowered.  On the other hand, if self-service tools are put into production to handle all the “easy” calls, talk time may increase.  The ability to comprehend what handle times are saying is key to understanding the nature of the call center agent’s culture.

Handle times are also a significant factor in the next metric, Schedule Adherence.

Schedule Adherence

Schedule adherence is a trickier metric because it can be open to many interpretations. The first step is to decide how to write the calculation:

Option 1:  Compare the total logged-in time to the entire scheduled time.  If someone was expected to work 8 hours, did they only log in for 7.5?

Option 2:  Compare the times logged in to the times they were supposed to be logged in. This is better than the first option because it combines the total login time and the arrival and departure times according to their schedule, plus it considers breaks, lunches, and other off-phone activities.

Option 3: Compare the timestamp someone logged in to the time when they were scheduled to be there, but add some leniency for centers with AHT over five minutes.  It’s even better than Option 2 because, after all, do you really want an employee hanging up on a customer in the middle of a call just because it’s time for their scheduled break? (Answer:  No!)

To determine the leniency factor here, I take my AHT and multiply it by 150% (x1.5).  That is the “grace period” to allow before dinging someone on their schedule adherence rate.  For example, if AHT is 8 minutes, and someone is scheduled to go to lunch at 1 pm, but they get stuck on a call, their schedule adherence does not begin to count against them until 1:12 pm. Using this 1.5x method reduced the number of manual exceptions I had to enter because they were still stuck on a call. Getting your WFM software to recognize these custom thresholds is another story.

The other tricky part of schedule adherence is that it should not always be treated as pure shrinkage. Here are three examples:

  1. If you are overstaffed to the point where schedule adherence does not move your service goal from passing to failing, you wouldn’t want to tack on schedule adherence as additional shrinkage to bulk up more staff. That would lead to excessive overstaffing.
  2. If your agents work on phones and email, or phones and other off-phone activities, you’d only want to consider schedule adherence to the workload portion related to phone work. Even though email can have a response time goal instead of a service level, schedule adherence doesn’t exactly apply in the same way to workloads with response time goals that are less than one hour.
  3. Schedule Adherence doesn’t always hurt service. For example, suppose someone shows up 30 minutes late (and it was during a time when calls were low and service wasn’t affected) and makes up that time during lunch or after work when volumes are higher. The net effect might actually end up helping the service level that day.

There’s still a behavioral issue to address with these three situations, but none of these warrant building in extra shrinkage, meaning additional headcount requirements, to compensate for the schedule adherence violation. We only want to include schedule adherence as a shrinkage factor when it negatively impacts service goals.

Abandons

Abandoned calls can be deceiving when they are completely absorbed into Calls Offered without any prior cleaning. A caller may abandon for any number of reasons:

  1. They accidentally misdialed and realized it once they heard the automated recording.
  2. They got impatient and tried another form of contact.
  3. They were interrupted while waiting for you to answer, so they hung up.
  4. They were forced to abandon because you didn’t have enough trunks open to handle the volume.
  5. They had a technical phone problem that disconnected them.
  6. They became impatient and decided to call back later.

Callers abandoning for reasons 1 & 2 should not show up again later as an answered call. But callers who hang up for reasons 3, 4, 5, & 6 are likely to show up again as a repeated call.  And depending on the urgency of getting through, they may have multiple repeat attempts before being answered.  There is a semi-linear effect of abandons vs. ASA (Average Speed of Answer). That abandon effect demonstrates that longer ASAs result in more abandons. When you determine the ratio of that effect for your specific call group, it will help you decide the best way to clean up abandons, resulting in a better and more accurate forecast.

I don’t know of any call reports distinguishing between Reasons for Abandoning, at least not yet. With all of the AI efforts in WFM software, I would hope that’s still to come. I’ve heard people say that it’s a bad practice to forecast abandons, but I don’t see it this way. I find it to be the key to getting the best forecast results. It is reasonable to expect that a small percentage of abandons are outside our control (#1 misdials, and #3 they were interrupted by their doorbell).  It’s even an accepted practice in some centers to exclude this small percentage of abandons from being measured in their Service Level goals. And once you know the abandonment rate effect, when someone asks you to tell them what happens if they take three people away, you have another way to mathematically show the full impact of staffing decisions (and the impact of making the wrong decision).

Tiffany LaReau is a Certified Workforce Manager and owner of Human Numbers, a firm that provides contracted forecasting and scheduling services. She may be reached at Tiffany@HumanNumbers.com or 770-609-6565.