Managing Email and Back Office (with AI)

By Ric Kosiba and Bayu Wicaksono, Genesys

The Problem of Email Modeling

We’ve discussed deferred work staffing algorithms in this column before. The gist is that planning for email and back-office work seems like it might be a simple thing. Emails are less dynamic than phone calls, it is not necessarily an immediate interaction, and it is work we can often perform during agent down times. But when modeling email or processing staffing, these “simplifying” features of email actually become problematic, because email does not adhere to the simplifications of agent scheduling assumed in all workforce management systems (interval requirements, short-ish handle times, no concurrency). But there are math models and metrics that work pretty well for deferred work (see “Using Service Level for Back Office”, On Target, Fall 2011).

I’d like to change gears, though, and discuss how improvements due to cloud computing and advancements in AI will take the management of deferred work queues. Note that I’ll discuss primarily emails, but please know that this discussion is applicable to back office work or other deferred work, such as canned documents, or transcribed voice interactions.

New Queuing Models

Our traditional work and call routing structures have been necessarily rigid. We set up queues and skills and priorities, and determine situational routing structures. Many an IT engineer finds his value in understanding how to organize a web of routing structures in different call center load scenarios. We have back-up groups and back-up groups backing up the back-up groups. Routing is prescribed.

That’s changing of late because of predictive routing systems. These AI-based systems predict which agents are most likely to produce the best customer result for each individual call and route each call appropriately. This shakes up the old queue-skill paradigm and delivers better customer (and agent) experiences.

Email and Back-office Work is Different

Today, emails are routed and handled in much the same way calls traditionally are. There are queues and skills. In the best case, deferred work can be routed via a predefined subject line or web page reason code to an agent skilled to handle the reason for the email or back-office item. Most often, however, these work items use generic first in/first out routing.

Service standards for deferred work are rough guesses (call service goals are as well), and are typically long, measured in hours or days. They are managed ad hoc, and usually not automated.

Improvements with Cloud Computing

I’ve been harping on this for a while, but once again, cloud computing is changing the way we normally do business. Cloud computing allows for more complex algorithms and allows us to store and analyze larger and larger data sets. The result is that we get to solve complex problems better and faster. We set up queues and skills because human beings could not allocate calls better or faster — it was a way to organize work in a predictable manner when computers were slow and data storage was expensive.

Rather than treat all work as the same, predictive routing can understand which agents are better at handling which calls. For example: I may have two agents trained on the very same skills, but one may be better on one skill than the other, and vice versa. Our new predictive routing technologies allow us to differentiate between the two agents, and on the margin, send more of the appropriate calls to each agent. It’s pretty neat and has real benefit.

But what about email? The same sort of thing applies, with one cool change.

Using Natural Language Processing

Natural Language Processing (NLP) is a technology newly available to everyone. Whether using Amazon Alexa or Google or a simple query bar in your favorite website, computers can listen to a voice command and translate that into a meaningful computer query. These same systems can learn from many other similar queries to elicit a response that has a high probability of answering the question well. These systems learn by noting the human reaction to the computer-generated response. In simple queries, the results of the query are measured for appropriateness. If a person clicks on one of the responses (hence it is “liked”), that response has a higher probability of being successful on the next similar query. Search results that elicit an additional, maybe more specific query from the consumer are indications that the first results were not acceptable. The systems learn, and the whole query-response-learn cycle is automatic. Systems like Google are constantly improving as more queries are made. This is a great example of “machine learning”.

Contact centers are using NLP to handle customers via chatbots, and some companies are using NLP to respond to email. But contact centers do not have the same scale that Google has, and the ability to learn and elicit a machine learning response may be limited by the lack of data.

But what can AI do, even without the volume of data that requires automatic responses? AI absolutely can be used to classify a specific email into a meaningful category. AI can read the email and determine which agent would be best to send the email to. Each email can come with suggested AI responses to help agents work faster. In effect, we get predictive routing for emails because AI can read the email before the human agent even gets to see the email. The same is true for back office items.

Queues, Priorities, and Service Standards

There is one other benefit to this new way of managing our work. We have a lot of data about how our actions are received by our customers. For example, they fill out satisfaction surveys, or they buy our products, or they don’t buy our products, or they call us back, or they try and communicate with us via another channel. Their reaction enables us to build a network of customer journey states (say, where the customer is in a sales process), our response strategies (how we communicate with them and with what urgency), and the customer’s reaction to our strategies (do they buy or do they go quiet).

One decision we make (that is not considered a communication strategy, but it is), is how we resource a channel. We can staff higher or lower and we can, by extension, determine a de facto service standard. But our data tells us a lot about the appropriateness of this staffing level. If customers become more uncommunicative the longer they wait, we need to staff up and shorten our service standards. If they search for alternative channels to communicate with us, we likely need to shorten our service standards. If they give us poor satisfaction scores, then we need to shorten our service standards. Handle time and staffing is a service strategy.

But is it at all illogical that we should tailor this strategy to why customers are contacting us? Certainly, there are reasons our customers reach out that have higher value to us or to our customers than other reasons, and we should react quicker based upon why customers are emailing us, or which stage in the processing journey a customer is in. If the sentiment expressed in the email (AI can determine this) is urgent or harried, then we know that email should jump the queue.

That’s where automatically reading the email first helps. We can customize an email strategy to the reason our customers are contacting us and the email’s mood. Important emails go to the front, less important emails can wait. Let the computer sort out the routing — there is a future where we don’t need IT to do it for us.

Ric Kosiba, Ph.D. is a charter member of SWPP and vice president of Genesys Workforce Systems. He can be reached at or (410) 224-9883.

Bayu Wicaksono is the director of Operations Research and Modeling at Genesys.