How to Manage a Virtual Agent Task Force
Ever since Alan Turing published the Turing Test as a measure of machine intelligence, the field of artificial intelligence has made significant advancements. The advancements in natural language processing, speech recognition and speech synthesis systems have led organizations to offer services that are not only able to detect user intent accurately, but also respond to the user in a conversational manner.
In the last few years, services such as speech recognition with intent detection and speech synthesis are being offered from the cloud by vendors such as Google, IBM and Microsoft. These developments have resulted in significant reduction in the cost of these services and made them readily available to a large number of users. Furthermore, cloud offerings allow data sets to be shared across the entire user base.
Cloud vendors can now offer NLP and speech recognition services with a much higher accuracy compared to premise-based systems, which has led to a rise in the number of Intelligent Virtual Agents being deployed. As virtual agents become more intelligent, they are used in contact centers for a wider range of activities. For example, in the banking domain, IVAs are being used to reset passwords and check account balances. In the hospitality domain, IVAs can assist with ordering merchandise, processing credit card payments in a fully compliant manner and answering general inquiries. And, according to a recent Gartner report, virtual agents will drive an estimated $1.2 trillion of business value by 2030.
As the use of IVA grows, businesses will need to manage and schedule their virtual agents just as they manage real agents.
How many IVAs do you need?
It is important to estimate the number of IVAs required to optimize contact center performance. For example, on a given day, a contact center may receive 1,000 customer calls, run an outbound campaign targeting 1,000 users and receive 500 text messages from customers. Additionally, back-office tasks may be required, such as generating the daily usage reports and sending emails that need to be completed at the end of the day. Based on this workload, the contact center needs to deploy enough IVAs so that all incoming calls and text messages are answered, the outbound campaign is completed within the defined dialing hours and the back-office tasks are completed on time.
Calculating this number is critical, as you need to account for the total number of expected calls on that day, the expected duration for each call and the arrival time in order to calculate the concurrency rate. The same must be done for text-based conversations. Finally, you need to factor in the virtual agent workload to process the back-office tasks. Not allocating enough IVAs would mean contact center performance will be affected during peak activity. Conversely, allocating too many IVAs would mean suboptimal contact center performance as the costs are not controlled effectively. Having the ability to reliably estimate the total number of virtual agents necessary for a contact center has the potential to significantly improve the overall contact center performance.
Effective IVA forecasting
Inference is developing a forecasting tool that can predict the number of IVAs required for any contact center for up to 30 days. The IVA Forecaster estimates how many IVAs would be required to handle all inbound and outbound calls without dropping any of them. The forecasting algorithm is based on a supervised machine learning model that uses a combination of actual regression along with weighted averages. Using available data on the number of calls answered, rejected, and dialed in the past, the model forecasts the number of virtual agents required to handle all the calls on a given day. Depending on the actual virtual agents assigned to the contact center on that day, you are also able to estimate how many calls would be rejected on that day during peak periods. The model learns as more training data is added, so the estimation accuracy gets better over time.
Figure 1 and Figure 2 show the forecasting results for two of our customers. You can see the IVAs that were estimated by the forecaster and the actual IVAs that would have been required to handle the peak load for an indicative 30-day period.
To demonstrate how the forecaster can be used to optimize IVAs, we have plotted the number of customer calls that were rejected each month because the virtual agents were busy during peak periods. Figure 3 shows the actual number of rejected calls per month for customer B for a 12-month period. Based on our forecaster’s prediction of IVAs required for the same 12-month period, an increase of two IVAs would have resulted in all the calls being answered without any rejections due to capacity limits. This demonstrates that if forecasting data is available, a small change to your IVA allocation can dramatically improve the contact center statistics in terms of call handling and call completion.
As the use of IVAs in contact centers increases, tools to accurately forecast the number of IVAs required become critical. Omni-channel contact centers that rely heavily on the use of virtual agents for a wide range of tasks would benefit from these tools. Such tools would assist contact centers operators in managing their IVA task force to optimize performance and increase contact center productivity.
To learn more about how our customers are using IVAs to improve call handling, check out our case study with Pizza Hut Australia.