Voice and Language Technology in Call Centres
An AI adoption case study
Many people are looking to answer the question of where and why AI/ML technology is used. One way to answer this question is by examining the barriers that lie in the way of using ML technology. Another perspective is to look for places where ML is being successfully adopted and explore the driving factors.
In 2018, Google gave a public demo of Duplex, a conversational AI system that’s able to make phone calls to do simple tasks like book appointments, with a human-like level of naturalness. The demo was met with mixed reactions — some saying that AI systems shouldn’t pretend to be human, while others were impressed at the capability on show. This demo neatly showed the capability of cutting edge voice and language technology, and hinted at the fact that call centres and customer service are a place where ML technology has relatively high adoption.
According to one recent study, 78% of call centre companies have intentions to use AI in the near future. In another study, it was reported that 46% of call centre interactions are already automated. McKinsey’s State of AI in 2021 report cites call centre automation as one of the top three use-cases for AI adoption. While these figures come from the US and are likely lower in the UK and other parts of the world, the direction is clear. So why are call centres a place that have adopted AI so readily?
Business case
Customer service is a big cost for many companies. And where companies spend significant money, they’re usually looking for efficiency gains to cut costs. In a call centre there’s a direct link between the amount of time that an agent spends on the phone with a customer and the cost to the company. Reducing call times or automating simpler calls has a direct impact on the company’s bottom line.
There’s also a direct link between call centre waiting times and customer satisfaction. By assisting agents or by automating some portion of calls, waiting times can be reduced and customers are happy, making them more likely to use the company again in the future.
For these two reasons, there’s a very strong business case for introducing automation in call centres.
Data
Call centres typically handle many calls and so have lots of data available to them. This data is very clearly lined up with the scenario in which any ML system will be used, making it ideal data for model training. Not every company looking to deploy ML has such easy access to such relevant data.
Further, the type of data that makes the bulk of call centre interactions is relatively simple — as far as speech and language can be — when compared to other candidate applications of ML. The call centre scenario of phone calls between strangers means callers and agents are likely to speak clearly and formally. The topics that people call about are limited in scope and mostly known in advance. These two factors make the data relatively easy to label and to model, at least for common call topics. As with all scenarios involving language though, there’s a long tail of more complex and difficult to handle requests which are far harder to automate.
Application
There are three modes in which ML is typically be implemented in a call centre:
- ML to analyse and assist operations – e.g. to analyse the kinds of call that are coming in and identify training opportunities or problems with products. Related to this is monitoring for quality control and regulatory compliance.
- Partial automation of call handling, where a conversational AI system answers simple questions and hands more complex queries to an agent.
- Full automated handling of all incoming calls.
The third category is viewed as the ‘holy grail’ of ML, but it’s difficult to achieve due to the large number of edge cases that occur when dealing with language. Hence most call centre providers have a mix of (1) and (2) which still provides significant value to the business. Furthermore, there’s a path to continuously expand the automated offering over time, handling more and more complex calls, incrementally adding value as the capability of the ML models grows.
Technology
Call centres use conversational AI or voice and language technology – automatic speech recognition (ASR), natural language processing (NLP) and text-to-speech (TTS).
In particular, ASR and TTS are relatively mature ML technology, but all the main ML components of call centres are available through APIs from large established technology companies. Many call centre businesses don’t try and build their own ML technology in-house, but build on top of these cloud services. Even further, there are entire cloud call centre offerings such as Amazon Connect, Google Contact Center, and Microsoft Dynamics 365.
Operating Environment
The history of ML in call centres goes back several decades. The first commercial interactive voice response (IVR) system was built in the 70s. The technology was limited then, and systems could only handle small vocabularies. Over time, particularly through the 90s and 2000s, the use of IVR systems grew as technology matured. This long history with ML means call centre companies have the institutional knowledge about working with this technology.
Additionally, the call centre market is large enough and stable enough that businesses are prepared to invest in ML and automation even if it may take time to come to fruition, because they are convinced of the utility.
Conclusion
Call centres have been using ML for many decades, and their capability for automation has grown alongside the technology. From a clear business case through access to data and a supportive environment within call centre companies, there are several factors which come together to make call centres an ideal place to adopt ML technology. By looking at the factors driving AI adoption in call centres, we can perhaps identify where other industries do not create the ideal environment to nurture the use of ML technology.
