EdgeTier
Nominated Award: Best Application of AI in a Startup
Website of Company: www.edgetier.com
EdgeTier is an Irish software company that is revolutionising customer service through the use of AI. EdgeTier’s innovative products boost the efficiency and performance of customer service teams by assisting agents and administrators with a combination of machine learning, automation, and intelligent software design.
From a societal perspective, poor customer service is something that impacts virtually everybody. Media articles and social media are littered with stories of poor customer service and frustrated consumers. Harvard Business Review reports that companies who score top in customer service scores grow at 2.5 times the rate of their competitors, so clearly delivering great customer service is important for more than just staying off Twitter. However, running a large-scale customer service team is extremely difficult, and many companies are struggling to deliver the high-quality and ultra-responsive customer service that is now demanded.
Through 2020 and 2021, EdgeTier interviewed over 35 customer service leaders. This market research unearthed key challenges common across modern contact centres and EdgeTier found a number of common themes:
Rising Expectations on Delivery: Contact centre leaders are tasked with becoming a feedback loop for customer issues and frustrations that lead to negative customer experience. However, most contact centres do not have adequate tooling or skills to distill real-time or historic contact centre data into useful insights in this manner.
Customer Service Agents Cannot Spot Trends: There is an expectation on customer service agents to collectively spot and report major issues to management so that they can be investigated; e.g. noticing that people are having login issues more than normal. Typically, major issues are spread across the entire team of perhaps hundreds of agents, making it impossible for individual agents to notice trends. Remote working adds to the difficulty of trend spotting since the ‘water-cooler’ discussions between agents don’t happen anymore.
Spotting Issues is Manual and Slow: Most contact centres have extremely manual processes for reporting on emerging issues, which involves manually searching through thousands of customer queries to find examples of issues. This process is slow, time-consuming and inaccurate. Sudden spikes in contact volumes cannot be understood and lead to panic throughout the customer service team.
As one contact centre leader put it “Say I got a spike in contacts last Friday afternoon, how would I know what the spike was about? Well, I’d speak to the agents and some would say they saw some queries about payments, others would say the customer portal wasn’t showing, others would talk about search results. But it’s all very anecdotal and I can’t say with any certainty what is driving it”.
To solve these issues, EdgeTier spent 18 months researching and building ‘WatchTower’ with the aim of both informing customer service managers in real-time exactly what issues their customers are having, while also informing them of what their customers are feeling. WatchTower leverages AI to analyse all of the activities in a contact centre, and allows customer service teams to move away from slow and inaccurate manual analysis, to proactive, quantified, real-time alerting.
Reason for Nomination:
NOTE: A PDF IS SUBMITTED TO ACCOMPANY THIS SUBMISSION WHICH INCLUDES A COPY OF THE TEXT HERE AS WELL AS SUPPORTING DIAGRAMS.
WatchTower ingests and analyses the text contents of emails and chats between customers and agents in real time, as well as metrics such as contact volumes, queue times etc. WatchTower’s AI then detects issues or unusual trends and informs contact center managers, often before they even know there is an issue.
WatchTower achieves four main goals:
1. Unsupervised Real-Time Alerting: WatchTower detects anomalous activity and informs contact centre managers in real-time. An anomalous activity may be an increase in people talking about some new topic or a change in contact patterns etc.
2. Understanding issues: WatchTower allows contact centre managers to easily understand an issue by locating the most representative customer messages that clearly explain the underlying reason behind an issue.
3. Quantification of issues: WatchTower automatically clusters queries together in order to accurately quantify the number of queries related to a specific issue.
4. Understanding customer emotions: Understanding what customers are thinking is of huge interest to customer service managers, many of whom are benchmarked on customer satisfaction scores. Yet, general NLP-based sentiment detection models are not designed for customer service data. EdgeTier has developed language-agnostic emotion detection models for detecting key customer-service related emotions such as ‘frustration’, ‘gratitude’ and ‘praise’.
With WatchTower, contact centres reduce manual effort, increase customer understanding, while also increasing the overall customer experience through quicker resolution of customer-impacting issues.
How does it work?
The WatchTower system uses a variety of machine learning techniques to automatically detect and summarise customer service issues. Over 18 months and numerous iterations, EdgeTier developed a new form of ‘anomaly detection’ that is specific to customer service and is capable of detecting deviations from expected behaviour. All technology was completely developed in-house by EdgeTier data scientists.
Fundamentally, the problem is posed as an unsupervised machine learning problem; the system does not, and can not, know what issues will arise in advance. Over a moving window of ingested data, WatchTower maintains a time-series model of the “expected behaviour”, including conversation topics, as well as customer and agent performance metrics.
For text-based communications, a large-scale trending algorithm is used. The system learns what the standard distribution of customer vocabulary is, and can identify when new topics of conversation are introduced, or when existing topics are occurring at statistically significant volumes. EdgeTier leverages word and sentence-level contextual text embeddings to encode customer communications and model variances in real time, achieving language-agnostic anomaly identification.
The WatchTower trend detection system identifies anomalies at a thematic level, rather than just being based on keyword detection. Thematic extraction is an important step to reduce false positives. EdgeTier’s ability to understand customer context at a thematic level is tested regularly in the real-world as some of EdgeTier’s customers operate in industries such as sports-betting where major sporting events may drive lots of similar, but normal, discussion. For example, during the European Football Championships, people naturally spoke more about certain teams or results on match-days. EdgeTier’s technology was able to filter through all the noise to still detect genuine customer issues about system failures, while not triggering an alert when, say, lots of people who contacted happen to mention the ‘euro finals’ in their messages.
EdgeTier’s emotion model uses attention-based and context-aware token-level embedding vectors with dense neural networks configured for text classification tasks. EdgeTier has built a custom dataset spanning multiple industries and languages to achieve >90% accuracy on detection of frustrated customers, allowing contact centre managers to instantly target the most affected customers during anomalous events.
For numeric inputs, WatchTower uses a Long-Short-Term-Memory (LSTM) based neural network model to predict the performance metrics of the centre every 5 minutes. The model bases predictions on its knowledge of the centre’s long-term seasonal fluctuations as well as more recent granular minute-by-minute performance. By comparing the predicted and real outputs from the model, anomalous periods can be identified, and the driving signal highlighted.
Additional Information:
Since its launch in mid 2021, the reception of WatchTower has been phenomenal. WatchTower is being used by market-leading companies in the travel, retail, delivery and gamblingindustries, ranging in size from 50 to 3000 agents. The companies currently actively using WatchTower include:
– Glovo – Spanish-based delivery company operating in ~23 countries.
– Freshly – US-based meal subscription service that ship ~ 1 million meals per week.
– Holiday Extras – UK travel company focusing on travel ancillary services.
– Tipico – Germany’s leading sports betting and casino company.
– CarTrawler – Irish travel technology provider.
– Love Holidays – UK travel agency.
– SuperBet – Romania’s leading sports betting and casino company.
– Swappie – European online phone retailer.
Senior customer service leaders noticed an immediate impact as soon as WatchTower was deployed. In one such example in CarTrawler, WatchTower allowed the team to detect and react to a system issue immediately. Tara Carmody from CarTrawler commented “We did have an outage today and the WatchTower alert was the reason we found out. This enabled me to pinpoint the time of the outage … not only did you help notify us, but you helped pinpoint when it went down”.
Similarly, UK lockdown restriction changes led to confusion amongst customers of Holiday Extras. WatchTower allowed them to understand their customer confusion and react to inform their customers better, as Luke Silcocks from Holiday Extras commented: “The WatchTower alert – SUPER useful. I have given it to PR here to spin up some guidance for the agents”.
Aside from the novel machine learning approach, a key technical challenge that EdgeTier had to overcome was the software development task of handling and analysing vast amounts of customer queries in real-time. EdgeTier designed a scalable architecture to handle hundreds of thousands of daily interactions, ingested from various customer service platform APIs.
EdgeTier are consistently working on new additions for WatchTower, and are working closely with our current and potential customers to identify the next stage of functionality. Functionality currently being implemented includes:
– Agent overview screens to help assess agent performance, for example showing which agents are driving frustration in customers.
– Product feedback screen to clearly show customer attitudes to products and services offered by the company.
– Multi-lingual phrase-based searching. A revolutionary approach to searching through customer contacts that allows contact centre managers to ask questions like “Find me customers in any language that said something similar to ‘You are treating me badly’”. When live, this will allow customer service managers to thoroughly investigate issues.
Over the coming months and years, EdgeTier will continue to innovate and apply next generation machine learning to solving complex problems within customer service teams. Overall, WatchTower represents what is possible when clear knowledge of the customer service domain is matched with deep expertise in AI and software development.