Two years ago, the team at Exinity, an online brokerage firm, was looking at its customer pain points. The business, which has over two million clients, in 150 countries, was handling a huge volume of live chats across its team of increasingly beleaguered customer service agents.
Agents were handling up to 16 chats concurrently, in multiple languages, and struggled to provide timely responses to the platform’s trading community.
At the time, the team was supported by two chatbots — one handling enquiries in Chinese and the other in English. The bots were “deflecting around 20% or so of the [chat] volume,” said Rachel Carliss, senior vice president, customer relations, Exinity.
Clearly, the team needed additional tools to improve its CX, so Exinity started investigating artificial intelligence (AI).“AI was just becoming a thing, and the decision was made to start investigating new types of bot,” added Carliss.
The fintech quickly signed Ebo.ai to develop two AI-chatbots that would handle a range of transactional and procedural enquiries, such as opening accounts, resetting passwords, and providing information around deposits and payment methods.
The project applied natural language processing (NPL) to understand over 90% of client enquiries and machine learning to improve language models.
One bot, or virtual agent, would handle transactions on the Alpari platform, and speak Farsi and Russian. The other would answer enquiries in Chinese and English for the FXTM side of the business.
Exinity and Ebo embarked on an in-depth conversation planning exercise. The brokerage platform provided a database of customer conversations that formed a baseline for its virtual agents. Workshops mapped processes, captured frequently asked questions and responses, and sketched out the personas for the bots.
“We sat with [Ebo] to craft how the VA should speak with the client. Should it be friendly? Should it be formal? Should it be informal? We have a lot of different clients from different parts of the world — some are very formal, and some are informal,” Carliss explained.
When the virtual agents were rolled out in July 2023, it was hoped they would increase the call deflection rate — the percentage of enquiries handled in their entirety by the bot — to 30%. But the AI-powered VAs rapidly pushed the call defection rate to 50%.
The bots currently handle “around 40,000 conversations live with customers,” said Carliss.
The virtual agents were implemented across multiple channels, including WhatsApp, Telegram and WeChat.
That omnichannel approach helped Exinity to improve customer satisfaction rates across its four core languages. CSAT scores have been consistent for the year.
The Farsi service for Alpari customers and the English language FXTM Virtual Agent both track around the 60% mark, while Russian language support in the Alpari VA sits in the mid-40% range.
Satisfaction with the Chinese VA lags further behind, partly due to a strong cultural bias against virtual agents. “Our Chinese clients just want to speak with a human. We’re trying to change that behaviour by proving to clients that the VA can answer quickly [and] efficiently,” commented Carliss.
Despite the success of the AI-chatbots, Exinity has no plans to reduce its headcount. Customer service agents now each handle six live chats simultaneously — still a high figure when compared with other markets. However, agents now focus their efforts on complex and higher value enquiries.
“Our chatbot initiative was really driven by cost optimisation, rather than cost reduction,” said Carliss.
“We didn’t reduce headcount despite the fact we’ve achieved a 50% deflection rate. We’ve repurposed those customer service agents in other roles in the business. We’ve given them career development,” she added.
Some of Exinity’s human agents have shifted into quality assurance (QA) work, where they check the responses of their human and virtual colleagues. Exinity is currently working to improve its English, Russian, Farsi and Chinese language support. It also plans to add Spanish, Portuguese, Arabic and possibly Vietnamese language support in the near future.
“We are currently offering a very simple conversational flow in three of these languages, but have not yet implemented a full end to end deflection programme,” explained Carliss.
Further AI work is also likely as the trading platform continually improves its virtual agents. “This is an ever-evolving project. There is always a need to feed more knowledge in. Whenever we change something in the business, the virtual agents have to be made aware,” said Carliss.