Telstra is working with IBM Research to make further enhancements to the backend of its Codi virtual assistant, coinciding with the telco’s adoption of an “asynchronous messaging” model for agent-assisted resolution.
Codi was launched in October 2017 and either answers support queries with auto-replies or escalates the question to a customer agent. [pdf]
Technical product owner for Codi, Nicole Hein, told last month’s IBM Think 2021 conference that in the past year, Telstra had shifted chat support “from a live chat [model] to an asynchronous messaging technology stack across” its app and website.
Async messaging is a support model where the agent is not exclusively answering questions for a single customer, but is likely responding to multiple support requests at once.
Hein said that “increased volume” in support requests via text-based channels during Covid - many traditional support channels were unavailable in that time - had pushed Telstra to uplift the AI that supports its people attending to customer support queries.
While the telco managed to scale up digital messaging services quickly for Covid, Hein noted there were concerns around maintaining its “brand and tone of voice” in all interactions.
“We know that the most efficient way of doing this is through pre-defined responses as it enables the agents to engage with customers and drive a really consistent experience,” Hein said.
“We undertook a six week project that looked at the predefined content that was available for agents to manually search and select, but found that it drove a higher agent effort and saw low takeup rates.
“So the question was how can we recommend responses for our agents to use that are contextual to the conversation, recommend documentation and links to enrich the conversations instead of using just text, [and] use preferred responses that are aligned to Telstra's tone of voice and brand? How can we keep it consistent irrespective of which agent is assisting the customer? And how can we continually adapt to Telstra's product changes?
“Each of these are important to ensure that every agent's responses are timely, accurate and personalised.”
Telstra worked with IBM Research and as a result has produced several enhancements to Codi, which is underpinned by IBM Watson and LivePerson technology.
“Since we partnered in 2017 with IBM's Watson technology, we've had an agreement to mask all the personal information but share our conversational transcripts with IBM Research,” Hein said.
“This partnership has enabled the rapid development of specific tooling for agents that aim to enhance every customer interaction.
“In processing the conversational transcripts document and website content, the research teams have created an advanced machine learning and information retrieval model to power the recommendation.
“The tooling enables our authenticated customers’ information to be available and automatically appear in the responses to further personalise the conversation.”
Hein showed off an ‘agent assist’ solution which presented the agent with a list of pre-written answers to questions.
The enhancement continually updates the list of pre-written options, based on what the customer types and the machine learning model suggests.
“The continual analysis of past conversations ensures that the right responses are provided in the conversation and reduces the maintenance overhead,” Hein said.
Agents can choose the right response with a single click and insert it into the chat window.
There is also a new “natural language summarisation” feature, which agents can use to automatically summarise the chat, which can then be copied and pasted into Telstra’s CRM system.
“There is a summary button, which the agent uses to generate notes which they're able to then copy and paste into CRM for the next agents to see the interaction,” she said.