GenAI in contact centres
GenAI is predicted to have a large influence on various fields, with its ability to generate content, code, images et al, with limited or no human intervention. One of the areas ripe for significant transformation is the contact centre operations. AWS provides GenAI technology, which can be effectively leveraged for this. Across the customer journey in a contact centre, there are various areas where GenAI solutions can be implemented for automation, improvement in efficiency of agents, better customer experience and lowering operational costs.
1) Chat and voice self service bots — As a customer reaches out to the contact centre, a self service automated bot can be introduced to interact with the customer on either chat or voice channels. While traditional AI/ML services from AWS like Lex and Polly can be used here, GenAI models will help increase the accuracy and range of the bots by being able to provide answers from unstructured data or past conversation transcripts between customers and agents.
2) Agent assist — GenAI technology can be used to provide predictive guidance to the agents on next steps or next response. Agents on chat can review the next response provided by GenAI and use it as such or make edits to it to respond to the customer. This will significantly reduce the time taken by the agents to revert to the customer and close interactions faster.
3) Transcript summary and insights — At the end of every customer interaction, GenAI can provide a summary of the entire transcript. Creation of the interaction summary used to be a manual job done by the agent at the end of every conversation. By automating this with GenAI, a significant amount of time is saved by the agent, improving operational efficiency. GenAI is also capable of providing extracts of specific aspects of the conversation like issue being discussed, action items agreed to and key outcomes of the discussion. This helps in maintaining more structured records of the conversations with the customer — enabling easier reviews and retrievals later.
4) Call analytics — The analysis can cover various aspects like compliance and data protection adherence, overall call performance, customer sentiment and abstractive call summaries.
5) Supervisor assist — Supervisors can be provided with pre-filled evaluation forms by GenAI, for each completed interaction between the customer and the agent. The evaluation forms can be auto filled to cover aspects like performance of the agent, compliance and data privacy checks, customer satisfaction and sentiment etc. This can also be used to identify coaching inputs for the agents. The other main advantage is that automation/GenAI solutions help cover the entire set of conversations between agent and customer as part of supervisor evaluation, instead of just a selected percentage as samples.
Let’s examine a solution architecture that provides some of these GenAI solutions on an AWS platform.
· Customer dials into the contact centre (1)
· The call is received on the IVR of Amazon Connect (2)
· Traditional AI/ML services from AWS (3) — Lex for voice to text and intent identification + Polly for text to voice — are used to provide the automated response.
· Lambda function (4) is activated which retrieves data or information required from backend systems (5) to complete the response to the customer
· If the Lex bot is not able to provide a response based on the intent identified, the customer query is passed onto a GenAI model trained on Sagemaker (8) or Bedrock.
· Call recordings are saved in an S3 bucket (6).
· Call recordings are retrieved with a Lambda function (7) and sent to GenAI models trained on Sagemaker (8) or Bedrock for subsequent analysis, call transcript summary etc.
· Call transcript summary, predictive guidance etc. is pushed to the agent screen (9)
· Call analysis report, transcript summary etc. is provided to the supervisor (10)
Like with all new and shiny technology, there is a tendency to force fit GenAI solutions even in areas where it is not strictly required. In some cases, other AI/ML services may be more appropriate to implement. In the above example, a two-step approach has been adopted, where Lex is used for intent identification and response initially, before passing onto GenAI models. Lex would be a better solution in cases where structured data and answers are readily available (eg: FAQ documents), as the responses can be pre-configured based on identified intent and accuracy will be higher. In this scenario, we are only using a probabilistic model to identify an intent, while response generation is deterministic. GenAI models would be a better solution, if the answer must be retrieved from massive amounts of unstructured data. The earlier approach will not be suitable here and we have to use probabilistic models to both identify intent and generate response. Identifying the appropriate use cases for GenAI implementation is therefore the key to successfully leveraging this technology to drive business outcomes.
