AI voice analytics for contact centre conduct risk detection and monitoring

On 24 November last year, Deloitte and Swiss Re held a combined Virtual Insights Session on the topic of Assistive Voice Analytics.

Brendon Fehon introduced the topic (twice in fact, as we had some technical challenges initially!). Brendon described the scenario of a remediation exercise that required the remediation team to review 900 thousand phone calls from 80 thousand customers. The complexity and time implications of doing this manually would have been prohibitive, and the project therefore resolved to use an AI-driven solution – Assistive Voice Analytics, or AVA.

The challenge

Mike Fowlds then set out some of the challenges of the task. The review needed to identify not only instances where the sales agent provided overtly incorrect or misleading information, but also cases where the agent overlooked or ignored customer hesitancy or vulnerability. The telephone records available were also not ideal, with multiple legacy telephony systems stretching over a period of 10 years, and with an absence of call metadata.


Elea Wurth then introduced the solution infrastructure, AI-driven natural language processing (NPL) which supported machine learning and voice data processing. The highlight of the session was her demonstration of the AVA tool. Elea gave an example using the tool of a voice recording between a sales agent and a potential customer. The call was flagged as high risk because the customer alluded to not being able to afford the insurance product on offer.


Mike then spoke about the favourable outcome of deploying the solution to production, which resulted in a doubling of the number of cases that each call assessor could review each day. This not only improved cost, accuracy and efficiency but just as importantly improved the customer experience.

Model iteration

Elea spoke briefly about other potential applications of AVA in a call centre environment, to improve operational efficiency, examine customer sentiment and for risk & compliance. She then handed over to Pratik Manhas to go a layer deeper into the technical elements of the solution. Pratik described the transcription functionality of AVA and then the iterative model development approach to identify high-risk calls. As the model iterated through later versions it became increasingly adept at identifying hesitancy and distraction from the customer, affordability concerns, misleading information, and pressure retention.

In summary, the AVA solution was not intended to replace human involvement. The final evaluation of each call needed to be made by a human. Its value came in allowing a vast quantity of sales recordings to be reviewed in a reasonable amount of time, in a consistent way, which ultimately benefited the customer.

CPD: Actuaries Institute Members can claim two CPD points for every hour of reading articles on Actuaries Digital.