Securex: Custom Speech-To-Text Agents for Customer Care

Securex partnered up with Faktion to integrate cutting-edge AI technologies across its business units. As a trusted leader in human resources and social administration, Securex serves organisations through a wide range of services. Known for its holistic approach to workplace well-being and productivity, Securex constantly seeks innovative solutions to optimise operations and enhance service delivery. Central to this mission is the Customer Care Team (CCT), which provides personalised advice, HR documentation, and support to businesses of all sizes.

Within the Customer Care Team, currently consisting of 110 employees, one of the critical operational bottlenecks was the manual process of documenting customer calls. Agents were required to manually summarize calls and log action items into Salesforce, a process that was:

• Time-consuming, reducing productivity and slowing response times.
• Prone to inconsistencies, leading to incomplete documentation and missed follow-ups.
• Difficult to scale, affecting the ability to manage increasing call volumes efficiently.

The process needed to be streamlined to improve customer satisfaction and service quality. Securex needed a more reliable system that would reduce human errors, improve response times, and streamline workflow efficiency.

In response, together with the AI Lab, we developed a custom Speech-to-Text agent tailored to Securex’ business context. This solution uses LLMs to automate the transcription of customer calls, generate concise summaries, and extract actionable insights. The tool integrates seamlessly with Salesforce, ensuring that all relevant information is captured and easily accessible to the CCT team.

The Speech-to-Text tool streamlines the entire workflow:

1. Call Upload: Customer Care Agents can upload audio files via the interactive UI, tagging them with relevant metadata such as agent and client details.
2. Transcription: Audio is processed through Whisper, which generates an initial transcription.
3. Optimisation: Transcriptions are further refined in an enhancement phase, which encompasses both optimisation - incorporating custom vocabulary and correcting inaccuracies - and diarisation to identify and label speakers
4. Summarisation & Action Items: Once optimised, the transcription is forwarded to the summarisation endpoint, where a structured summary and a list of action points is generated. The summary includes detailed information on speakers, questions asked, resolutions proposed, and subsequent action items, ensuring that critical insights are captured.
5. Translation: Outputs are translated into French and German to support multilingual operations.
6. Human-in-the-Loop: Customer Care Agents can review and validate the results through the UI, with options to provide feedback for further refinement.

Beyond simple transcription, the tool is designed to add significant value through features such as speaker diarisation, multilingual support, and action item generation. It enables the team to focus on delivering high-quality customer service while minimising administrative overhead.

The initial phase of the proof of concept focused on selecting the most suitable Speech-to-Text service. To do so, a thorough evaluation was conducted of three distinct models: Azure Speech API, OpenAI Whisper, and Deepgram API. Whisper was selected for higher transcription accuracy, lower cost compared to Deepgram and easy integration within the Azure environment.

Then, we started with Whisper, providing a foundation to build upon. and from there, we progressively added complexity to improve performance, focusing on optimising transcription accuracy by adding Securex-specific jargon and implementing speaker diarisation to distinguish between agents and customers. AI-generated summaries structured using predefined Securex templates and action items categorised based on task ownership (Securex, agent, client).

Throughout the project, feedback was our guiding force. We brought domain experts in the loop to not only review the results that were generated by the model to ensure accuracy, but we also used their feedback to continuously improve the user experience of the tool. This iterative feedback loop allowed us to refine the tool continuously, ensuring it meets the evolving needs of the CCT team and delivered meaningful improvements in efficiency and accuracy.

The results have demonstrated significant success across all key performance indicators (KPIs), surpassing the established benchmarks necessary for implementation in a production environment:

• Summaries were generated in under 5 minutes.
• Trustworthiness of summaries rated at 100% by Securex evaluators.
• The new workflow eliminated manual inefficiencies and improved Salesforce case tracking.

Securex AI Lab was impressed with the accuracy of the summaries, particularly for shorter calls.

The interactive UI has been particularly well-received, providing a seamless experience for accessing and validating call data.