Czech financial savings financial institution Česká spořitelna, a division of Austria’s Erste Group, not too long ago collaborated with AI resolution builder DataSentics to discover using GenAI in name facilities. Česká wished to enhance high quality management and optimize prices of their inbound name heart operations, which obtain round 2 million calls per yr. They selected the Databricks Information Intelligence Platform to experiment with each inside and exterior AI fashions to evaluate the effectiveness of name heart brokers.
Exploring a High quality Management System for Buyer Assist
The decision heart crew at Česká spořitelna wished to check a high quality management system powered by GenAI that may make sure that brokers adhere to scripted tips throughout buyer interactions. A important problem for Ceska was guaranteeing constant agent communication for routine buyer inquiries. When clients name about account balances, brokers have to direct them to on-line banking options, a key enterprise requirement that drives digital adoption and operational effectivity. The help crew wanted a scalable solution to confirm agent compliance and keep communication requirements throughout 1000’s of buyer interactions. To realize this, the crew started by utilizing Whisper, a speech-to-text mannequin from OpenAI, to transcribe conversations precisely. The problem was to supply human-readable textual content that precisely represented spoken phrases utilized by name heart brokers with out distorting their which means. The transcriptions wanted to make logical sense and replicate the intent of the dialog precisely for additional evaluation.
Following the transcription, the crew explored integrating each inside GPT fashions and open supply fashions akin to Mixtral to guage their effectiveness. GenAI fashions have been examined in a simulated QA position, the place they have been tasked with answering particular questions akin to “Did the agent redirect the client to on-line banking?”. The aim of this train was to evaluate how properly these fashions may mimic human understanding and decision-making when verifying compliance with established tips. By evaluating the efficiency of each the interior GPT mannequin and the open supply fashions, the crew aimed to seek out the simplest resolution for enhancing customer support via automated AI-driven high quality management.
Advantages of the Databricks Information Intelligence Platform for GenAI
The DataSentics crew evaluated a number of choices for this resolution, and finally selected to deploy the Databricks Information Intelligence Platform and Mosaic AI instruments at Česká spořitelna for a number of causes:
- Information Administration and Governance Advantages: Unity Catalog makes knowledge simply accessible for various fashions whereas preserving delicate knowledge beneath restricted entry.
- Complete Information Processing Capabilities: the Databricks Platform helps your entire workflow of preprocessing of name heart knowledge, from transcription to high quality management. This allows us to supply intermediate outcomes that may be leveraged for different fashions and tasks, akin to advertising and marketing, threat evaluation, regulatory compliance, and fraud detection.
- Mannequin Coaching and Assist: Databricks gives sturdy help and experience for GenAI, together with mannequin structure and coaching capabilities. This made it a perfect platform for testing and deploying open supply fashions rapidly, enabling us to experiment and iterate effectively.
- Ease of Cluster Creation: With Databricks, it’s easy to create clusters and deploy open-source fashions. This streamlines the experimentation course of and permits us to focus extra on mannequin efficiency and fewer on infrastructure administration.
Insights and Outcomes
All through the challenge, we experimented with numerous segmentation methods and gathered a number of worthwhile insights:
- High quality of Enter Information is Essential: The standard of the audio recordings diversified from shopper to shopper, with some talking quietly or from a distance, which may later have an effect on the accuracy of the transcription. Whisper or related methods might help remedy the issue.
- Class Definition is a Should: We realized that if classes can’t be simply outlined for people, it’s equally difficult for LLMs to know them. This bolstered the necessity for clear and exact class definitions to coach the fashions successfully.
- Open-Supply Fashions Ship Outcomes: Open-source fashions demonstrated that they may compete successfully with proprietary fashions like ChatGPT. This discovering is critical for companies trying to optimize prices whereas nonetheless reaching high-quality outcomes.
What’s Subsequent
With GenAI instruments powered by Databricks Mosaic AI, Česká spořitelna staff are actually in a position to acquire entry to solutions present in a spread of paperwork through “sensible search” performance. For instance, the buying crew might have to seek the advice of a whole bunch of pages of course of documentation on the best way to management and approve funds to completely different international locations. Earlier than leveraging Databricks, it could take staff hours to seek out the proper data they want. Now, RAG-powered search offers staff solutions inside seconds, together with citations and hyperlinks to the supply doc.
Wanting forward, there are many alternatives to discover extra GenAI workloads at Česká spořitelna. We purpose to create a strong integration between Databricks and Česká spořitelna’s inside database name heart recordings. This can unlock new use instances akin to churn detection, sentiment evaluation, and gross sales sign detection since Databricks is the go-to platform for streaming knowledge. These each day experiences will enable Česká spořitelna to react to modifications in actual time whereas reaching value reductions with improved high quality assurance of their name facilities.
This weblog put up was collectively authored by Petra Starmanova (Česká spořitelna), Tereza Mokrenova (DataSentics), Dalibor Karásek (DataSentics) and Joannis Paul Schweres (Databricks).