Balancing Act: Navigating the Tightrope Between User Experience and Security in Conversational Agents

Conversational agents powered by large language models (LLMs) are rapidly transforming how businesses interact with customers. These cutting-edge technologies offer personalised, engaging experiences that were once unimaginable.

However, as with any technological advancement, some inherent risks and challenges must be addressed, particularly regarding data privacy and security.

The Challenge: Balancing User Experience and Data Privacy

The power of LLMs lies in their ability to process vast amounts of data and generate human-like responses. However, this strength also threatens user privacy, as these models may inadvertently expose sensitive information during interactions. Imagine a scenario where a customer shares their credit card details or identification number during a conversational exchange, and this data becomes embedded within the LLM's training corpus. The consequences could devastate the individual customer and the business's reputation and compliance with data protection regulations.

The Solution: Privacy-Preserving Middleware

At Sanlam Studios, we recognised this challenge early on. We set out to develop a solution that would allow us to leverage the full potential of LLMs while maintaining the highest data privacy and security standards. Our approach involved implementing a privacy-preserving middleware, a sophisticated layer that acts as a gatekeeper between customer data and the LLM. All API calls to external services are initiated within this secure middleware, ensuring strict data flow and processing control.

This middleware intercepts customer input and meticulously removes personal identifiable information (PII) or sensitive data before forwarding it to the LLM. We utilise cutting-edge technology, such as Microsoft's Presidio, to identify and remove sensitive information like credit card numbers, identification numbers, and other personal identifiers. Doing so ensures that the LLM never directly interacts with this sensitive data, mitigating the risk of exposure or mishandling.

Anonymising Data for LLM Processing with Fidelity

However, simply removing sensitive data from the input is not enough. We must also ensure the LLM can still provide meaningful and relevant responses without compromising the user experience. This is where our approach truly shines. The anonymised data sent to the LLM is carefully crafted to preserve context and fidelity, allowing the model to perform its core functions and generate exceptional responses.

High-Level Conversational Flow

For instance, if a customer shares their credit card number during a conversation, our middleware identifies and removes sensitive information. However, it would replace the credit card number with a high-fidelity obfuscated placeholder instead of deleting the entire statement. This way, the LLM can still understand the context of the conversation and provide an appropriate response without access to the sensitive data.

Once the LLM generates its response, it is sent back to the middleware, where the previously removed sensitive information is meticulously reinjected into the reaction. This ensures that the final output makes perfect sense to the customer while maintaining the highest data privacy and security levels.

Robust Data Security Practices

Sanlam Studios takes its data security responsibility seriously. All sensitive data is stored and processed within our secure infrastructure, adhering to rigorous security protocols and data governance mechanisms. This alignment with industry best practices and regulatory requirements is a cornerstone of our approach, ensuring that we not only provide exceptional user experiences but also maintain the trust and confidence of our customers.

While hosting our own LLM could potentially mitigate third-party data privacy risks, it is not always feasible. Some LLMs require access to the supplier's infrastructure, and there are additional considerations to take into account. For instance, the supplier's data privacy practices might not meet the regulations of our region, or their legal framework might not align with our requirements. In such cases, we work closely with the suppliers to ensure their practices meet our stringent standards and mitigate potential risks.

Drawbacks and Considerations

While our privacy-preserving middleware solution offers robust data privacy and security, it is essential to acknowledge potential trade-offs and challenges that must be carefully considered and addressed.

The privacy middleware has a computational overhead as it needs to remove and reintroduce personal identifiable information. This may introduce latency into the conversation, affecting the user experience. It further represents an additional layer in the architecture, which increases complexity. It has its own infrastructure, which should be scalable to prevent any bottlenecks.

Conclusion

In conclusion, balancing user experience and data security in conversational agents is a delicate tightrope walk. At our company, we have embraced this challenge head-on, developing a robust privacy-preserving middleware solution that enables us to harness the full potential of LLMs while prioritising data privacy and security. By doing so, we are paving the way for customers to enjoy seamless, personalised interactions without compromising the sanctity of their sensitive information.