Conversational AI: Common Risks and How to Avoid Them

All Jane wanted to do was cancel her online subscription. Now, she’s on the phone with an AI agent asking her, for the sixth time, for the subscription number she received when she signed up for the service. She shares her name, date of birth, and ZIP code, but it’s futile. Her only remaining option is to write a lengthy email to customer support, cross her fingers, and wait. 

Meanwhile, Jason was excited about his company’s AI sales agent’s support in converting three prospects into paying customers. Little did he know that the AI agent misinterpreted the negotiated conditions. Rather than offering ten percent off the first year, it offered the discount for perpetuity. It’s a mistake that would come back to haunt him when his sales commission came due. 

Navigating the Potential and Pitfalls of Conversational AI

Across industries, companies are exploring the potential of Conversational AI to boost their operational efficiency and customer satisfaction. Investment has only been further spurred on by the meteoric rise of large language models and their promise to transform yesterday’s rigid interactive chat bots into the smooth talking conversational agents of tomorrow. 

But, for reasons both real and imagined, many companies have been reluctant to scale Conversational AI PoCs into operational deployments. According to McKinsey, only eleven percent of companies have adopted Gen AI at scale, and that, despite its potentially transformational value to business across sectors. 

To succeed in commercial deployments, Conversational AI needs to add value to its users, be they internal teams or external customers. While the McKinsey report outlines steps CIOs can take to capture its full value, this article outlines perceived risks associated with the technology and shows how they can be mitigated with the right technological components. 

Steering Clear of PoC Purgatory

As outlined by McKinsey, many Gen AI PoCs fail to scale into commercial deployments. Stuck in PoC purgatory, they fail to deliver the efficiency gains promised by the technology. But dig deeper and you’ll see that you’ll see that it promises to improve quality, employee satisfaction, customer satisfaction, competitive positioning, and, ultimately, revenue. 

To scale, McKinsey argues that CIOs need to move beyond experimental pilots. Instead, they need “rewire how they work, and putting in place a scalable technology foundation is a key part of that process.” This involves, among other key points, focusing on tasks, tech, and teams that add most value, while getting a handle on costs. 

Overcoming Precision Limitations of LLMs

While Large Language Models (LLMs) are probabilistic by design, many of the decisions we expect them to take are not. This is particularly true in mission-critical settings that require clear recommendations, such as in medical settings where decisions can have tangible consequences, affecting the health and wellbeing of patients. 

Knowledge graphs, on the other hand, are deterministic by design. Incorporating them into the technology stack powering GenAI balances out the probabilistic nature of LLMs. The additional context awareness they add in terms of knowledge and process flows increases the accuracy and reliability of conversational AIs, qualifying them for enterprise-grade GenAI deployments.  

Integrating Sensitive Information While Maintaining Security

Security looms large on the radar of businesses seeking to deploy Conversational AI at scale. How, for example, can large language models ingest sensitive documents and other information without potentially exposing secrets to unauthorized users? And what can be done to ensure that sensitive data isn’t intercepted in transit to and from the data center running the LLM? 

Rather than training large language models with sensitive corporate data (which would risk exposing sensitive data), Enhanced Retrieval Augmented Generation (Enhanced RAG) can add relevant data sources to the search query, in alignment with the users access rights as stipulated in access control lists (ACLs). 

Additionally, the Squirro Enhanced RAG platform caters to various deployment options, ranging from deployments on premises – meeting the most stringent security requirements – to deployments on public or private cloud servers. Either way, data can be securely encrypted at rest and in transit, keeping it safe from prying eyes.   

Ensuring Consistency across User Interactions

Ask an LLM the same question twice, and you’re likely to get two different answers. Again, this comes back to the probabilistic nature of the algorithms behind the technology. While such inconsistencies between user interactions can be a nuisance in everyday use, in professional settings, they can be detrimental to user satisfaction and lead to suboptimal performance.  

By giving the Conversational AI platform a deeper contextual knowledge of how topics and process steps relate to each other, knowledge graphs can ensure that all user interactions are based on the same underlying understanding of the task at hand, leading to more consistent, relevant, and accurate results. 

At the same time, AI Guardrails ensure consistency between user interactions by enforcing ethical and governance principles as well as data security, guiding AI behavior and preventing deviations from desired behaviors.   

Keeping a Handle on Costs to Speed Up ROI

Until they are convinced of the return on investment that their Conversational AI solutions provide through increased efficiency, enhanced productivity, and improved customer satisfaction, businesses may be reluctant to make investments into the solutions that go beyond the proof of concept scale. 

To keep a handle on costs and speed up ROI, businesses can adopt an alternative approach, starting small and scaling up where the technology promises to add most value. This allows them to acquire experience, smooth out wrinkles in their deployments, and expand further as the need arises. 

Avoiding Misinformation and Bias in Datasets

Foundational large language models are trained using data that is digitally available on the internet. As a result their outputs can be biased towards populations, opinions, and other perspectives that are overrepresented online. 

Because Enhanced RAG can incorporate internal data that closely reflects a company’s target group, it can reduce the impact of misinformation and bias found in the large language models. A business operating in a region that is culturally underrepresented online will, for example, typically be able to draw on customer insights gleaned not from the pre-trained LLM but from its own CRM system, delivering more accurate and relevant outcomes.  

Harness the Power of Advanced Conversational AI with Squirro

Jane and Jason, our heroes from the introduction, were both caught off guard by weaknesses in the technology behind their Conversational AI tools. Fortunately, innovation never stops, and today, Squirro’s state-of-the-art Enhanced RAG technology incorporates a series of technological components that mitigate the perceived risks often associated with Conversational AI. Now mature, our Enterprise GenAI platform is prepared to increase your business’s productivity, efficiency, and user satisfaction, even in the most challenging enterprise settings. 

GenAI’s honeymoon phase is over, says McKinsey. We’ve entered the technology’s maturing phase, during which businesses have the opportunity to transform the promise of GenAI’s promise into business value. Companies that implement successful deployments stand to benefit from a competitive advantage that will set them apart from competitors. 

To learn more about how the Squirro’s Enhanced RAG platform can transform your Conversational AI applications, head over to start.squirro.com to test the solution yourself, or book a demo for a deeper introduction into the platform and its capabilities.