It’s undeniable that hallucinations are one of the main sticking points when it comes to the mass rollout of AI technology.
This can be a roadblock for enterprises looking to implement AI, especially when there are constant reports about these hallucinations hitting the headlines.
While some are on the comical side, like when a Virgin Money customer was warned against hateful speech for using the word “virgin,” some are more serious.
Take Cursor, for example. The AI-powered software coding assistant from AI startup Anysphere went viral due to a chatbot hallucination in response to a recent customer service query.
When customers got logged out of their accounts and asked customer support for assistance, an AI chatbot named ‘Sam’ told them that this was “expected behavior” under a new policy — a policy the chatbot had simply invented all on its own.
This led to confusion and distrust among the company’s customers, with some even cancelling their accounts.
While no one is denying the benefits of AI-powered tools, the simple truth is Large Language Models (LLMs), which can be used to develop chatbots, are powerful, but they can also produce these types of frustrating hallucinations if proper guardrails are not put in place.
Speaking to CX Today, Nikola Mrkšić, CEO and Co-Founder of PolyAI, explained how his company has “been able to constrain the behavior of LLMs and use them in places where it makes the most sense to drive customer service conversations.”
The team at PolyAI helps enterprises speak with customers through voice AI agents, and it’s important that these agents not only say, but also do the right things, rather than hallucinate responses or tell you they’ve taken an action when they really haven’t.
Some of the guardrails in place with PolyAI’s agents are powered by retrieval-augmented generation, or RAG, a technique that enables AI agents to cross-reference knowledge from a generative model with a knowledge base.
This ensures that an AI agent checks its generated responses against information the enterprise has confirmed as factual.
In doing so, it prevents inaccurate, irrelevant, and inappropriate responses, and keeps customer conversations within established limits.
Where Enterprises Can Make a Difference Today with AI
In pursuing seamless CX, businesses must evaluate how AI and automation support accuracy, trust, transparency, operational costs, and efficiency.
According to PolyAI’s Mrkšić, enterprises considering where to start implementing AI for CX should consider sophisticated voice AI agents among their first real-world deployments.
Mrkšić said, “AI is such a big and monumental thing, and many people can’t resist mounting these large offensives. What they need is a lot of probing attacks on different front lines.
Large cloud providers or CCaaS vendors will tell enterprises to start with Agent Assist capabilities, and then down the road think about automation, but we take a different approach.
Mrkšić also noted that while AI assistance for human agents certainly brings business benefits, it does not always help address wider automation plans.
“The model proposed by CCaaS vendors will not lead to the future these enterprises need, where 90% of calls are automated.”
How PolyAI is Addressing the Challenges of Providing Good CX through Voice AI
Despite the benefits of CX over the phone, there are some considerations for deploying this type of technology, too.
Voice interactions remain central to CX but face obstacles such as:
Latency and speech recognition errors: frustrating delays and misinterpretations can degrade customer experiences.
Lack of contextual awareness: AI systems may struggle with complex queries requiring historical context.
Limited conversational flexibility: rigid AI scripts reduce adaptability in dynamic interactions.
However, PolyAI specializes in AI-driven voice assistants designed to address these challenges by maintaining natural, human-like conversations, providing accurate, contextual responses, and enhancing scalability.
Their AI agents can handle the complexities of real-world conversations, understanding diverse accents, navigating interruptions, and adapting to shifts in context.
PolyAI can resolve between 50% and 75% of inbound calls entirely autonomously.
At the core of PolyAI’s tech is a powerful blend of spoken language understanding, speech synthesis, and intelligent dialogue management.
By combining retrieval-based and generative AI models, their system delivers fast, accurate, and natural-sounding responses that adapt to each caller’s needs.
Integration is refreshingly straightforward: PolyAI’s platform works out of the box with systems like Salesforce, Twilio, and Amazon Connect, and its agents can be deployed in under six weeks, without the need to replace existing systems.
Designed for enterprise environments, PolyAI’s voice agents meet the highest standards of security and compliance, and they’re fluent in more than a dozen languages.
To see a demo of this technology in action, check out this video.