By Yaniv Masjedi, Chief Marketing Officer at Nextiva
Business leaders are discovering the many promises of voice AI, including unprecedented scale, zero initial hold times, and 24/7 availability. Yet “enterprise-scale deployment remains rare and difficult to sustain,” McKinsey reports. Why is that? Because many of these tools have not been designed for the reality of how humans operate.
AI receptionists are generally built to handle the most common use cases. That makes sense. The more an AI voice agent can provide basic information or answer common queries, the less time human staff needs to spend on those kinds of rote tasks.
But what happens if someone asks a question the tool doesn’t know how to handle, or communicates in a way the tool isn’t trained to understand? The solution may seem as simple as alerting a human to take over the conversation. But AI tools are complex. They’re built to try to be as helpful as possible, and not to give up too easily.
All too often, the result is customer experience (CX) failure. The AI receptionist keeps sending the caller down an unhelpful path, asks the same questions, repeats information, or creates unacceptable lag times.
In a recent report, McKinsey described two cases in which things went wrong. One was a global fintech company that brought on AI to automate customer services. “It encountered immediate and significant challenges, and as a result, customer satisfaction declined and complaints increased,” the consulting firm said. “Issues included a perceived lack of empathy, misunderstanding of callers’ intent, improper judgment, generic and repetitive responses, and a lack of nuanced support for edge inquiries.”
The other example involved a healthcare organization that was evaluating a plan to automate up to 90% of customer service interactions. “Unfortunately, the design focused only on high-level intents and business flows, neglecting edge cases, compliance requirements, historical logs, and back-end business constraints. The vendor had prepared the system to deal with simple queries, not the emotional, multi-topic calls it actually received. The system was not prepared for regional differences, had insufficient intent detection and context retention capabilities, and struggled (to put it mildly) with error recovery.”
The Power of Smart Escalation
Teaching an AI tool certain common conversation flows is important. But so is freeing the tool from the confines of those flows. That is what’s often lacking.
“Human speech and expression are reduced to a flow chart,” McKinsey says. “This technique is tempting but doomed. It leads to overly rigid paths and no recovery design for ambiguity, even though human interactions are full of ambiguity. AI voice agents do not fail because speech is imperfect but rather because they cannot handle the variability, unpredictability, and sophistication of human conversations.”
AI tools should be empowered with “conversation architecture” that's much more adaptive, McKinsey says. It prepares a tool to handle “interruption, ambiguity, recovery loops, and emotion. Live call shadowing and pilot cohorts are advisable before scaling.”
Also, people should be alerted when an interaction might be going wrong, and an AI tool could use human help getting back on the right track -- even before turning over the conversation to a human entirely. A comprehensive study published this year in the International Journal of Applied Resilience and Sustainability highlights the importance of human-in-the-loop methods. “The edge cases, the uncertain predictions or the delicate situations are escalated to human decision-makers, who possess the suitable expertise and authority,” the study says, adding that designers should make a habit of evaluating edge cases.
In designing our AI receptionist XBert, my team at Nextiva understood the importance of building in coverage for edge cases. We developed smart escalation rules that ensure human oversight and alert staff of possible problems. And we advise businesses to review edge cases monthly. Teach your AI tool what to do when similar cases arise in the future.
When evaluating the best AI voice answering service for enterprises and businesses of all sizes, edge cases are one of the most important differentiators. The best AI call handler will gracefully manage the unexpected, alert human experts to help guide it, learn from its mistakes, and keep evolving. It will quickly come to recognize the kinds of concerns people have, and understand how to guide them to the best possible resolution.
This kind of versatility is essential. After all, in the real world, humans don’t stick to a script.