Voice AI: Reducing risk and building trust

Everybody is talking about voice AI – and it's little wonder why. Promising conversational, low-latency and natural language interactions, voice AI can be automated and scaled to deliver huge efficiencies to the organizations using it. 

Voice AI is no longer a chatbot, it's an automated, conversational AI interface that customers can navigate as if talking to a human. It's replacing IVRs, it can be used to deliver onboarding, and it's also a channel, with customers now able to navigate their journey across written and spoken conversational interfaces. 

As Ebrahim Hyder, VP of customer care for Michael Kors and a CX Network Advisory Board member, says: "Conversational AI is no longer tied to a single channel or task. Analysts increasingly see it as an orchestration layer that connects systems, channels, and people around customer intent. In this model, a journey can start in voice, continue in chat, trigger backend actions, and involve a human, without losing context," he explains.  

However, as with all AI, voice is not risk free. This article looks at the importance of source verification, accuracy, transparency and safeguards, and explains how trust in voice AI can be maintained across internal teams, the wider organization, leadership and, of course, the customer base. 

Don't miss any news, updates or insider tips from CX Network by getting them delivered to your inbox. Sign up to our newsletter and join our community of experts. 

Learn More

Reducing the risk of failure with voice AI 

Like all AI systems, voice AI systems can hallucinate, and they are also susceptible to attacks that are difficult to detect. The primary risks include source verification and ensuring accuracy of outputs, but further risk exists in data misuse, unauthorized access, and impersonation scams. Organizations must also reduce the risk around data consent and transparency. As with all technology, risks such as these must be appropriately weighed and understood

"Organizations can't afford to govern only for what they can see coming," says Sue Duris, principal consultant at M4 Communications. "The risk sits on both sides – those deliberately exploiting the technology and the gaps the technology itself creates. This goes well beyond technology and starts impacting trust and reputation." 

Trusted sources

Organizations and customers must both be confident that a voice interaction originates from a legitimate, trusted source.

"This is complicated by the rise of voice cloning, where AI can replicate a person's voice with minimal input, making it increasingly difficult to distinguish authentic communication from synthetic impersonation," Duris explains. "When that capability is weaponized, it enables sophisticated phishing schemes designed to manipulate people into sharing sensitive information, passwords, or taking actions they otherwise wouldn't – attacks that are harder to detect precisely because the voice sounds credible and familiar."

Accuracy

Beyond deliberate misuse, Duris warns there are also inherent accuracy risks to tackle in the technology itself. 

"Voice AI systems can hallucinate – generating responses that are confident but factually incorrect – and organizations deploying them in customer-facing contexts must build in mechanisms to validate outputs before they cause harm," she says. "In high-stakes interactions involving financial, medical, or personal data, unchecked inaccuracies aren't just a quality issue; they carry real liability."

The right to know

Further risk is presented by consent and transparency, which comprises both ethical and regulatory risk. 

"Users have a right to know they are interacting with AI – and that disclosure becomes even more critical when the experience is degraded," Duris says. "Lag, unnatural pacing, and clunky conversation flow are tell-tale signs that the technology isn't ready, and when users don't know they're talking to AI, that frustration lands directly on the brand. If the experience isn't seamless, disclosure isn't optional." 

Essential technical safeguards for voice AI  

As Lauren Kiefer, head of enterprise agents go-to-market for ElevenLabs explains, there are technical safeguards businesses can implement to prevent risks. 

She says: "We think about this as a layered architecture where no single control catches everything, but together they make failures extremely rare."

Specifically referencing the guardrails feature offered by ElevenLabs, she adds: "Input validation detects manipulation attempts, every output is independently checked against your policies before it reaches the customer, and for high-stakes actions like refunds or account changes, deterministic controls gate the decision through verification from tools rather than LLM judgment."

At ElevenLabs, some enterprise users test their agents with hundreds of pre-launch simulations to catch edge cases before their own customers do. Kiefer says: "We believe customers should have full control over where data lives and how it's handled, so we also offer end-to-end encryption, flexible data residency options, PII redaction, and zero retention mode for sensitive data."

She concludes: "For enterprise teams, trust is key to adoption, and teams need to trust that their agents will be safe and reliable at scale."

Preserving trust and mitigating fraud  

Kiefer says trust operates at four levels: customers must trust the experience, internal teams must trust the workflow, security and legal teams must trust the infrastructure, and leadership must trust the business case. The organizations that solve trust first scale fastest.

1. Customer trust

This is rooted in the quality of the interaction, specifically voice quality, natural pacing, and seamless handoffs to humans when needed. 

"Customers are increasingly aware of when they're interacting with AI. Transparency and quality determine whether they engage or disengage," Kiefer says. "Consistency matters too: the agent should behave the same whether the customer calls or chats. Configurable guardrails play a critical role here, ensuring the agent stays on topic, follows brand guidelines, and escalates gracefully when it reaches the boundary of what it should handle autonomously," she adds.

2. Internal team trust

AI agents succeed when frontline teams see them as a tool that makes their work better. Kiefer says this requires organizations to design agent workflows around escalation and collaboration, routing complex or sensitive interactions to human agents with full context, "so the handoff feels seamless rather than like starting over". 

She adds: "Teams that use AI to handle routine volume find they can focus on higher-value interactions, reducing burnout while improving outcomes. Adoption depends on giving agents and supervisors visibility into what the AI is doing: what it resolved, what it escalated, and why."

3. Enterprise trust

Organizations need to determine what "enterprise-ready" means in practice, "not just checkbox certifications but architecture that delivers security by design", Kiefer says. 

Key considerations include regulatory compliance by industry, such as PCI DSS for financial services, HIPAA for healthcare, GDPR for European operations, as well as data residency options, end-to-end encryption, zero retention modes for regulated workflows, and monitoring and observability across every interaction. 

"Platforms with a vertically integrated stack – where all models run in a single pipeline rather than chaining third-party services – deliver fewer data handoff points and stronger privacy by design," she explains. 

"As the infrastructure matures, frameworks like AI Underwriting Certification (AIUC) are emerging as trust signals for procurement teams – quantifying risk to the point that AI agent interactions can now be insured," she adds.

4. Leadership trust

This is about connecting CX improvements to "business outcomes the C-suite cares about", Kiefer says, such as churn reduction, revenue impact, and lifetime value, not just average handle time. 

"Unified analytics that show agent performance across every channel give leadership the data to justify scaling," she explains. 

What customers think of voice AI

Excitement around voice AI in the business world means little if customers are not equally enthusiastic about the technology. After investing the time and money to develop and deploy a voice agent, the last thing an organization wants to see is a collapse in CSat, or containment rates that suggest customers are being blocked from reaching the resolutions they need. 

As with many things in CX, customers are satisfied when their experience is seamless, effortless, and efficient – and a series of recent studies from those developing and selling voice AI indicate it hits the mark in this respect. 

• The 2025 AI in customer service report from PolyAI found 87 percent of customers are happy or willing to use AI voice agents for customer service and more than three quarters of CX and customer service leaders believe that AI voice agents could one day replace human customer service representative. 
• Analysis published in 2026 by Tempo AI reported that approximately 25 percent of customers did not realize they were interacting with an AI voice agent until after completing the intake process, if at all. Among the 75 percent who recognized they were speaking with AI, only 12 percent requested to be transferred to a human representative.
• One way to get customers on board with voice AI is the use of celebrity voices. A small-scale 2024 study from 8x8 found customers would prefer to hear celebrity voice clones when calling customer service, with voices such as Taylor Swift's and Margot Robbie's in the highest demand. 

However, a 2025 survey by the global consultancy COPC found there are caveats in customer acceptance of voice AI: customers will accept limited empathy or a scripted tone if the interaction is effective, but they will not accept unresolved issues or repeated effort.
Globally, this research found that satisfaction soars above 90 percent when resolution occurs without further steps. But when AI fails to resolve the issue, the Net Promoter Score (NPS) drops by as much as 70 points.

Is voice AI right for your business? 

The ability to build a voice AI agent in your brand's voice is appealing for many reasons. 
As this whitepaper from ElevenLabs explains, the ideal AI voice depends heavily on the company's use case, brand identity and audience. It must be authentic, localized and capable of natural turn-taking without latency issues. 

Voice AI is also proving itself to have robust security provision, despite the scale and level of threat that exists at present. Enterprise grade platforms are designed with encryption, access controls, audit logs, and strict privacy settings to mitigate the risks set out in this article. But as with all cyber-security, threats evolve and what works today must be continually upgraded to ensure the same level of protection in future. 

To find out more about conversational voice AI, download 
the CX Network report, The new rules of conversational AI

Quick links

• Securing the contact center in the AI era: Prompt injection, consumer privacy, and data integrity
• Essential AI guardrails: How to do security testing, APIs, and logs
  
• Why voice fraud is CX's most underestimated problem
  
  

40906.007 - All Access: Future Contact Centers 2026

All Access: Future Contact Centers will explore the latest trends, from AI agents to advanced analytics, and discover the strategies you need to stay ahead and drive success in the years to come. 

Register Now