The Most Important Thing About Voice AI Is Not the Voice

The most important thing about voice AI is not the voice. It’s the experience voice can make possible.

Think about the best service experience you have ever had. You felt understood. You did not have to repeat yourself, learn how the organization worked or navigate its internal complexity. You explained what you needed, and someone used context, judgment and knowledge of the organization to help move you toward a successful outcome.

What if enterprises could provide more of those experiences, more consistently and at greater scale?

That’s the real promise of voice agents, not simply replacing typing with talking or automating every human interaction. Voice agents shift the experience from navigating systems to expressing intent. A customer or employee describes what they want to accomplish in their own words. The agent understands the context, coordinates the appropriate knowledge and systems behind the scenes, takes permitted actions, and brings in a real person when judgment, empathy or authority is required.

For decades, people have adapted themselves to technology. We learned where to click, which menu to open and how to translate a simple intention into a series of commands. Generative AI began to change that by letting us converse in everyday language. Voice takes the next step and begins to reverse the relationship entirely: the technology adapts to how people naturally communicate, while the complexity of the enterprise recedes into the background.

Why Voice Changes the Enterprise Experience

Voice matters because it removes friction for employees, customers and new hires alike. An employee calling a service desk should not need to know which system owns a problem or houses a solution. A customer should not have to repeat the same information as they move between channels. A new hire should not be limited to static training materials when they could practice a realistic conversation and receive feedback in the moment.

Voice also broadens access for people away from a keyboard, working in the field or simply more comfortable speaking than typing. Multilingual capabilities help organizations communicate consistently across geographies without turning every interaction into a separate production effort.

And yet voice is not a replacement for every screen or every human conversation. Visual and written mediums remain better for comparing dense information and reviewing precise details, and human judgment remains essential when a situation is ambiguous, sensitive or consequential. The opportunity is to use voice where conversation is the most natural path to an outcome, then design clear handoffs when it is not.

A Voice Interface Is Not Enough

A natural-sounding voice is not enough. The actual value comes from whether the agent can make the experience meaningfully better: removing friction, preserving context, accelerating resolution and helping someone reach the right outcome. That requires more than a compelling interface. It requires connected data and workflows, clear boundaries, trusted handoffs and accountability for what happens next.

So there’s a distinction between scripting and intent. A traditional workflow scripts every step in advance. A voice agent designed around outcomes, however, receives the intent, understands the rules and context, and adapts its approach until the goal is met or a person’s judgment is needed.

It helps to know where buyers actually are, because enterprise has moved through three phases in a short time. For a while we were in an adoption frenzy — adoption largely for its own sake. Then leaders found themselves accountable for what those investments returned, and they often could not articulate the value they had received. That produced a recalibration around outcomes. The question was no longer only whether a deployment produced an outcome, but whether achieving it compromised anything critical. There is a kind of generous skepticism — a real willingness to invest, paired with an expectation that both questions get answered.

The forecasts point the same way. Gartner expects agentic AI to autonomously resolve 80% of common customer service issues by 2029. Whether the exact number is 80% doesn’t matter. The important word here is resolve. The conversation begins with access, and the product is the resolution.

Customer Zero in a Very Public Setting

At DXC, one of our earliest voice AI deployments took place in an unusually visible environment: our earnings calls. Over three consecutive quarters, CEO Raul Fernandez wrote his prepared script while a custom AI voice model, built with ElevenLabs, delivered it in his voice.

The first call demonstrated the core capability. On the second, his remarks could be shared in Spanish, French, German, Portuguese and Arabic. By the third, the script went out simultaneously in six languages. Each time, Raul told listeners that he had written the words and that an AI model was delivering them.

An earnings call is not an easy venue: it is scrutinized, regulated and unsympathetic to anything that sounds slightly wrong. Four things came out of running the experiment there rather than somewhere safer.

• Authorship matters more than delivery. The remarks are his; the model simply reads them. Organizations can scale communication across languages without changing who wrote the message or who remains accountable for it.
• Disclosure strengthened the result rather than undermining it. It was initially unclear whether announcing the AI voice would cost us credibility. The opposite happened, and the alternative — hoping nobody noticed — would have been wrong on both ethical and practical grounds.
• Multilingual delivery provided real value. Shareholders who would otherwise wait for a translation now hear the same remarks, in the same voice, at the same time. That is a change in who gets equal access to information, not a production efficiency.
• Trust rests on accountability, not polish. People accept an AI-delivered message when they know who authored it, who stands behind it and what recourse they have. None of that is audible in the voice itself.

This reflects a principle we use in LabX, DXC’s AI-native product incubator: before asking customers to trust an emerging technology, we should be willing to test it inside our own environment. Being Customer Zero exposes the issues a polished demo cannot — where people hesitate, where trust breaks down and what guardrails are needed to operate responsibly at scale. The hardest enterprise AI challenges are usually human rather than technical.

From Demonstration to Transformed Experiences

Our strategic partnership with ElevenLabs, the AI company specializing in audio models and voice agents, builds on that Customer Zero experience and applies voice across three practical areas. In each case the starting point is a defined outcome, not simply a technology deployment.

• In service desks, voice agents can help employees describe an issue naturally, retrieve relevant knowledge, complete routine steps and pass full context to a specialist when escalation is needed. The measure is not how many calls were deflected. It’s whether the person who called reached a resolution, in one interaction, without describing the problem three times.
• In customer experience platforms, natural-sounding multilingual interfaces create more consistent interactions across phone and digital channels. Quality in most service organizations varies by who answers and when. There is an opportunity for real improvement — every interaction can be informed by the same context, with the hardest ones still routed to the people best equipped to handle them.
• In training, voice turns passive content into practice. Employees can rehearse a customer conversation, a safety scenario or a difficult management discussion, then receive tailored coaching. People get competent faster practicing the actual conversation than reading about it.

Industry-specific applications may be the most consequential. Insurers can explore guided intake for routine claims; manufacturers can give field technicians hands-free access to procedures; public-sector organizations can make information easier to reach in multiple languages. These are not generic agents placed on top of an organization. They have to be designed around the vocabulary, rules, data and risk profile of each industry.

Scale the Qualities of Great Service, Not a Scripted Interaction

The risk, of course, is that scale becomes sameness. Enterprises have made that mistake before: a decision tree applied uniformly, satisfying the average case and failing everyone else. What should actually scale are the qualities of a great service interaction — informed, responsive, oriented toward an outcome — while the interaction itself adapts to the person’s language, context and need. Consistency should mean reliably good, not identical.

Trust Is Part of the Experience

The more human a voice agent sounds, the more important transparency becomes. A person should know when they are interacting with AI. Organizations need explicit permission to create or use someone’s synthetic voice, controls over who can access it, strong authentication for sensitive actions, records of what the agent did and a clear route to a real human.

These are often treated as a compliance issue bolted on after the experience is designed, but they must be treated as part of the experience. Someone who can’t tell whether they are speaking with a person, can’t tell what the agent is authorized to do, or can’t reach someone when they need to is not having a good experience, however natural the conversation sounds.

We should be precise about where that trust comes from, because it’s a two-way street. A model has to be calibrated properly by the people who built it. The deployment around it — data handling, access control, logged actions, defined points where a person takes over — is a separate problem belonging to whoever puts the system into production. Neither substitutes for the other. Both matter more in voice, because the interaction is live, the agent may be acting while it talks, and there is no screen showing the person what just happened on their behalf.

External guidance helps here. The Federal Trade Commission has highlighted the potential harms of malicious voice cloning, and the NIST AI Risk Management Framework offers a practical structure for identifying and managing AI risk. Treat consent, disclosure, security, testing and ongoing monitoring as core product requirements.

Building an Open Path From Experiment to Scale

My expanded role as President of AI Innovation and Strategy & LabX brings together two sides of the same challenge: incubating AI-native products and developing the commercial side of DXC’s AI partner platform strategy. LabX remains a place to test frontier technologies against real customer problems, but its mandate now includes helping customers navigate an open ecosystem of partners such as Anthropic and ElevenLabs.

That open approach matters because no single model, platform or interface will be right for every enterprise and every use case. Our job is to combine the right technology with domain expertise, workflow design, governance and delivery capability, then validate that it produces a measurable outcome. Sometimes the right interface will be text. Sometimes it will be visual. Increasingly it will be voice — and often a thoughtful combination of all three.

The question is not where an enterprise can add voice. It is where conversation can produce a better experience and a better outcome, and whether the organization is prepared to redesign the work around it.