How voice AI is transforming customer service

Modularity over black boxes

Instead of relying on monolithic “black box” solutions for voice AI, Twilio works with customers to implement a highly modular, decoupled architecture that integrates AI with existing back-end systems.

Implementation typically follows a “pyramid” of complexity. Organisations begin with high-volume, highly deterministic workflows, such as triage processes for flight cancellations. Once these are refined, businesses work their way up to more complex conversational tasks that touch multiple back-end systems for actions such as returns or shipping.

A key aspect of Twilio’s implementation playbook is model neutrality. The company handles the complex communications layer – from speech recognition and generative AI voices to interruption handling – while allowing customers to choose and train their own AI models to interact with customer relationship management (CRM) systems and retail databases.

“We’re language-model agnostic, so our customers can plug and play various language models, whether they are small models, large models, locally deployed models for privacy concerns, or cloud-based models,” O’Dower said.

Voice AI agents can also be trained to resolve a customer’s issue over the phone and then extend the conversation to another channel – such as SMS, email or WhatsApp – to achieve a certain outcome. At the same time, the conversation can be enriched with contextual data, so the voice AI agent is aware of the customer’s loyalty status, preferences and any issues they’ve had in the past.

Take Philippine Airlines (PAL), for example. Following its 2021 restructuring, the flag carrier deployed the Twilio Flex contact centre platform to get a unified view of the passenger journey across voice, chat and online platforms, and to use AI to handle routine customer service tasks, such as flight status checks.

Since the deployment, PAL’s average contact centre wait time has plummeted to under a minute, while monthly customer service costs have fallen by around 30%. The airline is also aiming to reach what it calls a “super AI agent state” by April 2026, where 80% of tasks handled by live agents can be fully automated.

However, delivering similar experiences does not come easily for many enterprises, which often grapple with data silos and operate across diverse markets.

“The complexity of the fragmentation of Asia is one of the hardest things to figure out,” Woolfrey said, noting that a model that works in Thailand doesn’t necessarily work in Japan. Traditional industries, such as financial services, are also constrained by legacy infrastructure and data silos.

Furthermore, as voice AI becomes indistinguishable from human speech, the technology could raise concerns around deepfakes and scams. To prevent malicious use, Twilio implements infrastructure-level protections, such as know-your-customer (KYC) vetting, deepfake detection, and traffic monitoring and intelligence.

“We are a trusted platform, and that means that any customer on our platform needs to be trusted before they can send any type of outbound communication to any consumer,” O’Dower said, adding that Twilio also works with telcos, device manufacturers and messaging services to ensure voice AI calls are branded and trusted.

Achieving ROI

Historically, the business case for contact centre automation relied heavily on reducing human headcount. While the return on investment (ROI) of AI agents remains highest in expensive labour markets such as Australia or Singapore, Woolfrey noted that organisations in lower-cost markets are also pursuing the technology.

The reason lies in the inverted economics of voice AI. As O’Dower pointed out, an airline dealing with mass flight cancellations can instantly deploy the technology to handle 50,000 concurrent callers. This eliminates the need to overstaff for seasonal spikes or crises while preserving a high level of customer satisfaction.

Crucially, AI is enabling organisations to transform inbound support operations into revenue generators, Woolfrey argued, citing PAL as an example.

“Traditionally, we see the contact centre as a cost centre, but today, it’s somewhere where we can offer upsells,” he said, adding that PAL can identify passengers who haven’t purchased check-in baggage and trigger an outbound message, reminding them that it’s cheaper to buy an allowance now rather than at the airport. “We’re moving from a contact centre as a cost centre to a retention engine.”