An AI assistant that speaks naturally can still forget the customer, lose the thread of a complaint or misunderstand its authority to act. For enterprises deploying conversational AI, those failures have consequences that a more convincing voice cannot conceal.
Twilio’s 2026 customer insights series gives that problem a commercial dimension. Seven in 10 APAC consumer respondents reported abandoning an AI customer service interaction because it lacked basic information about them or their previous interactions. Five per cent said they ended their relationship with the brand as a result.
In an interview with Deeptech Times, Seth Familian, Twilio’s regional vice president of enterprise strategy, outlined a proposition that goes beyond replacing a human voice with a synthetic one. The harder work lies in connecting interactions to reliable customer data, retrieving useful context quickly and establishing permission to act.
Voice will not win every interaction
“Phone numbers don’t go away in an AI world. The voice interface doesn’t go away,” Familian says.
His argument is that conversational AI reinforces the need for communications infrastructure. Customers who become comfortable talking to an AI assistant may expect comparable ease when dealing with a business.
Nonetheless, Familian stops short of predicting that voice will dominate APAC customer engagement. Channel preference depends on circumstances: a noisy room, sensitive information or an address that is easier to type.
“I don’t think one channel is going to be predominant over another,” he says.
That qualification matters. An enterprise can invest in a capable voice agent while leaving customers frustrated by the transition between messaging, automated calls and human support. The business problem is continuity across those channels.
Twilio’s survey suggests that continuity is often missing. While 84 per cent of APAC brands believe their agents effectively recognise returning customers and their history, 70 per cent of consumer respondents say they often start from scratch in a new conversation. 65 per cent also report repeating themselves or filling information gaps after a handoff from AI to a human.
These findings cover AI customer service broadly, rather than voice alone. The survey included 7,652 consumers and 660 business leaders globally across 18 markets in April and May 2026. APAC coverage included Singapore and Malaysia. The release does not provide regional subgroup sizes.
IMAGE: Twilio
The competition moves into customer memory
Asked whether orchestration is becoming the next enterprise technology battleground, Familian distinguishes routing a message from knowing enough about the customer to route it intelligently.
The test, he says, includes whether a business has “the memory of what a customer did before”.
In his assessment, orchestration depends on the quality of the information behind it: who the customer is, what happened previously and what context should influence the next response. Moving a message between channels does not, on its own, supply that understanding.
“You have to have reliable identity resolved profiles for individuals built on trusted data,” he says. That data must also “represent business value, not just vanity metrics”.
For enterprises, this raises an uncomfortable possibility. A sophisticated conversational interface may simply expose unresolved weaknesses in customer records and organisational handoffs.
Buying a better model does not demonstrate that the right history will reach the right interaction.
Speed and cost introduce further constraints. Familian argues that profile information needs to be retrieved in tens of milliseconds for responsive exchanges. Feeding entire conversation histories into an LLM may preserve detail but can make each interaction slower and more expensive.
“If you hand that over to an LLM, you’re just gonna burn tokens,” he says. “It’s going to take both money and time to get the response you want.”
Twilio’s Conversation Memory extracts concise observations and summaries for subsequent retrieval. Familian presents that compression as a potential business differentiator: giving an agent enough context without repeatedly processing every earlier word.
The economics remain dependent on accuracy. An incorrect summary could become a persistent assumption; a mistaken identity match could attach one person’s history to another.
Twilio’s documentation warns that generated profiles can contain mistakes and should be reviewed. It also says Conversation Memory is not intended for sensitive personal information.
The question for enterprise buyers is whether less expensive memory remains sufficiently complete and correct for the decisions it supports.
Customer agents require a different kind of trust
The next complication is that the participant contacting a company may itself be an AI agent representing a customer.
Familian asks: “Does that thing that represents the person have permission to speak on behalf of that person?”
That changes the security problem. Recognising a customer or identifying suspicious traffic is not equivalent to establishing a representative’s authority. A legitimate agent still needs limits on the actions it can perform.
Twilio’s acquisition of identity platform Stytch, completed in November 2025, supports this strategic direction. Familian discusses device fingerprinting and ongoing signals of trustworthiness, alongside authentication and scoped permissions. Those signals can inform a risk decision, but do not alone prove consent to a particular transaction.
The study reveals a further gap in transparency: 70 per cent of APAC consumers believe agents should identify themselves at the start of every conversation, while only 22 per cent of businesses report doing so. Remembering customers does not settle whether customers understand how AI is participating.
Twilio’s proposition therefore reaches beyond carrying conversations. It aims to connect communications, customer memory and identity as agents enter business workflows.
Enterprise buyers should measure completed resolutions, repeat contacts, incorrect actions and safe escalation, alongside costs. A system that remembers customers but exceeds their instructions has not solved the trust problem. The business case becomes stronger when continuity and bounded authority can both be demonstrated in the interactions that matter.
