Voice AI's real revenue is in debt collection, not consumer apps

Debt collection is not the industry most observers would name when asked where voice AI is proving itself. Nathan Labenz thinks that is a mistake. He describes collections as a use case where voice AI is working “surprisingly well,” with AI agents making calls to people who are behind on debts delivering results that hold up in practice. These are not glamorous deployments, but they are the ones generating traction.

The behavioral case for debt collection is specific. Mati Staniszewski offers an explanation that fintech operators will recognize: shame is a real friction point in human-to-human collections calls, and AI removes it. “Frequently people would naturally feel ashamed of telling the real situation,” he says. “With AI people are much more open to share what actually happened.” That increased disclosure has direct operational value. A collections agent who gets an accurate picture of a debtor’s situation can negotiate more effectively than one who gets a defensive or incomplete account. The AI interaction changes what information surfaces, not just how fast the call gets completed.

The pattern is not limited to debt collection. Sam Parr, discussing the platform he is associated with, notes that it runs more than 70 percent voice by message volume. His comparison point is instructive: a platform like ChatGPT, he estimates, runs 90 to 95 percent text. The gap suggests that voice-first usage is not emerging uniformly across all AI deployment contexts. It is concentrating in specific products and verticals where voice is the natural or required interface, rather than spreading gradually across consumer AI as a whole.

The infrastructure supporting these B2B deployments is also shifting. Stefano Ermon, who works on diffusion-based language model technology at Inception, describes a customer called OpenCall that had been running its models on Cerebras custom chips to achieve the response speeds required for good call quality. That customer moved to Inception’s diffusion-based large language models and, as Ermon explains it, was able to get the same speed as running an autoregressive model on custom hardware, but on standard hardware instead. The economics of serving high-volume telephony workloads improve when the speed requirement no longer mandates expensive specialized chips.

What connects these observations is not a single technology or a single company. It is the match between voice AI’s current capabilities and the specific demands of B2B telephony: high call volume, structured interactions, known compliance requirements, and a user population that often responds better to an AI agent than to a human one. Collections callers are less guarded with an AI on the line. The use case fits the tool’s actual capability profile rather than the capability profile that consumer demos tend to showcase.

The consumer voice AI story remains mostly prospective. The B2B telephony story is operational. That gap between where the attention goes and where the revenue is concentrated is worth tracking, because it shapes which companies will have the data, the margins, and the customer relationships to define what voice AI infrastructure looks like at scale over the next several years.