Freya Thinks Your Bank's Hold Music Is a Product Failure

The Macro: The Call Center Problem Nobody Actually Fixed

I have called my bank six times this year. Each time I got a menu tree that felt like it was designed by someone who has never personally needed to call a bank. Press 1 for English. Press 3 for account services. Press 2 for existing accounts. Hold for 22 minutes. Get transferred. Explain everything again. This is not a technology problem anymore. This is a product problem.

The contact center AI market is growing fast. Multiple research firms project it will cross $10 billion within the next few years, driven by exactly the frustration I just described. Banks, insurers, mortgage lenders, and fintechs spend enormous sums staffing phone lines that their customers actively hate using. The economics are broken in an obvious way: the companies spending the most on customer service are often delivering the worst customer experience.

The first wave of “AI for call centers” gave us IVR systems that recognized maybe 40% of what you said and routed you to the wrong department the other 60%. The second wave gave us chatbots that could handle password resets and nothing else. Both waves failed at the same fundamental thing: having a real conversation about a real problem in real time.

What changed is that voice AI models got dramatically better in the last 18 months. Latency dropped. Comprehension improved. The ability to hold multi-turn conversations without losing context went from research demo to production-ready. The companies that matter in this space now are the ones building specifically for regulated industries where getting it wrong has consequences. Parloa is doing this in Europe. Cognigy has a presence. PolyAI targets restaurants and hospitality. But financial services, with its compliance burden and multilingual customer bases, is still wide open.

That is where Freya is aiming.

The Micro: Voice Agents That Actually Know What They Cannot Say

Tunga Bayrak and Tomas Nepala started Freya out of San Francisco as part of Y Combinator’s Summer 2025 batch. Bayrak is an AI research engineer with a background in audio transformers and vision-language models. He was a Math Olympiad competitor and studied at UPenn. Nepala comes from Wharton and previously worked at Gallagher, the second-largest insurance broker in the world. The team is five people.

That combination matters because enterprise voice AI is not just a modeling problem. It is a compliance problem, an integration problem, and a trust problem, all at once. Having someone on the founding team who has actually worked inside a large insurance operation changes what questions get asked during product development.

Freya builds voice AI agents for banks, mortgage lenders, fintechs, and insurers. The agents plug into existing CRMs, ticketing systems, and knowledge bases. They speak dozens of languages and dialects. They run 24/7. The company claims cost reductions of over 50% compared to traditional call centers.

The feature I find most interesting is compliance-aware reasoning. In financial services, there are things an agent literally cannot say. Regulatory disclosures that must be included. Phrases that trigger legal obligations. Most voice AI products treat compliance as a filter layer on top. Freya appears to be building it into the reasoning itself, so the agent understands policy manuals and regulatory documents as part of how it thinks about a conversation, not as a post-hoc check.

That is a meaningfully different architecture than what Bland AI or Retell offer for general-purpose voice agents. Those platforms are flexible but industry-agnostic. When you are handling someone’s mortgage application over the phone, agnostic is a liability.

The multilingual piece is the other differentiator worth watching. A bank in Miami needs English and Spanish at minimum. A mortgage lender in Toronto needs English and French. An insurer in New York needs everything. Most voice AI products handle multilingual support by running separate models for each language. Freya claims to handle it natively across languages and dialects, which, if it works well, is a real operational advantage for companies serving diverse customer bases.

I want to know the actual latency numbers. Voice conversations are unforgiving. A 500-millisecond delay feels natural. A two-second delay feels like the system crashed. And I want to know how the agents handle edge cases: the angry customer, the ambiguous question, the moment when the right answer is “let me transfer you to a human.” Getting the handoff right is where most voice AI products fall apart.

The Verdict

Freya is attacking a large, painful, underserved market with the right founding team composition. The compliance-first approach to financial services voice AI is smart because it creates a moat that horizontal voice platforms cannot easily cross. Parloa and Cognigy are the closest comparisons, and both are much further along, but neither has the same focus on the US financial services stack.

What I would watch at 30 days is whether any of their enterprise pilots are producing real call deflection numbers. Enterprise voice AI is full of impressive demos that collapse under production load. At 60 days, I want to know if the multilingual claims hold up across accents and dialects, not just languages. At 90 days, the question is whether they can close enterprise contracts fast enough to matter, because five-person teams do not survive long sales cycles without revenue.

The hold music industrial complex has been unchallenged for decades. I think Freya has the right thesis. Whether a five-person startup can sell into banks fast enough to prove it is the question I cannot answer yet.