Leaping AI Builds Voice Agents That Get Better While You Sleep

Every voice AI demo sounds amazing.

The agent greets the caller warmly, answers a question about a delayed order, and books a follow-up appointment without missing a beat. The buying committee is impressed. The pilot gets signed.

Then the agent meets real callers, and the gap between demo and production opens up. Someone asks about two orders in one sentence. Someone switches languages halfway through. Someone describes their problem in a way nobody anticipated, and the agent transfers the call to a human, which is exactly the outcome the company paid to avoid. So an engineer reads transcripts, adjusts prompts, redeploys, and waits to see whether this week's fix broke last week's. Most companies that buy a voice agent discover they've also bought a permanent tuning job.

Leaping AI attacks that second, less glamorous problem. The company, a Y Combinator W25 graduate, builds voice AI agents for call centers that monitor their own performance, rewrite their own prompts, and A/B test the changes against live traffic. Its agents handle more than 100,000 calls a month for customers like Eurowings and Hawesko, automating up to 70% of them end to end while keeping customer satisfaction above 90%. Investors noticed: Leaping AI raised a $4.7M seed round led by Nexus Venture Partners, with Paul Graham among the backers, a round the founders say came together in under a week.

The interesting claim has little to do with how human the agent sounds. Leaping AI is selling what happens after launch: an agent that gets good, and stays good, without a team of engineers babysitting it.

The maintenance problem nobody puts in the demo

Call centers are an enormous business built on repetitive conversations. Order status, appointment scheduling, cancellations, address changes, lead qualification: the same fifty conversations, thousands of times a day, staffed by people who burn out and churn at rates that make workforce planning a full-time discipline.

That repetition is why voice AI has become one of the most crowded categories in applied AI. The core technology finally works. Modern speech models handle interruptions, accents, and natural phrasing well enough that callers stop noticing they're talking to software.

But getting an agent to production quality on a messy, real use case is a different job from getting it to demo quality. Real call flows have edge cases that only surface at volume. A prompt tweak that fixes one failure mode quietly introduces another. And because the whole system is probabilistic, teams can't rely on the software instinct that a passing test today will pass tomorrow. The result is that many voice AI deployments plateau: the agent launches at 30 or 40% automation, and creeping improvements depend on how much engineering attention the vendor or the customer can spare.

Leaping AI's founders treat that plateau as the actual product problem. Their bet is that the winner in voice AI won't be whoever ships the most natural-sounding agent, but whoever closes the loop between what happened on yesterday's calls and how the agent behaves today.

Agents built as graphs, not one giant prompt

The foundation for that loop is architectural. On Leaping AI's platform, an agent isn't a single sprawling prompt. It's a multi-stage graph. Each stage handles one part of the conversation, described in plain English, with explicit transitions that define when the call moves from, say, identifying the caller to modifying their order to confirming the change.

The structure matters less for what it enables during the call than for what it enables afterward. When a conversation goes wrong somewhere inside a 20-turn call, a monolithic prompt gives you almost nothing to debug. A staged graph lets the platform evaluate each stage separately, so an error or a regression can be traced to the specific step that caused it. Testing gets tractable for the same reason: you can simulate traffic against a single stage and know what you're actually measuring.

Anyone who has debugged software recognizes the principle. You don't find bugs faster by staring harder at a big function; you find them faster by breaking the system into pieces that can be tested in isolation. Leaping AI applies that discipline to conversations.

The self-improvement loop

On top of that structure sits the company's signature feature, the reason its Launch HN post was titled simply "Self-Improving Voice AI."

After calls complete, the agent analyzes its own transcripts and looks for the places where it stumbled: the intents it misread, the moments callers got frustrated, the transfers to humans that a better response could have avoided. Then it proposes changes to its own stage prompts and runs them as A/B tests against live traffic. Changes that reduce transfer rates and keep satisfaction up survive. Changes that don't get rolled back. The founders describe watching the transfer rate to human agents fall day over day with no one touching the system.

"The AI improves itself" is a claim every vendor in the category gestures at, so the hard part deserves spelling out. Improvement requires knowing what better means, measuring it reliably, and changing the system without breaking what already works. That's an evals problem, and evals for audio conversations are much harder than for text: you're scoring tone, latency, interruptions, and task completion across thousands of unpredictable humans. The Leaping AI team is blunt that this is where they've concentrated their effort. Doing evals, simulation, and A/B testing exceptionally well is, in their view, the edge that lets them take on large and sensitive use cases that competitors can't safely touch.

The staged architecture and the self-improvement loop are really one design. Stage-level evaluation is what makes automated self-editing safe. An agent that rewrites one stage's prompt and measures the result is running a controlled experiment. An agent that rewrites its entire prompt is running a gamble.

What it looks like in production

The customer that best illustrates the model is Hawesko, Germany's largest wine merchant. Hawesko historically ran its customer service with more than 100 human agents. Today, Leaping AI answers 100% of incoming support calls for one of Hawesko's brands, over 1,000 calls a day, and resolves 70% of them with no human transfer at all. The remaining calls arrive at human agents pre-sorted, with context attached.

Eurowings, the Lufthansa Group's budget airline, uses Leaping AI for flight booking and cancellation requests. That detail deserves a pause: rebooking a flight is nobody's idea of a simple scripted call. It involves identity, fare rules, schedules, and payments, and it's exactly the category of "complex use case" that most voice AI deployments avoid in favor of store hours and order lookups. Thompson Creek Window Company, a US home services firm, rounds out a customer list that spans retail, travel, telecom, insurance, and real estate.

Across deployments, the company reports handling roughly 10,000 calls daily, hitting 70 to 80% automation on its best use cases, and holding customer satisfaction above 90%. The platform ships with the compliance coverage that enterprise call volume demands, including SOC 2, GDPR, and HIPAA, and supports multilingual calls, which matters when your flagship customers are German companies serving international callers.

Leaping AI aims squarely at businesses with real volume, roughly 100,000 calls a year and up. That focus follows from the thesis. A self-improving system needs calls to learn from, and the economics of automation compound with scale.

From Munich to San Francisco

The company's path ran through Germany before it ran through Silicon Valley. Kevin Wu studied computer science at TU Munich and spent three years at BCG before founding the company in 2023 with Arkadiy Telegin, whose background is in autonomous robotics across self-driving and space tech, and Shraey Bhatia. The team built its early customer base among German enterprises, which explains a customer list most seed-stage American startups would struggle to assemble: European companies are demanding buyers on data protection and process, and landing Hawesko and Eurowings early forced the product to grow up fast.

Acceptance into Y Combinator's Winter 2025 batch moved the team to San Francisco, and the combination of European enterprise traction and YC's network converted quickly into the $4.7M seed. When a round led by Nexus Venture Partners, with Paul Graham writing a personal check, closes in under a week, it usually says less about the pitch deck than about the revenue chart behind it.

The compounding argument

The voice AI market has a strange shape right now. The barrier to building a demo has collapsed; dozens of platforms let you wire speech recognition to a language model and get something that talks. That has convinced some observers the category will race to the bottom.

Leaping AI's counterargument is that the demo was never the product. The product is the operational system around the conversation: the evals that catch regressions, the simulations that test changes before callers experience them, and the improvement loop that turns every call into training signal. Those capabilities are unglamorous, hard to build, and nearly invisible in a demo, which is precisely what makes them defensible.

There's also a compounding logic to the position. A voice agent that improves itself gets more valuable the longer it runs and the more calls it handles, while a static agent depreciates as its edge cases accumulate. If that holds, the gap between self-improving deployments and hand-tuned ones won't stay constant. It will widen every week, on its own, which is a good property for a product and an uncomfortable one for the competition.

Call centers have spent decades as the textbook example of work that software should have automated long ago but couldn't, because the interface was human conversation. The conversation part is now solved. Leaping AI is betting the remaining race is about who improves fastest, and it built an agent that doesn't wait for its makers to improve it.