Leaping AI — Investment Memo · Orange Collective

Thesis

Voice AI is the largest labor-replacement opportunity software has ever pointed at — and the call center is the wedge. Most products demo the easy 30% (status checks, simple routing, FAQ deflection) and quietly bail on the complex 70% (claims, billing disputes, retention saves, multi-step workflows with CRM writes) where the labor savings actually compound.^[1] ^[2] Leaping is purpose-built for the hard calls — the integration-deep, workflow-heavy contact-center work the incumbents skip.

01
Voice is the deepest labor pool software has ever touched. ~2.85M customer service representatives in the US alone — a ~$100B/year wage bill at ~$40k fully loaded.^[11]Globally the customer service workforce is on the order of 17M. Every minute of automated talk-time is a minute of headcount returned to the P&L. The deflationary force here is more tangible than any other AI category.
02
The easy 30% has been won. The hard 70% is open. FAQ deflection and IVR replacement is a commodity category — Replicant, PolyAI, and the long tail of CCaaS retrofits have it covered.^[10] ^[13] The complex 70% — multi-turn workflows, escalations, claims, scheduling against inventory, retention saves with subscription writes — is where the labor savings actually compound. Leaping has been shipping into that 70% from day one. Hawesko isn't a wine FAQ; it's purchase calls. ACA pre-qualification isn't a survey; it's a regulated multi-step intake.^[2]
03
Enterprise integration is the moat. The model layer commoditizes. Whisper + GPT under the hood is a commodity input. The differentiation is how the agent operates inside the customer's stack — CRMs, billing systems, ticketing, scheduling, the back-office workflows incumbents skip because they don't scale to a generic API.^[4] That's the unsexy work that wins enterprise.
04
Self-improvement compounds the unit economics. Leaping's autonomous post-call analysis and A/B testing turns every deployment into a flywheel: more calls → better prompts → higher containment → more contracted volume. The labor savings widen with scale instead of plateauing.^[3] Founder Kevin Wu spent the company's first months hand-tuning prompts for customers; the product is what happens when that workflow is automated and turned on every customer at once.

Problem

The call center is the largest unautomated labor pool in software. Most voice AI products demo the easy 30% and skip the part that matters.

Running a contact center with humans is one of the most expensive line items in any consumer-facing business. ~2.85M customer service representatives in the US sit at ~$40k/year fully loaded — a $100B+ annual US wage bill that's structurally exposed the second voice agents cross the quality bar.^[11] Globally the number is closer to 17M workers.

The existing voice AI category handles the easy 30% — order status, store hours, password resets, simple routing. That's the IVR-replacement use case, and it's already crowded. The complex 70% — exceptions, escalations, claims, billing disputes, retention saves, multi-step workflows that touch the CRM and the billing system — is where the headcount actually lives, and it's where the incumbents have been the slowest to ship.

Leaping picked the hard end of the distribution as its wedge. Hawesko (Germany's largest wine merchant) runs >1,000 calls per day with 70% autonomously handled in wine purchasing.^[2] A US-bound Jamaican BPO running ACA pre-qualification processes >5,000 calls per day with one minute saved per call.^[2] Thompson Creek Window Homes — a $100M+ home services business — runs Leaping for AI-driven appointment scheduling.^[1] These aren't FAQ bots. They're full call flows that move money and time.

~17M

Customer service workers globally

The deepest labor pool software has ever pointed at

~2.85M

US customer service reps

~$100B annual US wage bill at ~$40k fully loaded

80%

Contact centers using some AI

76% plan to boost AI investment in next two years

BLS Occupational Employment^[11] · Deloitte 2024 Contact Center Survey^[6]

Why Now

Three preconditions converged in the same eighteen months.

LLM costs collapsed. TTS crossed the uncanny line. Buyers stopped saying no. Voice AI moved from experiment to budget line item in the same year.

Conversational AI is the most significant change to the way companies engage with their customers since the iPhone.

Bret Taylor^[9]

Co-CEO · Sierra · ex-Co-CEO Salesforce

Generative AI is going to reinvent virtually every customer experience we know, and enable altogether new ones we've only dreamed of.

Andy Jassy^[7]

CEO · Amazon

We measured Leaping against three other vendors on the same call distribution. They were the only one that didn't crater on the hard 30%. That's the whole game for us.

RC

OC reference call^[12]

Enterprise customer · Home services

The category just crossed the chasm.

LLM inference costs dropped 60–85% across 2023–24. For the first time, sub-second voice latency at gross margins that work for a SaaS business is structurally possible. The 2022 voice AI category was racing a cost curve; the 2025 category is on the cost curve's good side.

TTS crossed the uncanny line. Customers no longer hang up when they hear an AI voice. 61% of new buyers prefer faster AI responses to waiting for a human.^[6] The buyer-side resistance that capped category penetration for a decade is gone.

Customer expectations forced enterprise adoption. 80% of contact centers already use some form of AI; 76% plan to boost AI investments in the next two years.^[6] AWS, Salesforce, and the platform giants are now treating customer experience reinvention as the keynote of every earnings call.^[7] The category has moved from experiment to multi-year contract motion.

Conversational AI is the most significant change to the way companies engage with their customers since the iPhone.

— Bret Taylor, Co-CEO of Sierra and ex-Co-CEO of Salesforce^[9]

How It Works

Self-improving voice agents that handle full call flows, not just FAQs.

The flywheel is where the moat lives.

More calls → better prompts. Every customer interaction is structured training data. The autonomous A/B test loop chooses the better prompt variant on outcome metrics — containment, CSAT, transfer rate, resolution latency — without engineering involvement.^[3]

Better prompts → higher containment. Hawesko's deployment started at basic call handling and progressively improved to 70% autonomous purchase-call completion while maintaining 90% CSAT. That happened without the customer's team manually reviewing or updating prompts.^[5]

Higher containment → more contracted volume. Every percentage point of containment is a budget unlock for the customer. Leaping captures that unlock as expanded call volume — the textbook land-and-expand motion for voice AI. NRR scales with talk-time, not seats, which is the right shape for the category.

The Complex 70%

FAQ deflection is a commodity. Multi-step workflows that touch the CRM are not.

The category dividing line is whether the agent can read and write to the system of record. That's where the labor savings actually compound — and that's where Leaping has been building from day one.

Where the work actually lives. The complex 70% is integration-deep by definition.

Claims status that requires a policy lookup. Not a Q&A — a multi-step flow that reads the policy system, applies the deductible rules, and reports back in natural language. The integration is the product; the LLM is plumbing.

Appointment scheduling against real inventory. Thompson Creek doesn't need a calendar bot — it needs an agent that respects technician availability, parts inventory, and zip-code routing rules. Leaping ships into the integration layer of that business, not the front-end of it.^[1]

Retention saves that update the subscription. The agent applies the offer, writes the change to billing, and confirms with the customer. That's an end-to-end workflow that incumbents historically required a Tier-2 human agent to complete.

Warm transfers that carry context. When the agent does need to escalate, it hands the human seat the full call history, the customer's intent, and the recommended next action — no "tell me your story from the beginning."^[12]

Market

The voice category sits on top of the largest labor pool in software.

US customer service representatives: ~2.85M, fully loaded at ~$40k/year — a ~$100B annual US wage bill exposed to deflation.^[11] Globally the customer service workforce is on the order of 17M. The contact-center software market is already ~$48B and growing at 23%+ CAGR through 2032 — voice AI is the highest-growth wedge inside that envelope.^[8]

Adoption is past the experiment phase. 80% of contact centers already use some form of AI; 76% plan to boost AI investment in the next two years.^[6] The buyer side is not the bottleneck — the bottleneck is whether the product can actually do the complex 70% reliably. That's what Leaping is selling.

Competitive landscape

Four categories of competition. Leaping is positioned against all of them.

Each category has a structural limitation — sales motion, source segment, or call-distribution coverage. Leaping's complex-70% + autonomous-improvement + mid-market motion is the answer to all four.

The category-defining question for voice AI in 2025 isn't whether the agent sounds human. It's whether the agent can complete a multi-step workflow inside the customer's stack. Sierra answered it for the Fortune 100. Leaping is answering it for everyone else.

— Orange Collective

Founder deep dive

A consulting-trained operator, a real-time-systems engineer, and an NLP PhD walked into a call center.

Why Kevin built it. Three years at BCG Berlin taught Kevin Wu how mid-market and enterprise operations actually run — including how much money sits on the floor of a contact center. When Leaping launched, he spent the first months personally hand-tuning prompts for paying customers. He noticed every prompt improvement was a labor-savings unit; he automated the workflow that generated those improvements. The product is what happens when one founder's manual workflow becomes the entire customer base's default.

Why Arkadiy built it. Arkadiy Telegin came from autonomous robotics — self-driving cars and space-tech software where real-time orchestration is the entire job. Voice latency is the same problem in a different costume: sub-second response budgets, stateful conversation context, and a hostile real world that won't wait for a retry. He brings the systems depth that voice infrastructure demands.

Why Shraey built it. Shraey Bhatia holds a PhD in NLP / LLMs from the University of Melbourne and time at IBM Research. The academic depth anchors the autonomous prompt-optimization layer — turning what would be ad-hoc heuristics into a principled, measurable optimization loop.

Why this team is the right team. A consulting-trained operator who has lived enterprise procurement. A real-time-systems engineer who has shipped autonomous software. An NLP PhD who can build the optimization layer. Voice AI is a stack problem — telephony, real-time orchestration, model selection, workflow execution, integration depth — and Leaping has full-stack coverage from founders alone.

Why velocity is a feature. $500K → $930K ARR in eight weeks. Tens of thousands of calls per day in production. 90% CSAT across heterogeneous industries. The execution density matches the size of the opportunity — and reference calls came back with the same observation: the founders ship faster than the customer's internal team.^[12]

The long arc. Leaping becomes the platform that runs the complex 70% of contact-center work for thousands of mid-market businesses. The structured-transcript-plus-outcome dataset compounds into the operational system of record for voice. The incumbents that own the easy 30% become rails; Leaping owns the workflow layer where the labor savings actually live.

Founder & team

Kevin Wu

Co-founder & CEO

Three years as an Associate at BCG before founding Leaping. Studied Computer Science at TU Munich and at the Center for Digital Technology & Management (CDTM). Spent the first months of Leaping hand-optimizing prompts for paying customers — then turned that workflow into the company's autonomous prompt-improvement system.

Arkadiy Telegin

Co-founder & CTO

Background in autonomous robotics — self-driving cars, space-tech, consumer applications. Led the software division of Germany's largest space-tech student organization. Brings the real-time systems and orchestration depth that voice latency demands.

S

Shraey Bhatia

Founding AI Lead

PhD in Natural Language Processing from the University of Melbourne. Previously at IBM Research. Brings the academic NLP depth that anchors the autonomous prompt-optimization layer.

Risks & mitigations

Risk

Model commoditization — Whisper + GPT under the hood is a commodity. Why doesn't this collapse to a wrapper category?

Mitigation

The model layer is commoditizing, which is the bull case for the workflow layer. The differentiation is the enterprise integration depth — CRMs, billing systems, ticketing, knowledge bases — and the autonomous prompt-improvement loop that compounds with deployment volume. Leaping's wedge is exactly the work that doesn't generalize from a single API call. Customer reference calls confirm this is also why they picked Leaping over three other vendors on the same RFP.

Risk

Premium incumbents — Sierra ($4.5B), PolyAI, and Replicant have raised hundreds of millions and have years of enterprise depth.

Mitigation

The incumbents have segmented up — Sierra is sales-led at the Fortune 500 enterprise tier; Replicant and PolyAI optimize for the easy 30% IVR-replacement category. Leaping is winning the mid-market and high-volume BPO segment by handling the complex 70% the incumbents skip and pricing for technical buyers who don't want a six-figure procurement cycle. That gap will narrow but the head start in the workflow layer is real.

Risk

Regulatory exposure — FCC artificial-voice disclosure rules, state-level laws (CA, UT), GDPR/CCPA on call recordings, and emerging bias liability for voice AI.

Mitigation

These rules are tailwinds for compliance-ready platforms and headwinds for fly-by-night voice bots. Leaping is building the disclosure, consent, recording-retention, and bias-testing primitives into the platform from the start. The same regulation that creates friction creates a moat against the long tail of unserious vendors.

Risk

Generalization across verticals — does the self-improving system actually transfer from wine purchasing to insurance pre-qualification to home services scheduling?

Mitigation

The data so far says yes — Leaping is live across retail, travel, insurance, and home services with 90% CSAT across deployments. The architecture is industry-agnostic; the prompt optimization is per-deployment. Risk mitigation is doubling down on vertical playbooks (claims, scheduling, retention) where the workflow primitives transfer across customers in the same category.

What we're watching

ARR pace into Series A — the $500K → $930K → $1M+ trajectory at Demo Day is the proof point. Sustaining that velocity post-batch is the read on the wedge.
Net revenue retention by quarter — voice AI deployments expand by call volume, not seats. NRR > 150% would confirm the land-and-expand thesis holds.
First seven-figure logo — moving from mid-market BPO and home services into a named Fortune 1000 enterprise account.
The self-improvement curve — does autonomous prompt optimization compound containment rates 5%+ per quarter without human tuning? That's the bull case turning into a moat.

References