By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

This Startup Built AI That 80% of Callers Think Is Human

In this episode of Founder Firesides, YC General Partner Nicolas Dessaigne talks to Will Bodewes, founder and CEO of Phonely (S24), which just raised a $16M Series A led by Base10 Partners.

Nicolas DessaignehostWill Bodewesguest

Apr 16, 202616mWatch on YouTube ↗

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Phonely’s Series A and the pitch in one line: “Answer your phone with AI”
Nicolas introduces Will Bodewes and the $16M Series A led by Base10. Will frames Phonely as a voice AI platform that answers calls, setting up the broader theme of measurable performance improvement rather than novelty demos.
- •$16M Series A announcement and context for the conversation
- •Phonely’s core promise: businesses can answer calls with AI
- •Positioning: practical, production-ready voice AI rather than experiments
Beyond answering calls: an optimization platform for better business outcomes
Will explains that Phonely’s value is continuous optimization of voice agents toward outcomes customers care about. The product emphasizes measurement, experimentation, and iterative improvement based on real call data.
- •Voice AI as the “tip of the iceberg”; optimization is the real product
- •Continuous improvement loop driven by outcomes (e.g., conversions)
- •Data and tooling surfaced to customers to make informed changes
Scale and breadth: millions of calls per month across many verticals
The discussion moves to usage volume and the range of industries Phonely serves. Will emphasizes that Phonely isn’t limited to a single niche and has learned to deploy voice agents broadly.
- •Processing millions of calls every month
- •Operating across hundreds of verticals, not one industry
- •Experience running voice AI in production at real scale
Who buys Phonely: call centers, insurance, and home services lead qualification
Will describes core customer profiles and why they care about performance optimization. Many customers are qualifying inbound leads and booking appointments, and want assurance the AI performs well and improves over time.
- •Primary customers: call centers, insurance-related workflows, home services
- •Common job-to-be-done: qualify leads and book appointments
- •Need to monitor, capture performance, and improve statistically over time
Differentiation in a crowded voice AI boom: battle scars + owning the model stack
Nicolas asks how Phonely stands out amid many voice AI companies. Will points to early work before “voice AI” was a category, plus deep production experience and building their own LLMs to enable optimization.
- •Started before “voice AI” became a mainstream category
- •Production learnings (“battle scars”) as a moat
- •Custom LLM work to focus on performance, not just conversation quality
Measurable iteration: using call data to change scripts and lift outcomes
Will gives a concrete example of data-driven optimization—changing a single question and seeing a 5% outcome lift. The broader idea is that voice conversations can be optimized like online checkout funnels once you have the right instrumentation.
- •Phonely identifies what to change (e.g., specific questions/prompts)
- •Statistical measurement ties changes to business results
- •Analogy: optimizing voice flows like ecommerce conversion funnels
Founder backstory: athlete mindset, a failed first startup, then an AI PhD in Australia
Will shares his unusual path: college athlete, COVID-canceled NCAA moment, then entrepreneurship and a pivot into AI research via a funded PhD in Australia. The thread is perseverance and an early conviction that AI would be transformational.
- •Cross-country skiing at a high collegiate level; COVID disrupted the peak moment
- •First company didn’t work out; reinforced drive to prove himself
- •Full-ride AI PhD in Australia and early voice experimentation
Origin story: solving his dad’s phone-answering problem when no software existed
Will describes the spark for Phonely: his dad’s growing practice struggled to handle calls. When Will couldn’t find existing software to reliably answer phones, he began building it himself.
- •Dad’s business pain: missing/handling phone calls at scale
- •Market gap at the time—no adequate AI phone-answering solution
- •Personal problem became a company idea
From SMB receptionist to enterprise motion: fast feedback, then a call-center pivot
Phonely began as an affordable small-business product to gather feedback and iterate quickly. After 4–5 months, landing a call center that paid more than all SMBs combined pushed the company to focus on larger customers.
- •Early SMB pricing used to accelerate learning and iteration
- •Short iteration cycles mattered more than immediate enterprise access
- •Call center customer economics drove a decisive pivot to enterprise focus
Why build custom models: latency, control, and a modular architecture
Nicolas probes the shift from off-the-shelf closed models to Phonely’s own models. Will explains their work with Groq for fast inference and their decision to use smaller specialized models rather than one large model to reduce latency and cost while maintaining quality.
- •Many competitors rely on OpenAI/closed models; Phonely built differently
- •Groq partnership and fast inference focus to tackle latency
- •Modular “small models for tasks” approach improves speed, cost, and maintainability
How the modular system works: separating tasks like variable capture and context handling
Will explains that model switching is less about the question type and more about pipeline components. By isolating functions (e.g., capturing customer details), Phonely can update and optimize parts of the system independently.
- •Decompose voice AI into components (e.g., variable storage/capture)
- •Different models handle different tasks instead of one monolith
- •Isolation enables targeted updates and operational improvements
State of the art: latency is “good enough,” now quality and accuracy matter most
They assess whether latency still constrains phone agents. Will argues latency is largely solved across the industry, and the frontier is conversational quality plus transcription/accuracy in messy real-world audio.
- •Latency no longer feels like the primary bottleneck
- •Remaining gaps: conversational quality and accuracy
- •Real phone audio edge cases make transcription and understanding hard
Human-likeness and disclosure: 80% don’t realize it’s AI, and ethics differ for outbound
Will claims most callers don’t realize they’re talking to AI today, and expects that to approach ubiquity soon. The conversation turns to disclosure norms—Will favors disclosure for outbound calls and anticipates regulation, while also noting users may grow to prefer AI interactions.
- •Claim: ~80% of callers don’t detect the AI; expectation of further improvement
- •Outbound calling should disclose; regulation likely
- •Consumers may prefer AI for context, speed, and reduced social friction
Where voice AI delivers ROI first: inbound, revenue-critical calls and lead qualification
Will describes adoption focusing on revenue-driving phone calls rather than classic support. Inbound calls from ads and billboards create high stakes; AI helps sift leads, capture the best opportunities, and either schedule directly or hand off when required (e.g., licensed roles).
- •Early adoption is revenue-focused more than support-focused
- •Inbound calls dominate (ads/billboards → phone number → immediate response)
- •Handoff vs full automation depends on compliance and workflow (e.g., insurance licensing)
Series A story: ultra-endurance LinkedIn post leads to Base10 preemptive offer
Will explains how Base10 found them through a personal story about ultra-endurance cycling and founder mindset. After conversations, Base10 moved quickly with a preemptive Series A offer, and Will prioritized partner fit over running a broad process.
- •Ultra-endurance athletics background and public sharing on LinkedIn
- •Base10 outreach sparked relationship; culminated in preemptive offer
- •Choosing investors based on trust and long-term working relationship
What’s next: technical improvements, defensibility vs generic models, and scaling to 50M+ calls
Will outlines remaining technical work (interruptions, endpointing, transcription edge cases) and argues telephony complexity creates defensibility beyond generic models. He shares an ambitious scaling goal and closes with hiring plans and founder lessons about the daily grind and who should start companies.
- •Near-term roadmap: interruption handling, endpoint detection, transcription robustness
- •Moat thesis: deep telephony + execution makes “generic models” less threatening
- •Growth goals: from millions to 50M+ calls/month; hiring sales and engineers in SF
- •Founder lessons: success feels like constant battle; advice on founder fit and persistence

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.