This Startup Built AI That 80% of Callers Think Is Human
Nicolas Dessaigne (host), Will Bodewes (guest)
In this episode of YC Root Access, featuring Nicolas Dessaigne and Will Bodewes, This Startup Built AI That 80% of Callers Think Is Human explores phonely’s voice AI answers millions of calls, optimizes outcomes continuously Phonely provides AI phone agents for businesses and emphasizes continuous optimization toward measurable outcomes, not just conversational ability.
Phonely’s voice AI answers millions of calls, optimizes outcomes continuously
Phonely provides AI phone agents for businesses and emphasizes continuous optimization toward measurable outcomes, not just conversational ability.
The company handles millions of calls per month across many verticals, with strongest traction in call centers, insurance-adjacent workflows, and home services.
Phonely differentiates via production “battle scars,” analytics that recommend specific script changes, and an architecture using smaller specialized models to reduce latency and cost.
Bodewes traces the idea to his father’s need to answer phones, then describes iterating from SMB pricing/feedback loops into an enterprise call-center focus.
The conversation covers voice AI’s current realism (most callers don’t realize it’s AI), disclosure/ethics—especially for outbound—and remaining technical gaps like interruption handling and transcription edge cases.
Key Takeaways
Voice AI wins when it’s optimized for business outcomes, not demos.
Phonely frames success around metrics customers care about (e. ...
Early SMB customers can be a fast feedback engine even if they aren’t the end market.
Phonely started with low-cost SMB plans to tighten iteration cycles and learn what matters, then shifted focus after realizing one call-center customer could outspend the entire SMB base.
Modular, task-specific models can beat “one big model” for telephony constraints.
Instead of a single monolithic LLM, Phonely splits voice-agent functions (e. ...
Latency is becoming table stakes; accuracy and conversational handling are the frontier.
Bodewes argues latency is now “good enough” for most providers, while interruption handling, endpoint detection, and robust transcription in garbled phone audio remain key differentiators.
Real-world adoption is currently revenue-driven inbound calling more than pure support.
Many deployments focus on answering and qualifying inbound leads from ads/billboards where missed calls directly reduce revenue; handoffs still occur for regulated steps (e. ...
Disclosure norms may diverge between outbound and inbound use cases.
Bodewes supports disclosing AI—especially for outbound—and expects regulation, while also predicting consumers may prefer AI once it’s no longer associated with rigid “press 1/press 2” systems.
Fundraising and founder endurance are tightly linked to narrative and persistence.
A LinkedIn post about ultra-endurance cycling sparked the Base10 connection, reinforcing his broader view that founders ‘roll the dice’ daily and create luck through sustained effort.
Notable Quotes
“Phonely is a platform that allows you to answer your phone with AI.”
— Will Bodewes
“We do millions of calls every single month.”
— Will Bodewes
“We can statistically show them that changing one question can increase one of our customers' outcomes by 5%.”
— Will Bodewes
“For about 80% of our customers, they have no idea they're speaking with an AI agent.”
— Will Bodewes
“There’s a lot of people out there who think that they wanna be founders, and then there’s people out there who have no choice but to be founders.”
— Will Bodewes
Questions Answered in This Episode
What specific “outcomes” do Phonely customers optimize for most often (book rate, qualified leads, revenue per call, CSAT), and how are they measured end-to-end?
Phonely provides AI phone agents for businesses and emphasizes continuous optimization toward measurable outcomes, not just conversational ability.
Can you walk through a concrete example of the 5% lift: what question changed, what metric moved, and how you validated it statistically?
The company handles millions of calls per month across many verticals, with strongest traction in call centers, insurance-adjacent workflows, and home services.
How do you decide which parts of a call should be handled by specialized small models versus a larger general model?
Phonely differentiates via production “battle scars,” analytics that recommend specific script changes, and an architecture using smaller specialized models to reduce latency and cost.
What does your evaluation harness look like for voice agents—offline simulation, human review, live A/B tests, or all three?
Bodewes traces the idea to his father’s need to answer phones, then describes iterating from SMB pricing/feedback loops into an enterprise call-center focus.
You say latency is “oxygen” now—what latency thresholds actually change caller behavior (hang-ups, interruptions, perceived naturalness)?
The conversation covers voice AI’s current realism (most callers don’t realize it’s AI), disclosure/ethics—especially for outbound—and remaining technical gaps like interruption handling and transcription edge cases.
EVERY SPOKEN WORD
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome