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

Bodewes — This Startup Built AI That 80% of Callers Think Is Human (YC Root Access)

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.

Phonely is an AI-powered voice platform that answers phone calls for businesses — handling millions of calls a month across hundreds of verticals using custom LLMs that statistically optimize outcomes over time. They discuss how watching his dad struggle to answer phones sparked the idea, going from small business customers to enterprise call centers, and why by the end of this year most callers won’t know they’re talking to AI.

— channel description, YC Root Access

TL;DR

A ~16:17 YC Root Access Founder Firesides episode (published 2026-04-16), interviewer Nicolas Dessaigne (YC General Partner), founder Will Bodewes of Phonely (YC S24). The conversation is a vertical-AI-vendor founder fireside on:

1. The “80% don’t know it’s an AI” empirical claim. “For about 80% of our customers, they have no idea they’re speaking with an AI agent. By the end of this year, I would say it’s probably going to be close to 100% of people won’t know.” (~8:23.) Note Bodewes corrects the interviewer’s framing of most to people who answer the AI — the share of AI-handled calls overall is a separate (larger but unstated) number.
2. Multi-model modular architecture as a build-side differentiator. Most voice AI products run on OpenAI or other large general-purpose models under the hood (~6:08–7:54); Phonely built its own LLMs and routes different conversational components (variable storage, name capture, etc.) to different smaller models running on Groq’s fast inference chips. “Based on our architecture, it just made more sense to have smaller models doing different tasks rather than one big model doing everything… you save cost, reduce latency, and still get the same quality.”
3. Optimization-as-the-real-product. “What we do is we allow businesses to optimize their voice AI agents so that they can continuously get better on the outcomes that the customers care about.” The product surface is not “AI answers the phone” — it’s a statistical-optimization platform where Phonely surfaces back to the customer that, e.g., changing one specific question increased outcomes by 5% (~2:44–3:11). “Nobody’s realized that you can optimize that checkout conversion for an outcome, right?”
4. Inbound-revenue-driven adoption, not customer-support cost-cutting. “We work a lot in businesses trying to make more money. That’s where our adoption is right now is it’s not as much of the customer support and customer service time. It’s usually like businesses they have people that are calling in like every single billboard out there that has a phone number on it — those leads need to be sifted through.” (~10:10–10:47.) Mostly inbound (call-center / insurance / home services). Insurance and other licensed-professional verticals require human-handoff to a licensed agent; appointment-scheduling can be AI-only.
5. Disclosure ethics & regulation forecast. “For outbound calling, yes — I feel like there is going to be some regulations around that. I genuinely feel like the way that the world should be is you should disclose that this is an AI.” (~9:01.) Bodewes predicts consumers will come to prefer AI calls for psychological safety (no judgment when asking a “dumb question” about finance).
6. The Series A story. $16M Series A led by Base10 Partners (the ASR transcript repeatedly mis-renders this as “Bessemer” and “Best 10” — the channel description is the authoritative source). Lead investor Caroline at Base10 reached out after Bodewes posted on LinkedIn about an ultra-endurance cycling race; “a few months later we had a few chats and they decided to preemptively offer us a Series A.”
7. Founder origin story. Cross-country skier on Stanford-class NCAA track; race cancelled by COVID, leaving a “chip on the shoulder”; first company didn’t work out; took a full-ride AI PhD in Australia (~Melbourne); voice AI emerged from his father’s small practice — “if somebody could answer my phones, that would be amazing.”

Caveats. Marketing-channel founder fireside on YC’s own channel; every empirical claim is unaudited and presented through obvious commercial incentives. The “80% don’t know” number is a self-report with no methodology disclosed; “50M+ calls/month” (up from current 1M/month) is an aspirational two-year goal. ASR mis-transcribes Base10 as Bessemer throughout — the wiki uses the description’s authoritative attribution.

Why this matters in the corpus

The wiki holds a growing cluster of vertical-AI-vendor founder-vantage anchors — companies that built deep into one customer workflow and now report enterprise-scale traction:

• Emergent 2026 — natural-language-to-app for SMB / domain-expert; $100M ARR in 8 months.
• AnswerThis 2026 — academic research workflow agent; $2M ARR with 2 FTEs.
• Campfire 2026 — AI-native ERP for tech companies; doubled ARR every quarter since Q4 2024.
• CS153 2026 — names Salient (voice agents for loan servicing) as the canonical YC forward-deployed-engineer worked example.
• Bodewes / Phonely 2026 (this source) — voice agents for call centers / insurance / home services; millions of calls / month; $16M Series A.

Phonely is the wiki’s first ingest of a voice-AI-as-vertical-product founder vantage — the surface-form-adjacent companion to Hu’s Salient example (Salient does loan-servicing voice; Phonely does multi-vertical voice with a horizontal-platform-but-vertical-customer-base shape). The conversation also lands the multi-model routing architectural principle at the production-voice-AI altitude, complementing agent-harness’s software-side multi-model-allocation thesis with hardware-side fast-inference routing on Groq.

What was actually ingested

The full ~16:17 transcript was read end-to-end. The transcript is auto-generated and ASR-cleaned; speaker labels (>> markers) are inconsistent in the raw VTT and were not corrected. The Base10 → Bessemer ASR mis-transcription was caught and corrected against the channel description.

Topics not covered in this fireside (worth tracking if Phonely surfaces again): unit economics per call (pricing tier discussion was generic “30, 50, 100 bucks a month” for small business early days but enterprise pricing not disclosed); model-training pipeline specifics; how Phonely competes with the broader voice-AI category (Bodewes claims “battle scars about how to get voice AI agents up and running in production” but doesn’t name specific competitors); the founder’s PhD topic specifically.

Linked entities and concepts

Entities promoted by this source:

• Y Combinator — channel; already entity. Bumps source-count.
• Groq — fast inference chip provider; named as Phonely’s inference hardware partner “from the early days… mostly for latency.” First wiki mention; currently Dangling; promote on second-source mention.

Dangling — single-source mention, deferred:

• Phonely — YC S24; voice AI platform; this video is Bodewes’s wiki debut. Promote on second substantive source.
• Will Bodewes — founder/CEO of Phonely; Stanford-class NCAA cross-country skier; AI PhD in Australia (~Melbourne). First wiki mention.
• Nicolas Dessaigne — YC General Partner conducting the interview. First wiki mention; the Founder Firesides interviewer line-up at YC is plausibly multi-source over time.
• Base10 Partners — lead investor on Phonely’s Series A. First wiki mention.
• Caroline (Base10 Partners) — investor who initiated outreach. Surname not stated.

Concept pages touched:

• agent-harness — adds the multi-model-modular-routing on fast inference hardware architectural pattern at the production-voice-AI altitude (smaller specialised models on Groq, one per task component, vs one big general-purpose model). Companion to Tan’s ADHD-CEO-vs-autistic-CTO model-routing pattern at the coding-agent altitude.
• ai-employment-effects — Bodewes’s “by end of year 100% won’t know it’s an AI” claim is a Voice-AI-specific datapoint on AI-substitution in call-center work. Bodewes carefully distinguishes the share-of-calls-answered-by-AI (unstated) from the share-of-AI-answered-calls-where-the-caller-knows (80%, projected 100%) — a useful framing in itself for downstream employment-effects discussion.

Source quality

• Channel: YC Root Access — founder-marketing channel; Founder Firesides series.
• Format: ~16-minute conversational interview, single-camera, founder-led, no slides.
• Empirical anchors: $16M Series A from Base10 (verifiable from press release); ~millions of calls/month (founder self-report); 80%-don’t-know-it’s-AI (founder self-report, no methodology).
• Bias / motive: Phonely is announcing the Series A on this episode; YC is the accelerator and the channel; treat every numerical claim as motivated-but-falsifiable.
• Transcript provenance: yt-dlp VTT fallback after the wiki’s youtube-transcript-skill’s Playwright path failed to render the engagement panel within 90 seconds — the second of four YC Root Access videos in this ingest batch to require fallback. Rolling-caption duplicates de-duplicated by the build script’s suffix-overlap collapse pass.