Voice Agent Fails — Channel Concept (parked, in detail)

1. The one-liner

Kitchen Nightmares for business phone lines. A real engineer-operator builds the world's worst AI receptionist on purpose, lets you cringe at it, then rebuilds it organ by organ — and every episode ends with a phone number you can call to hear the real thing done right.

2. Why this lane is empty

• AI-fail compilation channels = entertainment, no expertise, no fix. Vendor channels = polished demos, no honesty about failure modes, no trust.
• Nobody occupies "the voice agent mechanic": shows the broken engine AND rebuilds it on camera, founder-led, with a live product behind it.
• The fear this content answers is the #1 sales objection for the entire category: "AI will embarrass my business on the phone." Showing failure modes — and naming them before prospects experience them — converts fear into trust. Same trust mechanic as Quill on SermonPrepDesk: dramatize the guardrail.
• Higher viral ceiling than SermonPrepDesk: zero-context comedy, broad audience, high absurdity ceiling.

3. Audience — dual track (same architecture as SermonPrepDesk)

Track	Who	Content	Job
Crossover (growth)	General tech-curious viewers, fail-content audience	Fail clips, villain-bot Shorts	Algorithm fuel, breakouts, channel velocity
Buyer (conversion)	Small-business owners (dental, HVAC, plumbing, vet, funeral home, salon…) searching "missed calls," "AI receptionist for X," "after-hours answering"	Rebuild episodes, before/after case studies, outcome-framed fixes	Demo-line calls → trials → contracts

Hard calibration rule: the buyer doesn't want a SIP-trunk diagram. Teaching is outcome-first ("here's why your bot hangs up on callers — listen to it fixed"), with at most ~1 in 6 videos as a deep-technical authority flex (those attract engineers and competitors, not buyers — useful for credibility, useless for conversion).

Channel separation: this CANNOT live on @ChurchWiseAI or @SermonPrepDesk (mixed personas poison channel-level distribution — established in sermonwise-youtube/strategy.md §2). It is its own channel, fronted by the local-business voice brand.

4. Format architecture

A. The Fail Lab (Shorts, 20–45 s) — discovery

The villain bot fails at one specific, named thing per Short. Failure modes scripted from real ones seen in the wild. Hook in 1.5 s = the fail itself, then John's one-line diagnosis. On-screen label always present: "Staged demo — I built this badly on purpose."

B. The Rebuild (longform, 8–15 min, chaptered) — the spine

Kitchen-Nightmares arc using the anatomy frame (the founder's ear/brain analogy, formalized):

Organ	System	Failure modes shown	Fix demonstrated
Ears	STT / endpointing	Mishears names, talks over callers, dies on accents/noise/hold-music	Proper STT tuning, interruption handling, the listening discipline (HEAR)
Brain	LLM + RAG + prompts	Hallucinates prices, argues about its own hours, knows nothing about the actual business	Per-business knowledge, grounding, "I don't know → here's a human"
Mouth	TTS	Robot voice, wrong pacing, no warmth, mispronounced local names	Voice selection, prosody, brand personality
Nervous system	Telephony	Dead air, latency, dropped transfers, no fallback when the AI dies	Carrier infra, failover, transfer logic, the unglamorous plumbing
Conscience	Escalation + guardrails	Handles a crisis call badly, takes a message it should never take	Human-bridge rules, escalation, what AI must refuse to do

Season 1 = rebuilding the villain one organ per episode. The season arc IS the curriculum.

C. The House Call (longform, occasional) — the converter

A consenting real business (customer or recruited prospect; signed release): their real missed-call pain, the rebuild, then a real before/after call. Case study + demo + episode in one. This is the highest-converting format; produce one per quarter minimum once the channel is live.

5. The villain bot — recurring character

• A deliberately, precisely terrible voice agent John builds with his own stack (LiveKit + intentionally bad config). Every flaw is a real failure mode from the wild — the comedy is documentary.
• It's a character with a redemption arc: the season rebuilds it organ by organ until the finale, where it takes a full call flawlessly… and the next season starts a new villain with new failure modes (new vertical, new disasters).
• Lessons from Quill apply wholesale: junior/controlled status, scripted not live, one overstep gag per appearance, name the character (naming = branding), simple card on screen rather than a humanlike avatar.
• Name TBD by founder — candidates: Patch (gets patched weekly — fits the redemption arc), Dial-Ann, Hold-On Harold, VEX. Pick one that's funny to say repeatedly on camera.
• Voice: pick a Cartesia voice that is the OPPOSITE of the production agents' warmth (something flat/over-chipper). NEVER reuse Carson/Cindy or any voice deployed for paying customers — the villain's voice must be nobody's production voice, ever.

6. Hard guardrails (the three traps, made into rules)

1. Never mock a real small business. No calling deployed agents of identifiable businesses for content. They are the ICP — prospects, not props. Punching up at Fortune-500 phone trees is permissible; punching down is never.
2. Staged is disclosed, always. Every villain-bot appearance carries the on-screen label and a description line: "The failing agent in this video is a demo I built deliberately. Every failure mode in it is real — I've heard them all in the wild." Never present a staged fail as a found one. (Honesty is also the better joke.)
3. Recording consent. Any real call used on camera = written consent from the business (House Call release). Calls to John's own agents are fine. Commercial use of recorded third-party calls without consent is off the table regardless of one-party-consent law.
4. Existing marketing rules apply unchanged: no negative competitor call-outs by name, no fabricated testimonials, claims verified against the product before scripted (Rule: the fix shown on camera must be a fix the product actually ships).

7. Funnel

Fail Short (label: staged demo)
  └─ related-video link → Rebuild episode (the organ that failed)
        └─ verbal + description CTA: "Call the real one yourself — right now: <demo line>"
              └─ live demo agent call (is_demo flag suppresses owner notify — already shipped, PR #832)
                    └─ demo → discovery call (Cal.com) → contract

• The killer CTA nobody else has: a phone number. Viewers can call the GOOD agent live, mid-video. The demo line IS the lead magnet. (Track demo-line call volume as the channel's primary conversion metric.)
• Secondary lead magnet: "The Missed-Call Calculator" or "10 questions to ask before buying an AI receptionist" (PDF) — email capture for not-ready-yet owners.
• Vertical landing pages (wiseaiagency.com / local-business funnel) linked per episode when the episode's example business matches a vertical.

8. Pilot batch (produce as ONE sprint before launching anything)

1. "I built the world's worst AI receptionist — on purpose" (longform, ~10 min): villain intro, ten failure modes in one disastrous call, the anatomy frame teased as the season map. This is the channel trailer.
2. The Ears (Rebuild ep): the bot that interrupts, mishears "Schmidt" five ways, and can't handle a kitchen-noise caller → fixed, with before/after audio.
3. The Brain (Rebuild ep): the bot that invents a $99 special and argues about its own business hours → RAG/grounding fixed; "an agent that says 'I don't know, let me get a human' is a feature."
4. The Nervous System (Rebuild ep): dead air, latency, the transfer to nowhere → telephony done right, with the fallback drill ("what happens when the AI dies mid-call? a human-reachable phone line, that's what").
5. The House Call #1 (converter): consenting real business before/after. Recruit from existing customers or a warm prospect — this one episode is also a sales asset for cold outreach forever.

• 12–15 Fail Lab Shorts cut from the villain's scripted disasters (each fail = one Short).

9. KPIs and kill criteria

• Primary: demo-line calls/month attributed to YouTube and discovery calls booked. Secondary: subs, engaged views, House Call → contract conversions.
• Day-60 gate after launch: if Fail Shorts get views but demo-line calls stay ~0, the crossover→buyer bridge is broken — fix the related-video chain and the CTA before making more fails.
• Kill criteria: 90 days, <relevant-thousand engaged views/mo AND <10 attributed demo calls/mo AND it's costing SermonPrepDesk its cadence → park it again. The hub-vs-shorts sustainability rule applies here too: the Rebuild spine outranks Fail volume.

10. Launch gate and sequencing (the founder-bandwidth rule)

• DO NOT START until SermonPrepDesk passes its Day-60 gate (mid-Sep 2026 per sermonwise-youtube/strategy.md §8) with the cadence intact. One person cannot debug two channel launches at once; SPD has first claim.
• Earliest sensible start: pilot-batch scripting in late Sep, recording in one Oct sprint, launch Nov 2026 — OR fold to Q1 2027 if Advent content load is heavy.
• Production reuses the SermonPrepDesk machine wholesale: production-process.md (OBS profiles, Descript pipeline, credit budget, checklists), same desk, same recording-day discipline. Marginal cost of channel #2 is scripting + one extra sprint, not a new system.
• Descript plan note: a second channel's footage roughly doubles media-hours and credit burn — Creator (30 hrs / 800 credits) likely still holds with the batching habits, but re-check the meter in the first dual-channel month; Business tier is the fallback.

11. Open decisions for the founder (decide at the gate, not now)

1. Channel name/handle — candidates: Hold, Please · The Voice Agent Mechanic · Don't Hang Up · Missed Call (check handle availability at decision time)
2. Villain bot name (see §5)
3. Fronting brand: wiseaiagency.com vs. a vertical (funeral/vet are highest-ticket; general local-business widest)
4. Whether House Call #1 recruit is an existing customer (easier) or a cold prospect (better story)
5. Demo line: dedicated number for the channel (clean attribution) vs. existing demo line