Voice Agent YouTube — Research & Recommendation

Founder's raw idea (preserved): Use @ChurchWiseAI and/or a new channel for the local-business market. Viral-leaning videos where John has live conversations with his own voice agent. Possibly visiting other businesses' AI agents and "showing the difference." Many SHORT videos, each a different conversation type. Hooks: churches — "wouldn't it be nice if someone handled all those benevolence requests?"; businesses — "wouldn't it be nice to have a receptionist that worked 24 hours straight and never got sick?" Show off deep voice-agent knowledge.

---

Executive Summary (decision-ready)

Verdict: the idea is right, the framing needs one honest correction, and it should be FOLDED INTO the parked fails-channel concept — not run beside it and not replace it.

1. The lane is still open. Mid-2026, nobody on YouTube is the founder-operator who takes live, unedited calls with a production voice agent on camera for buyers (church admins, funeral directors, vets, trades). The big voice-AI channels serve agency-builders, not end buyers — Liam Ottley (~400K subs, audience = aspiring AI-agency founders, not businesses buying receptionists — Automation Atlas), Jannis Moore / Voice AI Bootcamp (Vapi/Retell course-sellers — Whop review), AssemblyAI (~80K, engineers — LearnWithPath). Vendor channels (Vapi, Retell, Synthflow, Bland, ElevenLabs) post polished demos and dev tutorials. No one owns "watch a real AI receptionist take real calls, owner-to-owner." That matches the fails concept's §2 finding and the founder's instinct.

2. Honest viral calibration: demos convert; fails and novelty travel. What actually went viral in this category: AI failing (Taco Bell drive-thru fail clips, 21M+ views, forced a strategy change — mlq.ai), AI-to-AI novelty (Gibberlink, 13.7M+ views, shared by MKBHD — ElevenLabs), and adversarial comedy (Kitboga ~3M+ combined, Jolly Roger Telephone — NPR, The Register). A clean "watch my AI handle a call" demo is table stakes in 2026 — every vendor has one; none went viral. So: the founder's "wouldn't it be nice" conversation Shorts are the CONVERSION layer (excellent), and the parked fails concept supplies the GROWTH layer. Together they're one channel. Apart, each is half a channel. The single best new format this research surfaced is the hybrid: Fail→Fixed — villain bot botches a call, hard cut, the real agent takes the same call live (before/after inside one 40-second Short).

3. Channel architecture: two channels, phased — exactly as already decided. Evidence is unambiguous that mixing church + HVAC personas on one channel damages both: the 2026 algorithm clusters audiences at the channel level; multi-niche channels see split CTR and collapsed homepage distribution (Subscribr, TubeRanker) — same finding as sermonwise-youtube/strategy.md §2. Therefore: church conversation-Shorts go on @ChurchWiseAI starting now (they ARE in-funnel product demo content — the channel's existing mission); all local-business content waits for the new channel at the Day-60 gate (~mid-Sep 2026), launched per the fails concept with this research folded in.

4. "Visiting other businesses' AI agents" survives only in constrained form. Recording is legal in Canada one-party-consent terms (Criminal Code s.184(2)(a) — Recording Law), but ~11 US states require all-party consent, PIPEDA adds business obligations, and publishing a third party's call commercially is a separate risk tier (Quo). The fails concept's Rule 1 (never mock a real small business — they're the ICP) and Rule 3 (written consent for any real third-party call) stand. Allowed versions: consenting businesses (the House Call format), Fortune-500 phone trees (punch up), and re-enacted/anonymized fails (disclosed). Cold-calling a competitor's deployed agent for content is out.

5. Shorts-for-B2B reality check: Shorts alone don't close high-ticket B2B; they discover, longform converts (Stormy AI, Koanthic). But this business has the one CTA that collapses the funnel: a phone number you can call mid-video. The demo line (is_demo infra already shipped, PR #832) lets a Short convert directly — rare for B2B. Keep the hub-and-spoke spine anyway (the Rebuild longform), per the proven SermonPrepDesk machine.

Sequencing in one line: Phase 0 (now): church Shorts trickle on @ChurchWiseAI, ~half a recording day per month, zero new infrastructure. Phase 1 (post Day-60 gate, ~Oct–Nov 2026): launch the local-business channel with the fails-concept pilot batch plus the Fail→Fixed and "wouldn't it be nice" Short series. Phase 2 (quarterly): House Calls.

---

1. Competitive Landscape (mid-2026)

1a. Who makes AI-receptionist / voice-agent content, and for whom

Channel / creator	Approx. scale	Audience	Format	Threat / lesson
Liam Ottley (Morningside AI)	~400K subs Dec 2025, ~730K claimed 2026	Aspiring AI-agency founders, NOT business buyers	Business-model deep dives, build tutorials (Vapi etc.), 46-min "is it saturated" essays	Owns the builder audience. His pivot to voice agents confirms the category's heat. Zero overlap with our buyer ICP. (Automation Atlas, note.com analysis)
Jannis Moore (Integraticus / Voice AI Bootcamp)	Mid-tier; community of course buyers	Agency builders learning Vapi/Retell	Tutorials, bootcamp funnels	The "voice AI educator for builders" seat is taken and saturated. Don't compete there. (Whop)
Nate Herk	~600K in <2 yrs	Automation/n8n builders	Workflow tutorials	Proof the builder lane explodes fast — and that it's crowded. (Atak Interactive)
AssemblyAI	~80K+	Engineers	Voice-agent engineering, STT, function calling	Engineering-authority lane, taken. (LearnWithPath)
Vendor channels (Vapi, Retell, Bland, Synthflow, ElevenLabs)	Modest subs relative to product scale	Devs + prospects in-funnel	Polished demos, launch videos, docs-as-video	Polished demos with no honesty about failure modes = the trust gap the fails concept identified. None are founder-fronted for SMB buyers. (Softcery platform survey)
Jolly Roger Telephone	Niche but repeatedly viral clips	General public	AI bots wasting telemarketers' time, real call audio	Closest format ancestor to "live AI call as entertainment." Comedy + righteous target = shareable. (The Register, Fortune)
Kitboga	~1M+ combined Twitch/YT; now runs an AI bot army vs scam centers	General public	Character-driven recorded calls with scammers	The masterclass in making recorded calls watchable for years: recurring characters, a villain the audience roots against, escalating stunts. (NPR, Frank on Fraud)
Air.ai-style reseller hype	(collapsed)	Biz-opp seekers	"AI closes sales for you" hype	FTC secured an $18M judgment against Air AI for deceptive AI business-opportunity marketing. The hype lane isn't just off-brand — it's an enforcement target. Our no-hype rule is a legal moat, not just taste. (Lathrop GPM, FTC)

1b. Saturated vs open

• Saturated: "build an AI agency" tutorials; platform comparison reviews ("Vapi vs Retell 2026" — dozens of blog/video entries, e.g. Lindy, Retell's own comparison content); generic "AI will answer your phones" promo.
• Open: founder-operator content for the buyer — live calls, failure-mode honesty, vertical-specific scenarios (nobody does church; nobody does funeral home with dignity; vet triage content is thin); the call-it-yourself demo-line CTA (no competitor channel offers a number you can dial mid-video).
• Confirmation: review searches for "I tested AI receptionists on camera" surface blog listicles (Allo, Lunacal) but no established video creator — an open review-lane too, though one we should NOT take (reviewing competitors on camera conflicts with our disparagement rule, §5).

---

2. The "Live Conversation With an AI on Camera" Format

What actually went viral, and why

Case	Views	Why it traveled
Taco Bell AI drive-thru fails (18,000 waters, infinite drink prompts)	21M+ across clips	Schadenfreude + a giant brand + a relatable frustration. Fails travel. (mlq.ai, Cybernews)
Gibberlink (two AIs realize they're both AIs, switch to beeps)	13.7M+	Novelty + a tiny sci-fi frisson; zero context needed. Stunts travel. (ElevenLabs, TechCrunch)
Kitboga / Jolly Roger recorded calls	Years of sustained millions	Character + a villain you're allowed to laugh at. Adversarial comedy travels.
Vendor "watch our agent take a call" demos	Rarely break containment	Competent ≠ interesting. Plain demos don't travel — they convert viewers who arrived with intent.

Implication for the founder's idea: a live conversation with your own agent is the right conversion format and the right authenticity format, but on its own it has a low viral ceiling. It needs one of three amplifiers per video: a fail (villain bot), a stunt (AI-vs-AI, stress test), or a stake (a real business's real 2 AM problem).

Failure modes of the live format — and the mitigation

1. Latency dead-air on camera. Sub-second turnaround is the product's own bar (Bland markets 400–700ms as "conversational gold" — Softcery). Dead air in a demo reads as product failure even when it's network. Mitigation: record calls over the real phone line (what a customer hears), keep the take, and narrate any hiccup honestly — "that pause was real; here's why it happens" is an authority moment, not a blooper. Never silently edit latency out (see Gemini, below).
2. Cherry-picking accusations. The default audience assumption in 2026 is that AI demos are faked. Google set the precedent twice: the Duplex demo skepticism and the Gemini demo, which Google admitted was edited (latency reduced, prompts entered off-camera) and got hammered for it (CNBC, The Tech Portal). FTC guidance: if a demo could mislead about real performance (simulated results, edited outputs), disclosure is required and claims must be substantiated (kr.law).
3. Disclosure norms that exist mid-2026: staged/simulated content labels (our fails-concept Rule 2 already exceeds the norm); FTC AI-content disclosure expectations with per-violation penalties up to ~$53K (thestacc); YouTube's auto-labeling system (May 2026) covers AI-assisted editing without manual disclosure (per sermonwise-youtube/strategy.md §9), but a staged scenario presented as found is a trust decision, not a platform-policy one.

The authenticity play — turn skepticism into a format. Three mechanics, all cheap:

• "One take, no cuts" badge on every live-call video, with the running clock visible. The anti-Gemini.
• Viewer torture tests: pin a comment — "what should I throw at it next week?" — then take the meanest suggestions live on camera. This converts "it's staged" commenters into co-producers, feeds the comment engine (load-bearing in the 2026 recommendation model per house strategy), and produces the fails the growth layer needs with full honesty.
• Call it yourself: the demo line in the description. The viewer can replicate the video personally in 30 seconds. No other authenticity device in the niche comes close, and no competitor channel has it.

---

3. Calling Other Companies' AI / Phone Systems on Camera

Law

• Canada (John's jurisdiction): one-party consent — recording a call you participate in is legal (Criminal Code s.184(2)(a)) (Recording Law). BUT: PIPEDA applies to organizations recording in a commercial context (notification, purpose, consent alternatives), and publication of a recording is assessed separately from the recording itself — lawful-to-record does not mean safe-to-publish (Quo, Vilkhov Law).
• US: ~11 states require all-party consent (California, Florida, Washington, Pennsylvania, Illinois among them). Cross-border calls get judged by the stricter rule in practice (Avoma, Wikipedia summary). Calling a US business's AI line from Ontario and publishing it = multi-jurisdiction exposure for a clip.
• The "but it's a robot" theory doesn't save you: an AI line can hand off to a human mid-call; the business's prompts/voice may be protected expression; and the business (an identifiable party) bears the reputational impact, which is where defamation/disparagement lives, not wiretap law.

Defamation / disparagement when comparing named competitors

Truth, opinion, parody and hyperbole are defenses; the danger zone is provably false factual assertions that harm a business (Foundershield, Webupon). A real recorded call is "true" — but editing, framing, and an unrepresentative sample are exactly where "false impression" claims and platform takedown fights start. House rule already prohibits negative competitor call-outs by name (social-strategy.md do-not list, fails concept §6.4). Keep it. It's both brand (punch up, never down) and risk management.

YouTube platform posture

No bright-line "recorded calls" policy; enforcement runs through harassment and privacy complaints (YouTube harassment policy). A targeted small business filing a privacy complaint about a mocking call clip is a plausible strike vector. Channels that survive on recorded calls (Kitboga, Jolly Roger) target scammers — parties with no sympathetic claim and no standing they'd dare use. A Woodstock plumber's bad voicemail is the opposite case in every respect.

What "showing the difference" becomes (the survivable versions)

1. The House Call (already in the fails concept §4C): a consenting business, signed release, real before/after. Highest-converting, fully safe, slowest to produce.
2. Punch up: Fortune-500 phone trees / IVR mazes ("press 4 to be forgotten"). Fame defense, opinion framing, nobody's livelihood harmed. Use sparingly — entertainment, not the spine. Even here, prefer screen-recorded hold-time receipts and re-enactment over publishing raw audio of named companies.
3. Re-enact, disclose: "a real call I heard this month, recreated" — the villain bot performs failure modes John has genuinely encountered. This is the fails concept's entire architecture, and it's the right answer. The disclosure (Rule 2) is also the better joke.

Bottom line: the founder's "visiting other businesses' AI agents" idea is folded into House Call + punch-up + re-enactment. Cold-recording identifiable small businesses' agents for content stays banned (fails concept Rule 1/3 reaffirmed).

---

4. Shorts-Driven B2B Trust-Building — the 2026 Evidence

• Shorts discover; longform converts. B2B founders flooding channels with Shorts attract subscribers who never buy, and Shorts RPM ($0.06/1K) vs longform ($7/1K) reflects intent difference; 8–12 min longform is where high-ticket trust forms (Stormy AI, Koanthic). Some service businesses do book consultations from 30-second clips — but only with a deliberate system behind the content (Interact Marketing).
• Video is how B2B buyers de-risk in 2026: 81% trust video over text; 70% watch video during the purchase decision; founder-led educational video outperforms polished brand films on conversion metrics (Levitate Media, RiseCraft).
• House data already encodes the 2026 algorithm reality (engaged-view counting, the Sept-2025 Shorts decay, 2-Shorts/day sustained beats bursts, comment signals) — sermonwise-youtube/strategy.md §5/§8 applies wholesale.
• Our exception to "Shorts don't convert B2B": the demo-line phone number. A Short that ends "call it yourself, right now" turns a discovery surface into a conversion surface — the funnel is Short → dial instead of Short → hub → email → trial. Keep BOTH funnels: phone number for the impulsive owner, Rebuild-hub → Missed-Call-Calculator lead magnet for the cautious one.

---

5. The Two-Audience Question — Channel Architecture

Evidence: the 2026 algorithm clusters viewers by watch-history at the channel level; mixed niches produce split CTR, messy retention, and homepage-distribution collapse; best practice is one tight cluster per channel, spinning off when "why is this here?" comments appear (Subscribr, TubeRanker, SocialPilot). This independently confirms the channel-separation rationale already established in sermonwise-youtube/strategy.md §2 and fails concept §3.

Would HVAC-receptionist content damage @ChurchWiseAI? Yes, twice over: (a) algorithmically — the existing subscriber base is churches; HVAC uploads would tank early-engagement signals and re-cluster the channel's identity; (b) brand — social-strategy.md explicitly bans posting business content at church audiences and vice versa. Do not mix.

Recommended architecture (3 surfaces, 2 channels):

Surface	Role	Content	Timing
@ChurchWiseAI (existing)	In-funnel church buyers	Church "wouldn't it be nice" conversation Shorts + occasional longform walk-throughs. This is the channel's existing product-demo mission, upgraded from screencasts to live calls.	Phase 0 — start now (bandwidth-capped)
New local-business channel (the fails-concept channel)	Growth + conversion for wiseaiagency verticals	Fail Lab + Fail→Fixed + "wouldn't it be nice" buyer Shorts; Rebuild longform spine; House Calls	Phase 1 — at the Day-60 gate
@SermonPrepDesk	Untouched	—	cross-link only in About/featured

Relationship to the parked fails concept — explicit: FOLD, don't supersede. The fails concept remains the architecture of record for the new channel (villain bot, anatomy frame, Rebuild spine, House Call, guardrails, launch gate, KPIs). This research adds the founder's conversation-demo layer to it: (a) the Fail→Fixed Short becomes the flagship discovery format (it carries the fails concept's comedy AND the founder's live-demo proof in one unit); (b) the "wouldn't it be nice" series becomes the named buyer-track Short format the fails concept lacked; (c) live unedited calls + viewer torture tests become the channel's authenticity mechanic; (d) "visiting other businesses' AI" is absorbed into House Call / punch-up / re-enactment per §3. The fails concept's §10 launch gate is unchanged. Phase 0 (church Shorts on @ChurchWiseAI) is the only genuinely new commitment, and it lives outside the gated channel.

---

6. Format Recommendations (ranked)

Viral probability is stated honestly. "Conversion power" = demo-line calls / bookings per view.

#	Format	Length	Viral ceiling	Conversion power	Notes
1	Fail→Fixed — villain bot botches a call; hard cut; the real agent takes the same call, live, one take	30–60s Short	High (fail = hook, fix = payoff; rides the Taco-Bell-meme energy with a redemption beat nobody else has)	High (ends on the demo number)	New flagship. Staged-demo label on the fail half, "one take, no cuts" on the fix half.
2	"Wouldn't it be nice…" buyer-track Shorts — one pain point, one live conversation	30–45s Short	Low–medium (niche-relatable, not crossover)	Highest per view — speaks directly to the ICP's named pain	The founder's hook series. Concepts mapped below.
3	The Rebuild (fails concept §4B, unchanged)	8–15 min	Medium	High (the trust spine; where the longform B2B evidence points)	Monetizable watch hours; the channel's asset.
4	Viewer Torture Test — comment-sourced attack prompts thrown at the live agent	60s Short or 6–8 min	Medium–high (stunt energy + co-production)	Medium	The authenticity engine; manufactures honest fails for format #1.
5	The House Call (fails concept §4C, unchanged)	10–15 min	Low	Highest absolute (case study + sales asset forever)	1/quarter once live.
6	Punch-up phone-tree pieces ("I waited 47 minutes so you don't have to")	Short or mid	Medium (relatable rage)	Low–medium	Sparingly; legal care per §3; never the spine.
7	AI-meets-AI stunts (Gibberlink-style lottery tickets)	Short	Lottery (very high or nothing)	Low	1–2/quarter max; only when a genuinely novel idea exists. Never force it.
✗	Competitor review/teardown on camera	—	—	—	Rejected: disparagement rule, §3 risk, and it makes us a reviewer instead of an operator.

The "wouldn't it be nice" Short series — 16 concrete concepts mapped to USE_CASE_CATALOG

Each is one live conversation, one pain point, hook spoken in the first 1.5s, demo-line CTA. Church batch ships on @ChurchWiseAI in Phase 0; business batch ships on the new channel in Phase 1.

Church (@ChurchWiseAI — Phase 0):

#	Hook	Conversation shown	Catalog ref
C1	"Wouldn't it be nice if nobody had to walk into the office to ask for rent help?" (the founder's hook)	Caller quietly asks for benevolence help; agent handles it with dignity, confirms callback number, routes confidentially	3.4 submit_benevolence_request (+ the Phase-3 benevolence voice work just shipped)
C2	"Your congregation calls at 11 PM. Who answers?"	Late-night crisis-adjacent call → resources + pastor notification	1.1 request_callback, Safety Protocol; hook already approved in social-strategy.md
C3	"She asked for prayer at 2 AM. Nobody had to know."	Confidential prayer request, confidentiality flag explained in 5 seconds	8.3/11.5 submit_prayer_request + RBAC
C4	"The same five questions, 40 times a week."	Rapid-fire montage: service time, parking, kids check-in — agent answers all five in 30s	2.1 FAQ tools
C5	"It works on her rotary phone."	An elderly caller (real volunteer, on camera, consenting) just… calls the church number	10.4 — the anti-tech-fear demo; gold for skeptical boards
C6	"The board member who thinks AI is the mark of the beast — I let him call it."	A real skeptic (consenting) calls live; his face is the content	16.1/16.6
C7	"Mrs. Henderson's surgery — found out two weeks late. Never again."	Family member reports a hospital visit; agent captures hospital/room/urgency	1.4 request_pastoral_visit
C8	"New family, three kids, one question: is my toddler safe with you?"	Visitor asks about check-in, allergies, background checks	6.2 get_kids_info

Local business (new channel — Phase 1):

#	Hook	Conversation shown	Source
B1	"Wouldn't it be nice to have a receptionist who worked 24 hours straight and never got sick?" (the founder's hook)	Montage: 7 AM, noon rush, 11 PM — same warm agent, three captured leads landing in the inbox	AI Front Desk package, local-business-go-to-market.md §2
B2	"A family calls the funeral home at 2 AM. A voicemail is the last thing they need."	First-call handled with genuine care; director paged; nothing lost	Funeral vertical (launch-first per GTM §3); dignity rules from social-strategy.md
B3	"Saturday, 9 PM: 'my dog ate something.' Who picks up?"	Vet triage call — agent asks the right questions, escalates urgently	Vet vertical; poison-triage logic already modeled
B4	"You're under a sink. Your phone rings. That call was worth $4,000."	Plumber on a job; agent books the estimate; lead appears in inbox on screen	Trades; Missed Lead Recovery
B5	"Somebody tried the 18,000-waters prank on MY agent."	Caller attempts the Taco Bell prank; agent declines gracefully, stays useful	Rides the 21M-view meme; demonstrates guardrails — the trust sale
B6	"Ask it something it doesn't know. Watch what happens."	Agent says "I don't know — let me get a human" and routes	The AI Bridge principle as content; the #1 objection answered
B7	"Dinner rush. Phone rings. Your host has four tables waiting."	Restaurant reservation call captured during chaos	Restaurant / Review Engine entry vertical
B8	"It's flu season. Your clinic's phone doesn't get the flu."	Dental/clinic appointment overflow handled after hours	Second-wave verticals, GTM §3

Plus Fail→Fixed variants of B1–B6 (villain version first) once the villain bot exists — each concept yields two Shorts for one script.

---

7. Guardrails (consolidated — the fails concept's three rules, extended)

1. Never mock a real small business (fails Rule 1, reaffirmed). No cold-recording identifiable businesses' agents. Punch up only.
2. Staged is disclosed, always (fails Rule 2, reaffirmed) — and extended to demos: never edit latency/outputs out of a "live" call without an on-screen note (the Gemini lesson — CNBC); prefer the "one take, no cuts" badge; substantiate every capability shown against shipped product (FTC demo standard — kr.law).
3. Recording consent (fails Rule 3, reaffirmed): written release for any third-party call published, regardless of one-party-consent law; on-camera callers (the skeptic, the grandma) sign too; calls to our own agents with is_demo are clean. US-touching calls: assume all-party consent applies.
4. No named-competitor negativity (social-strategy.md), no hype claims (the FTC's Air AI judgment makes the no-hype rule an enforcement shield, not just brand — Lathrop GPM), HEAR voice, dignity rules for funeral/vet content, no luck/sacred-exclamation language.
5. Real callers' data: never show a real lead's name/number on the inbox screen — demo church / churchwiseai-demo and seeded demo data only (Cross-Project rule: no junk data in prod).

---

8. Production Reuse

• The SermonPrepDesk machine carries channel #2 wholesale (fails concept §10 already established this): sermonwise-youtube/production-process.md OBS profiles, Descript pipeline, sprint cadence, checklists. Marginal cost = scripting + ~one extra recording sprint/month. Watch the Descript credit meter in the first dual-channel month (Creator tier likely holds; Business is fallback).
• Phase-0 church Shorts are even cheaper: phone on speaker + OBS screen capture of the admin dashboard showing the captured lead/prayer request. No set, no villain bot, no new channel. Half a recording day/month yields 8–10 Shorts.
• Demo-line CTA: is_demo flag ships already (PR #832 — tags + suppresses owner notify). Recommend a dedicated demo number per channel for clean attribution (fails concept §11.5) — decide at the gate.
• Villain bot: own stack, deliberately bad config; Cartesia voice selection must follow Rule #21 (serial calls only) and never reuse any production customer voice (fails §5).
• Capture discipline now: every interesting call John has with his own agents from today forward gets recorded and archived — Phase-1 launch inventory builds itself for free.

9. Sequencing (honors the Day-60 gate and one-person bandwidth)

Phase	When	What	Bandwidth cost
0 — Church trickle	Now → mid-Sep 2026	8 church "wouldn't it be nice" Shorts (C1–C8) on @ChurchWiseAI, published ~2/week; record in one half-day batch/month; archive raw calls for Phase 1	~½ day/month. Must NOT touch SPD's sprint cadence — if a month is tight, Phase 0 skips, never SPD.
Gate check	Mid-Sep 2026	SermonPrepDesk Day-60 gate per sermonwise-youtube/strategy.md §8. Pass + cadence intact → proceed. Fail → Phase 1 slips to Q1 2027 (fails concept §10 fallback).	—
1 — Channel launch	Oct script / Nov 2026 launch	New channel per fails-concept pilot batch (§8) plus Fail→Fixed Shorts and B1–B8; demo-line CTA everywhere	One scripting block + one Oct recording sprint
2 — Converters	Q1 2027 →	House Call #1, quarterly cadence; torture-test episodes; scale per fails-concept KPIs	Steady-state per fails §10

10. KPIs

• Primary (both tracks): demo-line calls/month attributed to YouTube and discovery calls booked (Cal.com) — unchanged from fails concept §9.
• Church track (Phase 0): chat/voice trials or admin-token signups attributed to YouTube; engaged views per Short; comments from church staff (the qualitative gold signal, as with pastors on SPD).
• Business track (Phase 1): fails-concept KPIs and kill criteria adopted unchanged (Day-60 fail-Short-views-without-demo-calls check; 90-day kill criteria; Rebuild spine outranks Fail volume).
• Phase-0 mini-gate (Sep, alongside the SPD gate): if 8 church Shorts produce zero demo calls AND zero meaningful church comments, the "wouldn't it be nice" framing needs rework before Phase 1 builds 8 more on the business side — cheap lesson, bought early.

---

Decisions for the Founder

1. Approve the fold: your conversation-demo idea merges INTO the parked fails channel as its buyer track (Fail→Fixed + "wouldn't it be nice"), one channel for local business — not a separate demo channel. Yes/no?
2. Approve Phase 0: start the 8 church Shorts (C1–C8) on @ChurchWiseAI now, capped at ~½ recording day/month, with SPD cadence always taking priority. Yes/no?
3. Accept the constraint on "visiting other businesses' AI": consenting businesses (House Call), Fortune-500 punch-ups, and disclosed re-enactments only — never cold-recording an identifiable small business's agent. Yes/no?
4. Dedicated demo numbers: one per channel for clean attribution (vs reusing the existing demo line)? (Provisioning per knowledge/runbooks/voice-provisioning.md when approved.)
5. Phase-0 on-camera cameos: do you want C5 (the grandma) and C6 (the skeptic board member) — both need a willing real person + signed release — in the first batch, or hold them for after the format proves itself?

(Channel name, villain-bot name, and fronting brand remain gate-time decisions per fails concept §11 — nothing here forces them early.)