AI Starter Map — Reference cases

for content prep

Three personas, three tiers

The frictions are designed to be universally recognizable. Cases A, B, and C also represent three complexity tiers — useful for picking what to actually build live in S2.

Case A · Simple

Lina · Pilates studio

Customer-facing · inbound triage. Built live in S2.

Case B · Medium

Diego · Gym chain

Member retention · signal detection. Walk-through only.

Case C · Complex

Sofia · Omnichannel ops

Internal ops · data assembly. Reference only — too big for live.

casting note

The frictions stay the same — only surface details change. Lina (pilates) flexes to spa, yoga, PT, esthetician. Diego (gym chain) flexes to any multi-location member-based business. Sofia (omnichannel) flexes to any multi-channel SME. Read the audience composition in the chat or registration data and lightly recast as needed.

LinaCustomer-facing — inbound triage

Boutique pilates studio · 2 locations · ~120 members

SimplePRIMARY LIVE-BUILD CANDIDATE

“I spend 60–90 minutes a day responding to messages. Half the time I'm at a red light typing the same answer about pricing. The real prospects? They cool off because I take 4 hours to reply.”

Friction Map · top 5

★Repetitive inbound responses (90 min/day)
2.Scheduling first-class trials (15 min × 4–5/week)
3.Class no-show reminders (manual, often forgotten)
4.Monthly invoice reconciliation (Mindbody + Stripe)
5.Quarterly instructor feedback review

★ picked as the quick win

Flow

Trigger: WhatsApp / IG DM / web form

① LLM classify → {FAQ, lead, client, spam}

② FAQ → auto-reply from canned answer

③ Lead → qualifier reply + notify Lina

④ Client → route to ops manager

Output: 80% auto-handled, leads surfaced with context

Metric: time-to-first-response < 2 min · auto-resolved rate 60%+

Stack · no-code

Make.comWhatsApp Cloud APIIG Graph APIOpenAI gpt-4o-miniAirtableTelegram bot

Expected outcome

›~50–70 min/day saved → ~5 hr/week back
›Lead response: 4 hr median → < 2 min
›Trial conversions: estimated +20–30% (60-day measure)

What can go wrong (prototype stage)

!Classifier hallucinates intent → log every classification, Lina spot-checks first 200.
!Real client hits FAQ → include "Reply HUMAN to talk to Lina" escape hatch.
!Instagram API rate limits → poll every 2 min, not real-time.

How to use Lina in the webinar

›S1 Block 3 (Friction Map): open with Lina, audience mirrors.
›S1 Block 4 (Matrix): show her placement of the 5 frictions.
›S1 Block 5 (Flow): walk through line-by-line as the template.
›S2 Block 4 (Live build): primary candidate. ~22 min realistic with creds pre-loaded.

DiegoMember retention — signal detection

Multi-location gym chain · 4 locations · ~2,400 members

MediumBACKUP / WALK-THROUGH ONLY

“I see the churn number on the first of each month. I know half of those people would've stayed if someone had called them two weeks earlier. But nobody has the time to scan the booking system every day for ghosts.”

Friction Map · top 5

★At-risk member detection (zero today)
2.Personalized re-engagement (currently generic broadcast)
3.Trainer / class capacity rebalancing across locations
4.Member feedback collection and triage
5.Cross-location pricing exception approvals

★ picked as the quick win

Flow

Trigger: daily cron @ 8:00 local per location

① Query DB: members with last_visit 14–21 days ago + active

② Pull context: avg visits, preferred class, preferred instructor

③ LLM compose personalized message ("missed seeing you at Tuesday vinyasa with Maya…")

④ Send via WhatsApp + log delivery

⑤ If reply → route to retention specialist Slack thread

⑥ If no reply 7 days → flag for manual call

Output: at-risk member contacted before going dormant

Metric: rescue rate = % contacted who book within 30 days (baseline 0%)

Stack · no-code

n8n (self-hosted)Mindbody / ABC APIOpenAI gpt-4oWhatsApp Cloud APIPostgres / AirtableSlack webhooks

Expected outcome

›~150 at-risk members contacted/month across 4 locations
›15–25% rescue rate → ~25–35 members saved/month
›~$2k–$3k MRR rescued at $80 ARPU

What can go wrong (prototype stage)

!Tone mismatch — LLM reads as spam → A/B 3 prompt variants, Diego reviews first 50.
!False positives (member on vacation) → include opt-out, ask before assuming.
!Booking system data freshness → run cron 2hr after closing, not midnight.

How to use Diego in the webinar

›S1 Block 4 (Matrix): Diego is the operator tempted to pick strategic #2 first. Use to illustrate the "bigger feels more important" trap.
›S2 Block 5 (When to scale): "at 4 locations × 24/7, this prototype needs observability + retries + on-call" — bridge to the AI Blueprint conversation.
›DO NOT build live. Walk-through if audience flow matches this pattern.

SofiaInternal ops — data assembly

Wellness omnichannel ops · eCom + retail + 2 studios + marketplace

ComplexREFERENCE ONLY · TOO BIG FOR LIVE BUILD

“Every Sunday night and Monday morning, three of us spend 90 minutes pulling numbers from six tools. We assemble a doc. The founder reads it in four minutes. Eight person-hours a week to read four minutes.”

Friction Map · top 5

★Weekly metrics assembly (8 hrs/week)
2.Monthly board pack assembly (~16 hrs/month)
3.Daily inventory reconciliation (Shopify ↔ retail POS)
4.Customer service triage across email + WhatsApp + IG
5.Refund / exchange escalations across channels

★ picked as the quick win

Flow

Trigger: Sunday 23:00 cron

① Parallel fetch: Shopify, Stripe, Mindbody, Google Ads, Klaviyo, Meta Ads

② Compute deltas WoW + YoY

③ LLM narrates digest in plain English with action items

④ Send Monday 7am email + Slack #ops thread

⑤ Key callouts → tagged action items per owner

Output: one-pager + Slack thread auto-delivered before team logs in

Metric: hours saved ≥ 6/week · founder monthly satisfaction (single-question survey)

Stack · no-code

Make.comShopify APIStripe APIMindbody APIGoogle Ads APIKlaviyo APIOpenAI gpt-4oGmail + SlackGoogle Sheets (history)

Expected outcome

›Sofia's team gets 6+ hrs/week back
›Founder gets digest 4 hrs earlier with more analysis
›Compounding pattern detection from month 2 onward

What can go wrong (prototype stage)

!An API breaks → graceful fallback ("Shopify unavailable, retry") not full failure.
!Hallucinated stats → LLM only narrates pre-computed numbers, never generates them.
!Founder doesn't read it → ship lean, iterate format weekly.

How to use Sofia in the webinar

›S2 Block 5 (When to scale): canonical example of "prototype graduated to plumbing." Bridge directly to AI Blueprint Sprint conversation.
›NEVER attempt live build. Narrate only.

how to use these cases

Where each case shows up in the sessions

Block	Case A · Lina	Case B · Diego	Case C · Sofia
S1 · 3 Friction Map	Open with Lina, audience mirrors	—	—
S1 · 4 Time × Impact	Show her matrix placement	Use as "strategic trap" example	—
S1 · 5 Design the flow	Walk through line-by-line as the template	—	—
S2 · 2 Prototype mindset	Simple tier reference	Medium tier reference	Complex tier reference
S2 · 4 Live build	Primary build candidate	Walk-through if audience flow matches	Do NOT build live
S2 · 5 When to scale	—	Observability + on-call talking point	Canonical "graduated to plumbing" → AI Blueprint bridge