Zapier — Module 22.7

How This Lesson Fits the Module & Volume

This is the Vol. 22 capstone vendor. You catalogued frontier LLMs (22.1), image/video/voice (22.2–22.4), inference platforms (22.5), coding IDEs (22.6), and automation (22.7). Zapier is what most non-engineering teams actually buy: the widest SaaS connector catalog, the simplest Zap mental model (trigger → actions), and a growing AI layer (AI actions, copilot to build Zaps, chatbots/agents—names evolve). Vol. 21 workflow automation + Vol. 15 HITL still police AI steps. Vol. 18 FastAPI / Celery remain the escape hatch.

After Zapier, Vol. 23 starts building: ChatGPT Clone. The point of this volume was never to memorize logos—it was to pick substrates without inventing prices or skipping HITL. Zapier is the last catalog card before you implement.

Learning Objectives

By the end of this lesson, students should be able to:

Describe Zapier as SaaS task-based automation with the broadest app catalog in this trio.
Contrast Zapier vs Make (visual power) vs n8n (self-host) vs custom Vol. 18 workers.
Place AI by Zapier / agents / copilot as untrusted steps or builders—not unsupervised CRM gods.
Explain task metering qualitatively and when volume pushes you off Zapier.
Write an enterprise pick-list: residency, SSO, HITL, AI policy, exit to n8n/Celery.
Hand off to Vol. 23: clone products still need these glue platforms around them.

Definition

Zapier is a cloud automation platform. A Zap is a trigger plus one or more actions (with filters, paths, delays, and tables/interfaces on some plans). Zapier’s historic advantage is how many apps a non-engineer can connect with OAuth. AI by Zapier (and related Copilot / Central / Agents SKUs—re-read current product names) can draft, classify, or help author Zaps. Metering is typically tasks (successful action runs) inside plan tiers (Free → Enterprise). It is not a self-host workflow engine like n8n and not your customer-facing LLM app (that is Vol. 23).

The 22.7 Triangle + Vol. 18 Escape Hatch

If you optimize for…	Default pick	Watch-out
Widest SaaS catalog + simplest training	Zapier	Task volume; weaker VPC story
Visual routers/aggregators without self-host	Make	Operations explosion
Residency + Code node + AI Agent pockets	n8n self-host	You run Docker/HA
The graph is the product / custom SLA	FastAPI + Celery (Vol. 18)	You own all connectors

Zap anatomy

Trigger (app / webhook / schedule)
Filter / Paths / Delay
Action(s) + AI step
Optional Tables / Interfaces

AI roles (keep distinct)

Copilot: helps build the Zap
AI action: classify/draft inside a run
Agent SKU: bounded tool use (HITL)
Never: silent prod spend

Vol. 23 implication

Your ChatGPT clone is an app
Zapier may wrap it (webhook ↔ CRM)
Do not implement the clone inside Zapier

AI Steps, Privacy, and HITL

Every Zap that sends customer text to an AI action is a Vol. 20 privacy decision: data leaves your VPC into Zapier and the model vendor behind the action. Use dummy data in labs. Enterprise plans add SSO, SCIM, app admin, and (sometimes) data-retention controls—verify the current trust center; do not recite a 2019 blog. HITL: Paths to a human approval app, or delay + email confirm, before any send/pay. Same Vol. 15/21 rules as n8n and Make.

Catalog Snapshot (Qualitative)

Dimension	Zapier	Make	n8n
Pricing posture	Plan tiers + task metering (and AI add-on usage on some SKUs)—live sheet only	Plan + operations	Self-host infra or Cloud executions
Strengths	Largest connector catalog; non-engineer UX; enterprise familiarity; AI copilot to author Zaps; Tables/Interfaces ecosystem	Visual power (routers/aggregators)	Self-host; Code; AI Agent nodes
Weaknesses	Task cost at volume; SaaS data path; less code-native; easy to create shadow-IT Zaps	Steeper UX; ops metering surprises	Ops burden; smaller long-tail catalog
AI steps	AI actions + builder copilot + agent SKUs (verify names)	AI modules	LLM + AI Agent nodes
Self-host	No (classic SaaS)	No (classic SaaS)	Yes

Zap + AI Action Sketch (then Vol. 23)

# zap_blueprint.md — dummy-data only
Trigger: Webhook from your future Vol. 23 ChatGPT clone
         (HMAC secret in Zapier; never log raw PII)
Filter: payload.env == "lab"
Action: AI by Zapier — classify intent → JSON enum
         {billing, faq, human}
Paths:
  faq → Action: send *template* reply (not free-form model prose)
  billing → Action: create ticket + Delay/approval (HITL)
  human → Action: Slack #support-lab
Error: notify owner; do not retry money moves

ZAPIER_POLICY = {
    "ai_may_send_freeform_email": False,
    "hitl_before_crm_write": True,
    "task_budget_alert": True,          # qualitative ops hygiene
    "shadow_it_review": "quarterly",
    "pii_in_lab_zaps": False,
}

# vol23_handoff.py
def clone_vs_zapier(job: str) -> str:
    if job == "chat_ux_rag_eval":
        return "Vol. 23 ChatGPT Clone (you build it)"
    if job == "crm_ticket_glue":
        return "Zapier or n8n/Make (you configure it)"
    if job == "custom_sla_graph":
        return "Vol. 18 FastAPI + Celery"
    return "decide with residency + catalog + HITL"

When to Pick Zapier (Volume Close)

Pick Zapier when

The long-tail app is on Zapier and nowhere else.
Ops teams must self-serve with light training.
Enterprise SSO/SCIM + vendor familiarity unblock legal.
Task volume is modest and budgeted.

Pick something else when

VPC/self-host → n8n.
Heavy visual branching → Make.
Productized orchestration / SLAs → Vol. 18.
You are building the chat product itself → Vol. 23, not a Zap.

Vol. 22 Recap Map (Capstone)

Module	What you learned to pick
22.1 Frontier LLMs	OpenAI / Anthropic / Google / … as APIs
22.2–22.4 Media	Image / video / voice-music vendors
22.5 Platforms	HF / Replicate / Groq / OpenRouter / …
22.6 IDEs	Cursor → Replit: agentic vs Tab, privacy, when to pick
22.7 Automation	n8n self-host vs Make/Zapier SaaS; AI steps in graphs

Vol. 23 ChatGPT Clone will consume 22.1 models + 18.1 SDKs + 21.1 chatbot patterns. Zapier/n8n/Make will sit around that clone for ops glue—not inside the chat loop as the only architecture.

Related Lectures

Lecture	Role
n8n / Make	Self-host vs visual-SaaS peers
Workflow automation / chatbots	Product patterns Zapier wraps
Agentic workflow / HITL	AI actions vs free agents
FastAPI / Celery	When to stop using Zapier
ChatGPT Clone (Vol. 23)	Next volume: you build the app

Common Misconception

“Zapier Agents replace building a ChatGPT clone.” Different jobs: glue vs product UX/eval. Second: more tasks are always cheaper than n8n self-host—run the volume math (qualitative), do not invent $. Third: AI copilot that authors a Zap is the same as an AI action inside a run. Fourth: shadow-IT Zaps need no Vol. 20 review. Fifth: Zapier is self-host. Sixth: Vol. 22 ending means vendor choice is finished forever—re-read plan sheets each quarter.

Knowledge Check

Short Answer: What is a Zap, in one sentence? Answer: A cloud automation: trigger plus actions (filters/paths optional) metered mainly as tasks.
True/False: Zapier is the self-host option in the 22.7 trio. Answer: False—n8n is; Zapier is SaaS.
Multiple Choice: Zapier’s headline strength vs Make/n8n is: (a) widest connector catalog + simple UX, (b) CUDA training, (c) WebContainers. Answer: (a).
Short Answer: Name two distinct AI roles on Zapier. Answer: Any two of: copilot to build Zaps; AI action inside a run; agent SKU with tools (HITL).
True/False: Free-form AI email to customers should usually be replaced by templates + HITL. Answer: True.
Multiple Choice: Pricing posture is: (a) plan tiers + task metering (+ AI usage on some SKUs), (b) a memorized $19.99, (c) only self-host GPUs. Answer: (a).
Short Answer: When should Vol. 23 ChatGPT Clone use Zapier? Answer: As surrounding ops glue (CRM/tickets), not as the chat product itself.
True/False: This lecture invents official Zapier dollar prices. Answer: False—qualitative posture only.
Multiple Choice: Next volume lecture is: (a) ChatGPT Clone, (b) Midjourney, (c) Perplexity. Answer: (a).
Short Answer: Name the 22.7 pick triangle. Answer: Zapier (catalog/UX), Make (visual SaaS power), n8n (self-host/code)—plus Vol. 18 escape hatch.

Key Takeaways

Zapier closes Vol. 22: SaaS automation with unmatched catalog, task metering, and AI steps that still need HITL.
Choose Zapier vs Make vs n8n vs Celery using catalog, visual power, residency, and whether the graph is the product.
AI copilot (build time) ≠ AI action (run time) ≠ unbounded agent.
Your ChatGPT clone (Vol. 23) consumes 22.1 models; Zapier only wraps it for ops.
Next volume: ChatGPT Clone.

Trainer’s Guide

Capstone lab: Students produce a one-page vendor pick: one IDE from 22.6, one automation from 22.7, one LLM from 22.1, with qualitative pricing posture, privacy path, and HITL. No dollar figures. Dummy webhook Zap or sketch if accounts unavailable.

Handoff: Open Vol. 23 ChatGPT Clone. Ask: “Which 22.1 API + 18.1 SDK + 21.1 chatbot pattern do we implement first—and where would Zapier sit outside the chat loop?”

Recap: Zapier is Vol. 22’s capstone automation vendor—SaaS tasks, widest catalog, AI steps under HITL. The ecosystem catalog is complete. Continue to Vol. 23 ChatGPT Clone.