Zapier AI: AI by Zapier, Agents, Copilot and MCP

Video lecture

Zapier AI: AI by Zapier, Agents, Copilot and MCP

13 chapters · about 9 min · full transcript

Coming soon

Chapter 1 of 13

Zapier AI

AI by Zapier (AI step)
Copilot (builder)
Agents (goal-driven)
MCP + Human in the Loop

The narrated lecture is in production

Every chapter is scripted and ready. Browse the chapters and read the full transcript now — the video will appear here when it’s published.

Chapters

Transcript of the narration, chapter by chapter.

0:00 Zapier AI

What if your sales rep came in each morning to find every new lead already researched, summarized, checked against the CRM, and a tailored intro email drafted in Arabic and English, waiting for one click of approval? With Zapier's AI features, that's a few hours of setup, not a software project. In this lesson you'll learn AI by Zapier, Copilot, Zapier Agents, Zapier MCP and Human in the Loop, and how to combine them safely.

0:33 Lesson roadmap

Here's what we'll cover. First, why AI inside your existing app workflows matters. Then the lineup of Zapier's AI features and an analogy for choosing between them. Then how to configure AI by Zapier with structured outputs, how to set up agents and MCP safely, and two worked examples: an inbox assistant with human approval, and a lead research agent for a distributor in Riyadh.

1:01 Why it matters

Why does this matter? Because the hard part of AI in business isn't getting a model to write something. It's getting that output into the right app, at the right time, with the right checks. Zapier already connects thousands of apps. Adding AI steps, agents and approvals on top means you can put AI to work where your team already lives: the inbox, the CRM, Slack and spreadsheets.

1:31 The lineup

Here's the lineup. AI by Zapier is an AI step inside a Zap: instructions and inputs in, named output fields out. Zapier Copilot builds and edits Zaps from plain-language descriptions and can explain existing ones. Zapier Agents are AI agents with instructions that use your connected apps as tools, running on triggers or on demand. Zapier MCP exposes Zapier actions as tools to MCP-compatible AI clients. Human in the Loop pauses a Zap for approval or data collection. And there are chatbots, tables, interfaces and Canvas for process diagrams. Names change, so check the help center.

2:13 Analogy

An analogy for choosing between them. AI by Zapier is a specialist on an assembly line: one precise job at one station. An agent is a junior colleague with a goal and access to some tools, who decides the steps. MCP is giving your AI assistant a set of keys to specific rooms. And Human in the Loop is the manager's sign-off desk. For most business workflows, start with specialists on the line, add a sign-off desk, and only then consider junior colleagues with keys.

2:50 AI by Zapier, structured

Configure AI by Zapier with clear instructions, mapped inputs and named output fields with allowed values. For a Karachi web agency's inbox, the instruction might return a category from new project, support, billing, job application or spam; a service from website, SEO, social, branding or other; the language, English, Urdu or mixed; and a one-sentence summary with no personal data. Later steps map those fields directly, and filters and paths route on them. It's the same pattern you learned in n8n and Make.

3:26 Agents and MCP

Zapier Agents have instructions, tools, optional knowledge sources and triggers. Write instructions like a job description: the goal, the steps, the rules, what done looks like, and when to stop and ask. Give only the tools needed, prefer read actions, and route any write action through approval. Then test on real examples and review the agent's activity log. For Zapier MCP, you choose which actions to expose, connect your app accounts, and an AI client sees them as tools. Expose narrowly, confirm writes where the client supports it, and review usage. Every action uses tasks.

4:07 Example 1: inbox assistant

Example one, simple. An inbox assistant. A Gmail trigger catches messages to your hello address. AI by Zapier returns the category, service, language, a summary and a short polite reply draft with no promises on price or timelines. A filter keeps only new projects and support. Human in the Loop sends the summary and draft to a reviewer who can edit. If approved, Gmail sends the reply and HubSpot updates the contact. If declined, a task is created for a human to reply.

4:43 Example 2: Riyadh lead research agent

Example two, realistic. A B2B distributor in Riyadh triggers an agent whenever a new lead lands in HubSpot. The agent researches the company website, summarizes what they sell and their approximate size, checks whether they're already a customer, drafts a short Arabic and English intro email for their sector, and posts the summary and draft to the rep in Slack. It has web browsing, a read-only HubSpot search and Slack, but no email sending. After two weeks, the team compared agent summaries with rep research on thirty leads: sector was accurate, size less so, so they told the agent to mark size as unverified.

5:28 Common mistakes

Three mistakes to avoid. Letting agents send external emails directly in week one, before you've seen their work. Vague instructions like analyze this, instead of named fields with allowed values. And exposing every Zapier action through MCP just in case. Also remember governance: AI steps and agent actions consume usage, so watch the dashboards, keep personal data out of instructions where you can, and document each AI step's purpose, instruction version and owner.

6:00 Watch me do it: AI draft + approval

Watch me do it. I add the AI follow-up draft to our Zap. After the Paths I add AI by Zapier. Instructions: write a short follow-up email to this lead in their language, maximum one hundred and twenty words, reference their service interest and one detail from their message, no promises on price or timelines, sign off from the assigned owner. I define named outputs: subject, body and language. I map in the lead's name, message, service interest and owner name. The test output is decent but opens with a generic line, so I add an example of a good first line to the instructions and test again. Next, Human in the Loop, request approval. Reviewer: the assigned owner, notified in Slack. Data for review: the subject and body, editable, plus a context field with the lead summary. Timeout: if no response in two hours, send a reminder, and never auto-approve. On approval, a Gmail step sends the approved subject and body. On decline, a HubSpot task is created with the reviewer's reason. I run a live test, receive the approval card in Slack, edit one sentence, approve, and the email arrives in my test inbox. Finally, I add a Zapier Tables row for each draft recording whether it was edited, so we can track the edit rate.

7:35 Recap + try this now

Recap. Use AI by Zapier as a structured step, Copilot to build faster, Agents for goal-driven work with minimal tools, MCP to let AI assistants act through Zapier's catalog, and Human in the Loop to approve anything customer-facing. Try this now: build the inbox assistant from the lesson text with a Human in the Loop approval, run it for a week, and count how many drafts were approved without edits. That number tells you how close you are to safe automation.

8:10 Try this now

Try this now. Build the inbox assistant. Use a Gmail trigger for messages to your enquiries address. Add AI by Zapier with named outputs: category, service, language, summary and a short reply draft, each with allowed values where possible. Add a filter for new projects and support. Add Human in the Loop request approval, sending the summary and draft to a reviewer with edits allowed. On approval, send the reply and update your CRM; on decline, create a task. Run it for a week and track the approval-without-edit rate.

Zapier's AI layer

Zapier has added several AI capabilities on top of Zaps. Names and features change; verify in Zapier's help center.

Capability	What it does	Use for
AI by Zapier	An AI step inside a Zap: give instructions and inputs, get outputs (including structured fields)	Summarize, classify, extract, draft inside a Zap
Zapier Copilot	Builds and edits Zaps (and other Zapier assets) from natural-language descriptions	Faster building; explaining existing Zaps
Zapier Agents	AI agents with instructions that use connected apps as tools, run on triggers or on demand	Goal-driven work across apps (research leads, triage inbox)
Zapier MCP	Exposes Zapier actions across its app catalog as tools to MCP-compatible AI clients (for example Claude or ChatGPT)	Let AI assistants take real actions in your apps
Human in the Loop	Built-in tool that pauses a Zap for approval or data collection by reviewers	Review AI drafts before sending
Chatbots, Tables, Interfaces, Canvas	AI chatbots, data tables, forms and portals, process diagrams	Build small AI apps and document processes

AI by Zapier: a step with structure

Configure the step with clear instructions, the mapped inputs and named output fields (for example category, urgency, summary). Later steps map those fields directly, and Filter or Paths route on them. This mirrors the "AI as a step with a schema" pattern from n8n and Make.

Example instruction:

Classify this inbound message for a Karachi web agency.
Return:
- category: one of [new_project, support, billing, job_application, spam]
- service: one of [website, seo, social, branding, other]
- language: one of [en, ur, mixed]
- summary: one sentence, max 25 words, no personal data
Message: {{trigger.body}}

Zapier Agents

An agent has instructions, tools (actions in connected apps, searches, web browsing where enabled), optional knowledge sources, and triggers (run when a new lead arrives, on a schedule, or on demand). Good practice:

Write instructions like a job description: goal, steps, rules, what "done" looks like, when to stop and ask.
Give only the tools needed, and prefer read actions; route write actions through approval.
Test on real examples; review the agent's activity log.

Zapier MCP

Zapier MCP lets an MCP-compatible AI client call Zapier actions: "Create a HubSpot deal", "Send a Slack message", "Add a row to Google Sheets". You configure which actions are exposed and connect your app accounts; the AI client then sees those actions as tools. Benefits: huge app coverage without building integrations. Cautions: every exposed action is a capability an AI can use; expose narrowly, require confirmation for writes where the client supports it, and review usage logs. Actions consume Zapier tasks.

Worked example: a Riyadh B2B distributor's lead research agent

Trigger: new lead in HubSpot. Agent instructions: research the company website (web browsing tool), summarize what they sell and approximate size, check if they are an existing customer (HubSpot search), draft a short Arabic and English intro email tailored to their sector, and post the summary and draft to the sales rep in Slack. Tools: web browse, HubSpot search (read), Slack send message; no email send. The rep reviews and sends. After two weeks, the team compared agent summaries with rep research on 30 leads, found good accuracy on sector and weaker accuracy on size, and adjusted instructions to mark size as "unverified".

Hands-on: AI by Zapier + Human in the Loop

Trigger: Gmail "New email matching search" (for example to:hello@yourdomain.com).
AI by Zapier: instruction above; outputs category, service, language, summary, plus reply_draft (max 120 words, polite, no promises on price or timelines).
Filter: continue only if category is new_project or support.
Human in the Loop > Request approval: send summary and reply_draft to a reviewer; allow edits.
On approval: Gmail > Send email (reply to thread) with the approved text; HubSpot > Create or update contact.
On decline: create a task for a human to reply.

Costs and governance

AI steps and agent actions consume tasks or AI-specific usage depending on plan; monitor usage dashboards.
Keep personal data minimal in AI instructions; check Zapier's data handling and your plan's settings.
Document each AI step's purpose, instruction version and owner.

Pitfalls

Letting agents send external emails directly in week one.
Vague outputs ("analyze this") instead of named fields with allowed values.
Exposing every Zapier action through MCP "just in case".

AI by Zapier adds structured AI steps inside Zaps; Copilot builds and explains Zaps from natural language.
Zapier Agents use connected apps as tools; write instructions like a job description, give minimal read-first tools and approve writes.
Zapier MCP exposes selected Zapier actions to MCP-compatible AI clients; expose narrowly and review usage.
Human in the Loop pauses Zaps for approval or data collection, ideal for reviewing AI drafts before sending.

Build the inbox assistant (Gmail -> AI by Zapier -> Filter -> Human in the Loop -> Gmail + HubSpot). Run it for a week and record the share of drafts approved without edits.

Enrol for free to save your progress

Reading is always free. Enrol to keep your place, take the final assessment and earn a verifiable certificate.

Zapier AI: AI by Zapier, Agents, Copilot and MCP