Zapier

From PRIMO.ai
Jump to navigation Jump to search

YouTube ... Quora ...Google search ...Google News ...Bing News


Zapier is a no-code automation platform that connects web applications so that an event in one app can automatically start work in others. Founded in 2011 in Columbia, Missouri, the company now describes itself as an AI orchestration platform and advertises connections to more than 9,000 apps and tens of thousands of triggers and actions, used by millions of people. The basic unit of automation is the Zap, a workflow that begins with a trigger (for example, a new row in Google Sheets or a new form submission) and continues through one or more actions (for example, creating a CRM record and posting a Slack message). Builders assemble Zaps in a visual drag-and-drop editor and can customize them with webhooks, conditional logic through Filters and Paths, data cleanup through Formatter, Code steps, and custom apps, all without writing a full integration.

Zapier belongs to the integration-platform-as-a-service (iPaaS) category and overlaps with robotic process automation, although it works mainly through application APIs rather than by imitating clicks on a screen. Usage is measured in tasks: each successful action step normally counts as one task, triggers and built-in logic steps such as Filter, Paths, and Formatter do not count, and AI steps can consume several tasks per run. The Free plan is limited to two-step Zaps and 100 tasks per month, and paid plans started at $19.99 per month for the Professional tier as of early 2026, with larger task allowances priced in bands.

Over the past several years Zapier has layered generative AI throughout this platform. As of late 2026 its AI features fall into a few families: AI by Zapier adds model-powered steps inside Zaps; Zapier Copilot helps people build Zaps from plain-language descriptions; Zapier MCP, the Zapier SDK, and the Zapier CLI let outside AI agents call Zapier's app actions; and AI Guardrails by Zapier screens content for privacy and safety risks. The most important recent change is consolidation: Zapier is migrating its standalone Agents product into AI by Zapier, within the main automation editor, so that agent-style reasoning and tool use now run as an ordinary Zap step. This page focuses on AI by Zapier, explaining what it is, how to set it up, and how to use it well, and then places it alongside Zapier's other AI offerings.

Load video
YouTube
Load video
YouTube


AI by Zapier turns agentic AI into an ordinary, governed workflow step: the model reasons and calls tools, while triggers, filters, approvals, task caps, and run history keep its behavior predictable and auditable.


AI by Zapier

What AI by Zapier Is

AI by Zapier is a built-in Zapier app whose main action, Analyze and Return Data, sends a prompt to a large language model and returns the result as data that later steps in a Zap can use. Without it, a Zap runs on fixed if-then logic; with it, a workflow can read unstructured text, make a judgment, and pass structured answers downstream. Typical jobs include extracting fields from emails or documents, classifying incoming requests, summarizing long content, drafting replies, and analyzing images or files supplied by URL. Because the step is native, users do not need their own API keys: Zapier provides access to models from OpenAI, Anthropic, Google Gemini, and Azure OpenAI from Microsoft, while organizations that prefer their own accounts can bring their own key, including through Amazon Bedrock.

Zapier distinguishes between two ways of using the same step. A standard AI step reasons over its input and produces one output, much like a single request to a chatbot. An agentic AI step is given tools, such as Zapier app actions, connected knowledge sources, or web browsing, and decides for itself which tools to call and in what order before returning an answer, a pattern of model-driven tool use that research such as Toolformer helped popularize. Tools are optional, so the same step can behave as a simple completion or as a small agent depending on how it is configured. In Zapier's framing, the tool calling, reasoning, and autonomous action that once required the separate Agents product now live inside a single AI step.

Several features shape what the step can do. Output fields let a builder define named, typed results (such as category, sentiment score, or subject line) that arrive as separate mappable values, and Zapier can generate these fields automatically from the prompt. Knowledge sources ground the model in company material from services such as Google Drive, Google Docs, Notion, Confluence, Box, Dropbox, SharePoint, uploaded files, and Zapier Tables, an approach similar in spirit to retrieval-augmented generation; each step can reference up to 20 sources. The step is stateless between runs, and Zapier notes that it does not learn from earlier actions, so any memory has to be supplied through mapped data, tables, or knowledge sources. Content moderation is always on for Zapier-provided models, and a step can fail if its content violates the usage policies of Zapier's AI subprocessors.

From Zapier Agents to AI by Zapier

Zapier's earlier agent product, Zapier Agents, grew out of the Zapier Central beta and ran as a separate application at agents.zapier.com, where users described an agent's job in plain English, attached tools, and paid for usage in a separate unit called activities. Zapier has acknowledged that the separation became the problem: because Agents sat outside the core platform, an agent could not use Zap triggers, filters, paths, or run history. During 2026 Zapier began folding that product into AI by Zapier, and its migration timeline listed July 15, 2026 as the date AI by Zapier with looping tool calls became generally available. The migration guide, last updated October 1, 2026, and an October 6, 2026 announcement titled "Zapier Agents is now AI by Zapier" map each Agents capability to its new home in the Zap editor. For a broader comparison of agent-building platforms, see ...build an agent.

Migration is largely automatic. Zapier converts each eligible agent into a Zap containing a native trigger and a single AI by Zapier step that holds the original prompt, instructions, connected tools, and reasoning; tool calls are not split into separate native steps unless the builder moves them out by hand. Users open the converted Zap from their agent list, review the AI step, run a preview, test the whole Zap, and publish it, and Zapier offers to turn off the original agent at that point, since leaving both running can produce duplicate actions. A bulk option migrates all published agents at once, and super admins can publish migrated Zaps on behalf of users, in which case live actions such as sending email run under that user's connected accounts. Zapier has not set a shutdown date for the standalone product and says it will email affected users in advance.

Zapier Agents feature Equivalent in AI by Zapier
Agent A Zap with a native trigger and one AI by Zapier step
Instructions The prompt field of the AI step
Tools The tools picker inside the step (tools can optionally be moved into native steps)
Scheduled or webhook start Any Zap trigger across Zapier's app library
Chat-triggered (on-demand) agent A webhook trigger that any chat surface can call
One agent calling another A Sub-Zap by Zapier trigger on a child Zap
Human intervention "Require approval before running," set per tool
Model choice Standard, Advanced, or Premium tier, or a specific model
Run history Zap history, with per-run tool and task details
Activities (billing) Tasks, under Zapier's standard task-based pricing

Not everything has carried over yet. As of the October 2026 guide, knowledge sources from HubSpot, Asana, Zendesk, Zoom, and Jira were not supported (Zapier targeted Q4 2026), admin-level bring-your-own-model controls were listed for a future release, and on-demand agents had no direct trigger beyond the webhook workaround. Notion knowledge sources work with limits: sources with 1,000 or more pages can time out when added, older or long-tail pages can be hard to find, and the connected account needs access to every page. Teams that relied on any of these behaviors should test migrated Zaps carefully before turning off the original agents.

Setting Up an AI by Zapier Step

AI by Zapier requires a paid plan for production use. Professional, Team, and Enterprise plans include all model tiers and tools, while Free accounts can add the step and preview results with the Standard tier but cannot add tools or publish a Zap that contains it. The steps below follow Zapier's help center documentation as updated in October 2026; editor labels change from time to time, so the exact wording on screen may differ slightly.

  1. Build and test the trigger. Create a new Zap (or open an existing one), configure the trigger app and event, and test it so the editor has a real sample record to work with.
  2. Add the AI step. Click the + icon to add an action, search for AI by Zapier, and select it. The step opens with a Configure panel (prompt, model, and settings) and a Preview panel.
  3. Write the prompt. Start from a template (Summarize, Write, Classify, or Extract) or write instructions from scratch. Type / to insert fields from earlier steps so each run works on the current record, and mark an input as an image or file when the step should analyze content at a URL.
  4. Choose a model. Pick a tier (Standard, Advanced, or Premium) and either a specific model or Auto, which uses the tier's default model. Choose Bring Your Own Key to run on your own provider account.
  5. Define outputs. Under Settings, add output fields with a name, type, optional description, and optional required flag, or click Generate from prompt. Turn on "Return a list of results" when the step should produce line items; without output fields, the step returns one combined output.
  6. Add tools and knowledge (optional). Expand Tools and Knowledge, switch on Browse the web, or click + Add tool and choose an App action or a Knowledge source. For each app action, select the app, action, and connected account, then decide how the AI handles each input field.
  7. Set guardrails. Turn on "Require approval before running" for tools that create, update, or delete data, and review the step's task limit (75 tasks per run by default, adjustable up to 500).
  8. Preview, test, and publish. Click Preview, check the output and the Tools used list, refine the prompt with preview regeneration, then click Finish. Test the full Zap with realistic data before publishing, keeping in mind that full-Zap tests consume tasks just like live runs.

When a tool is an app action, each of its input fields can be handled in one of four ways. "Agent determines field" lets the model write the value from the prompt; "Set specific field" locks a fixed or mapped value; "Let agent choose from list" restricts the model to a defined set of options; and "Exclude field" hides an optional field from the model entirely (excluding a required field makes the tool call fail). These choices are the main lever for keeping an agentic step predictable, because a locked or constrained field cannot be steered by an unexpected prompt or input. A common pattern is to lock destinations, such as the Slack channel or CRM pipeline, and let the model fill in only the content, and builders who are unsure how to phrase a prompt or which tier to pick can click Ask Copilot inside the step for suggestions.

Model Tiers, Tool Calls, and Task Costs

Since June 15, 2026, AI by Zapier steps have been priced by model tier. The tier sets a multiplier on the step's base cost, and every successful tool call is charged at the same multiplier, so a tool-heavy step can consume many tasks in a single run. The table below summarizes the tiers as Zapier documented them in October 2026.

Tier Tasks per run (multiplier) Tools and knowledge Intended use
Standard 1 Not supported High-volume, simple prompts such as extraction or classification
Advanced 3 Supported Stronger reasoning for complex prompts and tool use
Premium 5 Supported Complex analysis and multi-step reasoning
Bring Your Own Key 1 Supported when an Advanced- or Premium-class model is connected Runs on your own provider account; content moderation can be turned off

The cost of a run follows a simple rule: tasks = tier multiplier × (1 + number of successful tool calls). A Premium step that calls two tools therefore uses 5 × 3 = 15 tasks, and an Advanced step with one tool call uses 3 × 2 = 6 tasks. The default cap of 75 tasks per run works out to a tool budget of about 24 calls on Advanced, 14 on Premium, and 74 with your own key before the run pauses for review. Zapier also changed the default tier for new steps during 2026 (its documentation has described both Advanced and, more recently, Premium as the default), so builders should confirm the selected tier on every new step rather than assume the cheapest option. Each run's breakdown by tool call, tier, and step appears in Zap history.

Approvals, Monitoring, and Safety

Because an agentic step can act on real systems, Zapier wraps it in several human-in-the-loop controls. Approval is set per tool rather than per agent, so read-only lookups can run on their own while actions that create, update, or delete data wait for a person. A step pauses when a tool that needs approval is about to run, when the model needs more information, or when the run reaches its task limit; the step then shows as Delayed on the canvas, Zapier sends an email, and a reviewer approves or rejects the action from the step's Activity tab in Zap history, with a rejection stopping the run. After any run, the same Activity tab lists the model tier used and each tool the step called, what it did, and whether it succeeded, which makes it the first place to look when results are unexpected.

Organizations also get account-level governance controls. Admins can require approval before any Zap containing AI is published, turn tool calling and agentic behavior on or off across the account, and restrict which models builders may use; organizations that had previously blocked Agents start with tool calling turned off in AI by Zapier. On the data side, prompts are processed by large language models from Zapier's AI subprocessors, and knowledge sources stay in the Zapier account until the content or the Zap is deleted. Prompt versions are kept only during an editing session, so teams that iterate heavily on prompts may want to save important versions elsewhere.

AI Guardrails by Zapier, introduced in early 2026 and free on all plans, adds a separate security layer that can sit before or after an AI step. Its actions detect more than 30 types of personally identifiable information, flag prompt injection and jailbreak attempts, screen for toxic language, and score sentiment, returning structured results that Paths and Filters can use to block, redact, or escalate content. A typical design checks untrusted input, such as an incoming email or form submission, for injection attempts before it reaches the AI step, and checks the AI's output for personal data before it is written to a CRM, an applicant tracking system for HR, or a customer message. Zapier describes Guardrails as one layer in a broader safety strategy rather than a complete compliance solution: detection can miss content or produce false positives, PII detection currently supports only English and Spanish, and checked data is still stored temporarily in Zapier's logs (7 days) and Zap runs (29 to 69 days).

How to Best Use AI by Zapier

Zapier's documentation, its product blog, and practitioners who build with the step converge on a consistent set of best practices. The common thread is to treat the AI step as one carefully scoped component in a deterministic workflow, not as a replacement for the workflow itself. The practices below run roughly from design decisions to day-to-day operations.

  • Use AI only where judgment is needed. Let the AI step handle reading, classifying, or reasoning, and use ordinary actions for predictable work such as sending a confirmation email or creating a record. This keeps costs down and makes failures easier to trace.
  • Give each AI step one job. Asking a single step to classify, summarize, and draft a reply at once tends to produce mixed results; separate steps (or separate output fields) are more reliable.
  • Ask for structured output. Define output fields such as category, priority, or confidence so downstream Paths, Filters, and app actions receive clean values instead of prose that must be parsed.
  • Right-size the model tier. Standard models are usually enough for extraction and classification; reserve Advanced or Premium for tool use and multi-step reasoning, and check the default tier on new steps.
  • Constrain tools. Add only the tools a step needs, lock fields that must not change, offer fixed choice lists where possible, and require approval for any tool that writes, updates, or deletes data.
  • Ground answers in your own material. Connect knowledge sources such as policy documents or product FAQs rather than relying on the model's general knowledge, and respect source limits (20 per step and roughly 96,000 synced words for most source types).
  • Treat outside text as untrusted. Emails, form entries, and web pages can carry instructions aimed at the model, so screen them with AI Guardrails and keep a human approval before irreversible actions.
  • Test with real data. Use preview regeneration to iterate on prompts quickly, then test the full Zap with a batch of real records (some practitioners suggest at least ten) before publishing.
  • Set budgets and monitor. Keep the per-run task cap at a level that matches the job, review per-run task breakdowns in Zap history, and check the Activity tab when outputs drift.
  • Let Copilot help. Zapier Copilot can draft Zaps from a plain-language description and suggest prompt wording or tier choices from inside the step.

Good prompt engineering matters as much here as anywhere else. Effective prompts state the goal clearly, map in only the fields the model needs, define the allowed labels or format for each output, and say what to do when information is missing (for example, return "unknown" rather than guess). Because each run is independent, prompts should carry any context the model needs every time instead of assuming it will remember earlier records.

Zapier's published examples show how tier choice should track task complexity. Simple extraction runs well on a Standard model, while workflows that must look things up, consult policies, or browse the web need an Advanced or Premium model with tools. The examples below combine Zapier's feature guides with typical business uses in marketing, sales, support, and finance.

Use case Typical flow Suggested tier
Data capture A status update posted in a team channel → AI extracts the project name, status summary, and owner → a row is added to a tracking spreadsheet Standard
Customer support New help desk ticket → AI classifies the request and drafts a reply from a knowledge base → the ticket is routed to the right team for human review Advanced
Deal intelligence Sales call transcript → AI extracts structured fields and logs them to the CRM → a follow-up summary is posted to a sales channel Premium
Refund handling Customer request → AI looks up the order, checks it against the refund policy, and drafts a response, with approval required before any refund action Advanced or Premium
Lead research New lead → AI browses the web for company information, finds the account in HubSpot, and posts a brief to Slack Advanced or Premium
Invoice matching New invoice → AI matches it to a purchase order in Airtable and posts mismatches to Slack Advanced

Limitations

AI by Zapier inherits the strengths and weaknesses of the models behind it. Outputs are probabilistic, so the same input can produce slightly different results across runs, and models can still misclassify, omit details, or state errors confidently; workflows that touch customers or money should keep human review in place. Costs are easy to underestimate, because tier multipliers apply to every tool call and a Premium step costs five times as much as a Standard one. The step also has no memory across runs, relies on the Browse the web tool (or the app's separate URL-extraction action) rather than its prompt to read web pages, and can analyze images only when they sit at a publicly accessible URL.

There are also open questions about the platform's direction. The migration from Agents is still in progress, with some knowledge sources and admin features pending, and Zapier is simultaneously piloting Next Gen Zaps, which let AI clients build entire workflows, so the boundaries between AI steps, agents, and AI-built workflows may keep shifting. Most detailed information about AI by Zapier comes from Zapier's own documentation and from consultants and partners, and independent, systematic evaluations of its accuracy and reliability are scarce, so teams should run their own verification and validation before relying on it for consequential decisions. Organizations with strict requirements should also review which AI subprocessors handle their data and confirm that their regulatory and compliance obligations are met.

Recent Developments

Zapier's AI features changed rapidly during 2025 and 2026, and older tutorials often describe products or screens that no longer exist. The milestones below are drawn from Zapier's announcements and help center. Readers following a video or blog guide should check its date against this timeline before assuming the interface will match.

  • March to April 2024: OpenAI retired ChatGPT plugins, ending the Zapier ChatGPT plugin; Zapier pointed users to AI Actions for custom GPTs (see Zapier ChatGPT Plugin below).
  • March 2025: Zapier launched Zapier MCP, letting MCP-compatible assistants run Zapier actions; it later became the recommended replacement for the deprecated AI Actions.
  • February to March 2026: Zapier added more than 80 free agent templates, model choice, and shared prompts, and introduced AI Guardrails by Zapier, which was formally announced at the end of March 2026.
  • June 15, 2026: AI by Zapier moved to model-tier pricing, and new steps gained tools (app actions, knowledge sources, and web browsing) with per-tool approval.
  • July 15, 2026: AI by Zapier with looping tool calls became generally available, according to Zapier's migration timeline.
  • September 23, 2026: At ZapConnect 2026, Zapier unveiled Next Gen Zaps in early access.
  • October 2026: Zapier updated its Agents migration guide (October 1) and published "Zapier Agents is now AI by Zapier" (October 6), describing the conversion of standalone agents into Zaps.

Related Zapier AI Offerings

AI by Zapier is one of several AI products that share Zapier's app connections, credentials, and task billing. Choosing among them depends mainly on where the model runs and who starts the work: inside a Zap on every trigger, inside an outside AI assistant on request, or inside code that an agent writes. The table below summarizes the main options before the sections that follow describe them in more detail.

Offering What it does Best fit
AI by Zapier AI step inside a Zap, optionally with tools and knowledge Repeatable, triggered workflows that need judgment
Zapier Copilot Assistant that drafts and edits Zaps from plain-language descriptions Building and troubleshooting automations
Zapier MCP Exposes Zapier actions as tools to MCP clients such as Claude, ChatGPT, and Cursor One-off actions requested in chat
Zapier SDK and CLI Code and terminal access to Zapier's app actions Developers and coding agents
Next Gen Zaps (early access) AI-built workflows with loops, unlimited paths, and code, managed in a Workflow Manager Complex automations built from AI tools
AI Guardrails by Zapier Detection of PII, prompt attacks, toxicity, and sentiment Safety checks around AI steps and MCP actions
Chatbots, Tables AI fields, Canvas Customer-facing chatbots, per-column AI prompts in Zapier's database product, and AI-assisted workflow diagrams Specialized AI features within Zapier's own apps

Zapier Copilot

Zapier Copilot is an AI assistant that builds Zaps from a description of the desired workflow. It outlines a trigger and actions, then helps configure, test, and refine each step, either automatically ("Auto-build," the default) or by asking for confirmation as it goes ("Ask as you build"). Copilot complements AI by Zapier rather than replacing it: Copilot helps a person build the automation once, while an AI by Zapier step runs a model every time the Zap is triggered.

Zapier MCP, SDK, and CLI

Zapier MCP is Zapier's implementation of the Model Context Protocol (MCP), the open standard that Anthropic introduced in November 2024 for connecting AI assistants to external tools. A user creates a server at mcp.zapier.com, chooses which apps and actions to expose, and connects an MCP client such as Claude, ChatGPT, or Cursor; the assistant can then search and act across the user's connected apps in conversation, while Zapier manages credentials, rate limits, and an action history. Zapier MCP is available on all plans and costs two tasks per tool call, and Zapier cites access to more than 9,000 apps and over 30,000 actions (some Zapier pages give higher counts). For developers, the Zapier SDK (in open beta) lets coding agents call Zapier actions with loops, conditionals, and error handling, and the Zapier CLI lets builders explore apps and run actions from a terminal.

Next Gen Zaps

Next Gen Zaps, unveiled at ZapConnect on September 23, 2026, extend the idea that AI tools can build automations rather than only run individual steps. Users describe a workflow inside an MCP-enabled assistant such as Claude, ChatGPT, or Cursor and deploy it to Zapier, or build directly in Zapier with natural language, code, or a visual canvas; the resulting workflows support multiple loops, unlimited paths, and code editing, and are monitored in a new Workflow Manager with step-level observability and governance. The feature is in early access on paid plans, tasks are free while it remains in early access, and Zapier describes it as having limited availability and active development. Third-party reviewers advise keeping sensitive production workflows on classic Zaps until general availability, because some enterprise and compliance features do not yet apply.

Zapier ChatGPT Plugin


Historical integration: This section documents Zapier's earlier ChatGPT plugin.
For current agentic automation capabilities and setup instructions, see AI by Zapier above.


Zapier ChatGPT plugin was a collaboration between OpenAI and Zapier, released in 2023 as one of the first ChatGPT plugins, that let users connect more than 5,000 apps (like Google Sheets, Gmail, or Slack) and run actions in them directly inside ChatGPT's interface. OpenAI wound down the plugin beta in 2024: users could not start new plugin conversations after March 19, 2024, and existing plugin conversations stopped working on April 9, 2024. Zapier directed users to AI Actions, which gave custom GPTs access to Zapier searches and actions, and later deprecated AI Actions in favor of Zapier MCP, which now serves the same purpose for ChatGPT, Claude, and other MCP clients. The plugin remains historically significant as one of the earliest mainstream examples of a chatbot taking real actions in business software, an idea that agent frameworks and Zapier's own MCP server have since generalized.

Load video
YouTube

Further Reading and Viewing

References

Additional Viewing

Load video
YouTube

In "Zapier AI Tutorial for Beginners: Automation Made Simple," technology educator Kevin Stratvert walks through creating a free account, using AI assistance to start a new Zap, connecting Google Forms and Gmail, cleaning data with Formatter, and adding an AI step that drafts automated email responses (the video is sponsored by Zapier). Its roughly 20-minute, chaptered format makes it a useful companion to the setup steps in this article, especially the trigger, field-mapping, and testing stages. Viewers should expect the AI step's configuration panel to differ from the current model tier and Tools and Knowledge layout described above.

Load video
YouTube

"Zapier AI Tutorial for Beginners (ChatGPT, Claude, Gemini)" demonstrates connecting OpenAI, Anthropic, and Google models to Zapier with API keys and building AI-powered workflows, with attention to prompting and cost. It complements the section on model tiers and task costs by showing the alternative of calling provider integrations directly, the same trade-off that the Bring Your Own Key tier now offers inside AI by Zapier. Comparing its approach with the built-in step helps clarify when a dedicated provider integration still makes sense.

Load video
YouTube

"Zapier MCP Demo: Connect AI Assistants to Nearly 8,000+ Apps Effortlessly!" shows how an AI assistant such as Claude, ChatGPT, or Cursor can be connected through Zapier MCP to everyday tools such as Google Calendar and Notion. It illustrates the chat-driven alternative to AI by Zapier described in the Zapier MCP section, where the model runs in an outside client and calls Zapier actions on request. The app count in its title reflects an earlier stage of Zapier's catalog, which the company now puts above 9,000.

Load video
YouTube

In "How to Use Zapier Next Gen Zaps (Step-by-Step Tutorial)," Kevin Stratvert demonstrates the early-access Next Gen Zaps experience announced at ZapConnect 2026, in which users describe what they want to happen and AI builds the automation. The video supports the Next Gen Zaps section by showing how AI-built workflows differ from adding an AI step to a classic Zap. Because the feature is still in early access, details shown may change before general availability.

Retrieved from "https://primo.ai/w/index.php?title=Zapier&oldid=53485"