Zapier AI Prompts: 22 Templates for AI by Zapier and Chat Steps

TL;DR: "Zapier AI prompts" today mostly means the Prompt field inside AI by Zapier, the built-in step that absorbed the standalone Zapier Agents product on July 15, 2026. Type / in that field to pull in data from an earlier step. Below: verified plan gating, the actual field-mapping mechanic, and 22 prompts built for a machine that runs the same instructions thousands of times without anyone reading the output first.

A Zap doesn't read your prompt once and move on. It reads it on run 1, run 400, and run 4,000, feeding whatever the previous step returned into the same instructions every time. A prompt that's charming in a chat window and vague about edge cases will eventually meet a record that breaks it, at 2am, with nobody watching. That's the design constraint every template below is built around: reliability under repetition, not a clever one-off answer.

This post covers what Zapier's own AI offering actually looks like as of September 2026 (its naming here has genuinely shifted more than once this year), exactly how a prompt receives data from an earlier step, when to reach for a Zapier-native AI step versus a dedicated ChatGPT or Claude action, and 22 templates organized by the job the Zap is doing.

What is "AI by Zapier," and where did Zapier Agents go?

AI by Zapier is Zapier's current built-in AI step: a prompt field, a model selector, and optional tool-calling, added like any other action inside a Zap. Zapier describes it plainly: "AI by Zapier is a built-in tool that lets you add AI-powered steps to your Zaps, powered by popular AI models. You can use a built-in connection or connect your own AI provider account."

If you've read about Zapier Agents as a separate product living at agents.zapier.com, that description is accurate for the past, not the present. Zapier's own migration guide, dated July 15, 2026, states the change directly: "Zapier is migrating Agents (agents.zapier.com) to AI by Zapier, a more powerful AI experience built directly into the Zap editor." The reasoning given: "Agents ran as a standalone product at agents.zapier.com. Because it lived outside the core Zapier platform, you could not combine your agents with the full library of Zap triggers, actions, filters, branching logic, and automation history."

The practical result is that tool-calling, which used to require a standalone agent, now lives inside a single AI by Zapier step. Zapier's phrasing: "When you add tools to an AI by Zapier step, you create agentic workflows that reason and act autonomously, the same capability that previously required standalone Agents." The migration guide's own timeline table lists July 15, 2026 as the date "AI by Zapier with looping tool calls is generally available", with Enterprise trial customers given until August 15, 2026 to move their agents over before those trials converted to a free Agents account.

That's a meaningful naming shift inside a single year. If a tutorial, forum post, or brief mentions "Zapier Agents" as the current product, check its date before trusting the terminology.

Which Zapier plans include AI by Zapier?

Free does not. Zapier's own plan-availability graphic marks the Free plan with a grey X, and the underlying markup's CSS class for that state is literally plan-free-ineligible, on the AI by Zapier doc, the tool-adding doc, and the separate Zapier Tables AI-enrichment doc: three different features, the same gate, all excluding Free and including Professional, Team, and Enterprise.

Per zapier.com/pricing, accessed September 3, 2026: Free is "$0/month", described as "Free forever", and includes "Zap workflows, Tables, and Forms" with two-step Zaps and Zapier Copilot, but no multi-step Zaps and, per the plan gate above, no AI by Zapier. Professional starts at $19.99/month and is described as unlocking "the full power of the Zapier platform with unlimited access to Zap workflows, Tables, and Forms", listing "AI fields" among its named features. Team starts at $69/month, aimed at teams that "collaborate... to build and manage complete AI-powered systems", and adds 25 users and shared workflows. Enterprise is priced by sales conversation.

How does a prompt pull in data from an earlier step?

Two different mechanics, depending on where you're typing.

Everywhere in a Zap, generally: you map a field by clicking the plus-sign icon on the destination field and choosing from a dropdown of fields the previous steps returned. Zapier's own guidance on why this beats typing a value directly: "When you select a mapped field (shown as a colored pill), the Zap uses the actual data from each trigger event, so every run processes the correct record." That's the standard mechanic across every action in Zapier, AI or not: a pill, not a string you type.

Inside the AI by Zapier Prompt field specifically, there's a second, faster option: "Type / to insert a field from a previous Zap step inline in your prompt." That opens the same kind of field picker without leaving the text you're writing, so you can build a paragraph of instructions with a step's output dropped in mid-sentence rather than only at the start or end of the field.

Both produce the same underlying reference; the forward slash is a shortcut for inserting it inline rather than a different kind of token. One structural detail worth knowing if you're building anything that inspects field names programmatically: Zapier's enhanced field mapping treats any field name containing a double underscore as a nested-field reference, and there's currently no way to tell that apart from a field that genuinely has a double underscore in its own name.

If you've read Airtable's documentation and expect a typed {Field Name} token you can see and edit as text, Zapier's default mechanic isn't that. The pill is a UI object, not literal text in the field, which is a meaningfully different mental model if you're moving between the two products.

AI by Zapier vs. a dedicated ChatGPT or Claude action step

Both live inside the same Zap editor, and both can run a prompt. The difference is what's holding the connection and what you get beyond a single completion.

AI by Zapier offers a built-in connection to OpenAI, Anthropic, Google Gemini, Azure OpenAI, or Amazon Bedrock with no separate account required, plus a Bring Your Own Key option. It comes in model tiers. Zapier's own description: "Standard (1x tasks): Best for high volume, simple tasks. Tools are not available on this tier." "Premium (5x tasks): Complex analysis, tool usage, and multi-step reasoning. This is the default selection for new steps." Only Advanced and Premium tiers can call tools or read knowledge sources.

A dedicated ChatGPT (OpenAI) action is a separate app step with its own named actions. Per Zapier's own integration page, the three current options are: "Send Prompt", described as "Sends a prompt to OpenAI and generates a completion" (the simplest, one-shot option); "Conversation", marked "(Recommended)" and described as sending "a chat to OpenAI, optionally storing messages for continuous conversation... Powered by Responses API", which supports web search, file search, and tool use; and "Conversation (Legacy)", which "Sends a Chat to OpenAI and generates a Completion... Powered by Chat Completions API" for anyone still on the older interface.

A dedicated Claude action is narrower: a single "Send Message" action, described as one that "Sends a message to Claude and the model generates the next message in the conversation, storing the messages as you go", with User Message, System, Model, and Memory Key fields. Unlike AI by Zapier's built-in connection, using Claude this way requires your own Anthropic API key; there's no Zapier-hosted Claude connection on this specific action.

The honest rule: reach for AI by Zapier when you want tool use, a knowledge source, or you don't want to manage a separate API key. Reach for the dedicated action when you specifically need a provider's own model lineup and are already paying for that account directly.

22 prompts for AI by Zapier and chat-action steps

Every template below assumes the field placeholders are mapped pills from an earlier step, shown here in Make/Zapier-style double-brace notation ({{previous_step.field}}) purely so the template is readable as text; in your actual Zap you'll select the real field from the dropdown, or type / if you're inside an AI by Zapier Prompt field. Every prompt also states what to do with empty or malformed input, because a Zap that fails silently is worse than one that fails loudly.

Classify and route

1. Support ticket triage

Classify this support ticket for routing. No human reads this response first.

TICKET
{{trigger.ticket_body}}

Return exactly one JSON object, no other text:
"category": one of "billing" | "technical" | "account" | "other"
"urgency": one of "high" | "normal" | "low"

If TICKET is empty or unreadable, return:
{"category":"other","urgency":"normal"}

2. Inbound lead qualification

Score this inbound lead for routing to sales.

Company: {{trigger.company}}
Role: {{trigger.role}}
Message: {{trigger.message}}

Return exactly one JSON object:
"tier": one of "enterprise" | "mid_market" | "self_serve" | "unknown"
"reason": string, maximum 100 characters

If all three fields are empty, return {"tier":"unknown","reason":"no data provided"}.

3. Content moderation gate

Decide whether this user-submitted comment can publish automatically.

Comment: {{trigger.comment_text}}

Return exactly one word: "approve", "hold", or "reject".
"reject" only for content that is abusive, spam, or illegal.
"hold" for anything ambiguous, sarcastic, or that needs human judgment.
Never explain your choice; return only the single word.

Summarize and extract

4. Meeting notes to action items

Extract action items from these raw meeting notes.

Notes: {{trigger.notes}}

Return a numbered list. Each item: the action, the owner if named in the
notes, and nothing invented. If no owner is named for an item, write
"unassigned" rather than guessing who it might be.

5. Customer email summary

Summarize this customer email in one sentence for a support queue view.

Email: {{trigger.email_body}}

If the email is empty or contains only a signature/greeting with no actual
request, return: "No actionable content detected."

6. Document key terms extractor

Extract from the attached document: the effective date, the party names,
and the termination clause if one exists.

Document: {{trigger.attachment}}

Return three labeled lines. If a term is not present in the document, write
"not specified" for that line rather than omitting it.

Draft and generate

7. First-response support draft

Draft a first-response reply to this ticket. Acknowledge the specific issue,
do not promise a resolution date, and ask one clarifying question only if
the request is incomplete.

Ticket: {{trigger.ticket_body}}
Customer name: {{trigger.customer_name}}

8. Personalized welcome email

Compose a personalized welcome email using only the fields below. Do not
invent details about the company or role that were not provided.

Name: {{trigger.name}}
Company: {{trigger.company}}
Product selected: {{trigger.product}}

9. Social caption from a blog title

Write two short social captions (under 200 characters each) for this post,
one direct and one question-led. Base both on the actual title and summary
below; do not invent statistics or claims not present in the summary.

Title: {{trigger.title}}
Summary: {{trigger.summary}}

Enrich and complete

10. Company research from a name

Research this company using its name and domain, and return industry,
approximate employee count, and one-sentence description. If you cannot
find reliable information, say so explicitly rather than guessing.

Company name: {{trigger.company_name}}
Domain: {{trigger.domain}}

11. Missing field completion

Given the fields below, infer a reasonable value for the missing job title
field only if the available context makes it unambiguous. If it is not
clearly inferable, return "unknown" rather than guessing.

Name: {{trigger.name}}
Company: {{trigger.company}}
LinkedIn bio: {{trigger.bio}}

12. Sentiment and theme tagging

Tag this feedback with sentiment and up to two themes from this closed list:
"pricing" | "performance" | "missing feature" | "bug" | "onboarding" | "praise" | "other"

Feedback: {{trigger.feedback_text}}

Return JSON: {"sentiment":"positive|neutral|negative","themes":["..."]}
If nothing in the list fits, use "other" rather than leaving themes empty.

Format and transform

13. Convert notes into a structured table row

Convert these raw notes into a single structured record.

Notes: {{trigger.raw_notes}}

Return JSON with exactly these keys: "date", "topic", "owner", "status".
Use "unspecified" for any field not present in the notes. Do not add keys.

14. Standardize a name field

Reformat this name into "Last, First" format. If the input already contains
a company name or title alongside the person's name, separate it out and
return it in a second field.

Input: {{trigger.raw_name}}

Return JSON: {"name":"Last, First","extra":"anything not part of the name, or empty string"}

15. Translate while preserving tone

Translate this message into {{trigger.target_language}}, preserving the
original tone (formal, casual, urgent; infer from the source text itself).
Do not add explanations or notes about the translation.

Message: {{trigger.source_text}}

Multi-step, tool-using prompts (Advanced/Premium tier)

16. Research and summarize with web search

Research the company below using web search, and return its most recent
funding round, industry, and headquarters city. If you cannot verify a
detail with reasonable confidence, write "unverified" for that field rather
than presenting a guess as fact.

Company: {{trigger.company_name}}

17. Pull the latest doc from a knowledge source and summarize

Using the attached knowledge source, find the most recently updated document
matching the topic below and summarize its key points in 3 bullet points.

Topic: {{trigger.topic}}

If no matching document exists in the knowledge source, say so explicitly;
do not summarize an unrelated document instead.

18. Create a row and notify, in one agentic step

When a new item matching {{trigger.criteria}} appears, create a record in
the connected table with fields "name", "status", and "flagged_reason", then
send a notification only if "flagged_reason" is non-empty. Do not send a
notification for routine, non-flagged items.

Guardrails and fallbacks

19. The universal empty-input guard

If the input below is blank, whitespace-only, or clearly not the expected
content type, return exactly: {"status":"no_input","value":null}
Otherwise, process normally and return {"status":"ok","value":"<your answer>"}.

Input: {{trigger.field}}

20. The universal fourth-category guard

Classify the input below into exactly one of: "A" | "B" | "C" | "other".
If the input plausibly fits two categories, or fits none of A/B/C, return
"other" rather than forcing a fit. Never invent a category outside this list.

Input: {{trigger.field}}

Chunking large input (Formatter + AI by Zapier)

21. Chunk-and-summarize for long documents

Summarize this chunk of a larger document in 2-3 sentences, focused only on
new information not likely covered in earlier chunks (assume this is chunk
{{trigger.chunk_index}} of {{trigger.chunk_total}}).

Chunk text: {{trigger.chunk_text}}

22. Final rollup after all chunks processed

Combine the chunk summaries below into one coherent summary of the full
document. Remove redundancy between chunks; do not simply concatenate them.

Chunk summaries: {{trigger.all_chunk_summaries}}

What actually limits reliability here?

A few constraints are worth planning around before you wire a prompt into something that runs unattended:

• No memory between runs. Zapier is explicit: "AI by Zapier does not learn from previous actions. Each time an AI by Zapier action runs, the AI knowledge is reset to its initial state." Anything the step needs has to be mapped in fresh or attached as a knowledge source for that run. It will not recall what happened on the previous record.
• No live web fetch by default. Zapier's own limitations list states it flatly: "AI by Zapier cannot search and extract information from websites or any URLs." Reaching a live page needs a tool added to the step, not the base prompt alone.
• Knowledge sources cap out. "You can include up to 20 knowledge sources in each AI by Zapier step", and most connected sources (Google Drive, Dropbox, SharePoint, Google Docs, uploaded files) share a "Maximum synced words: 96,000" ceiling. Notion is narrower still: only individually added pages, no databases, capped at "50 most recent pages" for Confluence Cloud specifically.
• Long input needs chunking. Zapier's Formatter app ships a beta action, Split Text into Chunks for AI Prompts, built for exactly this: breaking a large document into pieces sized to a chosen model and prompt before running each chunk separately, which is what templates 21 and 22 above are built around.

For the same discipline applied to a different automation platform, the Make.com scenario templates and the n8n node templates cover output contracts for the equivalent ground there. If your team's prompts currently live in a shared spreadsheet rather than inside the Zap itself, why prompt library spreadsheets break is worth reading before the next one gets built. And when a model returns a differently-shaped answer on run 400 after 399 clean ones, why you get a different answer every time covers why temperature alone won't fix it.

Sources

All verified September 3, 2026.

• Zapier Help Center, Use AI by Zapier to analyze and return data: plan gating, model tiers, prompt field mechanics, knowledge sources, limitations.
• Zapier Help Center, Add tools to your AI by Zapier step: tool-calling, model tier requirements.
• Zapier Help Center, Migrating from Agents to AI by Zapier: migration rationale, dates, task-usage guardrails.
• Zapier Help Center, Send data between steps by mapping fields: the pill-mapping mechanic, nested-field double-underscore behavior.
• Zapier Help Center, How to prompt AI in Zapier products: the full list of Zapier surfaces that take a prompt.
• Zapier Help Center, Modify large data for your AI prompts: Formatter chunking action.
• Zapier Help Center, Enrich records using AI in Zapier Tables: corroborating plan gate.
• Zapier, ChatGPT (OpenAI) Integrations and Anthropic (Claude) Integrations: exact action names and descriptions.
• Zapier, Pricing: current plan names and prices.