Zapier

I tested Zapier Copilot by asking it to build a real Gmail-to-Google Sheets workflow from a natural-language prompt. Copilot correctly chose Gmail's New Labeled Email trigger and Google Sheets' Create Spreadsheet Row action, then mapped the sender, subject, date, and plain-text email body to the appropriate spreadsheet columns. After connecting my accounts and selecting the spreadsheet, the workflow successfully wrote the test email to Google Sheets. I then asked Copilot to extend the existing workflow so that emails containing the word “collaboration” would also trigger an email notification. This exposed an important limitation. Copilot initially placed the filter before the Google Sheets step, which changed the original workflow logic and would have prevented non-collaboration emails from being saved. On the first correction attempt, Copilot described the correct order but did not actually update the Zap canvas. After a more explicit follow-up prompt, it correctly reordered the workflow to Gmail → Google Sheets → Filter → Email by Zapier. The final end-to-end test succeeded: Zapier found the labeled Gmail message, added a row to Google Sheets, passed the collaboration filter, and sent the notification email. Overall, Copilot was very effective at building a workflow from scratch, but more complex edits still required human review.

I tested AI by Zapier by asking it to analyze a real sample collaboration email pulled from Gmail. I wanted the AI step to return four structured outputs: a one-sentence summary, whether the message was a collaboration request, the company name, and a suggested professional reply. AI by Zapier automatically generated structured output fields from the prompt, including the summary, collaboration classification, company name, and suggested reply. My first preview failed because the Gmail subject and body references were not resolved correctly. After manually re-mapping the Subject and Body Plain fields from the Gmail trigger, the Standard model completed the task successfully. The final output was accurate. It correctly identified the message as a collaboration request, extracted NovaAI as the company name, summarized the email in one sentence, and produced a professional reply that was usable with minimal editing. I also tried to compare the Standard model with the Advanced and Premium tiers. However, even though I was using a Zapier Pro trial, Advanced and Premium still required upgrading to a paid plan. As a result, this hands-on test covers only the Standard tier.

I tested Zapier Agents by creating a custom research assistant for AItheMag. The agent was instructed to research an AI tool and return six sections: what the tool does, who it is best for, key features, main limitations, five article ideas, and five questions to investigate before a hands-on review. I also explicitly told it not to invent information and to flag uncertainty when necessary. For the first test, I gave the agent only one input: “Perplexity.” It followed the requested structure very well and used live Google Search during the research process. The output was detailed, readable, and particularly useful for generating article ideas and hands-on review questions. However, the first research brief contained several outdated or questionable claims. For example, it referred to Spaces as a current feature, said Perplexity had no voice mode, described its coding capability too narrowly, and included other limitations that conflicted with more current information available in its own search results. The activity log showed that the agent had used live web search, but it mixed official sources with third-party reviews and did not reliably resolve conflicts between them. I then asked the same agent to fact-check its previous answer again, prioritize official Perplexity sources, explicitly identify conflicting information, and label anything uncertain. The second pass was significantly better. It corrected the Spaces-to-Projects change, voice capabilities, coding claims, integrations, memory, and other outdated information. It also clearly separated claims into wrong, outdated, valid, partially outdated, and unconfirmed categories, and created a useful list of items that should still be manually verified before publication. Overall, Zapier Agents was strong at structuring research, using web search, generating editorial ideas, and correcting itself after detailed feedback. But its first-pass research was not reliable enough to publish without human fact-checking.

I tested Zapier Chatbots by building an AI Tool Finder for AItheMag and connecting five AItheMag tool pages as custom knowledge sources. The chatbot was instructed to recommend tools based on the user's goal, budget, experience level, and requirements, while avoiding unsupported claims or recommendations outside the connected knowledge base. In the first test, I asked for an easy, preferably free AI tool for creating university presentation slides. The chatbot correctly recommended Canva AI, explained why it was a good fit, included relevant pricing information and a limitation, and linked back to the AItheMag page. In the second test, I deliberately asked for an AI research tool even though the connected knowledge base did not contain a suitable research-focused tool. Instead of inventing a recommendation or relying on general model knowledge, the chatbot correctly returned the custom fallback message. The third test exposed a limitation. When I asked the broad question “What AI tool should I use for my business?”, the chatbot recognized that more context was needed and started asking about the user's needs, but it did not wait for an answer before recommending Zapier, Canva AI, and Make. This did not fully follow the instruction to ask clarification questions before making a recommendation. After I provided more context — a small online magazine, simple Google Sheets and email automations, low technical experience, and a budget under $20 per month — the chatbot improved significantly. It recommended Zapier as the best match, explained why it was easier for a non-technical user, compared it with the lower-cost but more complex Make alternative, and respected the stated budget. Overall, Zapier Chatbots was strong when answering from a controlled knowledge base and handling clearly defined user needs. Its main weakness was that it sometimes moved to recommendations too quickly instead of waiting for clarification on an ambiguous request.

I tested Zapier Tables' AI Field by creating an AItheMag content research table with structured fields for Tool Name, Category, Main Use Case, and Pricing. I then created an AI-powered Editorial Angle field that used those values to automatically generate a concise article angle for each record. I tested the field with Canva AI, Zapier, Make, and Clearscope. The AI Field generated an editorial angle automatically for every row without requiring a separate Generate action. The results were generally relevant to each tool and incorporated the supplied use case and pricing information. The strongest result was Canva AI, where the field turned the structured data into a useful Free vs. Pro editorial question without inventing specific features. Zapier and Make also produced relevant angles, although they occasionally introduced assumptions such as referring to small businesses even though that audience had not been included in the source fields. The Clearscope result remained usable even when the Main Use Case input contained a duplicated word, although the final wording became slightly more promotional than the prompt requested. I also tested whether the AI Field reacted to changes in source data. After changing Zapier's Main Use Case from a general automation description to a specific workflow involving form submissions, Google Sheets, and email notifications, the Editorial Angle automatically regenerated to reflect the new information. Finally, I removed Make's Pricing value to test how the field handled missing data. Instead of generating from the remaining fields, Zapier returned the error “All mapped fields must have a value.” This means every field referenced in an AI Field prompt must contain data before the AI output can run. Overall, Zapier Tables AI Fields were effective for turning structured records into automatically generated content, but they were less tolerant of incomplete data and occasionally added small assumptions beyond the source fields.

I tested Zapier MCP by connecting Zapier to Claude and asking the AI client to complete a controlled multi-app workflow across Gmail and Google Sheets. The initial ChatGPT setup path required ChatGPT Developer Mode and a custom MCP app, which I could not complete with my current ChatGPT Plus account. I therefore continued the hands-on test with Claude, where Zapier surfaced a native connector and completed the connection successfully. During the initial Zapier MCP onboarding flow, Claude used the connector to access Gmail and Google Sheets and generated a broad briefing from my inbox. This confirmed that the connection was working, but it also showed why narrow task instructions are important when giving an AI client access to connected apps. For the controlled test, I explicitly limited the workflow to one existing collaboration email with the subject “Potential collaboration with AItheMag.” Claude used Zapier MCP to find that specific Gmail message, create a Google Sheet named “AItheMag MCP Test,” add the requested Sender, Company, Subject, Date, and Summary columns, and insert one row containing the email information. It then created a Gmail draft reply on the original thread asking for product access, pricing details, and a press kit. The draft was not sent. Claude also confirmed that it did not substitute unsupported actions and that the controlled test only touched the specified email, the new sheet, and the draft. Overall, Zapier MCP successfully turned a single natural-language request into a real multi-app workflow with both read and write actions. The main caveats were the client-specific setup requirements and the need to define scope carefully when connected apps contain sensitive or broad datasets.