Add safety checks to your workflows

When everyone around you is declaring that AI is making and breaking businesses overnight, it’s hard not to feel like you need to adapt now—or risk getting left behind. And by all means, build with AI. Just don’t forget to protect your business against the risks that come with it, like sensitive data ending up where it shouldn’t, harmful content reaching your customers, or bad actors manipulating your AI workflows.

AI-generated content isn’t the only thing that needs screening. User-generated content can carry the same risks. Without a check in place, that content can move downstream unchallenged and create the kind of mess that’s hard to walk back.

That’s why we built AI Guardrails by Zapier. Add it as a step to any Zap, Agent, or Zapier MCP workflow to scan for personally identifiable information, toxic language, prompt injection attempts, or negative sentiment—then route, block, or escalate based on what it finds. To learn more about how this safeguard works, keep scrolling.

Table of contents

What is AI Guardrails by Zapier?

Before AI Guardrails by Zapier existed, if you wanted to protect your AI workflows, you may have tried to stitch together third-party anonymization apps or write custom code. With this built-in Zapier tool, you can just plug a safeguard right into your existing workflows without any of the extra fuss.

Here’s how it works: AI Guardrails analyzes text—usually from the output of an AI step, but it works for human-generated content, too—then returns a structured result you can act on. Each AI Guardrails action checks your content against a specific risk category and tells you what it found, so you can either pass the data forward, flag it for human review, filter it out, or send it somewhere else entirely. The checks happen in real time, so nothing gets held up waiting for a manual review (unless you specifically want to build that in with a Human in the Loop step).

There are two types of AI Guardrails actions. When you’re building, you’ll notice they either have ML or LLM in their name. That’s because different kinds of AI are working under the hood depending on what the action needs to do.

• Machine learning (ML) actions use a pattern-based classifier, like a well-trained scanner. You give it text, it looks for known patterns, and then it returns a clear label along with a confidence score. These work best for policy-style checks with well-defined categories, like detecting PII or classifying sentiment.

• Large language model (LLM) actions use a more reasoning-based AI, closer to what powers modern chatbots. These are better at catching things that are context-dependent or deliberately evasive, like a prompt injection attempt buried in otherwise normal-looking text.

It’s worth noting what AI Guardrails is and isn’t. It’s designed to be one layer in a broader safety strategy, not a standalone compliance solution. No AI detection system is totally failproof. False positives and false negatives can happen, especially with sarcasm, unusual formatting, or novel attack techniques.

Key features of AI Guardrails by Zapier include:

• Personally identifiable information (PII) detection: Scans AI-generated text for over 30 types of PII—including names, addresses, government IDs, and financial account numbers—and returns a pass/fail result with the specific types detected.

• Prompt injection detection: Analyzes input text for attempts to manipulate AI model behavior, like jailbreak attempts or instructions designed to override your system prompt. Returns a detection status and recommended action.

• Toxicity detection: Screens content for harmful or toxic language using either AWS Comprehend (ML-powered) or Amazon Bedrock (LLM-powered), returning a toxicity score and the specific label types found.

• Sentiment detection: Determines the emotional tone of a piece of text—positive, negative, neutral, or mixed—with confidence scores for each category, so you can route content based on how it reads.

Keep in mind that the PII detection action currently supports English and Spanish only. Other actions may work in other languages, but coverage and accuracy may change as the underlying AWS services and models evolve. We recommend testing with your own content before publishing your workflows.

What you can do with AI Guardrails by Zapier

Here are some ideas for putting AI Guardrails by Zapier to work:

Detect PII in form submissions before logging them

You want to prevent personally identifiable information from being written into a shared spreadsheet—and alert the right person when something gets flagged.

What this might look like:

1. You receive a form submission through Typeform.

2. AI by Zapier summarizes the submission content.

3. AI Guardrails scans the summary for PII, then returns a pass/fail result with the specific PII types found.

4. A path step splits the workflow into two branches based on the result:

   1. Path A (PII detected): Gmail sends an email to the form owner with details about what PII was found.

   2. Path B (No PII detected): Google Sheets logs the summary to the sheet.

Block prompt injection attempts in user-submitted inputs

You want to prevent users from submitting manipulative instructions through a public-facing form that feeds into an AI model.

What this might look like:

1. A user submits a message through your Zapier form.

2. AI Guardrails analyzes the submission for attempts to manipulate AI model behavior and returns a detection status. If an injection attempt was detected, the workflow stops and the next steps do not run.

3. If no issues were detected, ChatGPT summarizes the submission.

4. The safe AI-generated summary gets sent to your Slack channel of choice.

Route negative sentiment from calls to your team

You want to automatically flag calls where the prospect or customer sentiment skews negative, so your team can follow up before the relationship sours.

What this might look like:

1. Zoom finishes generating a transcript after a call.

2. AI Guardrails determines the emotional tone of the transcript and returns confidence scores for each sentiment category.

3. A filter step stops the Zap from proceeding if sentiment is anything other than negative or mixed.

4. If the sentiment was negative or mixed, Gmail notifies the call host.

Moderate user-generated content before it goes live in your community

You want to screen member comments for toxic language before they’re visible in your community, routing clean content through and flagging anything harmful for review.

What this might look like:

1. A member of your Circle community submits a new comment.

2. AI Guardrails screens the comment for harmful language and returns a toxicity score.

3. Paths by Zapier splits the workflow based on the result:

   1. Path A (toxicity detected): Your moderation team receives a Microsoft Teams channel message, informing them that a comment was flagged for review.

   2. Path B (no toxicity detected): A new Asana task is created for the community manager, so they know to accept and respond to the comment.

How to get started with AI Guardrails by Zapier

To build a Zap using AI Guardrails by Zapier, follow these steps. (You can learn more about adding steps to an Agent or Zapier MCP in each tool’s respective feature guide.)

1. Log in to Zapier and head to the Zap editor.

2. Set up your trigger. This is typically an app that produces or passes content you want to check, like a form submission tool, CRM, or ticketing platform. That content can be AI-generated or human-generated.

3. If you want to detect issues with AI-generated content, click the plus sign (+) to add an action step, then choose your AI app—like AI by Zapier, OpenAI (ChatGPT), or Anthropic (Claude). In this step, the AI will do something with your trigger data, like summarize a form submission, draft a reply to a support ticket, generate a help article, and so on. The output of this step is what AI Guardrails will screen.

   Connect your account, configure the step, and test it before moving on. If you’re checking human-generated content, skip to #4.

4. Click the plus sign (+) to add another action step and search for AI Guardrails by Zapier. Select it, then choose the action event that matches your use case: Check for Personally Identifiable Information (PII), Detect Prompt Injection, Detect Toxicity, or Detect Sentiment.

   

5. In the Configure tab, map the text you want checked to the Text to Check field (or Text to Analyze, if you chose Detect Sentiment as your action event). Mapping data is easy—just click the plus sign (+) inside the field, then click the data from your previous step in the modal that appears.

   

6. Under Throw Error if…, select True to stop the workflow when an issue is detected or False to let the workflow continue regardless. You’ll still see the results in the action’s output either way. Click Continue when you’re done.

7. Add any remaining steps to control what happens after AI Guardrails runs. Use a filter step to continue the workflow only under certain conditions or a path step to send the workflow in different directions depending on what was detected.

When you’re done, remember to test and turn on your Zap.

Automate confidently with AI Guardrails by Zapier

AI Guardrails by Zapier makes it straightforward to add a layer of safety to any workflow. Just drop it into an existing workflow and let it do the checking for you.

Ready to get started? Visit the AI Guardrails by Zapier integration page for inspiration or our help docs. Or if you want to jump right into building, go to: