Automate more with AI

Relay.app’s built-in AI actions, Custom Prompt steps and mini Agents make it easy to leverage AI in a predictable, controlled way.

We firmly believe that the easiest and most impactful way for most teams to leverage the power of AI in their work is by incorporating AI in repeated automations. The challenge is often that the output of AI is not always predictable. In Relay.app, you’ll find various ways of incorporating AI in your workflows in a predictable, controlled way.

Our AI features are particularly powerful when paired with Human-in-the-Loop steps. You get all the benefits of AI-powered automation while keeping the control and oversight you need to feel confident.

3 classes of AI features

To make it really easy to use AI in your workflows, we offer 3 classes of AI features in Relay.app. Each of them optimized for ease-of-use and predictability of the AI output, while increasing the amount of flexibility in each feature category.

  1. Steps for built-in AI Actions for common use cases like summarization and data extraction
  2. A ‘Custom Prompt’ step with structured output fields for total flexibility
  3. The ability to create mini AI Agents that semi-autonomously perform actions on your behalf

All of the above step types can be found under the ‘AI’ submenu in the step list, and can be used with the AI model of your choosing (learn more).

Screenshot of a Relay.app menu showing some of its AI-powered features

1. Built-in AI Actions for common use cases

There are many things AI does really well. To make the most common use cases easy to use, we created dedicated steps for various automation needs like text summarization, translation, and data extraction. Each of these actions is very easy to set up and results in output data that you can use anywhere else in your workflow.

  • Data Extraction (e.g. from emails, meeting notes, etc)
  • Summarization
  • Translation
  • Browse the web (to extract or look up information from the internet or a specific website)

Demo video | Tutorial

Screenshot of a built-in AI actions in Relay.app

2. ‘Custom Prompt’ step for total flexibility

To get the most out of the AI model of your choice, we created a ‘Custom Prompt’ step so you can provide your own prompt. Within your prompt, you can refer to data fields that are available in your workflow. You may also separately attach data objects (like files) to this step, giving the AI model more context to reason with.

To ensure the response of the AI model becomes usable in the remaining workflow steps, we made it very easy to specify the output you expect. You can specify the response in terms of typed data fields and make them either optional or required.

Demo video

3. Mini AI Agents

AI Agents are small programs that can perform actions on your behalf, “while you’re sleeping.” AI Agents are often used in the context of chatbots. In Relay.app, you can create mini AI Agents inside your workflows. AI Agents in workflows unlock the next level of automation flexibility by leveraging AI decisions and accompanying follow-up actions.

Zooming out, an Agent is described in two components:

  1. A prompt with a description of the role and responsibilities of the Agent (e.g., a customer support agent)
  2. A definition of one or more Actions that the Agent may autonomously perform.

Actions are sequences of steps, or “mini workflows.” For example, an Action named “File support ticket” might contain a single step to file a ticket in Jira. Based on your prompt and the context, the chosen AI model will decide which Actions will be called and how many times. This may vary across different workflow runs, of course.

Demo video

Screenshot of an AI agents configuration in Relay.app

Maximize confidence in AI through Human-in-the-Loop automations

For high-stakes use cases, you’ll want to maximize confidence in the output of AI. For example, when you’re using AI to generate a contract for a new customer, you’ll probably want to review it before it gets sent out.

Human-in-the-Loop steps in Relay.app let you insert checkpoints for human approvals or (missing) data inputs anywhere across your workflows, allowing you to get maximum value out of AI in your repeated automations!

Screenshot showing a human-in-the-loop approval step for the output of an AI action in Relay.app

Choose any AI model

Reasonable, free usage of OpenAI’s GPT-4o and GPT-3.5 models is included in all our plans (including our free plan). If you want to automate actions with another model or at a higher volume, you can connect your own account to use any model of your choosing from OpenAI, Google, Anthropic, Groq, or Perplexity.

Currently supported models:

  • OpenAI
    • GPT-3.5 (free reasonable use included!)
    • GPT-4o (free reasonable use included!)
    • GPT-4
  • Google
    • Gemini 1.0 Pro
    • Gemini 1.5 Flash
    • Gemini 1.5 Pro
  • Anthropic
    • Claude 3 Haiku
    • Claude 3 Sonnet
    • Claude 3 Opus
    • Claude 3.5 Sonnet
  • Groq
    • LLaMA 3 8b
    • LLaMA 3 70b
    • Mixtral 8x7b
  • Perplexity
    • LLaMA 3 Sonar - small
    • LLaMA 3 Sonar - large
    • LLaMA 3 8b
    • LLaMA 3 70b
    • Mixtral 8x7b
On this page

Frequently asked questions

  • What does free reasonable use of GPT-4o and GPT-3.5 mean?

    Reasonable usage of GPT-3.5, GPT-4o in every Relay.app plan. When you’ve exceeded usage for any of these, you’ll receive an email notification from us.

    We are still calibrating what the right free use limits are for GPT, but for any of our paid plans the free usage limits are 10x those of the free plan. We aim to be more explicit about these reasonable use limits in the near future. If you have any further questions on what this means for you, don’t hesitate to reach out to us via support@relay.app!

    Reasonable use resets every calendar month.

  • I’m getting a notification that I’ve exceeded free GPT reasonable use. Now what?

    Reasonable usage of GPT-3.5, GPT-4o in every Relay.app plan. When you’ve exceeded usage for any of these, you’ll receive an email notification from us. To continue using GPT, either upgrade to a paid Relay.app plan to get 10x more reasonable free use, or connect your own OpenAI account to pay for GPT usage with your own account credits. Learn how to connect your own OpenAI account here.

    Reasonable use resets every calendar month.

  • I’m getting a notification that I’ve exceeded usage of the free web browsing extension. Now what?

    Reasonable usage of the web browsing extension is included in every Relay.app plan. When you’ve exceeded usage for any of these, you’ll receive an email notification from us.

    We are still calibrating what the right free use limits are for web browsing, but for any of our paid plans the free usage limits are 10x those of the free plan.

    We aim to provide a more robust credit system for web browsing soon. In the meantime, if you hit limits please reach out to support@relay.app and we try to best support you!

    Reasonable use resets every calendar month.

  • How often do reasonable use counters reset?

    Reasonable use counters reset every calendar month.

  • I have a ChatGPT (Plus) account. Can I use it with the OpenAI models?

    No, you can’t use your ChatGPT (Plus) account, but you can use your own OpenAI API token (which is strictly speaking not coupled to a ChatGPT account at all). Learn more here.

  • Can I use AI to analyze images?

    Definitely! The easiest way to do so is by adding File objects containing images as context to your prompt.

  • Which AI models are best for data extraction?

    All models except for Perplexity’s work for data extraction.

    In our experience, the larger models like GPT-4, GPT-4o, Gemini 1.5 and Claude 3 perform best for more complex tasks. Smaller models will usually do fine for unambiguous input data where they have little opportunity to get confused

  • Which AI models do you support?

    Currently supported models:

    • OpenAI
      • GPT-3.5 (free reasonable use included!)
      • GPT-4 (free reasonable use included!)
      • GPT-4o
    • Google
      • Gemini 1.0 Pro
      • Gemini 1.5 Flash
      • Gemini 1.5 Pro
    • Anthropic
      • Claude 3 Haiku
      • Claude 3 Sonnet
      • Claude 3 Opus
      • Claude 3.5 Sonnet
    • Groq
      • LLaMA 3 8b
      • LLaMA 3 70b
      • Mixtral 8x7b
    • Perplexity
      • LLaMA 3 Sonar - small
      • LLaMA 3 Sonar - large
      • LLaMA 3 8b
      • LLaMA 3 70b
      • Mixtral 8x7b

  • What are Actions for my AI Agents?

    Actions are sequences of steps or "mini workflows" that the AI model might invoke in order to answer your prompt. Based on your prompt and the provided context, the chosen AI model will decide which Actions will be called, and how many times.

  • Can Actions be called multiple times?

    Yes, Actions may be called once, multiple times or never for every run. The AI model will decide based on the provided context and prompt.

  • How are Actions different from Paths?

    Actions are different from Paths. Actions are not mutually exclusive: some, all or none of the defined Actions may be invoked by the AI model during a workflow run. For Paths, in every workflow run always exactly 1 Path will be taken (if no Path rules match, there will be an error).

  • When should I use Actions, and when Paths?

    Use Actions when your AI model needs them in order to complete the prompt that you gave it. For example:

    “For each person with action items mentioned in the Email/Document, look up if that person exists as a user in Asana. For each of them that is an Asana user, assign a task for each of their action items. For all persons that don't have an Asana account, notify me.”

    • Action 1: Look up person in Asana (inputs: person name)
    • Action 2: Create task in Asana (inputs: asana user, task name)
    • Action 3: Notify me (inputs: person name)

    If the continuation of your workflow depends on an AI decision/extraction, use Paths after your AI step instead of Actions within it. Why?

    • Paths are deterministic, which makes it much easier to debug when things go unexpectedly
    • Paths are mutually exclusive, so your AI step might not [inadvertently] call multiple Actions when it should have only done one
    • Paths are only ever taken once, whereas the AI model might decide to call one and the same Action multiple times if it deems needed

  • Which AI models are best for Agent Actions?

    In our experience, GPT-4(o) or Claude 3 based models perform best for Agents.

    Groq's LLaMa-based models support Actions, but they'll not perform as well as GPT4 of Claude 3.

    Perplexity’s models don’t support Actions, so are not a fit for Agents.

Author picture (Garrett Clawson) Small business owner

“I've used several automation tools in my day, but Relay.app is by far the most intuitive to use. I now use Relay.app almost exclusively, and I find it integrates with almost every tool I use. I use Relay.app several times a week and would recommend it to anyone looking for an easy way to start automating their work.”

Garrett Clawson
Small business owner

What will you automate?

Sign up and get started with your first Relay.app workflow today.
Try Relay.app for free
Background imageBackground image