Best practices for using AI in automations

To improve the results when using AI in automations, apply the following best practices:

• Best practices when leveraging AI
• Best practices when building apps using AI
• Best practices for inputs to AI tasks

Best practices when leveraging AI

The following lists some of the general, high-level best practices to use when leveraging AI. 

• Set expectations and measure impact
• Should I use AI in automations?
• Have a human review the results and handle unexpected behavior

Set expectations and measure impact

When using AI in automations, it is important to set expectations and measure its impact. That’s because AI doesn’t always get it right. Even when there are errors here and there, you’ll want to make sure it works in your favor overall and be able to tell others how it’s helping you,

As first steps, consider the following:

• What are you trying to accomplish using AI in automations?
• What are the expected outcomes?
• How can you measure the impact of AI? For example, you might track the cost and time savings, productivity gains, and so on.  
• What is the impact when AI gets it right?
• What is the impact when AI does not get it right?
• How often does AI get it right? How often does it not?

Should I use AI in automations?

As an app creator, you can incorporate predefined AI capabilities as tasks in your automations. These AI tasks rely on Google’s Gemini models. You don’t need to train them, but that means they aren't tailored to your data.

To figure out if AI tasks are a good fit, we encourage you to try them out, first on a few examples and then on more data.

As a general guideline, consider the following:

• If your AI task does not require obscure knowledge, it might help you
• If the required knowledge to properly get the right response is difficult to find online, then your AI task might not return good results; for example, tasks that require industry-specific knowledge might impact the quality of the results

Have a human review the results and handle unexpected behavior

Set up your automations to use AI and human expertise in a collaborative manner to ensure the accuracy of the results, by doing the following:

1. Make sure humans can quickly review AI-generated results in your app.
2. Add steps in the automation flow to handle unexpected behavior from AI tasks.

For example, if you've defined an Enum value named "Other" to catch any undefined values, you might send yourself a notification when the value is set so that you can update the Enum column to include the missing value.

Best practices when building apps using AI

Use the following best practices when building apps using AI:

• Add additional details
• Add column descriptions

Add additional details

Each AI task has an Additional details field. AppSheet recommends that you use this field to provide more information about the task and the expected results. 

The following provide guidelines specific to each AI task type:

Add column descriptions

For the columns that are used to save or temporarily store output from an AI task, AppSheet recommends that you add column descriptions to provide more information for Gemini to use in its data evaluation. See Configure column properties.

Best practices for inputs to AI tasks

Use the following best practices for input to AI tasks:

• Use quality images
• Optimize PDF documents

Use quality images

The following provides guidelines to apply when using images with AI in automations:

• Use higher resolution images to improve result; avoid blurry images
  Consider setting the Image upload size to High, as described in Control image size
• If possible, modify the image as follows:
  • Rotate and crop images to include only the relevant subject matter
  • Mark up the image to highlight the relevant subject matter
• Images that require industry-specific knowledge might impact the quality of the results

See also limitations and known issues with images.

Optimize PDF documents

The following provides guidelines to apply when using PDF documents  with AI in automations:

• Consider splitting long PDF documents into multiple files
• Use PDFs created with text rendered as text instead of using text in scanned images
• Don't include low-resolution or small images in PDF documents

See also limitations and known issues with PDF documents.