Airtable AI assist field overview

In an open base navigate to the table/view where you want to add a new AI assist field. Then click + to add a new field. From here, you'll find a list of field types that you can either search for by name or scroll through. Choose the AI assist option and move on to the next step.

This will open the configuration window for the new AI assist field. Your first choice will be to Start with a template or Start from scratch. For our example, we will use a template.

After clicking Start with a template, another configuration window will appear. Airtable has created many templates for you to choose from that we have identified as common workflows where generative AI can be helpful. In this example, we will choose the Feedback categorization option under the product tab.

In order to have the AI assist field generate content, you'll need to configure a few pieces of information to help the AI models know where to look in your Airtable base as well as some additional instructions. This will vary by template, but you will one or more of the following:

• Define the starting point - You'll need to choose the field or fields that will be utilized by AI assist. In certain cases, there will also be a box to add a comma-separated list of static values. 
• Include additional fields - Sometimes, the template that Airtable provided might need more depth to allow the AI to process more information. In these cases, you'll want to add more fields of your choosing.
• Add your own examples - Adding examples helps to structure and format the result that the generative AI produces. Good examples will help the AI models align more closely with your expectations of what the output should be.
• Give any other instructions - All templates contain the option to give other instructions. This is your opportunity to add additional context that can further help the AI models with structure, formatting, style, language, and set boundaries for what the model should not do.

In our example, we chose the Feedback categorization template. This template only requires that we Define the starting point by choosing a field that contains the feedback and adding a comma-separated list of categories that the AI model will use for outputting content. You can see in the image below that:

• The "Additional notes" long-text field is being used for the Feedback to be categorized
• A comma-separated list of categories has been entered in the text box to help the AI assist field generate a category for the feedback.
• There's an additional option to Give any other instructions but, for our purpose, this is not needed.

Once you've configured the template to your liking, click Use template. You can come back and edit the template later as discussed in the next step.

Once you have chosen to use a template, you'll notice two new options appear in the Prompt section of the field configuration window:

1. Pencil icon - Use this to go back and edit the existing template's configuration. You can even try out another template if you are unsatisfied with the existing one.
2. Break icon - CAUTION! Clicking this option will transform the existing template into a text box prompt. At this point, the prompt will function the same way as a prompt written from scratch. It's important to note that this option cannot be reversed. That is to say, you can't convert the text box back to the template configuration. It is, however, a good way to better understand how it is that a prompt might be written. We encourage you to experiment with this while you are testing out the AI assist field, but not in live bases with potential downstream effects.

For our example, we don't need to use either of these options at the moment.

Once you've built out your prompt or template, you'll want to check the preview to see if the result is looking desirable. If not, it's time to go back and make some edits to the prompt or template to help the AI assist field achieve a more expected output. In some cases, the generated output might be something fairly simple, and in others, the output could be quite complex. It all depends on the workflow you are trying to achieve. 

The preview box can be opened or collapsed by clicking on the caret icon (> or v) next to the Preview section. In our example of categorizing feedback, you can see that the first record in the table has been categorized as "Mobile app" in the preview. This looks like a good categorization after comparing it to the notes

There are some additional field settings that you can configure by clicking on the Settings tab in the AI assist field configuration window:

• Model - There are two generative AI models to choose from, GPT-3.5 and GPT-4. We recommend you consult other resources if you want to dig into the differences between the two models, but briefly:
  • GPT-3.5 - OpenAI's first released model, and the model that currently powers ChatGPT. This model is faster, less expensive, and may be less accurate than GPT-4.
  • GPT-4 - OpenAI's most advanced model, which is more accurate than GPT-3.5. This model is slower, meaning that responses will take longer to generate. It is also more expensive and therefore much more limited in the number of calls you can make per day. GPT-4 is recommended for long-form writing such as marketing briefs, PRDs, and job descriptions.
• Randomness (Temperature) - Randomness can help give certain outputs more variety on a scale from 0 to 1.0. The number can be adjusted in increments of .1 depending on your needs. Generally, you can expect the following behavior in records that contain the same or similar data:
  • 0 - Minimal to no variation in outputs. A good setting to use when you want to have expected, linear results in the output. 
  • .5 - Moderate variation in outputs. A good setting to use when you want some variation to the output of the AI assist field.
  • 1.0 - Maximum variation in outputs. The best setting for situations that can use more variation between output, even given the same inputs. This can be helpful when using the AI assist field to gain inspiration for writing.
• Generate automatically 
  • This setting is only available when selecting the GPT-3.5 model setting and when the table that contains the AI assist field has 300 records or less.
  • When this option is turned on, AI assist will generate new outputs when the AI assist field's prompt sees that changes have been made in the base. To clarify, only changes that affect field(s) that the prompt is monitoring will cause a new output to be generated.
  • When this option is turned off a base user with Editor permissions or higher can Generate outputs manually. More on this in the next section.

With the prompt written and settings configured, it's time to create the new AI assist field. Click the blue Create field button.

• If the Generate automatically field setting is enabled, then you will begin to see each record in the table begin to populate with AI assist created content. Depending on the complexity of the prompt and the model you chose to use this may take a bit of time.
• If the Generate automatically field setting is not enabled, then each record in the table will contain a button with the label Generate. This allows base users with Editor permissions or higher to either:
  • Generate outputs record by record on an individual basis, by clicking the Generate button.
  • Or generate outputs for every possible record (bulk generate) within the view that the user has open. To do this, click on the dropdown arrow next to the name of the AI field and then click the Generate content for xx records. If your table has more than 300 rows, you can do this generation in batches of 300 until all the rows in the table have been generated.
    

There are 5 categories of templates to choose from. We recommend testing out all of the different templates, especially if you are unfamiliar with AI and how prompt writing works. This will allow you to become more familiar with the differences and similarities between prompts to help you better understand what it is that AI models need in order to have a good starting point to work off of in Airtable.

• General
  • Translation - Output text from a field in Airtable as translated text in the AI assist field.
  • Meeting summary - Output longer meeting notes as a succinct summary in the AI assist field.
  • Categorization - Configure a set of static, pre-defined categories and have the AI assist field categorize the content within another field in your table.
• Product
  • PRD - Create formatted text for a product requirements document from multiple fields of information stored in Airtable.
  • Product launch brief - Create formatted text for a product launch brief from multiple fields of information stored in Airtable.
  • Feedback categorization - Configure a set of static, pre-defined categories and have the AI assist field categorize the feedback stored in another field in your table.
• Marketing
  • Marketing brief - Use fields containing the product/feature name and the PRD for that feature to output a marketing brief in the AI assist field.
  • Campaign brief - Use fields containing the campaign name, description, and target audience or marketing brief to output a campaign brief in the AI assist field.
  • SEO keywords - Have the AI assist field create a list of keywords based on copy help within a field. You will set the number of words for the AI assist field to generate.
  • Tweet - Use a field containing a longer set of content that you want the AI assist field to summarize into a more succinct output of text.
• UX Research
  • Sentiment analysis - Have the AI assist field use a field containing feedback as a source to be categorized by sentiment (positive, neutral, negative).
  • Research plan - Use fields in Airtable containing the research questions and objectives and then set a research type (Qualitative, Quantitative, Hybrid) to have the AI assist field output a research plan.
  • Feedback summary - Have the AI assist field summarize feedback from notes stored in Airtable during a research session by extracting specific insights.
  • Interview guide - Give the AI assist field a field containing a research plan and a length of time for the interview to output an interview guide.
• Recruiting
  • Job description - Use fields in Airtable containing the job title, level/experience requirement, location, and any additional custom fields to create a job description in the AI assist field.
  • Outreach email - Have the AI assist field create a draft outreach email by using fields in Airtable that contain the job title, the candidate's current company, and their current role.
  • Search query - Give the AI assist field the job title, level, and location from other AIrtable fields and have it generate a list of keyword terms related to that job.
  • Candidate assessment - Have the AI assist field assess a candidate based on fields in AIrtable that contain their resume and the description of the job they are applying for.

Writing a prompt from scratch requires a few things in order to be successful with the AI assist field in Airtable:

• Giving the AI model a role is helpful (ex. "You are a product manager...")
• Include Airtable fields in the prompt as needed to provide context for the request. When you include a field, provide the context of what kind of content the field is storing, such as: 
  • "Write an email that addresses the following customer feedback: {customer_feedback}"
  • or "Create a meeting summary based on the following notes: {Meeting Notes}"
• You can use the blue + sign in the prompt window to add fields or begin typing the field name with a curly bracket {  in front of the name. Either option will open up a list of the fields you can use in the prompt. The field name will appear as a token in the prompt and will automatically update, even if the name of the field is changed later.
• There are limits to be aware of as discussed in the next section below.
• Give the AI model examples to work off of. When providing examples, make it clear at the end of the examples section what you want the model to do by prompting it specifically with the content you want it to ingest and what you want it to do.

  Below are some examples of feedback and related sentiments that you'd output.
  
  The feedback statements below are examples only.
  
  Feedback: "This is the worst tent I've ever used."
  Sentiment: Negative
  
  Feedback: "This backpack is alright."
  Sentiment: Neutral
  
  Feedback: "I really love this sleeping pad."
  Sentiment: Positive
  
  Feedback: {customer_feedback}
  Sentiment:

• Define expected formats and length of the output if that is important for the result you want. 
• Include "Do" and "Do NOT" statements to set boundaries for the AI model.
• Remember that more specific prompts typically yield more specific outputs, while less specific prompts might result in more variation. You can use either method to acheive different results, but it's important to be intentional with your approach.
• You can always go back and edit a prompt, but we strongly recommend that you save a previous prompt in case you want to revert back to it (paste it back into the prompt text box) while you are testing the AI assist field's output.

One "use" is counted when one cell is generating a response using AI.

• GPT-3.5: 2,000 uses of the AI field with the model type set to GPT-3.5
• GPT-4: 100 uses of the AI field with the model type set to GPT-4

The total number of words in your query and response must be under a certain amount set by OpenAI. This includes the words in the fields referenced by your query. If you exceed these limits in your query, you will see an error message. If the AI's response exceeds the limits, the text will be truncated.

• GPT-3.5: Up to ~3,000 words
• GPT-4: Up to ~6,000 words

The auto-generation setting on the AI assist field can only be used in tables with fewer than 300 records.

In fields with auto-generation turned off, batches of 300 records can be generated at a time.

• The workspace ID(s) that you submitted in your Beta enrollment form will automatically be enabled with the feature.
  • If you want to turn the feature off, the individual who enrolled the workspace in the Beta will need to go to their account page, choose the workspace, and then locate the Workspace AI permissions section.

• Global Admin Panel settings: To adjust this setting, admins should navigate to Settings > Integrations & development. Then look for the Block AI integrationssection.
  • Block AI integrations
    • Leaving this setting ON (default) will globally prevent usage of AI models in all organization-owned workspaces.
    • Turning this setting OFF will allow admins to enable AI features to be turned on in one or more workspaces.
  • Allow AI integrations in organization-owned workspaces? 
    • Allow in all workspaces - Choosing this option will allow AI features to be enabled in all workspaces owned by the organization.
    • Allow only in selected workspaces - If this option is chosen, you can choose specific workspaces where AI features may be enabled.
  • For Business and Enterprise Scale customers, the settings below will only be visible if the Block AI integrations setting is set to OFF.
    • Individual workspace setting: Workspace owners will need to enable the AI feature in their workspaces once the global setting to block AI is turned OFF. From your account page, choose a workspace on the left side of the page where you have Owner permissions. Find the Workspace AI permissions section and then choose to either enable or disable AI features for all bases within the workspace.

The easiest way to check whether Airtable AI is enabled in a given workspace is to try to create a new field in a base in that workspace. If you see the AI Assist [NEW] field type from the dropdown menu, then the feature has been enabled. If you do not see it, check the following:

• Check if you received a confirmation email that your company or account has been enrolled in the Beta program.
• Enterprise users should ask their admin if the OpenAI integration is enabled in Admin Panel and if it's enabled for your workspace. If it is but you still don't see the feature, ask a workspace owner to check their workspace settings page to see if the feature is enabled in the workspace.
• Team plan users should ask a workspace owner to check their workspace settings page to see if the feature is enabled in the workspace.

AI responses are designed to be reviewed and edited by users and may contain inaccuracies or biases. 

Based on the terms of Airtable’s agreement with our AI vendors, the data you provide as input into the AI field will not be used to train their large language models. However, some companies have policies that restrict personally identifiable information (PII) such as emails, addresses, and telephone numbers from being sent to external vendors. If this is the case, refrain from using columns containing PII as inputs into an AI field.

As a technology, AI has improved rapidly in its ability to understand language, answer questions, and carry out tasks. However, there are known limitations to the technology:

• The models that Airtable AI uses were trained over a year ago and do not have access to the Internet, so they don't know any current information like the weather today.
• These models are not good at doing math and may return nonsensical information if asked to do a task that requires calculation.
• The models are not good at representing their confidence in an answer and may give an authoritative-sounding answer that is actually invented or inaccurate. Always check the accuracy of answers provided by AI.

Absolutely! You can provide feedback by clicking Give feedback or report issue within the AI field configuration menu. If you’re an Enterprise customer, you can also reach out to your account team and share feedback.

Translation capabilities might be limited on OpenAI’s end, but in general, you should expect that most common languages are available for translation.