An AI field turns a recording into a value you can filter, report on, and sort by. Four properties control that extraction: Prompt tells the AI what to look for, Allowed Values lists what it may answer, Other Values decides whether it can answer with anything else, and Not Applicable Values decides what it says when it finds nothing. Configure them once at Settings → Fields, then map the field in an automation.
Prompt
The prompt is your instruction to the AI. It names the exact information to pull out of the audio, video, or text:
- Extract an entity: “Identify all company names mentioned in the conversation”
- Classify content: “Determine the sentiment of the speaker (positive, negative, neutral)”
- Extract structured data: “Find all dates and times mentioned in the transcript”
- Identify topics: “List all product names or services discussed”
Be specific rather than vague, say what format you expect the answer in, and test the prompt against a sample file before you run it across a library.
When you select the field in an automation, its prompt auto-populates, which keeps every run consistent. Editing the prompt on the field definition does not update automations that already use it. Update those automations yourself, or create new ones.
Allowed Values
Allowed Values is a list of the answers you want back. It standardizes extraction, so you get “Support” every time instead of “support”, “support request”, and “help desk” across three files. Use values that do not overlap in meaning, name them consistently, and keep the list manageable. Five to fifteen values usually works well. Think about how the values will read later in filters, reports, and dashboards.
Typical lists look like:
- Department: [“Sales”, “Support”, “Marketing”, “Product”]
- Workflow status: [“Active”, “Pending”, “Completed”, “Cancelled”]
- Priority: [“Low”, “Medium”, “High”, “Urgent”]
- Product type: [“Software”, “Hardware”, “Service”, “Consulting”]
Single and multiple selection
Selection mode controls how many of those values the AI may return:
- Single: the AI picks the one value that best matches, and returns it on its own, for example
value1. - Multiple: the AI picks every value that applies and returns them comma separated, for example
value1, value2, value3.
Other Values
Other Values is a toggle that decides whether the AI may answer with something outside your list.
Turn it on and Allowed Values becomes a list of preferred answers. The AI can return a value that is not on the list when that value is more accurate. If your list is [“Sales”, “Support”] and a file is clearly about customer success, it can answer “Customer Success”. This handles synonyms and terminology you did not anticipate, and it is the right setting while you are still discovering what your data contains.
Turn it off and the list becomes mandatory. The AI answers only with values from Allowed Values, and nothing else gets through. Use this once you know the values you want, and whenever reporting or compliance depends on a fixed vocabulary.
A good sequence is to run with Other Values on at first, review what comes back, promote the recurring answers into Allowed Values, then turn it off.
Not Applicable Values
This is the answer the AI gives when it finds nothing relevant, either nothing matching Allowed Values or, if you left that empty, nothing worth extracting at all. It keeps missing data consistent instead of scattered across blanks and improvised phrasings.
If you leave it empty, the field falls back to N/A. The AI returns that value alone and nothing
else. Pick something that clearly reads as absent data, keep it distinct from your Allowed Values,
and use the same value across similar fields.
| Field type | Common values |
|---|---|
| Boolean | “false”, “No”, “N/A” |
| Text | “Unknown”, “Not Specified”, “N/A”, “-” |
| Categorical | “Uncategorized”, “Other”, “None” |
| Status | “Not Applicable”, “Pending”, “Unknown” |
A field with all four set
Take a field that classifies the type of a call:
- Prompt: “Identify the primary purpose or type of this call from the conversation”
- Allowed Values: [“Sales Call”, “Support Request”, “Product Demo”, “Follow-up”]
- Selection mode: Multiple, because one call can serve more than one purpose
- Other Values: on, to catch types you have not listed yet, such as “Training” or “Onboarding”
- Not Applicable Values: “Uncategorized”
Run that field and the AI looks for the four listed types, returns any additional type it finds, lists several comma separated when a call covers several, and answers “Uncategorized” when the recording gives it nothing to go on.
Using fields in automations
When you build an automation with an AI Chat action, selecting a field brings its configured properties into the automation. You can override them there for a one-off case, and the extracted value is written back to that field on the media in your library. Setting the properties on the field is what keeps results consistent; the override is the escape hatch.
When results look wrong
- The AI returns values that are not in your list. Turn Other Values off to enforce the list, or add the returned values to Allowed Values.
- The AI returns “N/A” too often. Check that the prompt is clear, that Allowed Values is not too narrow, and that the content really does contain what you are asking for.
- You get several values when you expected one. Set selection mode to single, and make the prompt ask for one answer.
- A prompt edit did not reach an existing automation. That is expected. Update the automation, or rebuild it against the field.
Want a hand setting this up? Book a free consult and we’ll do it together on your account.
Related: Insights · Categories