Airtable's AI fields bill by length, not a flat per-field price

The setup

Airtable's spreadsheet-database hybrid already stored a team's row-by-row work before generative AI arrived, so adding AI to a cell meant fitting a model call into an existing table rather than opening a separate app. The vendor's own help documentation on using Airtable AI in fields describes these as field agents: AI-powered fields that retrieve, analyze, or generate a value at the cell level, added to a formula, long text, linked-record, select, or number field. Field agents require any paid plan, plus AI credits allocated to the workspace and AI turned on at the workspace level by an owner.

What the documents show

The documentation states credit usage is not a flat per-field price. It varies with the length of input and output, and a preview window shows estimated consumption for the first record in view before a field is even created, with a warning that other records may cost more depending on their own content. Airtable's separate product page for Airtable AI quotes one precise figure for a different feature: asking Omni, the platform's app builder, an analysis question about a base costs 10 credits per response, while asking it to build or edit an app carries no stated additional cost. Only one of those two features has a fixed number.

The friction

The documentation discloses a specific batch limit: turning on automatic generation for a long text field across a table of thousands of records is still bounded by the credits available to the workspace, so some records may generate while others error out once a limit is hit mid-run. Word limits also apply by model tier, roughly 12,000 words of combined prompt and response on lower-powered models and about 90,000 on higher-powered ones, and enabling internet search inside a field agent often costs additional credits on top of the base action.

What changed in the work

For a team that previously exported rows to a separate summarization tool, the documented change is that the generated value now lives in the same table already used to track the work, referenced from other fields by inserting a token for that column. That is a real before-and-after in where output sits. Reading it as saving time would go beyond the documentation; the narrower, editorial reading is that review now happens in the same grid view as the record, not in a separate export.

• How many AI credits does the workspace currently have allocated, and who can see that balance before a bulk run starts?
• Does a field agent's custom instructions reference other fields that might carry information an admin would rather keep out of a model call?
• What happens to records left unprocessed if a bulk automatic-generation job exhausts the credit balance partway through?

Airtable's documentation treats field agents as a metered capability layered onto the base plan's existing cost, rather than a separate line item with a posted price, a deliberate design choice worth naming rather than assuming away. A team turning on automatic generation across a large table is, by the vendor's own account, taking on an open-ended credit exposure rather than a fixed add-on fee.