How to Add AI Fields to an Airtable Base (Step by Step)
You add an AI field in Airtable by clicking the plus next to your last column, choosing AI, and picking an operation like Classify, Summarize or Extract. It runs once per record and costs AI credits — roughly 10-60 credits per run depending on prompt length and output size. A 2,000-record table can burn a month of credits in one click.
Airtable AI is available on Team and Business plans; Free and Starter plans see the field but cannot run it. That gate is the first thing to check before planning anything around it.
What can an AI field actually do?
Five operations, each with a defined output shape:
| Operation | Input | Output |
|---|---|---|
| Classify | Free text | One label from your list |
| Summarize | Long text | Short paragraph |
| Extract | Unstructured text | Structured field values |
| Categorize | Text | Multiple categories |
| Translate | Text | Same text, target language |
Extract is the one worth building a workflow around. Point it at a column of messy inbound emails and pull company name, budget and deadline into three clean fields you can then filter and sort.
How do you build your first AI field?
- Open the table and click the plus at the far right of the field row.
- Choose AI as the field type, then pick an operation.
- Point it at an input field — usually a long text or attachment column.
- Write the prompt with the label list or extraction schema spelled out.
- Test on one record and check the output before enabling the full column.
- Enable the field, or use a conditional so it only runs when the input changes.
Step 5 gets skipped constantly. Testing one record costs a few credits; running 2,000 records with a bad prompt costs a chunk of your monthly allowance and produces a column you then delete.
How should the prompt be written?
State the allowed values and the fallback. The model will invent a category when none fits, and an invented label breaks every filter downstream.
A classify prompt that holds up:
Classify this support email as exactly one of: billing, login, shipping, product defect, other. If none apply, respond “other”. Use only those five strings, lowercase.
For extract, define the schema explicitly:
Extract company name, budget as a number in USD, and deadline as YYYY-MM-DD. If a value is absent, return empty. Do not infer.
The “do not infer” line is the difference between Airtable AI saving you time and generating plausible garbage you have to audit anyway.
How do AI credits work?
Airtable bills AI usage in credits pooled at the workspace level. A short classification might be 10 credits; a long summarization with a big output could be 60 or more. Business plans include a larger monthly pool than Team, and you can buy more.
Practical math: a 1,000-record classify run at 15 credits each is 15,000 credits. If your plan includes 20,000, that is one run and a little left over. Plan the field before you turn it on, because there is no undo for spent credits.
Watch for re-runs. Any edit that touches the input field re-triggers the AI field. Import 500 rows from a CSV and the AI field fires 500 times without you asking.
How do you stop the accidental re-runs?
Three controls, all worth using together:
- Convert to static values after the run: turn the AI field into a plain text field so it stops recalculating
- Add a run-condition so the AI field only fills when a checkbox like
needs_aiis ticked - Import first, enable the AI field second rather than the other way around
The checkbox pattern is the safest for teams. A small automation ticks the box for new rows, and the AI field only runs for ticked rows, so imports and bulk edits never trigger a storm.
What are the real limitations?
No AI field can read an attachment directly without text extraction, so a scanned PDF column produces empty results. Run the PDF through an extraction step first.
AI fields also cannot reference other records’ AI fields in a chain reliably; a classify that depends on a previous summarize produces race conditions when both recalculate. Split the pipeline into separate runs.
And the output is never structured for filtering until you trust it. Always build a validation view that shows records where the AI output is empty or outside your allowed set.
Who is Airtable AI for?
Ops teams already running Airtable as their source of truth, where adding a classification column avoids an export-and-reimport loop. Recruiting pipelines, support triage, inbound lead scoring.
Skip it if you are on Free or Starter, since the feature is locked. Skip it if your volume is large and price-sensitive: below about 5,000 records a month it is convenient, above that a script calling an LLM API directly is far cheaper.
What is the honest comparison to a script?
A Python script with an API key costs a fraction per record and handles any schema. It also needs a developer, a scheduler, and error handling when the API rate limits.
Airtable AI wins on setup time — five minutes versus half a day. It loses on cost and flexibility above a few thousand records per month. Pick based on whether your bottleneck is engineering time or budget.
Quick Summary
- AI fields live on Team and Business plans; Free and Starter cannot run them.
- Explore the five operations, but Extract is the one that changes workflows.
- Spend is in credits, roughly 10-60 per record; a 1,000-row run is a real chunk of the monthly pool.
- Convert the AI column to static values after the run to prevent re-charge on edits.
- Above ~5,000 records a month, a direct API script is much cheaper.
