How to Use Airtable AI Fields for Record Classification

Short answer: Airtable AI fields are column types that run a model over each record’s text and return a value — a category, a summary, a sentiment label. They are available on Business ($45/user/month) and Enterprise, billed on AI credits, and they do not run retroactively for free. Each record you generate costs a credit.

On a 5,000-record base, a careless AI field can consume your month’s credits in one afternoon.

What can an Airtable AI field do?

Airtable is a cloud database that looks like a spreadsheet. Every base is a set of tables with typed fields. An AI field sits beside your normal fields and writes a value derived from another field’s text.

Four field types cover most needs:

Field type	Output	Good for
Classify	One of your defined options	Support ticket category, lead source
Extract	Structured values	Company name, deal size from email text
Summarise	Short paragraph	Meeting notes, call transcripts
Translate	Text	Localising a description across markets

The critical difference from a formula field: AI fields are non-deterministic. Run the same record twice and you can get different classifications. That rules them out for anything that drives money or access.

How do you set up a classification field properly?

Steps, in the order that avoids wasted credits:

1. Add a field, choose AI, pick Classify.
2. Choose the source field. Only one — if the signal is split across two fields, make a formula field that concatenates them first.
3. Define the output options yourself. Five to seven categories is the working range. Letting the model invent labels produces 40 near-duplicates within a week.
4. Add brief instructions: “Use the category the customer would choose, not the internal team name.” One sentence of instruction changes output more than any prompt-tuning later.
5. Test on 20 records with Run on selected records, not the whole column.
6. Only after manual review, run the whole column.

Step 5 is the one people skip. Twenty records on a five-category task will show you the misclassification pattern, and the fix is usually a missing category rather than a better prompt.

How much do AI credits actually cost?

Plans matter more than usage here. AI fields are on Business at $45 per user per month (annual), which includes a monthly AI credit pool per workspace; Enterprise adds more. Pro does not include AI fields at all, and there is no add-on to get them.

A classification run on a short text field typically costs one credit. A summary of a long transcript costs more, and repeating a summary on every record edit is how pools disappear.

Two guardrails worth setting immediately:

• Turn off auto-run on new records until the field has proven itself. Airtable will otherwise generate on every import.
• Use a “needs classification” checkbox and run in batches weekly, so you can see the cost per batch.

Why does the same record get classified differently over time?

Because the source text changes, or the options change, or the model changes. Airtable has updated the underlying model without notice before, and older classifications are not re-evaluated unless you force a re-run.

That creates drift. Six months in, half your records are classified by one model and half by another. If the field feeds reporting, the numbers shift without anyone editing data.

Mitigation: add a last classified date field, and re-run anything older than 90 days once a quarter. Also keep a manual override field with a formula that prefers the human value. Your team’s judgement should win, always.

Does the AI agent feature change this?

Partly. Airtable agents chain searches and model calls for multi-step work — qualifying a lead list, drafting briefs across records. They are priced separately and need careful scoping.

Where agents beat AI fields: tasks that need lookup across tables before deciding. Where fields beat agents: anything you want visible in a column, sortable and filterable. Most teams should start with fields and only add an agent when the manual process clearly has three or four dependent steps.

Where this breaks down

Three honest limits.

1. Non-English text. Classification quality drops noticeably on short Chinese or Japanese strings, and mixed-language columns produce inconsistent labels.
2. Long source fields. Text over roughly 3,000 characters gets truncated, and the truncated tail is often where the answer was.
3. No audit trail on output. You get the value, not the reasoning. When a classification is contested, you have nothing to show.

None of these are disqualifying. All mean you should not put an AI field in front of a customer-facing workflow without a human review step.

Who should use AI fields

Use them if you already pay for Business, you have more than 500 records of free-text that nobody reads, and someone will own the classification schema. Support triage and lead enrichment are the two cases with clear payback.

Do not use them on Pro — you cannot. Do not use them for fields that gate billing, permissions or SLA timers, because non-determinism will bite. And do not build six AI fields on day one; build one, live with it for two weeks, then decide whether the second is worth it.

Quick Summary

• AI fields run a model per record: classify, extract, summarise or translate. Available on Business ($45/user/month) and above.
• Define five to seven output options yourself; open-ended labels fragment within a week.
• Test on 20 records with “run on selected records” before filling a column of 5,000.
• Keep auto-run off until the field proves itself, and add a manual override formula.
• AI output drifts when models change — date-stamp classifications and re-run quarterly.