Airtable AI Field Agents: Enrich and Classify Records Automatically

An Airtable AI field agent is a field that runs a prompt against each record and writes the model's answer back into the base. Point it at an inbound inquiry and it can classify the request, pull out the company name, score the lead, and draft a first reply, automatically, record by record. This tutorial sets one up on a lead-triage table, covers the prompt patterns that hold up in production, and explains the two things most teams skip: cost control and quality sampling.

What you need

• An Airtable base on a plan with Airtable AI enabled (field agents consume AI credits from your workspace allowance).
• A table with some real records. We will use Inquiries with fields Name, Email, Company, Message (long text), and Source (single select).
• Editor or higher permission on the base.

The exact menu labels have shifted a few times as Airtable's AI features matured, so if a screenshot you find online looks different, follow the concept rather than the pixels.

Step 1: decide what the agent writes

Field agents write to fields. Before you write a prompt, add the output fields you want populated, with the right types:

• Request type: single select with options Sales, Support, Partnership, Recruiting, Spam.
• Lead score: number (0–10).
• Company domain: single line text.
• Summary: long text.
• AI reviewed: checkbox (for the sampling workflow later).

Adding the output fields first matters because the best prompts reference the exact select options the field allows. An agent asked to "categorise the request" will invent categories; an agent told to choose exactly one of five named options will mostly comply.

Step 2: create the field agent

Add a new field, choose the AI field type, and open the prompt editor. Pick the output type that matches your field (single select, number, text). Then write the prompt. Insert fields into the prompt using the editor's field picker rather than typing their names, so the agent receives the record's actual values.

A prompt for Request type:

You are triaging inbound website inquiries for an Airtable consultancy.
Read the message and classify it as exactly one of:
Sales, Support, Partnership, Recruiting, Spam.

Rules:
- Sales: the sender wants to hire us, asks about pricing, or describes a project.
- Support: an existing client asking for help with a base we built.
- Partnership: agencies, tool vendors, or referral proposals.
- Recruiting: job seekers or staffing agencies offering candidates.
- Spam: SEO offers, link exchanges, unrelated products, or nonsense.

Message: {Message}
Sender email: {Email}
Company: {Company}

Answer with only the category name.

For Lead score:

Score this inbound inquiry from 0 to 10 for how likely it is to become a
paid Airtable consulting engagement. Consider: a specific project described
(+), a business email domain rather than a free mail provider (+), budget
or timeline mentioned (+), vague or generic message (-), asks for free
help (-). Return only an integer.

Request type: {Request type}
Message: {Message}
Email: {Email}

Note that the second prompt reads the first agent's output. Airtable runs field agents when their input fields change, so chaining works, but it also doubles your credit spend per record. Chain only when the second decision genuinely needs the first.

Step 3: choose when it runs

Field agents can run automatically whenever inputs change or only when triggered. For a triage table fed by a form, automatic is right: each new submission gets classified once. For a table people edit constantly, automatic can mean the agent reruns on every keystroke to a long-text field, which is both expensive and confusing when a value flips back and forth. In that case run it on demand, or gate it behind a checkbox or status that a person sets when the record is ready.

Prompt patterns that survive contact with real data

Constrain the output. Name the allowed values, ask for "only" the answer, and set the field type to match. A single-select output type will reject anything outside the options; text outputs will not.

Give the negative cases. Models over-classify into the interesting categories. Spelling out what Spam and Recruiting look like cuts the false "Sales" rate noticeably.

Handle empty input. Add a line: "If the message is empty or under ten words, answer Spam." Otherwise the agent will guess.

Keep the prompt about one decision. One agent per output field. A prompt that asks for a category, a score, and a summary in one go produces a blob you then have to parse.

Do not send what you do not need. Every field you insert into the prompt leaves the base and goes to a model. If a record contains a phone number or a contract value that does not affect the decision, leave it out of the prompt.

Cost and credits

Airtable AI features draw down a pool of AI credits allocated to your workspace by plan, with add-ons available. Three things drive spend:

1. Records processed. A backfill over 20,000 historical records costs the same as 20,000 new ones. Run backfills deliberately, on a filtered view, not by accident when you create the field.
2. Reruns. Automatic triggers on frequently edited fields are the most common source of surprise usage.
3. Prompt size. Long text fields with full email threads in them cost more per run than a short message field.

Estimate before you enable: records per month multiplied by agents per record, then check the number against your plan's allowance in the workspace's AI usage settings. Put a monthly check on the calendar.

Quality: sample, do not assume

An agent that is 92% accurate will look perfect on the ten records you checked and quietly mislabel 8% of your pipeline forever. Build the sampling loop on day one:

1. Create a view AI – needs review filtered to AI reviewed unchecked, sorted by newest, limited by a date filter to the last seven days.
2. Each week, someone reviews 20–30 records in that view, corrects any wrong category or score by hand, and ticks AI reviewed.
3. Keep a tally. When a category is consistently wrong, fix the prompt, not the records, and re-test on the misclassified examples.

For the lead-score field, compare scores against what actually happened sixty days later. If 9s and 4s convert at the same rate, the prompt is producing noise and you should simplify it.

Wiring it into a workflow

Once classification is trustworthy, use it. An automation with trigger "When record matches conditions" (Request type is Sales and Lead score ≥ 7) can notify the sales channel and create a linked record in your CRM table. Another routes Support to the delivery team. Spam gets a view of its own and never interrupts anyone. The agent's job is to make those conditions possible; the automations do the rest. See our guide on choosing between Zapier, Make, and native automations for the plumbing.

When a formula is the better agent

If the rule can be written down, write it down. "Email domain is gmail.com → low score" is a formula, runs for free, and is deterministic. Use field agents for the part of the decision that requires reading prose, and formulas for everything else.

BaseBrainers designs and governs AI field agents for clients as part of our Airtable AI and Omni consulting. If you want a second pair of eyes on your prompts or your credit budget, get in touch.