Airtable AI Prompts: 20 Field Agent Templates for Every Base

TL;DR: Airtable's AI field is officially a "field agent" now, and it comes in five flavors (Long text, Formula, Linked record, Select, and Number). Reference another field with {Field Name} by typing the curly brace and picking from the list. Below: verified plan gating, the exact syntax, and 20 field agent prompts you can paste into Custom Instructions today.

If you searched "airtable ai prompts" expecting a single feature called "AI field," you'll find one, technically. It's just not what Airtable calls it anymore, and the field type you pick changes what the AI is allowed to output. A Long text field can write you a paragraph. A Select field has to pick one option from a closed list. A Number field has to return, well, a number. Get the field type wrong for what you're asking and the agent either can't do the job or has to squeeze a subjective answer into a structure that doesn't fit it.

This post covers the terminology Airtable actually uses in its own help center as of September 2026, exactly how the field-reference syntax works, which plans include it (the two most relevant Airtable pages don't fully agree, so both get quoted), and 20 copy-paste prompts organized by the kind of base you're probably running.

What does Airtable call its AI field in 2026?

Field agent. Airtable's own help article is direct about it: "In Airtable, we refer to most AI-enabled fields as Field agents." That's the umbrella term for five different field types that can run AI, and the plain single-cell text generator most people picture when they hear "Airtable AI field" is one member of that family, labelled "Long text field (a.k.a. AI text)" in the docs.

The distinction matters because the other four field types don't behave like a chat window at all:

• Formula field: AI helps you write the formula itself, via a "Generate formula" button in the field editor. The AI runs once, at setup, not per record.
• Long text field (AI text): the classic one. Custom instructions run per record and write into a text cell.
• Linked record field: toggle "Show AI-suggested records" and the agent recommends which records to link, based on other fields you choose to expose as context.
• Select fields (single or multiple): the agent has to output one option (or several) from your existing choices, not free text.
• Number-based fields (Number, Currency, Percent): the agent returns a numeric value derived from other fields.

Each type accepts the same kind of custom-instructions prompt, but only the Long text type gives you unconstrained prose. Everything else is closer to a classifier or a calculator wearing a chat interface, and the prompts you write for them should read that way.

Which Airtable plans include AI field agents?

Here's where Airtable's own documentation gets genuinely inconsistent, and you should see both sides before you plan around either one.

The field-agent article states plan availability plainly: "All paid plans". Read literally, that excludes Free.

The separate AI-billing article opens differently: "All Airtable plans include AI functionality with a set number of free AI credits each month, offering great flexibility and value." It then lists a Free-plan allocation explicitly: "Free plans - 500 credits per user with Editor permission and above". Team gets 15,000 credits per billable collaborator, self-serve Business gets 20,000 per paid user, and Enterprise Scale gets 25,000 per paid user at list price.

Those two statements don't reconcile cleanly. One reasonable read: the "All paid plans" note on the fields article may be stale relative to the billing article, which is more recent and more specific about the Free tier's 500-credit allotment. I'm not resolving it further here; if you're on a Free workspace and field agents don't appear where you expect, that's the actual account behavior to test, not something either doc alone will settle for you.

What both sources agree on: Airtable's plan names as of today are Free, Team, Business, and Enterprise Scale (the renamed successor to the old "Enterprise" tier). Per airtable.com/pricing, accessed September 3, 2026, Team starts at $20/user/month billed annually and Business at $45/user/month billed annually, with Enterprise Scale priced by sales conversation. Free-plan customers who need more AI credits are directed to upgrade to Team; there's no way to buy more credits while staying on Free.

How do you reference another field inside an Airtable AI prompt?

This is the part worth getting exactly right, because it's the difference between a field agent that adapts per record and one that's silently reading the same static text on every row.

Type a curly brace directly in front of the field's name inside the Custom Instructions box, and a list of your table's fields pops up to choose from, or skip typing and click the + Insert field button to get the same list. Either way, Airtable's docs confirm what happens next: "The field name will appear as a token in the prompt and will automatically update, even if the name of the field is changed later."

A separate FAQ entry in the same doc gives the token's exact shape: "The field will appear as a token (e.g., {Campaign Name}) in the custom instructions. If the field name changes later, the token will automatically update." So for a field literally named Campaign Name, the token in your prompt reads {Campaign Name}, spaces and all, no underscores required.

Airtable's own example prompts use this pattern directly:

Write an email that addresses the following customer feedback: {customer_feedback}

Create a meeting summary based on the following notes: {Meeting Notes}

Two behaviors worth knowing before you build anything real on top of this:

1. The token survives a rename. Change the field's name in the table and the prompt updates automatically, because the token is bound to the field itself, not to the text string you typed.
2. A field can be made optional. Click the field's dropdown inside the prompt to reveal "Make optional", which lets the agent run even when that field is empty, and you can then add conditional instructions for what to do when it does have a value.

If you're picturing something like a mail-merge system, that's a fair comparison, though Airtable's version resolves inside one base rather than across any tool you happen to be using. Prompt Architects' Global Variables solve a version of the same problem with a different token shape, {{name}} in double braces instead of single, and the resolved value works in whatever AI chat you paste the prompt into rather than one base. More on that comparison further down.

20 Airtable field agent prompts, by base type

Every template names the field type it belongs in and, where the output has to satisfy a closed list, spells out the fallback. Swap the bracketed field names for whatever your table actually calls those columns.

CRM and sales bases

1. Deal summary (Long text field)

Summarize the current state of this deal in 2-3 sentences for a sales manager
skimming a pipeline view. Do not repeat raw field values verbatim; synthesize them.

Company: {Company Name}
Stage: {Deal Stage}
Last activity: {Last Activity Notes}
Deal value: {Deal Value}

If Last Activity Notes is empty, say so explicitly rather than inventing activity.

2. Renewal risk flag (Select field, single select)

Classify this account's renewal risk based on the fields below.
Choose exactly one: "low risk" | "watch" | "high risk" | "unknown"

Usage trend: {Usage Trend}
Support tickets (90 days): {Support Ticket Count}
NPS score: {Latest NPS Score}
Contract end date: {Contract End Date}

If two or more inputs are empty, choose "unknown" rather than guessing.

3. Suggested next step (Long text field)

Given this deal's stage and notes, suggest one concrete next action for the
account owner. One sentence. Do not suggest generic advice like "follow up soon."
Name a specific action tied to what {Deal Notes} actually says.

Stage: {Deal Stage}
Notes: {Deal Notes}

4. Renewal likelihood score (Number field, 0-100)

Estimate this account's renewal likelihood as a whole number from 0 to 100,
based only on the fields below. Return the number alone.

Usage trend: {Usage Trend}
Support sentiment: {Support Sentiment}
Champion still employed: {Champion Status}

If Champion Status is "departed," cap the score at 40 regardless of other signals.

Support and feedback bases

5. Ticket sentiment (Select field, single select)

Read this support ticket and classify the customer's sentiment.
Choose exactly one: "frustrated" | "neutral" | "satisfied" | "unclear"

Ticket body: {Ticket Body}

If Ticket Body is empty or unreadable, return "unclear."

6. Draft support reply (Long text field)

Draft a first-response reply to this support ticket. Acknowledge the specific
issue in {Ticket Body}, do not promise a resolution timeline, and close by asking
one clarifying question if the issue description is incomplete.

Ticket body: {Ticket Body}
Customer name: {Customer Name}
Product area: {Product Area}

7. Feedback theme tags (Multiple select field)

Tag this piece of feedback with every theme that applies, from this closed list only:
"pricing" | "performance" | "missing feature" | "bug" | "onboarding" | "praise" | "other"

Feedback: {Feedback Text}

If nothing in the list fits, tag it "other" rather than leaving it untagged.

Content and marketing bases

8. Meta description generator (Long text field)

Write a meta description for this piece, 140-155 characters, that includes the
target keyword naturally and reflects what the piece actually argues, not a
generic summary.

Title: {Title}
Target keyword: {Target Keyword}
Body draft: {Body Draft}

9. Social caption variants (Long text field)

Write three short social captions (under 220 characters each) for this piece,
one direct, one question-led, one number-led. Base every caption on a real
detail from {Body Draft}; do not invent statistics.

Title: {Title}
Body draft: {Body Draft}

10. Ad copy angle (Long text field)

Given this product's positioning notes, write one ad headline (under 40 characters)
and one supporting line (under 90 characters). Ground the angle in {Positioning Notes}
rather than generic category claims.

Positioning notes: {Positioning Notes}
Audience: {Target Audience}

Recruiting and hiring bases

11. Resume summary from attachment (Long text field, document extraction)

Extract from the attached resume: total years of experience, most recent title,
and the top 3 skills relevant to {Role Title}. Return as three labeled lines,
no extra commentary.

Resume attachment: {Resume File}
Role being hired for: {Role Title}

If the attachment cannot be read, return: "Resume unreadable - manual review needed."

12. Candidate fit score (Number field, 0-10)

Score this candidate's fit for {Role Title} from 0 to 10, based only on the
listed requirements versus the resume summary. Return the number alone.

Role requirements: {Role Requirements}
Resume summary: {Resume Summary}

13. Interview question generator (Long text field)

Generate 4 interview questions for this candidate that probe specifically at
gaps between {Role Requirements} and {Resume Summary}. Do not generate generic
behavioral questions unrelated to this candidate's actual background.

Role requirements: {Role Requirements}
Resume summary: {Resume Summary}

Research and knowledge bases

14. Meeting notes summary (Long text field)

Summarize these raw meeting notes into: decisions made, open questions, and
owners for each open question. Use the exact names mentioned in the notes;
do not invent an owner if none was assigned.

Raw notes: {Raw Meeting Notes}

15. Contract key terms extractor (Long text field, document extraction)

Extract from the attached contract: effective date, renewal terms, and
termination notice period. Return as three labeled lines. Quote the exact
contract language for the termination notice period rather than paraphrasing it.

Contract attachment: {Contract File}

If a term is not present in the document, write "not specified" for that line.

16. Citation formatter (Long text field)

Reformat this citation into APA 7th edition style. If required fields (author,
year, title, source) are missing from {Raw Citation}, list which fields are
missing instead of guessing a value for them.

Raw citation: {Raw Citation}

Operations, real estate, and inventory bases

17. Listing description (Long text field)

Write a property listing description, 3-4 sentences, using only the facts in
the fields below. Do not invent amenities that are not listed.

Address: {Property Address}
Bedrooms: {Bedrooms}
Key features: {Key Features}
Price: {List Price}

18. Reorder flag (Select field, single select)

Based on current stock and weekly usage, classify this SKU's reorder status.
Choose exactly one: "reorder now" | "reorder soon" | "sufficient" | "unknown"

Current stock: {Current Stock}
Weekly usage average: {Weekly Usage}
Lead time (days): {Supplier Lead Time}

If Weekly Usage is empty, return "unknown" rather than assuming zero usage.

19. Maintenance ticket priority (Select field, single select)

Classify the priority of this maintenance request.
Choose exactly one: "emergency" | "urgent" | "routine"

Request description: {Request Description}
Property type: {Property Type}

Treat anything mentioning water, gas, fire, or no-heat-in-winter as "emergency"
regardless of how the request is phrased.

Finance bases

20. Expense category (Select field, single select)

Categorize this expense line item.
Choose exactly one: "travel" | "software" | "meals" | "office" | "other"

Line item description: {Expense Description}
Vendor: {Vendor Name}

If the description is ambiguous between two categories, choose "other" and
note the ambiguity is not requested here - just the category.

What limits should you plan around?

Airtable documents a combined word budget across your instructions, the fields you reference, and the agent's response: "Lower powered AI models: Up to ~12,000 words" and "Higher powered AI models: Up to ~90,000 words". Cross that limit on the input side and you get an error; cross it on the output side and the response gets truncated rather than rejected.

Two more constraints worth knowing before you design around them:

• Forms are a hard no. Airtable's FAQ states it without qualification: "No, field agents cannot be used in Airtable forms." A field with AI configuration still renders as a normal fillable field to whoever submits the form; the AI side just never runs during that submission.
• Document extraction has a plan split of its own. The supported attachment formats are, per Airtable's docs, "PDF, DOCX, PPTX, JPG, PNG, TIFF, BMP, XLSX, or WEBP." But image formats specifically carry a separate gate: "PNG, JPG, and WebP file type support is only available on Business and Enterprise Scale plans after opting into Airtable Labs", while the same article separately notes "PNG, JPG, and WebP support is available by default on Free and Team plans" for a different context in the same FAQ. Read that as: check your own base's behavior rather than assuming from either sentence alone.

If your field agent needs to reach outside the base, two features exist for exactly that, both opt-in per field rather than always-on: Enable internet search, which Airtable's docs describe as supporting "both general internet search... as well as search that specifies a URL", and a Google Drive connection that can be scoped to specific folders using a tag format in the prompt itself: <gdrive_folder> {Insert your identifier(s) here} </gdrive_folder>, capped at 100 folders per run.

Airtable field agent vs Prompt Architects Global Variables

Both features solve the same underlying problem, a value you don't want to retype, but they solve it in different places.

Neither is strictly better; they're built for different moments. A field agent belongs where the data already lives, recalculating itself as records change without anyone opening a chat window. A Global Variable belongs where the person does the writing, carrying one saved fact (a product name, a brand voice line, a technical spec) into whatever tool they happen to be prompting that day.

Common mistakes with Airtable field agent prompts

The hallucination risk here isn't the model inventing facts from nowhere; it's the model filling gaps you didn't tell it to leave alone. A few patterns show up repeatedly:

• No instruction for empty fields. If {Notes} is blank, an ungoverned prompt will often write plausible-sounding filler rather than flagging the gap. Every template above tells the agent explicitly what to do when an input is missing.
• An open-ended Select prompt. If you don't spell out the exact allowed values inside the instructions (not just the field's configured options), the agent can occasionally propose language close to but not matching an existing option, and Airtable will reject or mismatch it.
• Forgetting who can see the prompt. Airtable's own security note is specific here: any data referenced in a field agent's prompt can be visible to anyone who can view that field, even inside an interface that intentionally hides the source field from certain collaborators. Don't reference a field you wouldn't want the reader of the output field to see.
• Picking a low-cost model for a tool-heavy prompt. Airtable's own troubleshooting guidance for field agents notes that pairing a lower-powered model with internet search or multi-step reasoning can throw errors; if a prompt uses search or Google Drive, budget for a stronger model tier.

For the same discipline applied to a different kind of automation, the Make.com scenario templates and the n8n node templates cover output contracts for a pipeline rather than a database cell. If your prompt library currently lives in a spreadsheet rather than a base, why prompt library spreadsheets break is worth reading before you build the next one. And if you're feeding a field agent's output into anything downstream that expects strict structured output, the free JSON prompt generator covers the schema side of that problem.

Sources

All verified September 3, 2026.

• Airtable Help Center, Using Airtable AI in fields: field-agent terminology, field types, reference syntax, word limits, forms limitation, security note, FAQ.
• Airtable Help Center, Airtable AI billing: per-plan AI credit allocations and reset cadence.
• Airtable Help Center, AI Field agent - Build prototype: confirms "field agent" terminology and plan-availability wording on a second article.
• Airtable, Pricing: current plan names and per-seat prices.