Airtable for Startups: Build an AI-Assisted Operations Base

Build one base with three linked tables (Customers, Tasks, Invoices), add three AI fields that only write text (a note summary, a task category and an invoice-chase draft), and connect two automations. Check the plan first: Free allows 1,000 records per base and 500 AI credits per editor each month; Team lifts those to 50,000 and 15,000.

Most startups get stuck on structure, not on AI. A base built as one long table can't answer "which customers owe us money and what tasks are open for them?", and an AI field on top of a messy table only writes messy text faster. So the order below is tables first, links second, AI third, automations last.

Follow me on Instagram@sagnikteaches

The startup this base is built for

Take a four-person electrical contractor that started trading this year. Two people are on site, one handles quotes and scheduling, and the founder does everything else. In a normal month the business has about 35 active customers, 60 tasks (call back, order parts, book an inspection, chase a certificate) and 30 invoices. Everything currently lives in a shared spreadsheet, a phone's notes app and the founder's head. The aim is a single place where the question "what needs doing today, and who owes us money?" takes ten seconds to answer.

Connect on LinkedInSagnik Bhattacharya

The same shape fits a plumbing firm, a roofing company or a small insurance broker: swap "job" for "policy renewal" and the tables barely change. What matters is that every record belongs to a customer, every task belongs to a customer or an invoice, and every invoice belongs to a customer.

Subscribe on YouTube@sagnikteaches

Check the plan limits before you build anything

Airtable changes plan limits and AI allowances often, so read Airtable's pricing page before you commit. These are the figures on the page when I checked in November 2026, in USD list prices:

Three details matter for a startup. First, on Team and Business you pay for every person who has edit permission on at least one base, so a base with six editors on Team costs six seats. Second, AI credits are pooled across the workspace, and what a single run uses depends on the field and the model behind it, so the only reliable way to budget is to run a test on ten records and read the usage. Third, Airtable's help page on AI in fields gives its plan availability as all paid plans, while the pricing page still lists 500 AI credits per editor on Free, so check what your Free workspace actually offers before you design around it. The companion tutorial on estimating what AI fields cost a month goes through that test in detail.

A quick record sum for the electrician

Per year, the contractor creates about 360 invoices (30 a month), 720 tasks (60 a month) and perhaps 120 new customers. That is 1,200 records a year before attachments or notes. The Free cap of 1,000 records per base is reached around month ten. If the founder also logs every call as a record (say 100 calls a month), it arrives in month five. The practical reading: start on Free to prove the structure, and budget for Team before the busy season, not after the base stops taking new records on a busy Monday. Airtable's pricing FAQ says that at the cap you keep your bases and data but cannot add records until you upgrade.

Three tables and how they connect

Create a new base and delete the default table. Then make three tables. Use plain, singular field names; the AI fields will refer to them in prompts later, and short names stop prompts getting confusing.

Customers

• Name (single line text, the primary field)
• Type (single select: Homeowner, Landlord, Builder, Commercial)
• Status (single select: Lead, Active, Dormant)
• Notes (long text, where the team dumps raw call notes)
• Tasks and Invoices (linked-record fields, created automatically by the links below)
• Outstanding (a rollup that sums the Unpaid amount field on linked invoices)
• Last contact (a rollup of the latest Completed on date across linked tasks)

Tasks

• Task (primary field, a short sentence)
• Customer (link to Customers)
• Owner (a User field)
• Due (date) and Done (checkbox)
• Completed on (date, filled in when Done is ticked)
• Details (long text, the messy message or voicemail transcript)
• Category (the AI field, covered below)

Invoices

• Invoice number (primary field)
• Customer (link to Customers)
• Amount (currency)
• Issued and Due (dates)
• Status (single select: Draft, Sent, Paid, Disputed)
• Days overdue (formula)
• Unpaid amount (formula: IF({Status}!="Paid", {Amount}, 0), which the Outstanding rollup adds up)
• Chase draft (the AI field, covered below)

The Days overdue formula keeps arithmetic out of the AI. Something like IF(AND({Status}="Sent", IS_AFTER(TODAY(), {Due})), DATETIME_DIFF(TODAY(), {Due}, "days"), 0) returns a plain number for each unpaid invoice, and zero otherwise. Check the field names in your own base match the formula exactly, including capital letters, because Airtable rejects a formula that refers to a field it can't find.

Fill in ten real customers, ten tasks and ten invoices by hand before you add anything clever. The point is to find out whether the structure fits how the team actually talks about the work. If the team keeps wanting a field that doesn't exist (a job address, a site contact), add it now. Retro-fitting after the AI fields are built means rewriting prompts.

AI field one: a note summary on each customer

Airtable now calls most of its AI-enabled fields field agents. To add one, click the + after the last field; the menu that opens has a Field agents section at the top, and Build a field agent opens a Long text field with a Custom instructions box, where typing { inserts a reference to another field. Menu labels move between releases, so if yours differ, look for the agent option. The idea is the same under any name: a field whose value is written by a model from other fields in the same record, using instructions you write. Only owners and creators can set one up; editors can run it.

The first use is the least risky. A Summary field on Customers reads the long Notes field and produces three lines the founder can read before ringing someone. Raw notes after six months look like this (illustrative):

12 Mar - rang re faulty consumer unit, said Tuesday ok. Dog in garden.
19 Mar - did the swap, signed off. Wants quote for outside sockets.
2 Apr - sent quote 840. Said wait till after holiday.
30 Apr - rang, still thinking. Prefers texts not calls.

The prompt for the Summary field:

You are helping the owner of a small electrical contractor prepare for a call.
Read the notes below and write exactly three short lines:
Line 1: what we have done for this customer.
Line 2: what is open or waiting on them, with any quote amount.
Line 3: how they prefer to be contacted and anything to remember.
Use only facts that appear in the notes. If something is not in the notes, write "not recorded". Do not guess dates or amounts.

Notes: {Notes}

Sample output (illustrative):

Replaced a faulty consumer unit in March; signed off 19 Mar.
Open: quote of 840 for outside sockets sent 2 Apr, customer waiting until after the holiday.
Prefers text messages to calls; has a dog in the garden.

Notice the guard in the prompt: "use only facts that appear in the notes" and the "not recorded" fallback. Without them, a model asked for "preferred contact" on a customer whose notes never mention it will often invent a plausible answer. Test the field on five customers with thin notes and confirm it writes "not recorded" instead of filling the gap. If it doesn't, tighten the prompt and test again.

Two settings decide when it runs at all. A new field agent runs only when someone clicks Run agent in a cell, until you turn on Run automatically under the field's Settings. And a field you reference in the instructions needs a value before the agent will run, unless you open that reference's menu and choose Make optional. A customer with an empty Notes field may therefore get a blank Summary rather than "not recorded", which is fine, provided the team knows what a blank means.

AI field two: sorting incoming tasks

The second field sits on the Tasks table. Staff paste messages or voicemail transcripts into Details, and the field writes a Category from a fixed list so views and reminders can sort by it. A fixed list matters: if you let the model invent categories, you get "Callback", "Call back", "Follow-up call" and a dozen more that break every filter. Build Category as a Single select field agent with the seven options below already set up as choices, so the output can only ever be one of them, and keep the list in the instructions too.

You are sorting jobs for a small electrical contractor.
Choose exactly one category for the message below from this list:
Emergency, Quote request, Booking change, Parts or supplier, Paperwork, Payment query, Other.
Emergency means a safety risk such as sparking, burning smell, no power to a whole property or water near electrics.
Reply with the category name only.

Examples:
Message: "Socket in the kitchen is hot and smells of burning" -> Emergency
Message: "Can you move Thursday to Friday morning?" -> Booking change
Message: "Do you do outside lighting? Roughly what would it cost?" -> Quote request

Message: {Details}

Run it over the ten tasks you entered by hand and compare each result with what you would have chosen. In a test of this kind, expect most of the clear messages to land correctly and the ambiguous ones ("sorry about last week, can we sort it") to wobble. That is the sign to add an example for the case the model got wrong, not to switch models. After three or four example additions, a short list like this usually settles.

One rule keeps this field safe: an Emergency category must trigger a human, not a silent filter. A view of Emergency tasks that nobody opens is worse than no sorting at all, so pair the category with an automation (below) that messages the on-call person, and read the Emergency view each morning anyway.

Where the sorted tasks go

Make four views of the Tasks table:

1. Today: grid view, filter Done is unchecked and Due is today or earlier, sorted by Category so Emergency rows sit first.
2. By owner: kanban grouped by Owner, so each person sees their own column.
3. Waiting on customer: filter on a Waiting checkbox, so tasks that are blocked don't clutter Today.
4. Calendar: calendar view on Due, useful for the person scheduling visits.

AI field three: invoice chase drafts, with no figures from the AI

The third field writes the first draft of a payment reminder. The sums come from ordinary fields and the formula above; the AI only phrases them. That split is the most useful habit in the whole base: figures come from formulas, wording comes from AI.

Create a Chase draft field on Invoices that reads Customer name, Invoice number, Amount, Due and Days overdue:

Write a short, polite payment reminder from a small electrical contractor to a customer.
Facts to use exactly as given, without changing any figure:
Customer name: {Customer}
Invoice number: {Invoice number}
Amount: {Amount}
Due date: {Due}
Days overdue: {Days overdue}

Tone by days overdue: under 8 days = friendly nudge; 8 to 21 days = polite and direct; over 21 days = firm but courteous, and mention that we are happy to talk if there is a problem.
If Days overdue is 0, write "Not overdue".
Do not add late fees, legal threats or any promise that is not stated above.
Sign off with "[your name]".

Sample output for an invoice 12 days overdue (illustrative):

Hello [first name],

Invoice 1042 for 840.00 was due on 3 April and is now 12 days overdue. Could you let us know when we can expect payment, or tell us if there is anything about the invoice we should look at?

If it has already been paid, thank you, and please ignore this note.

Kind regards,
[your name]

Three things to fix whenever you see them. The field may drop the currency symbol or write the amount in a different format, so state the currency in the prompt or make the Amount reference a text version of the figure. Customer is a linked-record field, and Airtable's help page warns that linked, lookup and rollup tokens can reach the model as a list rather than the readable name you see in the cell; add a rollup on Customer that returns the name with ARRAYJOIN(values, ", ") and reference that field in the prompt instead. And "[first name]" will appear literally, which is deliberate: a person replaces it before sending. Never let a chase draft go out automatically. A drafted reminder sent to a customer who has already paid, or who is disputing the work, costs more goodwill than the invoice is worth. The tutorial on summarising and tagging quote requests shows the same pattern for the front end of the pipeline.

Two automations that earn their runs

Automations are what turn a tidy database into an operations system, but each run counts against your plan's allowance (100 on Free, 25,000 on Team), so build two that matter rather than ten that don't.

Automation one: an Emergency alert

• Trigger: when a record in Tasks matches conditions, Category is "Emergency" and Done is unchecked.
• Action: send a message to the on-call person (Send email, or the Slack Send a message action if the team uses Slack), including the task text and customer name.

Airtable's help page warns that an automation reading a field agent's output can run before the agent has written anything, and still report success. Two choices avoid that here: turn on Run automatically for the Category field, and base the trigger on Category itself, as above, never on Details, the field the agent reads. Then test it by creating a dummy task with the text "burning smell from socket" and timing how long the alert takes to arrive.

Automation two: an overdue invoice task

• Trigger: when a record in Invoices matches conditions, Days overdue is greater than 7 and Status is "Sent".
• Action: create a record in Tasks: "Review and send chase for invoice [number]", linked to the same customer, owner set to the founder, due today.

Days overdue changes because the calendar moves on, not because anyone edits the invoice. Airtable's formula reference says TODAY() recalculates roughly every 15 minutes while the base is open and, when it is closed, about hourly and only if the base has time-dependent automation triggers or actions, so the task may appear some time after the invoice tips over rather than at midnight. Watch the first few invoices tip past seven days and confirm a task appears each time. If one is missed, rebuild the trigger as At a scheduled time each weekday morning, followed by a Find records step for Sent invoices more than seven days overdue.

The person who opens the task reads the Chase draft, edits it, sends it from their own email and ticks Done. A human sends every message. The automation's job is to make sure nobody forgets. For more patterns at this scale, six Airtable automations to copy covers the rest.

What the first month of AI credits looks like

Airtable does not publish one credits-per-run figure for every field and model. Its AI billing help page lists typical costs per field agent run instead: about 1 credit to categorise a piece of feedback, 10 for an agent that searches the web and 200 for one that analyses a document. The Preview under each field's Custom instructions shows what the current set-up costs, and a model marked "Low cost" uses fewer credits. On the Team plan the pool is 15,000 credits per paid user a month, per the pricing page, and credits reset at the start of your billing cycle.

An illustrative sum. Say Category costs 1 credit a run, and assume, for the method only, that Summary and Chase draft cost 5 each because they read and write more text. The electrician's month is 60 categorisations (60 credits), 35 summaries (175) and 30 chase drafts (150): 385 credits. That fits inside one Team seat's 15,000 with room to spare, but it uses most of a single Free editor's 500, and a Free workspace cannot buy more. Replace the 5 with the figure your own Preview shows.

The bigger risk is not the monthly total but re-runs. With Run automatically on, a field agent generates again whenever the fields it reads change, so a customer whose notes are edited daily may use many more credits than one touched weekly. Leave Run automatically off on Summary and click Run agent before a call; keep it on for Category, because the Emergency alert depends on it. Only workspace owners can see credit usage, in the workspace settings, so the weekly check belongs to the founder. Airtable also emails a Team workspace when it passes 75% of its credit limit, which is the signal to look at which field is re-running.

A week of running the base: what the founder does

1. Monday, 10 minutes: open the Today view, read the Emergency and Payment query rows first, assign owners to anything unassigned.
2. Daily, 5 minutes: staff paste new messages into Tasks, check the Category the AI chose and correct any wrong ones.
3. Thursday, 10 minutes: open the overdue invoices view grouped by customer, read each Chase draft, edit and send.
4. Friday, 5 minutes: the founder, as workspace owner, checks AI credit usage in the workspace settings and the automation run count, and notes both in a "Base health" record.

Correcting wrong categories is more valuable than it looks: each correction tells you which example to add to the prompt. Keep a running list of the misses and add the three most frequent as examples once a month.

When the chase draft got the customer wrong

A failure worth planning for, described as an illustration: a plumbing firm links invoices to customers by name, and two customers share a surname at different addresses. An invoice is linked to the wrong "Smith", the overdue automation fires, and the draft goes to a customer who owes nothing. The AI did exactly what it was told; the data was wrong.

Two habits prevent it. Give every customer a unique Customer ID (an autonumber field) and show it in the draft's header so the sender sees it, and add an address or site field to the primary identification. Second, never send a draft you haven't read. The founder's Thursday ten minutes exist for exactly this. The tutorial on cleaning customer records before adding AI goes through duplicates and near-matches in detail.

Checking the base works before you rely on it

Run five tests, each with a deliberately awkward record:

• A customer with only one line of Notes: does the Summary say "not recorded" for the missing lines rather than inventing history? And with Notes empty, does the field stay blank as expected?
• A task with two issues in one message ("quote for lights and the fuse board is buzzing"): does the Category choose Emergency?
• An invoice with a due date today: does Days overdue show 0 and the Chase draft say "Not overdue"?
• A paid invoice marked Sent by mistake: does the overdue automation create a task? If it does, add a rule to your Thursday review to check Paid status first.
• A record created by a form, not typed in: do the AI fields still run? Airtable's help page says automatic generation picks up changes while the base is loaded in a browser, so test an entry made while nobody has the base open and note how long the Category takes to appear.

Write down what you saw. A base that passes five awkward tests is far more trustworthy than one that passed ten neat ones.

Adding people without losing control

Give editor rights to the people who create and change records daily, and read-only rights to anyone who just needs to look. With Free capped at five editors and Team billing per editor, access is also a cost decision. Before inviting a sixth person to Free, decide whether the base has earned the Team plan. The permission-levels checklist helps decide who should be what. Keep one person, usually the founder, responsible for changing field names and formulas; a renamed field breaks prompts and formulas quietly.

Finally, write a half-page "how this base works" note and keep it in the base as a record in a Notes table, or in a text field on a Base health record. It should list each AI field, what it reads, what it writes and who checks it. When someone leaves, or the founder is ill, that note is what keeps the base useful. The wider question of which workflow to automate next sits in seven AI workflows for startups ranked by set-up effort.

Questions startups ask before building the base

Can a startup run on the free Airtable plan?

For a few months, yes. The Free plan allows 1,000 records per base, up to five editors and 100 automation runs, plus 500 AI credits per editor a month. Airtable's help pages list field agents under the paid plans, so confirm the AI fields work in your Free workspace. A team logging 40 jobs and 30 invoices a month adds about 840 records a year before counting customers, so it reaches the record cap in roughly a year; plan the move to Team early.

Do AI fields change my numbers or invoice totals?

They should never touch figures. Keep amounts, dates and statuses in ordinary fields and formulas, and let AI fields read them and write text only: summaries, categories, draft wording. If a field output contains a number, check it against the source field.

Who should be allowed to edit the base?

Give editor rights to people who add or change records daily and read-only rights to everyone else. On paid plans you are charged for every user with edit permission on at least one base, so a lighter set of editors also costs less.

What happens when AI credits run out?

The base keeps working, but expect AI fields to stop generating new results until credits reset at the start of your billing cycle. Team and self-serve Business plans can buy add-on credit packs (listed from $20 a month for 10,000 credits); Free has to upgrade to Team. Only workspace owners can see usage, in the workspace settings, so the founder should check it weekly during the first month.

Is it safe to put customer details into an AI field?

Read Airtable's current AI and data-handling terms before you do, and keep the AI fields to the columns they need. A summary field needs the notes, not the customer's bank details. Leave payment and ID details out of any field that feeds AI.

Further reads

• Airtable vs Google Sheets for Business Data AI Can Use — Check whether a spreadsheet would do before you build.
• Airtable Team vs Business Plan: Records, AI Credits and Admin Controls — Know when Team stops being enough.
• AI Accounting for Startups: Monthly Close, Burn and Runway — Connect the invoice table to your monthly close.
• Startup AI Tools for Growth: 8 That Earn Their Subscription — See which other tools earn a place beside Airtable.
• How to Stop Zapier and Make Automations Breaking Silently — Catch automations that fail without telling anyone.
• AI Tools for Operations Management: Weekly Exceptions Report From Exports — A worked example with three exports, a review prompt, a sample exceptions list and the checks that catch a wrong join before it misleads you.
• AI for Startup Networking: Find and Follow Up With the Right People — A founder's practical system for event shortlists, research, individual follow-ups and a simple relationship tracker.
• Airtable vs AppSheet for a Job Tracker: Cost, Offline Use and AI — A six-person trade firm's job tracker priced on Airtable and AppSheet, with offline behaviour, record limits and AI features compared.
• Airtable vs Notion as a Business Database: Which Handles AI Better? — Airtable's AI fields and credits against Notion's agents and Business plan, compared on price, structure and upkeep for a small team.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Airtable pricing page (plans, records, automation runs, AI credit allowances and billing FAQ), Airtable help-centre pages on field agents, AI billing, the When record matches conditions trigger, the formula reference and field types, checked November 2026.