Home/Blog/Letting AI Agents Into Your Airtable Base: Permissions, Audit Logs and What to Keep Read-Only
Data IntegrationNovember 19, 2026· Team Nuevexa

Letting AI Agents Into Your Airtable Base: Permissions, Audit Logs and What to Keep Read-Only

Claude, ChatGPT and Gemini can now read and write your Airtable records. Here's whose permissions they use, what audit logs show, and which tables to keep read-only.

Letting AI Agents Into Your Airtable Base: Permissions, Audit Logs and What to Keep Read-Only - Nuevexa

For a lot of operations and nonprofit teams, Airtable is not just another app. It is where the clients, volunteers, donors, projects, suppliers and grant deadlines live. Every other tool reads from it or writes to it. So when someone asks whether they can connect Claude or ChatGPT to the base, it is a bigger question than it looks.

The appeal is obvious. An assistant that can find the stale records, update a status, draft a form or build an interface page saves real hours. The worry is just as obvious. The same connection can change hundreds of records, read fields that were never meant to leave the team, or set up automations nobody reviewed.

Airtable's Fall Release, announced on 21 October 2026, makes this question practical rather than theoretical. Outside AI tools can now read and write your records through Airtable's MCP server, and admins get better visibility of what agents did. The controls that decide what an agent can reach in the first place are still arriving. This guide covers what you can lock down today, what is still coming, and how to decide table by table.

What changed in Airtable's Fall Release for AI agents?

Airtable's Fall Release article lists several pieces, and they are not all at the same stage.

  • Omni 2.0, Airtable's built-in agent, is generally available on all plans. It can import data, set up AI fields, configure automations and write scripts.
  • Airtable MCP is generally available and open to all customers. It now connects to Claude, ChatGPT, Gemini, Amazon Quick and Cursor, with Copilot and Muse listed as coming soon. Connected tools can read and write records, submit forms, and build automations and interfaces.
  • Agent audit logs are generally available to admins. They sit in the enterprise audit log, which is an Enterprise Scale feature, and the new agent activity events are still rolling out on Enterprise plans.
  • Agent permissions, which control what an agent can read or write down to the field or record, are coming soon, with a waitlist for beta access.
  • Bring your own agent is also coming soon. Airtable Canvas, for building apps and dashboards on live data, is in open beta.

The short version: agents can already get in, and you can already see more of what they did. Fine-grained control over what they can read is not here yet. That gap shapes every decision below.

What can a connected AI tool actually do in your base?

Quite a lot. Airtable's MCP help article lists searching and analysing data, creating and updating records, creating bases, building interface pages and forms, submitting forms on your behalf, and working with automations.

There are some built-in brakes. New and updated automations created through MCP are saved as drafts, and someone has to turn them on in the Airtable interface. Automations containing a Script action cannot be created or edited through MCP. Record creation is limited to 10 records per request, and every call counts against your normal API rate limits.

Those brakes help, but they are not a review process. An agent can still update a large number of existing records, one request at a time, if the person who connected it has permission to.

Whose permissions does an Airtable agent use?

The person who connected it. Airtable says MCP permissions mirror the permission level that person has been granted. Owners, Creators and Editors can read and update records they can access. Owners and Creators can also change the base structure. Commenters and Read-only users can read data but not change it. Interface-only users can work only with the pages and data those interfaces expose.

Omni follows the same rule. Airtable's Omni guide says it mirrors the permissions of the person using it, and field and table permissions can stop it making changes. One setting is worth checking: "Allow Omni to create, update, and delete data for interface-only collaborators" is on by default.

This has a practical consequence. If the base owner connects ChatGPT with full access, that connection can do whatever the owner can do. The agent does not get its own, smaller job description. It inherits yours.

Can you stop an agent reading certain fields today?

Not directly, yet. This is the part most teams assume is already solved.

Airtable's field and table editing permissions, on Team, Business and Enterprise Scale plans, control who can edit values in a field and who can create or delete records in a table. They do not hide a field from someone who can open the base. An agent connected by that person can read it too.

Until agent permissions arrive, you have four practical options:

  1. Scope the connection to specific bases. During the OAuth step you can choose custom access to specific bases, interfaces or workspaces rather than everything you can see.
  2. Give the agent an interface, not the base. Interface-only access limits what comes back to the data those pages show, on supported page layouts.
  3. Move sensitive fields out. Keep contact details, payment data or case notes in a separate base, and use sync, on Team plans and above, to share only the fields other teams and tools need.
  4. Use a dedicated account. On Enterprise Scale, service accounts let you connect integrations with access an admin sets, independent of any one person.

When agent permissions ship, Airtable says you will be able to let an AI tool work on, for example, customer feedback without reading customer contact information. Plan for that, but do not design today's setup as if it already exists.

How do you make an Airtable connection read-only?

Leave out the write scopes. Airtable's MCP controls page says you can make access read-only by omitting data.records:write and schema.bases:write when you set up a personal access token or OAuth consent. Its help article gives the read-only set: data.records:read, schema.bases:read, data.recordComments:read and workspacesAndBases:read.

Personal access tokens are also scoped to specific bases or workspaces. If two agents need different data, Airtable suggests a separate token for each, rather than one token that reaches everything.

The other route is the account itself. If the person or service account that connects the agent only has Read-only access to a base, the agent cannot write to it, whatever scopes it asks for. Doing both is safer than relying on either.

On Enterprise Scale, admins can also block MCP access across the organisation, block third-party API integrations, and allowlist only the AI tools they approve. On other plans, these choices sit with each person who connects a tool, which is a good reason to agree them as a team.

What do Airtable's audit logs show about agent activity?

It depends heavily on your plan.

The enterprise audit log is available on the Enterprise Scale plan only, and only to admins. With the Fall Release, it now tells you whether a person or an agent took an action, and in which base. Omni's actions are attributed directly. Actions from third-party AI tools trace back to the integration and the person who set it up. You can filter by actor, base, event type, org unit and date, and export or stream events through the audit log API, which by default looks back 180 days.

On other plans, your main trail is record revision history. Airtable's MCP controls page says it shows the user who authorised the MCP connection and the name of the app or token that made the change. It is useful for checking one record. It is less useful for reviewing a week of agent activity, because it is view-only, cannot be exported, and how far back it goes depends on your plan. On the Free plan it is two weeks.

If your plan has no audit log, decide before you connect anything how you will notice a bad bulk change. A "last changed by" field, a small change-log table, or a daily view of recently modified records all help.

Which tables should an agent read, write or never touch?

Decide per table, not per base. Most bases mix low-risk reference data with records that would be painful to get wrong. Use this as a starting point and adjust for your own data.

Type of dataTypical tablesAgent accessWhy
Reference dataServices, products, locations, programme listsReadUseful context, rarely changes, low risk if read
Work queuesTasks, requests, intake, ticket statusWrite with reviewAgents save time here, and mistakes are easy to spot and reverse
Core operational recordsProjects, cases, bookings, volunteer shiftsWrite through an interface, small batchesOther tools and people depend on these, so limit what changes at once
Contacts and personal dataDonors, clients, beneficiaries, staffKeep read-only, ideally through an interface without contact fieldsField-level read controls are not available yet
Money and compliancePayments, gifts, invoices, grant reportingKeep read-only or keep outErrors here reach finance, funders or regulators
Sensitive recordsSafeguarding, health, HR, case notesKeep out, in a separate baseToday the only reliable way to stop an agent reading a field is not to give it access
Base structureTables, fields, automations, interfacesHumans approveAgents can draft, but schema changes affect every connected tool

If a table does not clearly fit one row, treat it as the stricter one until agent permissions give you finer controls.

How do you pilot an AI agent on one shared base?

Start with one question you already ask every week, such as "which requests have been waiting more than a few days?" or "which volunteers have no shift this month?". Answer it read-only first.

Then work through these steps:

  1. List the tables in the base and place each one in a row of the table above.
  2. Choose who connects. Ideally a person or service account with only the access the pilot needs, not the base owner.
  3. Scope the connection to the pilot base or an interface, with read-only scopes to begin with.
  4. Run it for a couple of weeks and check the answers against what your team already knows.
  5. Add write access to one work-queue table, keep batches small, and review the changes in revision history or the audit log.
  6. Write down the rules: which tables, which actions, who approves automations, and who to tell if something looks wrong.

For a wider view of which systems to connect first and where approvals belong, see our guide to always-on AI agents for small business. If the agent will do more than answer questions, our AI agents service page covers how we scope that work.

So, should you let AI agents into your Airtable base?

Yes, for the right tables, with the right connection. Airtable has made it straightforward to connect Claude, ChatGPT, Gemini and others, and its built-in brakes, such as draft automations and permission mirroring, are sensible. Read-only connections on reference data and work queues are low risk and quickly useful.

The honest limit is reading. Today an agent can usually read whatever the person who connected it can read. Field-level agent permissions are coming, but they are not here, and the richest audit trail is on Enterprise Scale only. So the decision is less about which AI tool you use and more about how your base is structured: what sits in which table, who connects, and what you can see afterwards.

That structural work, splitting sensitive data out, setting up scoped connections and building a simple change log, is what Nuevexa does as a done-for-you implementation partner through our CRM and data integration service, using the Airtable plan and AI tools you already pay for.

If you want to see where an agent would actually save your team time, start with an automation scan.

Ready to automate?

Book a free strategy call and leave with a clear roadmap for your first automation build.

Book a call

Tags

AI AgentsCRM IntegrationAI Automation
TN

Team Nuevexa

Building intelligent systems that replace manual work.

The 90-second scan

Ninety seconds to see where your leads are leaking.

We build and run the lead capture and follow-up system for accounting, advisory, legal and trades firms with 10 to 200 staff. Fixed price. Your accounts. Live in about three weeks.

Take the 90-second scan

Frequently Asked Questions

If manual work is your ceiling, let's remove it.

We reply within 1 business day.