# AI and Airtable. URL: https://levystreet.com/integrations/airtable/ Updated: 2026-10-19 An AI Airtable integration guide for NZ: Omni and AI credits, the Web API, webhooks and MCP server, plus record enrichment with validation and approval. ## Short answer An AI Airtable integration uses a model to fill in, check or classify fields in your bases, through Airtable's Omni assistant and field agents, the Web API, or Airtable's MCP server. For New Zealand businesses the reliable pattern is enrichment into separate suggested fields, validated against your field types and options, with a person approving before anything overwrites the master record. - Every Airtable plan includes monthly AI credits; Team includes 15,000 per paid user. - The Web API allows five requests per second per base; excess calls wait 30 seconds. - Airtable's MCP server works on all plans and mirrors each user's base permissions. - Webhooks created with tokens expire after seven days unless refreshed. - Enriching records about people from other sources brings IPP 3A notice duties since 1 May 2026. ## What Airtable's own AI does (October 2026) Airtable's assistant is [Omni](https://support.airtable.com/articles/1744327578-using-omni-ai-in-airtable): it builds apps, researches the web, analyses data and documents, and creates and updates records. Field agents run AI per record inside a field. Both draw on AI credits. As listed on 19 October 2026, every plan includes monthly credits (Free 500 per editor, Team 15,000 per paid user, Business 20,000, Enterprise Scale 25,000), pooled across the workspace or organisation and expiring at month end. Airtable's prices are shown in USD to NZ visitors: Team is US$20 per seat per month billed yearly and Business US$45. Credit use varies with input size and the model selected, and web content pulled into an answer is charged too. One setting to check early: "Allow Omni to create, update, and delete data for interface-only collaborators" defaults to on. ## Integration routes and when to use each | Route | What it gives you | Use it when | |---|---|---| | Omni and field agents | AI inside bases, billed in AI credits | Simple per-record enrichment at small volumes | | Web API | Records, fields and schema through personal access tokens or OAuth | You need validation, evidence and an audit trail | | Webhooks | A ping, then your code fetches change payloads; signed with an HMAC header | You want to react to new or changed records without polling | | MCP server (`mcp.airtable.com/mcp`) | Read, create and update records from Claude, ChatGPT or Cursor, up to 10 records per create request | Individuals doing analysis or small edits in their assistant | | Airtable automations, n8n, Make or Zapier | Triggers and HTTP steps | Light glue between Airtable and other systems | ## Automation 1: product record enrichment with validation - **Inputs**: New rows in a Products table with SKU, supplier and a spec-sheet PDF or supplier URL. - **Model step**: Extract category, dimensions, weight, material, country of origin and a short description into a fixed JSON schema, each value with a supporting quote. - **Validation**: Code checks types, units and ranges, and that the category matches an existing single-select option. Failures go to "Needs human" with the reason. - **Human approval**: A reviewer sees suggested and current values side by side in an interface, with the quote, and ticks Approved. - **Write-back**: An automation copies approved values into master fields and stamps Approved by and Approved at. - **Failure and recovery**: Previous values sit in their own fields and every write is logged with the record ID, so a bad batch can be reversed in one pass. | Model output | Airtable field | Type | Rule | |---|---|---|---| | `category` | AI Category, then Category | Single select | Must match an existing option; never add one | | `weight_kg` | AI Weight, then Weight (kg) | Number | Between 0 and 5,000 | | `evidence` | AI Evidence | Long text | Required for every value | | `status` | Enrichment status | Single select | Suggested, Needs human or Approved | This is the same pattern as our [product catalogue enrichment](/industries/retail/product-catalogue-enrichment/) workflow, and the validation rules follow our guide to [structured outputs and validation](/engineering/structured-outputs-and-validation/). ## Automation 2: enriching contacts without breaching IPP 3A Since 1 May 2026, [IPP 3A](https://www.privacy.org.nz/resources-and-learning/a-z-topics/ipp3a/) requires reasonable steps to tell people when you collect their personal information from another source, such as a website or a data vendor, as soon as reasonably practicable. 1. **Default to company data.** Enrich organisation fields (industry, size, website) rather than individuals' details. 2. **Record provenance.** Any person-level field keeps its source URL and collection date in adjacent fields. 3. **Trigger the notice.** New person-level data adds the contact to your notification process before anyone uses it for outreach. 4. **Approve first use.** A person checks enriched contacts before they enter a campaign. ## Permissions and approval design Run integrations from a service account, not a staff member's token, and limit its personal access token to the bases and scopes it needs (`data.records:read`, `data.records:write`, `schema.bases:read`, plus `webhook:manage` for webhooks). The [MCP server](https://support.airtable.com/articles/9897799762-using-the-airtable-mcp-server) mirrors each user's role, and users can choose which bases it reaches. Keep master fields editable only by people, and let AI write only to suggested fields. Our [AI workflow automation](/services/ai-workflow-automation/) work uses this split as standard, with reviews designed as in [human review queues](/engineering/human-review-queues/). ## Costs and limits (as listed 19 October 2026) - **Rate limits**: 5 requests per second per base and 50 per second for all token traffic from one user or service account. A 429 means waiting 30 seconds. - **Monthly API calls**: 1,000 on Free and 100,000 on Team; Business and Enterprise Scale are high or unlimited. - **Worked example**: polling one table every minute costs 43,200 calls a month (60 × 24 × 30), close to half Team's allowance before any real work. Webhooks avoid that. Enriching 2,000 products at 4,000 input and 500 output tokens on Claude Sonnet 5.5 costs about 2,000 × US$0.013 = US$26. ## Where it goes wrong - **Expired webhooks.** Token-created webhooks stop after seven days unless refreshed. Schedule the refresh and alert on failure. - **Duplicate events.** Pings are at-least-once and may be coalesced; track the transaction number. - **Invented options.** Models propose categories that do not exist. Validate, never auto-create. - **Credit burn.** Field agents re-running over large tables drain the shared pool, and unused credits do not roll over. For project tracking rather than records, see our [monday.com guide](/integrations/monday/). ## Questions people ask ### Is Airtable AI included in our plan? Yes, within limits. As listed on 19 October 2026, every plan includes monthly AI credits: 500 per editor on Free, 15,000 per paid user on Team, 20,000 on Business and 25,000 on Enterprise Scale. Credits are pooled and expire at month end. Extra packs start at 10,000 credits for US$20 a month. Building apps and agents with Omni does not use credits. ### Can Claude or ChatGPT edit our Airtable bases? Yes, through Airtable's MCP server, which works on all plans with OAuth or a personal access token. It mirrors your Airtable role: editors and above can create and update records, commenters and read-only users can only read. It can draft automations but not switch them on. Enterprise admins who restrict third-party integrations must allow-list it first. ### Why did our Airtable webhook stop working? The most common cause is expiry. Webhooks created with personal access tokens or OAuth tokens are disabled after seven days unless your code calls refresh or lists payloads, which extends them by another seven days. Notifications are also switched off after repeated failed deliveries over about a day, although payloads keep accumulating so you can catch up after re-enabling. ### Should we use field agents or an external model? Field agents suit simple per-record jobs, such as tagging feedback or drafting short descriptions, where a wrong answer is cheap to spot. Use an external pipeline when you need strict validation, evidence for each value, data from outside Airtable, an approval step, or predictable costs across thousands of records each month. ## Sources - [Airtable pricing](https://airtable.com/pricing) - [Airtable AI billing](https://support.airtable.com/articles/3378106230-airtable-ai-billing) - [Using Omni AI in Airtable](https://support.airtable.com/articles/1744327578-using-omni-ai-in-airtable) - [Using the Airtable MCP server](https://support.airtable.com/articles/9897799762-using-the-airtable-mcp-server) - [Rate limits](https://airtable.com/developers/web/api/rate-limits) - [Managing API call limits in Airtable](https://support.airtable.com/docs/managing-api-call-limits-in-airtable) - [Webhooks overview](https://airtable.com/developers/web/api/webhooks-overview) - [Collection of personal information from another source (IPP 3A)](https://www.privacy.org.nz/resources-and-learning/a-z-topics/ipp3a/) - [Pricing](https://platform.claude.com/docs/en/about-claude/pricing) Published by Levy Street (https://levystreet.com), Wellington, New Zealand.