Every Airtable base starts clean. The fields make sense, the statuses are clear, and everyone knows where data belongs. Then real life happens.
People start entering data differently. Duplicate records appear. Statuses go unused. Fields get repurposed for something they were never designed to hold. Slowly, quietly, your source of truth stops being true.
This post walks through a practical method for auditing and cleaning up a messy Airtable database using Claude's MCP integration — without blindly letting AI edit your live data.
Disclosure: The link above is an affiliate link. We only recommend tools we use ourselves.
The breakdown rarely happens in one dramatic moment. It accumulates. A field that was meant for one purpose becomes the catchall for anything that doesn't fit elsewhere. A status option nobody uses anymore never gets removed. A table that was created for a project two years ago still sits there, half-populated.
By the time someone notices the base feels off, there are dozens of these small degradations layered on top of each other. Reporting numbers don't match. Automations misfire. New team members can't figure out where to put things.
The good news: a messy Airtable base is fixable. The key is auditing before touching anything.
Connecting Claude to Airtable via the Model Context Protocol gives the AI read access to your base structure — tables, fields, records, and relationships. This is different from asking Claude to guess what might be wrong. It can actually look at your data.
The critical constraint: Claude should audit first, never edit first. The goal of the initial pass is to build a picture of what's wrong, not to start fixing it. Any AI that touches live data without a human review step is a liability, not an asset.
Once Claude is connected, the audit has three components: database structure, field-level issues, and record-level data quality.
The structural audit starts with the tables themselves. Claude looks for:
Deprecated tables — tables that haven't been touched in months, contain records that map to nothing active, or clearly belong to a project or process that no longer exists. These are candidates for archiving, not immediate deletion.
Legacy data — fields or tables that were created to support an old workflow and are now disconnected from the current system. Claude can identify these by looking at what fields are populated, what fields are consistently empty, and whether the data in a table is actually referenced anywhere.
Field overload — tables where a single field is doing too many jobs. The "Notes" field that contains financial figures, contact details, and project status is a common culprit. These fields are impossible to report on reliably.
Before Claude touches any records, it is worth being explicit about what it should never read or change. Employee compensation data, personal identifying information, client financial details — these categories should be explicitly off-limits, not just hoped to be respected.
Set these rules in your prompt before Claude runs any record-level queries. Make the boundaries explicit: "Do not read or suggest changes to fields containing [X]." This is not about distrusting AI — it is about having a documented, auditable policy for what AI can and cannot touch in your system.
Once the structural audit is done, Claude digs into individual fields across all tables. The field-level audit surfaces:
Important fields — fields that are core to reporting, automations, or integrations. These need to be documented, not touched without careful consideration.
Risky fields — fields where a change would break downstream automations or connected views. These require human approval before any modification.
Unused fields — fields that exist but are empty across most or all records. These are candidates for removal, but only after confirming nothing is keyed off them.
The cleanup risks with automations deserve specific attention. An automation that triggers on a field change can fire thousands of times if that field is bulk-edited incorrectly. Before touching any field that an automation depends on, the automation should be turned off, the edit should be made and validated, and the automation should be turned back on with test verification.
After the structural and field-level passes, Claude moves to individual records. This is where the day-to-day messiness lives:
Duplicates — records that represent the same entity but were entered separately. Claude can identify likely duplicates by looking for matching names, emails, or other key identifiers across records. It flags them for human review; it does not merge them automatically.
Dirty data — inconsistent formatting in fields that should follow a standard. Phone numbers with and without dashes. Company names that are sometimes abbreviated and sometimes spelled out. Dates entered as text strings instead of date fields. These inconsistencies make reporting unreliable and automations fragile.
Formatting issues — fields where the data type doesn't match what was intended. A field designed for URLs that contains narrative text. A number field used to store phone numbers. These are often the result of the base growing organically rather than being designed.
Even with a perfect audit, Claude's suggestions are not automatically correct. The AI can identify patterns that look like problems — but sometimes a "duplicate" is intentional, a "deprecated" table is still needed, or an "inconsistent" value is deliberately different.
This is why the audit output should be a structured review list, not an execution list. Claude generates the findings; a human decides which ones to act on, and in what order.
Not all cleanup tasks carry the same risk. A three-tier categorization makes the work manageable:
Safe to automate — changes that are unambiguously correct and carry no risk of data loss or automation disruption. Standardizing a phone number format. Removing whitespace from text fields. These can be batched and executed without individual review.
Requires human approval — changes that are probably right but need a second set of eyes before execution. Potential duplicate merges. Field type corrections. Status consolidations. These get added to a review queue.
Do not change — anything that touches core business logic, financial data, or fields tied to external integrations. These are flagged for awareness, not action.
The structured cleanup plan prioritizes issues in this order:
The most durable part of this process is creating a dedicated audit tracking table in Airtable itself. Each flagged record or field issue becomes a record in the audit table, with fields for:
Claude creates these tracking records as the final step of the audit. The result is a human-readable review queue that anyone on the team can work through — approving safe changes for automation and handling higher-risk items manually.
This approach keeps AI in the analysis role and humans in the decision role. It also creates an audit trail that documents every change made during the cleanup, which is valuable if something breaks downstream.
The single most important rule in this entire process is also the simplest: never let AI edit your data before you have reviewed what it intends to do.
AI that is connected to live data can move fast. That is exactly why it needs guardrails. The audit-first, approve-second workflow is what makes AI a useful tool in Airtable maintenance rather than a source of new problems.
If your Airtable base has been accumulating entropy for a while, this is the process that gets it back under control. Start with a free Airtable account here and build the system that earns the cleanup.