Using Sanity as the Safety Layer for AI Driven Marketing Automation

How the Loop Works

A Vercel cron job fires every night at 3am ET. It pulls recent visitor session data (chat logs, edit actions, what content people interacted with) and runs it through Claude Haiku to surface patterns. Things like: visitors are abandoning after a certain step, a section headline is not resonating, a call to action label is getting ignored.

From those patterns, the AI proposes changes. Not all proposed changes are equal in risk, so we built a three tier system to match the safety bar to the scope of the change.

Tier A: Sanity copy, automatically applied. Low risk changes like headline rewrites, subheadline tweaks, and button label updates go directly into Sanity. No approval needed. Sanity’s built in version history acts as the rollback mechanism. If something looks off, we revert from the Studio in seconds. Up to 10 of these apply automatically per day.

Tier B: System behavior, Discord approval required. Changes that affect how the AI assistant inside the tool responds get posted to a Discord channel as a proposal. We review it and can approve (commits to main), refine (Haiku rewrites the proposal), or decline. There is a 30 minute revert button on approvals.

Tier C: Code level features, GitHub Issue created. When the AI identifies something that requires new functionality it creates a GitHub Issue rather than attempting to ship anything. Those get reviewed and implemented manually.

Sanity’s Role: Kill Switch, Content Layer, and Version History

Sanity does three distinct jobs in this system.

1. Kill switch. The entire loop is controlled by a boolean field on the page document in Sanity. When that field is false, the cron reads data but applies nothing and sends nothing to Discord. One toggle in the Studio stops the loop. No deploys, no environment variables, no code changes. Give yourself a no code kill switch that lives in the CMS, not in the code.

2. Content layer. Tier A changes write directly to the Sanity document that backs the demo page. The AI edits content through the same data model, the same field structure, the same API as a human editor would. There is no special path for AI generated content vs. human written content. That consistency makes the system far simpler to debug.

3. Version history as rollback. Every Tier A change creates a new Sanity document revision. If a headline change makes something worse, we roll it back from the document history panel. This is version history used as an operational safeguard, not just a recovery tool.

Why This Pattern Scales

The three tier model is not specific to a marketing demo. It is a general framework for any AI automation that touches production content. Low risk reversible changes auto apply with version history as the safety net. Medium risk behavioral changes go through an approval gate before commit. High risk code level changes become tickets for human implementation.

The system does not try to automate everything. It automates what is safe to automate and creates the right artifact for what is not. We have been running this on our portal demo and the improvement rate is meaningfully faster than watching session recordings and making manual updates.

One More Pattern: Deduplication

The cron runs a 14 day dedup check before proposing anything. If the same friction pattern was proposed in the last two weeks, regardless of whether it was approved or declined, it does not get proposed again. This prevents the system from repeating the same suggestion after you have already decided it is not worth acting on.

If you are building a similar loop, dedup is essential. Without it, you will stop trusting the proposals because they feel repetitive, and you will start ignoring the whole channel.

Getting Started

Add a boolean kill switch field to your Sanity schema. Map your proposed changes to tiers based on risk and reversibility. For Tier A, write directly to Sanity via the API and lean on version history as your rollback. For Tier B and C, use a webhook or notification service and only commit or create issues after human approval. Ship a dedup check early, before you have had a chance to build the habit of ignoring noisy proposals.

We write in depth about how we build these systems on our journal at levelupdigitalmarketinggroup.com/journal if you want to follow along.

Carlynn Espinoza is the founder of Level Up Digital Marketing Group (levelupdigitalmarketinggroup.com), an AI driven marketing agency. She builds and ships the entire Level Up platform solo using Claude Code, Next.js, Sanity, and Vercel.

Sanity – The Content Operating System that ends your CMS nightmares

Sanity replaces rigid content systems with a developer-first operating system. Define schemas in TypeScript, customize the editor with React, and deliver content anywhere with GROQ. Your team ships in minutes while you focus on building features, not maintaining infrastructure.

Sanity scales from weekend projects to enterprise needs and is used by companies like Puma, AT&T, Burger King, Tata, and Figma.