Product Data Integration and a Review-First Workflow

Introduction

The one thing you absolutely must not get wrong in a newsletter is the product price. You wrote "20% off" but the real number was different; someone clicked a link and landed on the wrong product — mistakes like these aren't merely embarrassing, they directly erode customer trust.

The earlier articles covered how to have AI write email-safe HTML. But there are things you must not leave to AI: prices, image URLs, and product links — data that has to be exact. This article explains the product data integration that has the system guarantee these, and the production flow that inserts a human review before sending safely.

Don't let AI invent prices or URLs

LLMs will happily "invent" a price

A well-known weakness of LLMs is fabricating facts in a plausible way. In the newsletter context this shows up in its worst form with prices: ask AI to "introduce this product," and it may casually write a price that doesn't exist. Because the sentence reads naturally, it's easy to miss even in review.

So as a design principle, we decided that prices, image URLs, and product links are never touched by AI at all. What AI is good at is copy and structure — not guaranteeing facts.

Fetch accurately from the product master DB

So how do you get accurate data in? The answer is to have the system insert values fetched from the product master DB (Neon Postgres). Specify the products to feature, and the system pulls each product's current price, image URL, and product page link from the DB, and fits them into the vessel of copy AI generated.

This way, the copy draws on AI's creativity while the numbers and links are guaranteed by the database's accuracy.

Insertion keeps things consistent, too

Having the system insert the data buys more than accuracy. Prices stay current because it consults the DB even after a price change, and the system normalizes image URL and link formats, so breakage and mislinks are far less likely. It's faster and more reliable than a human bouncing around admin screens copy-pasting.

Insert a review before sending

Why not fully automate

Even with accurate product data, newsletters aren't sent fully automatically. Is the pitch's tone on-brand? Is the campaign wording appropriate? These are judgment calls that belong to a human. AI's copy scores high on average, but occasionally carries a "so close" phrasing. Passing it under a human's eyes at the end reliably stamps that out.

"Faster" and "error-free" can coexist, but "letting go of judgment too" is a different matter. That part we deliberately left to people.

Generate → review → correct → send

In practice, production runs as follows: AI prepares the groundwork HTML, a human reviews it, gives correction instructions if needed, and once it's fine, pastes it into the MA tool (HubSpot) to send. The human's role is not to "build" but to "finish and ship."

Share progress via Slack

Production progress is posted to Slack. When generation finishes, the reviewer is notified, and delivery-side activity is shared too, so the team can see "where things stand right now." Newsletter production often involves several people, so this visibility quietly pays off.

Summary

The key to automating newsletter production safely is drawing a clear line between "what AI handles" and "what the system or humans guarantee."

• Prices, image URLs, and product links are never invented by AI — they're inserted accurately from the product master DB.
• Copy from AI, facts from the database: this division of labor achieves speed and accuracy at once.
• Rather than full automation, a generate → review → correct → send flow keeps the final judgment with people.
• Slack notifications share progress so the team can run it with confidence.

The mechanism for having AI generate email HTML is in Getting AI to Write Email HTML That Doesn't Break, and the overall system in AI Newsletter Creation Overview.