How We Unlocked AI Order Prep for Interior Design Projects

Before you build anything with AI, or for AI, you have to ask: why are we building this? That question led us to one of the more counterintuitive fixes we've made for a design business. The answer wasn't AI. It was the data underneath it.

The problem behind the build

The design business I'm the fractional AI officer for had a fulfillment problem. The lead designer had been working in Google Slides for years and had a solid process for building design boards. She'd screenshot the vendor page she was shopping, drop it into the slide, and move on. It worked, but it created a downstream mess for ordering.

When it came time to place orders, someone had to go into every slide, look at each screenshot, pull out the URL by hand, and repeat that for every item across every board. On top of that, items sometimes moved between rooms, which meant they moved between boards, and things got hard to track. Ordering was slowing down fulfillment, costing time and money across the business.

Why throwing AI at the current process didn't work

The instinct was to ask: can AI handle this? Technically, yes. AI can go into Google Drive with built-in connectors, look through slides, use optical character recognition to identify items from screenshots, and extract URLs. It would also be slow and expensive. Every screenshot is a vision task. That eats tokens and adds latency. The verdict: not worth it on the current data.

Instead of building AI on top of a messy process, we went upstream and fixed the process so the job would become easy for AI. Two problems needed solving. The sourcing data was hard for AI to read (screenshots instead of text), and there was no reliable link between that data and the item in the slide.

What we built

Google Slides has fields on each image element you can write to via API. We used those fields to attach structured sourcing data, including the product URL, directly to the image of each item. Then we built a small iOS app that wraps around the designer's existing shopping behavior.

Instead of screenshotting a product page, she hits Share, taps the app, picks the images she wants, and the app handles the rest. It cuts out the background using Apple's built-in tools, picks the right slide in the right deck, inserts the image, and attaches the sourcing data. She didn't have to change her process. She swapped one gesture, screenshot, for another: share.

What the workflow unlocks

Now that sourcing data lives as text attached to each item, AI can read it instantly through standard connectors. No vision, no OCR, no token waste. Automated order preparation is now viable. AI can pull an item list from a design board, follow the links, and add items to cart. You can run that while you're working on something else.

The build took upfront work. Long-term, it's faster for the designer to build boards, sourcing data no longer gets lost, and AI has everything it needs to handle order prep. None of that was possible when the data lived in screenshots.

The lesson

This is AI infrastructure: the work that supports AI but isn't AI itself. It's often the most important part of your AI systems. Getting good results from AI is less about training models and more about giving AI access to data it can actually read. If you want AI to do something useful in your business, the first question is usually not "what model should I use?" It's "is my data in a shape AI can work with?" Sometimes the thing you should build isn't AI at all.