E-commerce / Product information

Catalog helper

Shows AI-generated catalog drafts from product photos, with every record awaiting human review.

Recorded AI workflow · Human review

The idea

This project explores a review workflow for creating more complete product content from images while preserving the original ERP records.

How it works

The original information is organized into proposed catalog fields. Original records remain intact, and a person reviews the proposed content before it is used.

What’s in the project

44 fictional products sent to review: 40 complete drafts, four incomplete records and an unsupported material claim kept visible.

Scope & limitations

This is an independent portfolio workflow. Review and approval are part of the process, and the demonstration does not claim real production results.

My role

Workflow design, data setup, AI instructions, testing and review rules

Built with

Make.com · Google Sheets · Google Drive

Scope

44 fictional products · image-based content · human review

Fringed Patchwork Utility Jacket photographed outdoors beside a vehicle

The starting point

The product existed. The story was missing.

The ERP held SKUs, prices, stock and variants. Most customer-facing fields were blank. The challenge was to draft useful content without turning a visual guess into a product fact.

Draft what can be seen. Flag what cannot.

01 / My decisions

Define the boundaries before the automation.

Keep the source intact

SKU, barcode, price, stock, logistics and variants remain controlled by the ERP.

Limit the claims

Images support visible details, not fabric composition or hidden construction. The instructions prohibit unsupported technical claims, but generated descriptions still need human review to catch errors.

Make review explicit

Every processed row moves to REVIEW. No product is automatically published to a store.

02 / Explore the workflow

From a ready row to a reviewable draft.

Select a step to see what happens, which tool is involved and where the boundary sits. This is an explainer, not a live execution.

See the Make.com scenario

Make.com scenario connecting Google Sheets, Google Drive and the AI agent

03 / Inspect the output

Same product. More useful information.

Browse the 44 records and switch between the original fields and the enriched draft. Product content stays in its original English so you can inspect the evidence as produced.

Unchanged: SKU, price, barcode, inventory, logistics and variants. This preview is not a live Shopify integration.

04 / Prototype evidence

Completed drafts. Visible exceptions.

44products sent to review

40 / 44with all enrichment fields filled

4records with missing fields

220size variants retained

The nine successful final Make.com runs used 374.08 credits and took 21 minutes 31 seconds in total. That is about 8.5 credits per product and 2.4 minutes per batch. Early tests and failed runs are excluded. These are prototype observations, not measured production savings.

I kept four incomplete records visible in the portfolio to show the exception path. In a live catalog, they would remain offline until a person completed the review.

A claim caught during review

In P026, the generated description called the coat “wool” and suggested warmth, even though Material and Fabric Composition were blank in the original ERP. This was an unsupported inference, not a verified product fact.

Original AI draft, preserved as evidence

A refined camel-toned sleeveless wool coat layers beautifully over a soft knit turtleneck, creating a streamlined, season-ready silhouette. The oversized lapels and bold front feel quietly confident, with roomy pockets adding a practical polish. Move with ease as the woolly texture gives this piece a tactile, luxe feel that drapes elegantly to mid-calf length. This is the coat to reach for when you want a polished, city-ready look with cozy warmth. Pair with the knit dress underneath or a tailored sweater and sleek boots for a contemporary, chic finish.

Highlighted phrases make claims about material, feel or warmth that the source does not confirm.

Editorial correction for human review

A camel-toned sleeveless coat with wide lapels, visible pockets and a long silhouette, shown over a ribbed high-neck layer. Pair it with tailored trousers and boots for a coordinated look. Fabric composition, feel and warmth require confirmation from a reliable product source.

This correction addresses the full description only. It is not publication approval or a complete audit of this product or the 44-record catalog. Original data and downloadable spreadsheets remain unchanged.

Inspect the sanitized run history

Final run	Status	Credits	Duration
01	Success	47.69	2:28
02	Success	40.03	2:47
03	Success	39.81	2:41
04	Success	42.49	2:44
05	Success	38.37	2:07
06	Success	41.89	2:12
07	Success	41.61	2:10
08	Success	40.01	2:07
09	Success	42.18	2:16
Total	9 successful	374.08	21:31

05 / Iteration

The status column mattered as much as the prompt.

A smaller batch made the workflow usable

The first attempt to process all 44 products timed out. I reduced each batch to five READY rows. Completed records move to REVIEW, so later runs can pick up remaining work without repeating it.

Better instructions made the copy more natural

The first descriptions sounded stiff. I separated visible facts, styling suggestions and details needing confirmation. This made the drafts clearer without treating uncertain attributes as facts.

How I created the synthetic catalog

I used ChatGPT to develop fashion sketches and turned the collection into realistic product images. This dataset was created before, and separately from, the Make.com enrichment workflow.

Hover over a sketch, or tap it, to reveal its product image.