Can AI Agents Scale a Physical-Product Business?

The research hypothesis

An AI agent could expand a product catalogue from a fixed specification, leaving people to set the rules and handle exceptions.

Each product needs a subject, description, price, customer image, manufacturing file and fulfilment record. These must agree as the catalogue grows.

World Patches was the test. A location brief goes into the generator; code produces a supplier file, store preview and catalogue record. The 59 products exposed repeatability problems and work that still needs people.

The guide image

The manufacturing file needs room around the finished patch for cutting. To show the image model where that room belongs, every request includes a 985 × 985 px guide. Three colours mark the safe area, bleed ring and outer area that will be cut away.

Reference inputThe production artwork must replace every marker colour. The guide is geometry and is not retained in the print file.

Marker	Meaning	Instruction sent with the image
Green	Safe area	Keep the patch face, landmark, location name, and subtext inside this circle.
Orange	Bleed ring	Continue decorative artwork through this ring. Do not place critical text here.
Red	Cut-away area	Fill the square to every edge so a crop cannot expose an empty corner.

Design decision

Marker colours should sit outside the intended product palette. The current patch prompt forbids green, orange, and red so residue is obvious. That also prevents legitimate use of those colours. A general generator should select a temporary marker palette per job and test that it has been removed.

The request sent to the model

Gemini receives the guide PNG and the product brief in one request. The brief supplies the location, landmark, visible labels, border colour, permitted palette, and explicit rules for each zone. Google documents image-plus-text input for Gemini image generation and provides a batch API for larger jobs. Batch capacity does not remove the need for acceptance tests.

contents = [zone_guide_png, product_prompt]

image_config = {
  "aspect_ratio": "1:1",
  "image_size": "1K"
}

candidate = generate(model, contents, image_config)
manufacturing = resize(candidate, 984, 984)
save_png(manufacturing, dpi=300)

Model output is treated as a candidate. Deterministic code resizes the first returned image to 984 × 984 px and writes 300 dpi metadata. A second pass masks the square image into a circular store preview and overlays the supplier border. The square manufacturing file keeps artwork outside the visible circle because those pixels are needed at the cut line.

Square Machu Picchu manufacturing artwork with the circular patch face in the center and decorative artwork continuing to every edge. — Manufacturing fileFull square. Decorative pixels continue beyond the finished round patch.

Circular customer preview of the Machu Picchu patch with a black overlocked edge. — Store previewCircular mask and physical-edge overlay. This file is not sent to print.

Results from the first 59 products

We audited the repository on 10 July 2026. File presence and raster geometry are consistent across the catalog. Visual acceptance is still a human step.

Catalog records: 59
Manufacturing files present: 59 / 59
Store previews present: 59 / 59
Manufacturing dimensions: 59 / 59 at 984 × 984 px
Manufacturing density metadata: 300 dpi
Automated visual acceptance: Not implemented

Audit scope: catalog JSON, generated files, raster dimensions, and density metadata in the World Patches repository. It does not include a new print run or an end-to-end paid order.

Failures observed during generation

Earlier prompts produced errors that led to more specific instructions. Lofoten returned a tiny inner scene, buildings over the ring, a green ring, and incorrect inferred subtext. Faroe and Geiranger also needed individual reruns for boundary violations and scene clutter.

Observed candidate	Current response	Control still missing
The center scene came back too small.	The prompt now tells the model to fill the inner medallion edge to edge.	Measure occupied area automatically; the current check is visual.
Houses, mountains, and boats crossed into the text ring.	The prompt names the ring as a hard boundary and tells the model to omit objects that do not fit.	Segment the ring and reject foreground detail inside it.
Guide colours survived in some candidates.	Green, orange, and red are explicitly forbidden in the requested palette.	Run a tolerant colour-residue test before accepting the file.
Long names and inferred subtext were unreliable.	Visible labels and subtext are supplied as separate bounded inputs; individual failures are rerun.	OCR the two text arcs and compare them with the product record.

Material appearance

A valid file can still describe the wrong product.

The generator asks for raised embroidery texture. The supplier makes this SKU by dye sublimation on spun polyester. The bleed geometry may pass while the preview sets the wrong material expectation. The next catalog regeneration must be tuned against physical samples for a flat printed appearance.

There is also a supplier-spec question to close. Prodigi’s patch page lists JPG or PDF at 300 dpi; its general Print API documentation accepts JPG, PNG, or PDF. The current pipeline submits PNG. A sandbox order and supplier confirmation should settle the product-specific contract before unattended generation.

What we would add before 10,000 assets

Batch requests can generate files faster. At catalogue scale, each file also needs a recorded acceptance check.

Acceptance check	Machine test	Current implementation
File contract	Dimensions, density metadata, colour mode, file type, and readable asset URL.	Dimensions and density are normalized. The rest needs a formal validator.
Safe-area occupancy	Critical content stays inside the inner circle; decorative pixels reach the outer edge.	Prompted and reviewed by eye.
Text fidelity	OCR result matches the requested location and subtext.	Manual review only.
Guide removal	No marker-colour residue remains after generation.	Prompted and reviewed by eye.
Physical match	Approved sample links prompt version, supplier SKU, source asset, and received product.	Samples exist; the art direction still needs retuning against them.
Reproducibility	Immutable manifest records model, prompt, guide, inputs, output hash, review, and reruns.	Catalog identity exists. Generation history is not yet complete enough for unattended scale.

Physical samples remain part of the test plan. Pixel checks can prove that a file matches a template. They cannot prove legibility on fabric, colour under the supplier process, edge placement after cutting, or whether the rendered material matches what arrives.

From research pipeline to venture

We chose patches to test catalogue expansion with one supplier product, a fixed print template and repeatable product fields. The resulting 59-product catalogue provided the inventory for World Patches.

The storefront uses Next.js, Stripe Checkout and Prodigi. Each accepted asset must retain its product identity through publication, payment and the supplier order.

At the 10 July 2026 review, the order workflow had a retry defect. A supplier-order failure was acknowledged to Stripe with HTTP 200, which marks the webhook delivery as successful. Stripe documents retries after non-2xx responses. Durable retry and reconciliation must move outside this request before the order path can be trusted.

Commercial experiment: World Patches

Primary references

Google AI for Developers: Gemini image generation
Image-plus-text inputs, output sizing, model limits, and batch generation.
Prodigi: round fabric patch specification
PATCH-ROUND size, material, print process, file format, and 300 dpi guidance.
Prodigi Print API reference
Asset sizing modes, order payload, and idempotency behavior.
Stripe: receive webhook events
Signature verification, delivery status, and retry behavior.

Repository and source review completed 10 July 2026.