Agentic Commerce for Print-on-Demand Goods

Abstract

AI e-commerce is usually discussed as a marketing problem: better product copy, faster ads, prettier images, lower content cost. That misses the more important engineering question.

The hard question is whether an agentic workflow can turn an open-ended market surface into a physical product catalog without breaking manufacturing constraints, customer trust, legal boundaries, or fulfillment operations.

This paper uses a Lyon Industries travel-patch commerce system as the case study. The system takes a location-linked product idea, identifies the landmark and visual brief, generates a constrained patch image, produces both customer and manufacturing assets, publishes the product into a storefront, collects payment through Stripe Checkout, and sends a print-on-demand order to the fulfillment partner.

The lesson is not that generative AI can make novelty merchandise. The lesson is that an agent becomes commercially useful only when it can honor a product contract.

The Decision This Paper Informs

The immediate decision is whether agentic product-generation workflows are worth building for physical e-commerce.

The recommendation is yes, but only when the product category has three properties:

Without those gates, the system becomes content churn. With them, agentic AI can operate like a catalog factory.

The Short Version

The useful architecture has two loops.

The first loop creates sellable products:

1. Research the location, landmark, region text, product category, and coordinates.
2. Convert that research into an image specification.
3. Generate artwork with a model that can accept detailed image instructions.
4. Post-process the result into customer-facing and manufacturing-facing files.
5. Attach price, copy, category, geography, and product metadata.

The second loop fulfills orders:

1. Let customers discover products through search, collection pages, and a map.
2. Quote shipping based on country and cart contents.
3. Collect payment and shipping details through Stripe Checkout.
4. Use the completed checkout event to identify the ordered artwork.
5. Create the print-on-demand order with the correct SKU, quantity, recipient, and artwork URL.

Most AI-commerce demos stop after step three of the first loop. The business does not start until the second loop works.

What Was Actually Built

The case system is a direct-to-customer store for made-to-order travel patches. The consumer-facing product is deliberately narrow: circular souvenir-style patches tied to places.

That narrowness is the point. A constrained product makes it possible to write a strong generation contract.

Each product record carries:

This is where the term "agentic" earns or loses its meaning. The agent is not just generating a picture. It is assembling a product record that has to survive design review, production, checkout, and delivery.

The Generation Contract

The image-generation system uses a visual template with three conceptual zones:

The prompt also specifies dimensions, resolution, safe text placement, border color, forbidden marker colors, and the difference between decorative texture and manufacturing stitching. The workflow then post-processes the generated image into a customer image and a manufacturing image.

That distinction matters. Generative models are good at producing plausible imagery. They are much less naturally reliable at respecting physical production rules unless those rules are made explicit, repeated, checked, and encoded into the pipeline.

Google's Gemini image-generation documentation describes native image generation and editing as a multimodal interaction: text and images can be used together to create or alter images.^{1Google AI for Developers, "Image generation with Gemini", accessed 2026-07-07.} That is the useful capability here. The model can take a reference image and a detailed product brief, then produce an image that is evaluated against the template.

The model is not the product system. The contract around the model is.

Why Manufacturing Constraints Matter

For physical goods, the artwork is not finished when it looks good in a browser. It has to survive:

The patch category is a good test because it forces discipline. Text follows a curved layout. Landmarks must be recognizable at small sizes. Decorative detail has to stop before it damages readability. The manufacturing file must be different from the preview file. A weak pipeline produces attractive images that are not sellable products.

The Commerce Loop

The commercial system uses Stripe Checkout for payment and shipping collection. Stripe's Checkout Session API supports line items, shipping-address collection, and metadata on the session.^{2Stripe, "Create a Checkout Session", accessed 2026-07-07.} Stripe metadata is useful because a checkout session can carry the product identifiers needed after payment without exposing internal fulfillment state to the customer.^{3Stripe, "Metadata", accessed 2026-07-07.}

After the checkout completes, a webhook validates the event, retrieves the shipping details and line items, maps product identifiers to manufacturing artwork URLs, and creates the print-on-demand order.

The print partner matters because scale should not require pre-purchased inventory. Prodigi describes its platform as print-on-demand software with global dropshipping and a catalog of customizable products.^{4Prodigi, "Print on demand for businesses", accessed 2026-07-07.} Its product catalog positions one-off fulfillment as a way to pay only for what sells.^{5Prodigi, "Print on demand products", accessed 2026-07-07.}

The resulting loop is:

The business model is not "AI makes pictures." It is "AI helps produce a reviewable catalog, and the commerce stack converts that catalog into physical orders without holding inventory."

What Makes It Scalable

The market surface is large because locations are composable. A catalog can expand across landmarks, cities, national parks, ski resorts, universities, events, expeditions, museums, ferry routes, airports, trails, teams, and seasonal collections.

The system becomes more interesting when those surfaces are treated as queues:

That is where the "millions of avenues" idea has to be translated into operations. The addressable concept space can be very large. The sellable catalog is the subset that passes research, legal, image, manufacturing, and support gates.

Where The System Breaks

The failures are predictable.

These risks do not argue against agentic commerce. They define the operating system it needs.

What This Teaches About Agents

The strongest agentic workflows are not open-ended. They are bounded by a contract.

For physical e-commerce, the contract has five layers:

Agents are most useful when they move work across those layers without hiding the uncertainty. A good agent should recommend concepts, assemble briefs, generate candidate assets, flag weak outputs, and prepare product records. It should not be trusted as the final authority on trademark risk, factual accuracy, manufacturing suitability, or customer promise.

The Next Build

The next version should not add more products blindly. It should improve the factory.

The invalidator is simple: if the workflow cannot produce products that pass quality review faster than a human can manually design and list them, the agentic system is theater. If it can, the business becomes a repeatable catalog engine for physical goods.

Footnotes

1. Google AI for Developers, "Image generation with Gemini", accessed 2026-07-07. ↩

2. Stripe, "Create a Checkout Session", accessed 2026-07-07. ↩

3. Stripe, "Metadata", accessed 2026-07-07. ↩

4. Prodigi, "Print on demand for businesses", accessed 2026-07-07. ↩

5. Prodigi, "Print on demand products", accessed 2026-07-07. ↩