Vertical AI wins in commerce where the work is high-volume, judgment-heavy, and data-rich but still done by hand: merchandising, pricing, catalog and content, fraud and returns, post-purchase support, and demand planning. The companies worth co-founding pick one such workflow in a specific commerce category, own the transaction data it generates, and sell the outcome rather than a tool.
Commerce looks saturated from the outside. There is a tool for every task, a Shopify app for every workflow, and a decade of retail software already sold. That surface hides the real picture: underneath the tooling, an enormous amount of commerce work is still done manually by merchandisers, buyers, category managers, support agents, and analysts who spend their days on repetitive, high-stakes judgment calls. That is exactly the terrain where vertical AI wins, and it is why commerce is one of the three sectors gAI Ventures co-founds companies in. Here is where the value actually concentrates, and what separates a defensible commerce AI company from a feature.
Why commerce is a vertical AI problem, not a horizontal one
A horizontal AI tool treats every product like every other product. Commerce does not work that way. The judgment that makes a merchandiser good at apparel does not transfer to auto parts, grocery, or industrial supply, because the signals differ: a return in fashion often means fit, a return in electronics often means a defect, and a return in grocery often means spoilage. Pricing elasticity, catalog attributes, seasonality, and fraud patterns are all category-specific. A model that does not encode that context produces plausible, generic output that a category expert immediately distrusts.
That is the core of the vertical thesis, and it is the same reason vertical beats horizontal in the enterprise, laid out in what vertical AI is and why it beats horizontal AI. In commerce the effect is sharper because the volume is enormous and the errors are visible: a mispriced hero product or a catalog full of wrong attributes shows up in the numbers the same week. The buyer does not want a general assistant. They want a system that already knows their category.
Where the value concentrates: the commerce workflows worth building around
The workflows worth co-founding a company around share three traits. They are high-volume, so automation compounds. They require judgment, so they were not solved by rules-based software. And they generate proprietary data as a byproduct, so the company gets more defensible with every transaction.
| Commerce workflow | Why it is done by hand today | Why AI fits, and where the data moat forms |
|---|---|---|
| Merchandising and assortment | Buyers rely on experience and spreadsheets to pick what to stock | Demand signals and sell-through data train a category-specific model no generalist has |
| Pricing and markdowns | Elasticity varies by item, season, and channel | Continuous repricing on owned transaction data outperforms static rules |
| Catalog and product content | Descriptions and attributes are written and tagged manually at scale | Generation and enrichment against a category schema, improved by conversion feedback |
| Fraud, returns, and abuse | Rules miss new patterns; manual review does not scale | Category-specific behavior data flags what generic fraud tools cannot |
| Post-purchase support | Agents answer the same order and returns questions repeatedly | Resolution on the merchant's own order and policy data, sold as deflection, not seats |
| Demand and inventory planning | Planners forecast in spreadsheets against noisy signals | Forecasting on owned sales history reduces stockouts and overstock |
Each of these is a business, not a feature, when a founder goes deep in one commerce vertical rather than shallow across all of them. Andreessen Horowitz's breakdown of AI and commerce makes a related point: purchases differ enormously by level of consideration, from impulse buys to researched, high-consideration ones, and the AI opportunity looks different in each. A company that owns the researched-purchase workflow in one category is building something a horizontal shopping assistant cannot easily copy.
Agentic commerce is moving the buyer and the interface
The ground is also shifting in a way that reopens workflows that felt settled. AI agents are starting to research options, compare, and transact on a shopper's behalf, which changes both who the buyer is and where the transaction happens. Adobe's analysis of retail traffic found that AI-driven revenue per visit rose 84 percent over seven months as shoppers grew comfortable buying after an AI-assisted session, detailed in Adobe's report on generative-AI shopping. When an agent, not a human, is evaluating a product, the catalog, the pricing, and the structured data behind a listing become the interface. That is a new surface to build for, and it rewards companies that already own clean, category-specific product and transaction data.
What makes a commerce AI company defensible
The uncomfortable truth is that the model is not the moat. Any founder can call the same foundation models, so a commerce AI company that is only a prompt on top of a general model is a feature waiting to be absorbed. Defensibility in commerce comes from two places. The first is proprietary transaction data: the sell-through, returns, pricing, and behavior data a company accumulates by running a real workflow, which a foundation model provider does not have and cannot buy. The mechanics of that advantage are covered in the data moat in vertical AI. The second is workflow depth: being so embedded in how a category actually operates that switching means re-plumbing the business.
There is a related shift in how these companies are priced. The emerging pattern, described in Foundation Capital's work on services as software, is selling the outcome rather than the tool: more sell-through, fewer fraudulent returns, higher forecast accuracy. Commerce is well suited to this because the outcomes are measurable in the merchant's own numbers, which makes the value provable and the company harder to displace.
How gAI Ventures co-founds in commerce
gAI Ventures is a venture builder and pre-seed fund that co-founds vertical AI companies, and commerce is one of its three focus sectors alongside financial services and enterprise productivity. The model is not to fund a founder and wait. It is to co-found the company from before it exists, taking an expert operator from minus one to one: validating the specific commerce workflow, then standing up a production-grade founding engineering team from day zero so the operator is not stuck searching for a technical cofounder while the opportunity moves. The full sequence is described across the gAI Ventures manifesto and our vertical AI investment theses, and the companies built this way are on the gAI Ventures portfolio.
For a commerce operator, the practical implication is that the hard question is not which model to use. It is which workflow, in which category, generates the proprietary data and measurable outcome worth building a company around. That is the question worth answering before writing a line of code, and it is the one a technical venture builder is built to answer with you. If you are an operator with a commerce insight, the gAI Ventures team and more thinking on the gAI Ventures blog are where that conversation starts.
Frequently asked questions
- Is commerce not already saturated with AI tools?
- The storefront layer is crowded, but the manual work behind the storefront is not. Merchandising, pricing, catalog enrichment, fraud and returns review, and demand planning are still done largely by people using spreadsheets and experience, because rules-based software could not handle the judgment involved. That judgment-heavy, high-volume, data-rich work is exactly where vertical AI wins, and most of it has not been genuinely automated, only lightly assisted.
- Why does commerce AI need to be vertical rather than a general tool?
- Because the right answer is category-specific. A return in apparel usually signals fit, in electronics a defect, and in grocery spoilage, and pricing elasticity, catalog attributes, seasonality, and fraud patterns all vary by category. A horizontal model that ignores this produces generic output a category expert distrusts. A vertical system that encodes one category's context, and improves on that category's own data, consistently outperforms a general one on the work that matters.
- What makes a commerce AI company defensible if anyone can use the same models?
- Proprietary transaction data and workflow depth, not the model. The sell-through, returns, pricing, and behavior data a company accumulates by running a real workflow is something foundation model providers do not have, and being deeply embedded in how a category operates makes switching costly. A company that is only a prompt on a general model has neither, which is why it tends to be absorbed as a feature rather than growing into a business.
- What is agentic commerce and why does it matter for builders?
- Agentic commerce is shopping where AI agents research, compare, and transact on a person's behalf rather than a human clicking through a store. It matters because when an agent evaluates products, the catalog, pricing, and structured data behind a listing become the interface, and clean, category-specific data becomes a competitive advantage. It reopens workflows that felt settled, which is where new companies get built.
- How does gAI Ventures work with a commerce operator?
- gAI Ventures co-founds the company rather than passively backing it. It works with an expert operator to validate a specific commerce workflow, then provides a production-grade founding engineering team from day zero and milestone capital, so the operator is building rather than recruiting a technical cofounder. This is education about the model, not an offer of investment; the point is that a commerce insight paired with a day-zero technical team gets to a defensible product faster than either does alone.
End of article · #011