John Collison, Stripe’s co-founder and president, recently called keyword search a “ridiculous” way to find things to buy. He’s right. The way customers discover and decide is moving away from traditional search and towards SEO to agents. Shoppers already research purchases in ChatGPT and Claude, ask an AI assistant to compare options or complete a look, and increasingly delegate the purchase itself. Shopify storefronts now connect to ChatGPT, Copilot, and Gemini. Similarly, Stripe, OpenAI, and Meta back an Agentic Commerce Protocol now live with Etsy, Walmart, and a million-plus Shopify merchants; and Google launched a Universal Commerce Protocol.
The distance between intent and purchase is collapsing and becoming faster, more personal, and more context-aware than e-commerce has ever been. But most retailers and ecommerce companies aren’t architected for this change.
What will help win the next wave? Not bolting on another chatbot, search widget or recommendation engine. Those create incremental lifts, but they don’t create advantage. The winners will own the intelligence beneath the experience and the layer that understands intent, connects it to product and customer data, orchestrates action across systems, and learns from every interaction. With agentic commerce, as a new intelligence layer it is finally possible to deliver on a “seamless customer experience.”
But the walls between search and commerce took twenty years to build, and they’re now shifting unpredictably. Customer ownership is a new battleground, and the winner will be whoever delivers the best experience which increasingly begins in ChatGPT, Claude, and Gemini, sending a flood of traffic back to retailers’ sites. You can’t forecast where this lands, so the only durable strategy is adaptability.
Agentic commerce isn’t just a smarter interface. It’s a new operating model, and it runs on four layers worth understanding:
- Experience is everything the customer touches, including the storefront. It’s where retailers have traditionally competed, using what they know about their customers and the context around them to remove friction, creating moments when finding and buying feels effortless, even a little magical. The catch is, much of this “seamlessness” results from a Frankenstein of easy-to-buy and fast-to-ship point-solutions optimized for a single moment, making true end-to-end personalization impossible.
- Intelligence is where intent gets interpreted, and decisions get made. It’s the models and logic that turn data into real judgment – making decisions that run the full range: lower risk calls like ranking, recommendations, personalization, and product understanding, all the way up to bigger ones around logistics, returns, and resolving payment issues. For regulated, high-precision decisions, this is also where governance belongs including deterministic guardrails embedded in the decision logic.
- Orchestration is the coordination layer. It’s about choosing the right model, tool, or workflow for the job, keeping track of state, enforcing permissions and governance, and routing work across agents and commerce systems.
- Context makes enterprise data usable by agents by organizing product, customer, session, and business data such that an agent can retrieve the right facts at the right moment. This is the raw material the intelligence layer then reasons over. We use “context” narrowly as the governed, retrievable data an agent draws on instead of a catch-all for everything it needs to act. This can help you accomplish two things at once: token costs drop sharply (the agent pulls only the slice of data it needs for each query instead of stuffing the entire catalog into every prompt), and innovation speeds up, all without having to rebuild existing data layers.
Retailers know the top layer well, but the three beneath it often go undervalued as companies have been limited based on what their ecommerce platform-of-choice can deliver. Taking advantage of the flywheel effect these layers can have on agentic commerce requires retailers take four key actions:
1. Don’t wait for perfect data. Build an agent-ready context layer today.
Most retailers treat messy data as a reason to wait. They know their catalogs are incomplete, attributes thin, images inconsistent; reviews, inventory, promotions, and order history sit in disconnected systems, and so on. But a multi-year cleanup before you touch AI is expensive and creates massive delays. The real cost isn’t in the data work, but in the customers you lose, the acquisition spend that climbs, and the high-intent signals you never capture while you wait.
You don’t need to rip out the costly PIM, OMS, CDP, or commerce platforms you’ve already invested in. You can build an agentic context layer on top of the data estate you already have that works in parallel. It still takes time but dramatically accelerates AI innovation.
A strong context layer spans four things: product understanding (enriched attributes, embeddings, fit, compatibility, use cases, tradeoffs), customer context (profile, behavior, loyalty, order and service history, stated intent), session context (what the shopper is doing right now — what they compared, asked, hesitated on, left in the cart), and business context (inventory, margin, promotions, shipping windows, merchandising and compliance rules). That’s what lets an agent move from answering a question to helping a customer finish a goal.
In these early days of agentic commerce, if you build the context layer now the advantages compound: every search, question, image, and hesitation becomes structured intent, and that intent makes ranking, recommendations, content, and service smarter.
It’s important to note that a context layer isn’t a substitute for fixing your data, nor will it clean it. Keep raising data quality in parallel while the layer compounds intent on top of it. The upside is that the layer makes your weak spots visible, because every gap shows up as a question the agent can’t answer well. The experience improves because the business gets smarter, and token costs drop sharply because the agent pulls only the slice of data it needs for each query.
2. Don’t rent your advantage. Own the intelligence layer.
Many retailers are implementing AI point solutions: a search vendor, a chatbot, a recommendation engine, a generative content tool, a visual search plugin. Each solves a narrow problem. Together they create a fragmented stack where the retailer doesn’t own the customer’s intent and data. The sizing tool learns the customer’s body, the search vendor learns their intent, and the chatbot learns their questions. The conversion lift might be immediate, but the learning compounds inside someone else’s product. That’s the hidden cost of the bolt-on economy.
Own what teaches you about your customer and be careful about renting the layers that capture intent and define the experience. For most retailers, the unfair advantage lives in how they understand customers, products, occasions, brand voice, and service moments, and that is precisely the intelligence layer.
3. Build a composable architecture for a market that is changing every day.
The pace of change today is unlike any prior emerging technology wave in speed and volatility. Frontier models ship every few weeks, prices fall, and the best model for a given task changes constantly. Against this backdrop, adaptability must be architected.
Companies need an orchestration layer that decouples models and platforms from the application, so they’re configurable rather than hard-wired (or contractually stuck). With clean interfaces and strong data contracts, any model can plug in, draw on your portable context, and be proven against your own data and quality bar through a standing evaluation harness before you switch.
When no one can predict which AI vendor will be the rising star, locking into long-term commitments, like multi-year ERP and AI-agent vendor contracts, sacrifices the adaptability you need, and if you later realize you’ve backed the wrong horse, switching becomes a cost nightmare. What do you do in a situation like this? You build an architecture that anticipates this pace of change and establish an intelligence backbone that keeps learning while you keep replacing the parts around it.
4. Reopen the build-vs-buy decision. The economics of SaaS is changing.
For decades, retailers leveraged SaaS platforms because building was too expensive. Custom software meant big teams, long timelines, and maintenance most couldn’t justify. But buying meant choosing from the same platforms and configuring them the same way.
Agentic software engineering is breaking that math. Development is increasingly happening through agents across a new AI-driven software development lifecycle (AI SDLC). Roadmaps that once took years are now delivered in months, and that pace will continue to accelerate. As delivery cost, timeline, and maintenance effort compress, the old argument against custom capability starts to break down.
Now that the cost and time of owning differentiated capability has dropped, retailers must ask themselves where does it now make sense to own rather than rent? Agentic software engineering is lowering the cost of ownership for custom systems, especially in areas where differentiation, data access, interoperability, and customer experience matter most. We’re already seeing retailers remove large AI SaaS products off their sites because it didn’t deliver meaningful value, and the cost was too high.
This doesn’t mean the answer is to build everything. Point solutions will still win parts of the customer journey. But retailers should also invest in a broader AI commerce operating model including shared infrastructure, reusable intelligence, interoperable systems, and continuous learning across the organization.
Building that operating model requires retailers to take more control of their AI investment portfolio, wherever they are in their AI journey.
Moving from product detail pages to learning systems
The future of commerce isn’t a better product page, it’s a learning system, an AI flywheel, that understands intent across the whole journey and puts that intelligence everywhere while getting smarter with every interaction.
This shift only works if retailers stop treating AI as disconnected tools and start treating it as architecture they own. The shift is already underway, and you can see it in the dramatic changes in SEO volume and the rise of AI search traffic. Yet adoption isn’t uniform. Though many shoppers are comfortable discovering and taking recommendations through AI, many still hesitate to hand an agent their credit card. In this interim period of adoption, the cost of delay is a compounding data problem. In it, a month of delay is a month of training data your competitors are generating that you are not. To stay competitive, retailers need to prepare for agentic commerce and consider using AI as a competitive advantage immediately.
