How one sentence becomes a transaction
A person says to their AI: "I want my favourite coffee." The agent knows their preference — flat white, oat milk, from the same place they ordered last week. It looks up that store. Checks whether the store has an agent-readable endpoint. Finds the product. Evaluates the price against what they paid before. Places the order. Handles payment through Stripe. Sends a confirmation.
The person is still mid-conversation. They did not open a browser. They did not type into a search bar. They did not browse anything. They issued one instruction to an AI that had the infrastructure it needed to complete a real transaction — and the store got a sale it never touched.
That is agentic commerce. Not a better checkout page. Not a smarter recommendation widget. A complete purchase, initiated and completed by an AI acting on a user's behalf, against infrastructure the store controls. The stores that have that infrastructure get the sale. The stores that don't are invisible to this traffic channel entirely.
The word "agentic" is doing real work
An AI that answers is reactive. It waits for a question. An agent is proactive — it monitors signals, makes decisions, and takes actions. The coffee example shows the difference clearly: the AI did not answer "where can I get coffee?" with a list of options. It completed the purchase. That is a fundamentally different capability, and it is what stores either have infrastructure for or do not.
ChatGPT, Operator, and Perplexity Comet are already executing purchases on behalf of users. They use the Agentic Commerce Protocol (ACP) — a joint specification from OpenAI and Stripe — to identify which stores can serve them. Stores without ACP infrastructure are invisible to this traffic channel the same way a store with no sitemap is invisible to Google.
The four places agents act in a commerce stack
The commerce lifecycle has four stages where agents generate measurable return when built correctly:
1. Product discovery. Most stores rely on navigation menus, search bars, and manually curated related-products widgets. These are built for the average customer. Agents replace them with a decision engine that responds to the specific customer's session signal — time on page, scroll depth, previous purchases, referral source — and surfaces the product most likely to convert at that moment for that person. Not a recommendation. A decision.
2. Cart abandonment. The industry standard is a three-email sequence sent to everyone who abandons a cart. The problem: abandonment triggers are different for every customer — price friction, shipping uncertainty, product hesitation, a phone call that interrupted the session. An agent identifies the trigger from session behavior and responds to that specific signal. Two customers who abandoned the same product might receive completely different messages at completely different times because the signal was different.
3. Checkout intelligence. Checkout is where stores lose the most revenue and have the least visibility. An agent instruments the checkout flow — which fields cause hesitation, which price display triggers back-navigation, which shipping option surfaces the most abandonment — and either addresses those moments in real time or feeds them back as actionable data the owner can act on.
4. Post-purchase automation. After the sale: order confirmation, access delivery for digital products, onboarding sequences, review prompts, and LTV expansion. All of this can be agent-driven with zero manual touch per order. For a $2,000 product with 50 sales per month, eliminating two hours of post-purchase support work per sale is a measurable outcome before you count the revenue effects.
What makes this different from a SaaS tool
The standard approach to each of these problems is to buy a SaaS tool — Klaviyo for email, Rebuy for recommendations, CartHook for checkout, Gorgias for support. Each tool does its specific job passably. None of them talk to each other. The recovery email does not know what the checkout agent learned. The recommendation widget does not know what the recovery sequence triggered. The post-purchase flow does not know the customer's session history from before the sale.
Agents built for a specific store share context across the lifecycle. The discovery agent and the checkout agent run on the same session data. The recovery agent knows what the checkout agent logged. The post-purchase agent knows what the customer's full journey looked like before they converted. That shared context is what makes the actions intelligent rather than just automated.
The second difference is ownership. SaaS tools are rented. If the vendor raises prices, deprecates a feature, or goes under, the capability leaves with them. Agentic commerce built as custom code is owned — it runs on infrastructure you control, it is in your repository, and it continues to work regardless of what happens to any third-party vendor.
What agentic commerce is not
It is not a chatbot. A chatbot is a reactive interface that waits for input and responds. An agent monitors, decides, and acts without waiting to be asked.
It is not plug-and-play. Agents that work are trained on your catalog, your customer data, and your specific checkout flow. A generic model dropped into a store produces generic results. The investment is the specificity.
It is not a replacement for a store that does not work. Agents are a performance layer on top of existing commerce infrastructure. If the product is wrong, the pricing is wrong, or the traffic is the wrong audience, agents will execute against a broken premise more efficiently — which is not an improvement.
ACP and AP2 — the two emerging standards
Two protocols now define whether a store is visible to AI shopping agents. The Agentic Commerce Protocol (ACP) is the standard OpenAI and Stripe deployed in 2025. It defines how AI shopping agents discover and transact with online stores. When ChatGPT's Operator looks for a product, it follows ACP to find stores that have declared themselves agent-ready — via a discovery file, structured data, and a purchase endpoint. Stores without ACP compliance are invisible to this traffic channel, the same way a store with no sitemap is invisible to Google.
The Agent Payments Protocol (AP2) is Google's forthcoming standard for the same problem — a protocol for AI agents to authorize and complete payments directly, without a human-operated checkout. AP2 is on the roadmap, not yet deployed, but its direction is clear. Stores already built to ACP will have a structural head start on AP2 compliance when it ships, because the underlying infrastructure — agent-ready endpoints, structured purchasing data, machine-readable store declarations — is the same foundation both protocols build on.
Who is building it and who needs it
The businesses that benefit most from agentic commerce are stores with real transaction volume — $10,000/month and up — where the gap between current revenue and achievable revenue is measurable, and where the bottleneck is not traffic or product but conversion, retention, or post-purchase economics.
The stores that do not benefit yet are pre-launch or early-stage — there is no behavioral data to work with, no abandonment patterns to analyze, no checkout friction to instrument. Agents need signal to act on. Build the store first.
If your store is generating revenue and you know specifically where the leak is, that is the problem agentic commerce solves. The build starts with a clear use case and produces working infrastructure at the end of it.
READY TO BUILD
Agentic commerce,
built for your store.
Foundation builds start at $5,000 — no payment until your agents are live and working on your server. Full Suite at $20,000. Both Together at $25,000. Applications take 10 minutes. Response within 48 hours.
Apply for a build → ← Back to Agentic Commerce overview