Agentic commerce is buying and selling where an AI agent does the purchasing on someone's behalf. A person says what they want. The agent finds the options, compares them, and, with permission, pays. The seller still delivers the thing. What changes is who does the shopping.
That is the short answer to "what is agentic commerce," and almost every explanation you will find stops at retail: shoes, groceries, flights, a checkout inside a chat window. If you run a B2B software company, very little of it maps to what you sell. You do not sell items out of a cart.
So this post covers the retail version honestly and briefly, then makes the argument those explanations leave out. For software, what an agent buys is not a product or a seat. It is a service, run inside the customer's account, as that customer.
What agentic commerce is today, and how it works
How does an agentic commerce agent work? The buyer states an intent. The agent reads structured product data instead of a storefront, picks something, and asks the seller to start a checkout. Payment moves as a token scoped to that one purchase, so the agent never holds the card. The seller accepts or declines, charges, and fulfils the order.
Stripe and OpenAI wrote that flow down in September 2025. The Agentic Commerce Protocol is open source under Apache 2.0, and it launched with Instant Checkout in ChatGPT, where US users could buy from Etsy sellers without leaving the chat. The payment piece is a Shared Payment Token, which Stripe describes as "scoped to a specific merchant and cart total." The seller stays the merchant of record.
The rest of the payments industry moved quickly. If you searched for agentic commerce platforms, or the best providers for agentic commerce, this is the honest list today:
- Stripe and OpenAI. The Agentic Commerce Protocol, plus Stripe's Machine Payments Protocol for agents that pay a service directly.
- Google. The Universal Commerce Protocol, announced in January 2026 and developed with Shopify, Etsy, Wayfair, Target, and Walmart.
- Mastercard. Agent Pay, announced in April 2025, built on Mastercard Agentic Tokens, an extension of its existing tokenization.
- Visa. Intelligent Commerce Connect, announced in April 2026: one integration that accepts agent-initiated payments across the major protocols, ACP and UCP included.
In September 2026, Visa, Mastercard, and Ant International announced a shared Know-Your-Agent framework, so an agent can be traced back to the person or business operating it. That tells you where the real work is. Who is this agent, what was it allowed to spend, and who answers for it afterward?
Two honest notes on adoption. The forecasts are large: McKinsey puts agentic commerce at up to $1 trillion in US retail revenue and $3 trillion to $5 trillion globally by 2030, as reported by Digital Commerce 360. The reality so far is smaller. In March 2026, OpenAI said "the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide," and that merchants could use their own checkout while it focused on product discovery, according to TechCrunch.
The lesson the industry took from that is a good one. Checkout.com summed it up in a line: "The merchant owns the checkout, not the AI platform." Keep that sentence in mind. It matters even more for software.
Agentic commerce vs traditional e-commerce
Less changes than the headlines suggest. In traditional e-commerce, a person browses a storefront you designed, and your page does the persuading. In agentic commerce, the agent reads your data, not your design, and the buyer shows up with the decision half made. Discovery moves to the agent. Trust moves to tokens and agent identity.
The product itself does not change. A pair of running shoes is the same pair no matter who clicked buy. It is identical for every buyer, it is delivered once, and the sale is over when the box arrives. That is why retail is the easy case, and why the card networks are doing most of the work. The hard parts are payment and trust, not the thing being sold.
A software company has no cart
Now sit in the chair of someone running a B2B SaaS product. What would an agent buy from you?
Not a SKU. Probably not a seat either, because an agent does not log in, and one agent can do the work that used to justify several logins. The closest thing on offer is payment per call. Stripe's Machine Payments Protocol lets an agent pay a service per request, and it is real: Browserbase sells headless browser sessions to agents that way, and Parallel charges per API call for web access.
If your product is the data or the compute, that fits. For most software it puts the price on the wrong thing, and I wrote up why in how to price your software when the customer is an AI agent.
The advice aimed at software companies is thin. Stripe's own guide to agentic commerce gives SaaS one sentence: "building your own MCP ensures your software can be discoverable and interoperable with the new generation of AI tools." That is true, and it is step one. But an MCP server makes your product callable, not sellable. A tool list has no price, no guarantee, and nobody who owns the result.
The forecasts skip you too. Gartner's widely quoted prediction, as reported by Digital Commerce 360, is that 90% of B2B purchases will be handled by AI agents by 2028, more than $15 trillion in spending. Read the context and it is procurement: agents that negotiate, contract, and place orders. Agentic commerce for B2B, in that telling, is still a purchase order. Nobody is describing what happens when the thing being bought is work done inside a software product.
What an agent buys from software is a service
Think about what your customer wants when they hand a job to an agent. They do not want access, they already have it. They want the monthly report built and sent. They want the account audited every Monday and the drift fixed. They want something watching for the failed sync at 3am and acting on it.
Each of those is an outcome defined on top of your product. That is the unit: a service. It differs from a retail item in three ways, and each one changes what commerce has to mean.
It is configured per customer. The Monday audit for one account checks different things than the audit for the next. It runs inside that customer's account, on their data, so it has to run as that customer, with their permissions and nothing more. I covered why that rule matters in what to expose to your customers' AI agents.
It is standing work. A retail purchase ends at delivery. A service keeps running, on a schedule or when an event fires, usually when neither the customer nor their agent is awake. So what gets sold looks less like a transaction and more like a subscription to work: run limits, entitlements, a log of every run, a way to cancel.
And the vendor is the merchant. This is the lesson retail just learned, applied to software. The price is yours to set. The checkout runs on your Stripe account, and receipts, refunds, and disputes stay with you.
A few agentic commerce examples in that shape: a monthly executive report at $199 a month, a one-time configuration audit that returns what is wrong in ranked order, a monitor that watches a threshold and acts within rules the vendor wrote. The prices are illustrative. The shape is the point: a named outcome, a price, a run log.
Who is the buyer? There are two answers, one for now and one for next.
Now, it is the customer talking to the vendor's own customer agent. Picture someone inside an invoicing product asking for a monthly summary for their leadership team. The agent offers the service and its price, checkout opens on Stripe, and the service unlocks when the payment clears. The first run happens as that user, and the log says so: "acting as matthew@hartwell.io · list_invoices · build_report · done."
Next, it is the customer's own agent, arriving from Claude or ChatGPT, reading your catalog of services and buying one as that customer. The groundwork is going in, and I described the inbound side in the Agent Gateway post. I want to be plain that this part is next, not live. The catalog is the same in both cases, which is the reason to build it now.
Where Aimdoc sits
Aimdoc is the customer agent platform for B2B SaaS, and this is the version of agentic commerce we built. You connect your product's MCP, Aimdoc authenticates as each customer so every call runs as that user with their permissions, and the agent and every service do real work in your app, over email, and on your site. Services are the outcomes you define on top, configured to each customer and kept running. Connect your Stripe account once, choose per service whether it is included with a plan, requestable, or sold one-time or as a subscription, and the agent sells it in context, with checkout on Stripe and payouts landing in your Stripe account. Your own agent selling services inside your product is live now, and external agents buying from the same catalog is next.
If agentic commerce for your product looks like a service rather than a cart, selling services is the page to read, and you can start free when you are ready.