What Is Agentic Commerce? Two Sides, One Foundation

Agentic commerce is commerce in which a software agent searches, compares and buys on the user's behalf, and runs the merchant's order, pricing and inventory flows without every step waiting for human permission. The term has two sides: the buyer's agent represents the customer, the merchant's agent represents the store. Both turn out to rest on the same foundation: machine-readable data and a working integration layer.

The buyer's agent: what the market is talking about

When someone says agentic commerce in 2026, they usually mean the buyer's agent: software that searches for products, compares prices and terms, and completes a purchase on the user's behalf, sometimes without the user ever visiting a storefront.

This is no longer a vision. In September 2025, Stripe and OpenAI launched the Agentic Commerce Protocol (ACP), an open standard powering Instant Checkout in ChatGPT: US users could begin buying products directly from the conversation. Google published the Agent Payments Protocol (AP2) in the same month, an open payments protocol developed with more than 60 organisations from the payments and technology sectors. Visa Intelligent Commerce and Mastercard Agent Pay arrived in spring 2025: the first gives agents payment tools and agent-specific tokens, the second brings Agentic Tokens to the network that bind a purchase to a named agent and a consent policy.

The summary is simple: the rails for agent purchases exist. The question is no longer whether buyer's agents will start buying. The question is whose store they buy from.

The buyer's agent does not see your website the way a person does

A buyer's agent reads data: product names, attributes, prices, availability, delivery terms and payment interfaces. If these are in order and machine-readable, you are findable and buyable in agentic commerce. If your product information lives only in a beautiful user interface or a PDF price list, the agent does not see you.

Our experience across more than 15 large-scale ERP and e-commerce integrations shows that the machine-readability gap is rarely a deliberate decision. It is a leftover: systems were built to be read by people, not by agents. PIM, pricing rules, inventory and carrier interfaces and APIs are precisely the layer that decides whether a buyer's agent can complete a purchase in your store at all.

The market's most visible test to date confirms this. In March 2026, OpenAI scaled back native checkout inside ChatGPT and began routing purchases to retailers' own environments; ACP continues at the application level. The right conclusion is not that agentic commerce failed. It is more precise: no single checkout form is the durable value. What endures across every form is machine-readable data and working order fulfilment.

The merchant's agent: what already runs without you

Every store at scale is already a partially autonomous business. The order passes payment and stock control, the ERP updates inventory, the carrier interface creates the shipping label, the customer receives confirmation, the invoice is generated and cash flow is reconciled with accounting. None of these steps waits for human permission.

A real purchase path looks roughly like this: the customer pays, the payment provider confirms, stock is reserved, the order rule checks minimum order value and delivery terms, the pricing rule validates campaign prices, the carrier interface selects the service, the shipment is dispatched, the invoice is generated and accounting is synchronised. Nobody needs to be awake for this chain to work.

Most commerce performance already arises machine to machine, unnoticed and around the clock. On the merchant side, agentic commerce does not start from zero. It starts from this working foundation, which is missing one layer: interpretation and decision-making in the cases fixed rules do not cover.

What makes agentic different from automation

Conventional automation is deterministic: if condition A holds, action B follows. The same input produces the same output, every time. An agentic system adds interpretation: it reads an incoming document, understands the customer's actual question, matches an order line to the right product even when the supplier's item name does not match your catalogue.

The difference is the same as between an invoicing program and an accountant: the first follows rules, the second also reads between them. A well-built agentic commerce business needs both, because interpretation and dependable execution serve different cases.

The four levels of autonomy

In commerce, autonomy is not a switch but a ladder. Every process climbs it step by step, and not every process needs to reach the top. We wrote the ladder out in our article on AI in business processes: assist, recommend, act with controls, coordinate autonomously. What is worth repeating here is its conclusion: the level of autonomy is not a goal but a risk measure. McKinsey's 2025 research shows that the companies getting the most value from AI are almost three times more likely to have fundamentally redesigned workflows rather than simply added a new tool to existing work.

Where the boundary belongs

According to Gartner's June 2025 forecast, more than 40% of agentic AI projects will be cancelled by the end of 2027, for escalating costs, unclear business value and inadequate risk controls. That is not an argument against agents. It is a reminder: software that acts on its own initiative needs clearer boundaries than software that only calculates.

For the merchant's agent, the trust boundaries we wrote out in Can AI send your invoices? apply: the model interprets, a deterministic system validates, and people decide the cases that carry commercial weight. The six questions leadership must answer before granting autonomy are not repeated here; they are written out there.

On the buyer's side, protocols attest payment consent: ACP, AP2 and the payment networks' tokens are built to make consent explicit and provable. The merchant's agent must set its own boundaries. That responsibility stays with you.

Three prerequisites, without which an agent stays a pilot

An integration layer. The agent needs access to business context: storefront, ERP, PIM, CRM, warehouse and carrier interfaces. Without it, an agent is an expensive version of a notebook. If this layer is missing, your problem is orchestration, not intelligence.

Data quality and permissions. An agent is as good as the systems it reads from. Read and write access must be designed separately, because interpretation and modification are different risks.

An owner and a metric. Every workflow must belong to someone and measure something leadership wants to know: cycle time, exception rate, cost per case.

With all three in place, agents are an addition. Without them, an agent is a step backwards.

Two common mistakes

The first mistake is buying the agent before the workflow. The market sells ready-made agent pilots, but a tool does not make a process measurable. Start from the workflow, then choose the tool that serves it.

The second mistake is starting at the top of the ladder. Autonomy looks more impressive than assistance, but assistance delivers faster, cheaper learning: first prove interpretation accuracy on inputs, then allow action. Companies that reverse this order pay for an expensive pilot and create distrust that outlasts the project.

Five steps

Step one: list what already runs on its own. Every workflow that already runs machine to machine is your foundation and the baseline every future decision must improve on.

Step two: assess how machine-readable your data is. Are product information, prices and inventory available through APIs so that external agent services can read them? That is what decides whether you get found in agentic commerce.

Step three: choose one merchant workflow where errors are already expensive and the rules fit on a single page. Inbound purchase orders, first-line returns handling and customer-data updates are usually the right first candidates.

Step four: map the workflow as it actually runs, not as the process document claims. This exposes the gaps no ready-made tool can solve: uneven data quality, unwritten rules and missing integrations.

Step five: separate interpretation from execution. AI reads the incoming order; the rule engine checks contract prices, stock and customer data; a person decides the exceptions. Start with assistance, move to controlled action only once accuracy is proven, and measure every decision against the baseline.

Conclusion

Agentic commerce is two movements happening at once: the buyer's agent begins purchasing on the user's behalf, and the merchant's agent begins fulfilling orders without a person in the loop. Both require the same foundation: machine-readable product data, clear pricing rules and a working integration layer.

The buyer's agent buys based on data, wherever it finds it. The merchant's agent fulfils where the workflows are in order. Both sides begin with your data. Zaproo helps assess and build both: making product data machine-readable, pricing rules clear and merchant workflows dependable. Talk to an engineer.