E-commerce
August 26, 2026
Agentic commerce promises delegated purchases to AI agents that read your catalogs, compare your products, and place orders without the buyer ever seeing the product sheet. Launch announcements are multiplying, forecasts are optimistic, but almost no actual adoption data is published. Between technical demonstrations and production deployment, four gray areas persist: the lack of a unified protocol, the question of legal liability in case of error, the reliability of the match between customer intent and the ordered item, and the operational management of returns and disputes. This article separates what is already working from what remains a gamble, to help you prepare your store without betting blindly on a technology that is still unstable.
Summary
What is already working in agentic commerce
A software agent can today read a product catalog, compare items based on defined criteria, assemble a shopping cart, and submit a payment identifier to a merchant checkout tunnel. These five steps (catalog reading, selection, cart construction, payment authorization, order writing) are demonstrated in a controlled environment.
What distinguishes an agentic purchase from a classic purchase is that the person never sees the product page. The instruction was given upstream, often in the form of a delegated mandate: "buy me coffee beans every three weeks, maximum budget 25 euros". The agent executes this instruction autonomously.
The necessary infrastructure exists: Adyen launched Adyen Agentic in June 2026, Google published tools for agentic commerce in January of the same year. But these announcements are vendor launches, not measured adoption reports. No public transactional volume accompanies these product releases.
Convert over 2,000 customers on average per month with Qstomy.
The world’s 1st Shopify AI dedicated to customer conversion
Empowering 200+ e-commerce merchants
Four unsolved problems you must anticipate
Four operational areas remain open and represent risks to your cost structure. First, no unified protocol has emerged to orchestrate the exchange between agent and merchant. Second, legal liability in the event of a purchase that does not comply with the customer's actual intention has not been decided. Third, the accuracy of the match between the formulated instruction and the item actually ordered remains uncertain.
Fourth, the management of returns and disputes in this context is completely unprecedented, with no established framework. Each of these four points lands directly in your operations, not in the forecasting slides of consulting firms. McKinsey and Deloitte have published on the subject, but their documents describe the opportunity, not the limits.
This article addresses these four limitations in the order in which they affect your cost base, then separates the preparation work that is useful regardless of the winning protocol from the commitments that remain pure gambles.
Fragmented protocols: every integration is a gamble
More than one agent-merchant protocol coexists today. The clearest proof is commercial: a product category now exists to translate between these protocols. The launch of Adyen Agentic in June 2026 explicitly positions the tool as this translation layer. No one builds a translator for an already dominant standard.
Each protocol must resolve five responsibilities: catalog discovery and reading price and availability, proof of agent identity and client mandate, transfer of the cart to the merchant checkout, payment authorization for an absent buyer, and sending the order confirmation back to the agent. Protocols that diverge on agent identity or payment authorization create the most rework, as these two blocks touch systems that you cannot easily rebuild.
The merchant cost of this fragmentation is expressed in parallel integrations, order reconciliation by channel, and engineering time spent on work that may potentially be thrown away when one protocol eventually prevails. This cost is not publicly quantified by the players selling in this category.
Who decides and who pays for the redesign?
Because no single protocol has won out, the choice is often made by the team that integrates first, usually the technical team. This means that a decision with multi-year consequences on the redesign is driven by a development roadmap rather than by the person holding the revision budget.
This ownership problem is silent and costly. You can find yourself migrating twice in eighteen months because the team chose a protocol that did not survive. To mitigate this risk, document which protocol you are integrating, who chose it, and based on what criteria. Include a budgeted exit clause in your infrastructure roadmap.
If you are exploring agentic commerce, product recommendation based on purchase history will provide you with a behavioral database usable by an agent, regardless of the chosen protocol.
Legal liability: who is responsible when the agent makes a mistake?
Dispute and chargeback rules were built around a cardholder who saw what they were buying. An agentic purchase removes this assumption. No published regulatory framework has filled this gap yet, so the merchant holds an unquantified liability until a framework is published.
The mechanism matters here. In an agentic purchase, the person does not authorize a transaction; they authorize a mandate: an instruction with a scope, a budget, and often a time window. This mandate produces a bounded or tokenized identifier, which the agent then presents to the checkout funnel. The individual transaction is authorized by the agent working within the scope of this mandate, so no human looks at an amount and approves it.
This breaks dispute arbitration in a specific way: the claim "I did not authorize this purchase" is usually defended with evidence of intent—namely, the person's session, the pages viewed, the choices made. When the instruction was a standing intent expressed to an agent, this file contains an abstract mandate rather than a human browsing trail.
Who bears the risk of disputes and chargebacks?
As long as no rules are published, the risk of dispute and chargeback rests on the merchant, not the agent, not the protocol provider, and not the card issuer. You must therefore provision this liability in your operational forecasts.
A defensive strategy consists of keeping a structured record of the initial mandate, the timestamp of each transaction, and the agent selection logic. These three elements constitute your representation file in the event of a dispute. If you outsource order management, verify that your order management system can store this metadata per order.
The real cost of this uncertainty is not technical, it is financial: either you provision a higher dispute rate for the agentic channel, or you accept a budget surprise when the first disputes arrive.
Accuracy of the match between intent and purchase
An agent may purchase an item that is technically compliant with the mandate but does not correspond to the client's actual intention. Example: mandate "buy me running shoes size 42", the agent selects a road model while the client runs in the mountains. The mandate was imprecise, not the agent.
This problem is not a bug, it is a formulation limit. Current agents interpret instructions, they do not read minds. The more ambiguous the mandate, the wider the margin of interpretation, and the higher the risk of mismatch. This risk translates into return rates.
To reduce this mismatch, three levers exist: refining the granularity of product attributes in the catalog, structuring mandates with explicit criteria, and offering a confirmation before execution for purchases exceeding a certain amount. The third lever partially breaks the agent's autonomy, but it limits operational risk.
Returns management: a process to be rebuilt
Agentic returns raise a new question: who initiates the return, the agent or the customer? If it is the agent, on what criteria does it decide that an item must be returned? If it is the customer, how do they signal to the agent that the purchase is not suitable? Neither of these two flows is standardized.
Your current return process probably relies on a customer portal or an email. A software agent cannot fill out a standard web form without specific integration. You will therefore either need to expose an agent-readable return API or manually process agentic returns, which negates some of the automation gain.
The hidden cost here is operational: processing time per return, possibly a higher return rate due to the mismatch between intent and purchase, and the need to build an agent-compatible return interface. If you sell products with a high return rate, test the agentic channel on a low-return category before generalizing.
Product catalog: your data becomes your real storefront
In agentic purchasing, the structured product catalog is your only storefront. The agent does not see your photos, does not read your marketing descriptions, it parses your product attributes: title, price, availability, technical characteristics. If these attributes are incomplete, ambiguous, or obsolete, the agent will select poorly or not at all.
This reverses the usual hierarchy of product content. In traditional commerce, the photo and description sell, while structured attributes are used for filtering. In agentic commerce, structured attributes are the sole sales interface. Your product page becomes a machine-readable file before it is a human page.
To prepare for this change, audit the completeness and consistency of your product attributes. Verify that each product has a standardized title, an up-to-date price, real-time stock, and comprehensive technical attributes. This discipline also benefits your e-commerce SEO and your Google Shopping feeds, so the effort is not wasted if agentic commerce takes time to arrive.
Real-time availability and pricing: a new technical requirement
An agent reading a catalog expects current price and availability, not a static copy updated once a day. If your product feed is regenerated every six hours, an agent may order an out-of-stock item or display an outdated price, leading to a cancellation or dissatisfaction.
Therefore, you will either need to expose a real-time availability and price API, or accept a higher cancellation rate on the agentic channel. The first option requires a synchronized infrastructure between inventory, pricing, and catalog. The second option degrades the experience and increases the operational cost of processing cancellations.
If you are already managing your average e-commerce shopping cart and tracking your performance in real time, you likely have a reliable inventory database. Expose it via API instead of multiplying batch exports; you will gain in precision and responsiveness.
Qstomy: a conversational agent that remains anchored in your current funnel
Qstomy is a Shopify AI agent that guides your visitors toward purchase through product recommendation, upsell, cross-sell, cart assistance, order tracking, and after-sales service. Unlike autonomous agentic systems, Qstomy intervenes within your existing checkout funnel, without replacing human interaction or removing final validation by the customer.
More than 100 merchants use Qstomy to convert traffic into dialogue. The agent answers product questions, offers alternatives in case of stockouts, follows up on abandoned carts, and handles common customer service requests. The customer remains in control of the final transaction, which preserves the proof of intent needed in case of a dispute and maintains your current return process.
While waiting for agentic protocols to stabilize and liability rules to clarify, Qstomy allows you to capture the gains of automation and personalization without carrying the operational risk of a fully delegated purchase. You improve conversion, e-commerce marketing performance, and customer satisfaction, all while keeping control of your funnel and the traceability of each order.
Checklist, in brief and FAQ
Agentic Commerce Preparation Checklist
Audit your product attributes: completeness, consistency, machine-readable format.
Expose price and stock in real time via API rather than batch export.
Document the integrated protocol: which one, who decides, budgeted exit clause.
Provision an agentic channel dispute rate separate from your usual rate.
Prepare an agent-readable return API or a dedicated manual processing workflow.
In brief
Agentic commerce is proving its technical feasibility, but four operational limits persist: fragmentation of protocols, unclear legal liability, accuracy of intent-to-purchase matching, and return management. Prepare your catalog and infrastructure for an agentic future, but provision for risks and document your integration choices. While waiting for the market to stabilize, prioritize conversational agents like Qstomy that automate without removing final human validation.
FAQ
Does a unified agentic protocol exist in 2026?
No. Several protocols coexist, which is why translation layers like Adyen Agentic are appearing. Each integration remains a gamble until a standard is established.
Who is responsible if an agent buys the wrong product?
No published rule decides this. By default, the merchant bears the risk of dispute and chargeback until a regulatory framework is established.
Does agentic commerce increase return rates?
Probably, due to the potential mismatch between the formulated mandate and actual intent. Test on a low-return category before scaling.
Do you need to rebuild your product catalog for agents?
Not rebuild, but structure. Complete your product attributes, expose price and stock in real time, and normalize your titles. This work also benefits your e-commerce SEO and your advertising feeds.
Enzo
August 26, 2026
