Agentic commerce: when AI shops for the customer

AI assistants increasingly search, compare options against a shopper’s constraints, and in some cases complete the checkout on the customer’s behalf. The shopper states the intent; the agent does the legwork.

For most of the internet’s history, shopping meant typing a query, scanning a results page, and opening several sites to compare. That sequence is starting to change. AI assistants increasingly search, compare options against a shopper’s constraints, and in some cases complete the checkout on the customer’s behalf. The shopper states the intent; the agent does the legwork. During 2025, OpenAI and Stripe, Google, Visa, and Mastercard each published standards or programmes aimed at this kind of commerce [1][2][3][4][5].

For retailers, the practical question is what we must expose (product data, pricing, inventory, and returns policies) so that an agent can find our products, understand them, and buy them. For omnichannel businesses the answer reaches beyond the website, into catalogues, pricing engines, inventory systems, and store operations. This article sets out what that means in practice, then tests the argument against its weakest points, because there are several.

• $3–5 trillion Potential global retail revenue orchestrated through agentic commerce by 2030. McKinsey projection, October 2025 [6].
• $900 billion–$1 trillion Potential US B2C retail revenue orchestrated through agentic commerce by 2030. McKinsey projection, October 2025 [6].
• 60+ organisations Organisations Google said it was collaborating with when it announced AP2 in September 2025. Google launch announcement, September 2025 [3].

Figures indicate projected opportunity and industry participation, not measured consumer adoption. McKinsey’s estimates cover goods and exclude services and B2B commerce.

From search results to shortlists

When a person browses, they tolerate ambiguity. They open several tabs, read reviews, and decide for themselves what “lightweight jacket” means. An agent works differently. It reads structured data, filters options against explicit constraints such as size, budget, delivery date, and return terms, and then returns a shortlist or acts on it. Much of the comparison can now happen before a shopper reaches any retailer’s website.

That makes data quality a commercial issue. A beautifully photographed product with a vague description may never reach the shortlist, and neither may one whose price or stock status is out of date. The retailer loses those sales without ever seeing the comparison that excluded it. Discoverability becomes a property of the data, not only of the storefront.

What retailers must expose

Four categories of information largely determine whether an agent can find a retailer’s products and complete a purchase. The table sets out what the agent needs to know and the kind of question it will realistically ask.

These answers must agree across the website, app, stores, and agent-facing services. If price or availability changes during the journey, the customer needs an accurate answer before the purchase is confirmed.

The integration problem

Exposing this information is an integration project before it is a product feature. It depends on consistent product identifiers across merchandising, order management, and store systems. It requires event-driven updates so that price and stock changes reach every channel quickly, and APIs that serve the same answer to a website, a mobile app, and an agent. Many retailers discover that their systems were built channel by channel, and reconciling them is the real work.

The standards landscape is fragmented, as the table shows [1][2][3][4][5]. Retailers should avoid deep integration with any single protocol until the market settles. The durable approach is one governed commerce data layer that feeds many channels, so each new assistant becomes an extension rather than a rebuild. For a group with several business units, that layer can serve every unit’s storefront and partner feeds.

Trust, consent, and control

Agentic commerce changes the trust equation. The retailer now persuades not only the shopper but also the agent acting for them, and the shopper must still feel in control. What may an agent buy without confirmation? A low-value repeat purchase may reasonably be automated; a high-value or unusual one should need explicit approval. How does a customer see and override the agent’s choices? And which personal data is shared with a third-party assistant, for what purpose, and for how long? Clear consent flows, auditable transaction logs, and data-handling rules that hold across every channel are easier to defend with customers and regulators when they are designed in from the start.

Where this argument is weak

It would be easy to present agentic commerce as a simple case for investment. It is not, and several parts of the argument deserve challenge.

Readable data is also scrapable. This is the most serious objection. The openness that lets a shopping agent compare our products also lets a competitor record our live prices and stock across every store and reprice within hours. A blanket ban would block real customers and their assistive tools, so it is not the answer.

What we do: Keep public product attributes open. Put live prices, store stock, and promotional terms behind authenticated APIs, and rate-limit and monitor for unusual query patterns [9]. What we accept: Determined scrapers will still get through, and aggressive bot defences will catch some real shoppers. We assume some price visibility will leak and compete on service and availability rather than hidden prices. Check your logs now, because harvesting may already be happening.

A practical readiness test

What this means for technology teams

Agentic commerce is not a reason to rebuild the storefront. It is a reason to treat commerce data as a product: owned, versioned, observable, secured, and designed for consumers that are not human. Prioritise unified inventory, accurate pricing, clean product data, and well-governed APIs. These investments pay off on today’s website and app, and they position the business for agent-led journeys whenever they arrive.

A useful test: if an AI agent asked a complete question about one of our products, could we give it a fast, accurate, and trustworthy answer, and would we be comfortable if a competitor asked the same question?

Sources

1. OpenAI, “Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol,” September 2025. openai.com
2. Stripe, announcement of the Agentic Commerce Protocol co-developed with OpenAI, September 2025. stripe.com/newsroom
3. Google Cloud, “Agent Payments Protocol (AP2)” announcement, September 2025. cloud.google.com/blog
4. Visa, “Visa Intelligent Commerce” announcement, 2025. usa.visa.com newsroom
5. Mastercard, “Agent Pay” announcement, 2025. mastercard.com newsroom
6. McKinsey & Company, “The agentic commerce opportunity: How AI agents are ushering in a new era for consumers and merchants,” October 2025. mckinsey.com
7. OWASP, “Top 10 for Large Language Model Applications,” LLM01: Prompt Injection. owasp.org
8. Gartner, press release on agentic AI project cancellations, June 2025. gartner.com/newsroom
9. OWASP, “API Security Top 10 (2023),” API6: Unrestricted Access to Sensitive Business Flows. owasp.org