When AI Agents Start to Shop and Pay
For years, online shopping meant searching, comparing, clicking “Buy” and paying. Now AI agents are starting to do more of this work for us.
You can give an agent a simple goal: “Find running shoes under $150 that can arrive by Friday.” The agent can search, compare products, build a cart and increasingly help complete the purchase. This is called agentic commerce.
What does an AI agent need to do?
A shopping agent needs several capabilities:
- Find products and compare offers.
- Build a cart and place an order.
- Pay securely with the user's permission.
Different protocols are being developed for these tasks. They are not all direct competitors. They often work at different levels.
- MCP connects AI agents to external tools and data.
- UCP supports agentic commerce between shoppers, businesses and payment providers.
- ACP supports commerce and checkout flows between agents, users and businesses.
- AP2 focuses on secure and authorized agent payments. It was originally introduced by Google and moved to the FIDO Alliance in 2026.
The big idea: agentic commerce is becoming a stack, not a single protocol.
A reality check
In September 2025, OpenAI launched Instant Checkout in ChatGPT. It showed that an AI assistant could move beyond recommendations and help users buy products directly.
But the first version also showed how difficult agentic commerce is. A good shopping experience needs more than a “Buy” button:
- Multi-item carts
- Inventory and delivery
- Returns
- Merchant integration
- Payment security and user authorization
The lesson is simple: a protocol can work well while the overall product still fails to offer enough value.
From discovery to the full shopping journey
The next step is not only about recommending products. AI platforms are starting to cover more of the journey:
Discovery → comparison → cart → checkout → payment
Google’s Universal Cart is an example of this direction. It lets users collect products from different merchants across services such as Search, Gemini, YouTube and Gmail.
The goal is clear: the AI interface can become a place not only to ask what to buy, but also to prepare and execute the purchase.
The real challenge: trust
The hardest question is not “Can an AI agent shop?” It is “Can we trust it to act for us?”
- Did the user really authorize the purchase?
- Did the agent respect the budget and other limits?
- What if the price changes?
- Who is responsible if the agent makes a mistake?
This is why secure authorization and traceable payments are becoming so important. The system must connect the user's intention with the agent's action and the final transaction.
What should online businesses do?
- Make product information clean, accurate and machine-readable.
- Keep prices, stock, variants and delivery information up to date.
- Do not depend on only one protocol or AI platform.
- Make products easy for agents to understand and compare.
- Protect the customer relationship, even when an agent handles the transaction.
- Build strong trust, security and customer service.
The strategic question
Traditional e-commerce was built around people visiting websites. Search, mobile and marketplaces already changed that model. AI agents may change it again.
If an agent can compare thousands of products, build a cart and eventually pay, the customer may no longer need to visit a merchant's website.
So the real question is not only: “Will AI agents shop for us?”
It is: “Who will control the customer when the customer stops shopping directly?”
The bottom line
Shopping is moving from clicks to goals. We describe what we want, and agents increasingly handle the work.
The winners will not necessarily be the companies with the most impressive AI demos. They will be the businesses whose products are easy for agents to find, understand, trust and buy.
