# Machines of Buying and Selling Grace

Adam Behrens, New Generation | AI Engineer World's Fair 2025 | 19:37

Source: https://www.youtube.com/watch?v=zlZz0mDF2eg
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/machines-of-buying-and-selling-grace
Published: 2025-07-23
Tags: chatbots, design, mcp, workflows

## TL;DR
- AI commerce will move from websites and human browsing toward merchant agents, consumer agents, and infrastructure that handles intent.
- Agentic commerce needs shared product APIs, richer buyer and seller preferences, and systems that can reason about conflicts and negotiate outcomes.
- Retail brands should expose their product, brand, and payment systems through interfaces that can work inside many chat products and other surfaces.

## Summary
Adam Behrens defines a store as a location and protocol that lets buyers and sellers complete transactions. E-commerce digitized merchandise and distribution, while AI can digitize the participants and their interactions. Consumers may shop through agents that browse websites, or through programmatic connections such as APIs and MCP servers. Behrens describes the engineering problems this creates: authenticating agent payments, turning vague requests into specific products, finding inventory across merchants, and representing changing buyer and seller preferences. He argues that agentic commerce needs intelligence at every stage, including reasoning, coordination, and negotiation. New Generation's approach is to give brands consistent product APIs, connect product data to brand systems, create experimental generative interfaces, and support transactions from AI surfaces. Behrens expects stores to remain, but to return toward their original form as conversations. In the questions, he discusses embedded shopping, credit cards, stablecoins, super-app behavior, and possible affiliate or data revenue for model providers.

## Key ideas
### A store is a system that lets buyers and sellers transact
[00:03](https://www.youtube.com/watch?v=zlZz0mDF2eg&t=3s)
Behrens starts by comparing three forms of retail. In the older store, inventory sat in the back and a clerk fetched what the customer requested. Information systems enabled large retailers, moving inventory to the front and supporting browsing. The internet then moved that store online, combining broad merchandise with broad distribution. His definition is more general: a store is both a location and a protocol that facilitates transactions between merchants and buyers. This definition lets him ask what changes when AI enters commerce, rather than treating a website as the whole store.

### AI can digitize the participants and their interactions
[02:08](https://www.youtube.com/watch?v=zlZz0mDF2eg&t=128s)
Behrens says e-commerce digitized merchandise and distribution, while AI digitizes the participants and their interactions. Static websites can become merchant agents. Consumer browsing can become consumer agents. Payment infrastructure can move toward infrastructure for expressing and acting on intent. He describes two possible paths. An agent might visit a website designed for agents, use the catalog and brand guidelines, and complete checkout there. Or an agent might use an API or MCP server to access merchants programmatically and receive dynamically generated interface elements inside a chat product.

### Agent payments need delegated authority or an intermediary
[04:44](https://www.youtube.com/watch?v=zlZz0mDF2eg&t=284s)
The immediate transaction problem is that software may be the thing pressing the buy button. Behrens says Operator mostly errors on current e-commerce sites. One existing approach is for the chat or software provider to handle checkout, create a virtual card, and purchase on the user's behalf. He calls Visa's delegated authentication approach more elegant because the agent can use the user's actual credit card through the checkout flow. Payment can therefore work without forcing every merchant to accept a new intermediary payment method.

### Agents must turn vague intent into a specific product
[06:17](https://www.youtube.com/watch?v=zlZz0mDF2eg&t=377s)
People often express a need such as wanting running shoes rather than naming a specific SKU. Today, systems commonly force the user or agent to provide a product detail page URL. Behrens argues that conversational data gives a richer view of intent because a system can ask what the user is trying to do instead of inferring intent from keywords, clicks, and site metrics. He says merchants working with New Generation are seeing users from AI channels show higher conversion, higher dollar value, and higher lifetime value. That could change how merchants think about fulfillment and returns.

### Merchant offers will become dynamic and contextual
[07:51](https://www.youtube.com/watch?v=zlZz0mDF2eg&t=471s)
Behrens describes the current product detail page as mostly static, with a buy button, price, discounts, and perhaps bundles. An agentic seller interface could show product availability in real time, offer contextual pricing and discounts, and create bundles across multiple merchants. This changes the seller's product representation from a fixed page into something that responds to the user's situation and the merchant's current conditions. The merchant needs systems that can express what it can sell and how it wants to sell it at the moment of interaction.

### A unified product-data layer can replace feeds and scraping
[08:30](https://www.youtube.com/watch?v=zlZz0mDF2eg&t=510s)
Finding a particular item across thousands of stores is another infrastructure problem. Existing product-feed systems require a chat product to work separately with each merchant. Scraping creates repeated work and sends bot traffic to websites. Behrens describes New Generation's longer-term approach as a unified API for product data across merchants. He compares the idea to Plaid, except that the aggregation is over merchants rather than financial institutions. A common access layer would give agents a consistent way to find products and availability.

### Agentic markets need honest, changing preferences and coordination
[09:33](https://www.youtube.com/watch?v=zlZz0mDF2eg&t=573s)
Current commerce data is narrow and one-sided. Users have siloed accounts, transaction data, and limited model memory. Businesses reveal little beyond periodic reporting. Behrens imagines richer user context and businesses expressing real-time goals such as low inventory, target users, or changes caused by tariffs. He is direct about the difficulty: preferences change, conflict between buyers and sellers, and participants have reasons to misreport them. Drawing on his work at Bridgewater, he says finance uses third-party institutions and market makers to manage these differences. Agentic commerce may need similar institutions.

### Brands can prepare by exposing data, design, and payments
[12:16](https://www.youtube.com/watch?v=zlZz0mDF2eg&t=736s)
For a company such as Samsung, New Generation starts by creating an API and MCP server that chat clients can use. It abstracts complex internal product systems into consistent endpoints. The next step connects product data with other company data, beginning with the brand and design system so products appear in the way the company wants. New Generation then provides an AI subdomain for experimenting with generative interfaces that combine product and brand data. The final piece enables payment flows on this new surface, where bot traffic becomes a desired form of customer access.

### The store may return as a conversation
[15:03](https://www.youtube.com/watch?v=zlZz0mDF2eg&t=903s)
Behrens does not predict that stores disappear. He says they evolve toward their original form, which was conversational. In the question period, he says shopping may initially continue to link from ChatGPT or other chat products to websites. Brands still want to own a surface, so he expects web-like interfaces that can transport their data and components into many applications without rebuilding everything. He describes this as an inversion: instead of users going to a website, the website's data and components go to many places.

## Notable quotes
- "A store is a location for and a protocol that facilitates transactions." (01:42)
- "If e-commerce digitized the merchandise and the distribution, AI digitizes the participants and their interactions." (02:08)
- "The more elegant long-term solution that we're working on is to actually create a unified API to access product data across every merchant." (09:13)
- "We don't think stores go away. We just think that they evolve." (14:55)
- "We think that form is actually a conversation." (15:00)

## Tools & references mentioned
- New Generation
- Walmart
- Home Depot
- Adidas
- Reebok
- Brooks
- ChatGPT
- Claude
- Operator
- MCP
- Stripe
- Visa
- Bridgewater
- Samsung
- Plaid
- Google
- Gartner
- stablecoins
- crypto

## Who should watch
- You are building commerce infrastructure and need to decide how agents will discover products, express intent, and complete payment.
- You work for a retailer or brand with separate product, inventory, design, and checkout systems that may need to work inside chat products.
- You are evaluating whether APIs, MCP servers, delegated authentication, or richer market coordination are needed for machine customers.

## Related talks

- [The Agentic Commerce Stack](https://aietalks.com/talks/the-agentic-commerce-stack) (Ahnaf Prio, Best Buy, 20:38)
- [Ionic Launch: Opening the economy to AI agents](https://aietalks.com/talks/ionic-launch-opening-the-economy-to-ai-agents) (Justin, Ionic Commerce, 05:30)
- [Multimodal Collaborative Agents for Next-Gen Commerce](https://aietalks.com/talks/multimodal-collaborative-agents-for-next-gen-commerce) (Nidhi Kaushik Vyas, Google DeepMind, 21:08)
- [Teaching agents to pay](https://aietalks.com/talks/teaching-agents-to-pay) (Anna Spysz, Stripe, 19:10)
- [When AI Agents Pay and Sellers Monetize: Building x402 Apps on AWS](https://aietalks.com/talks/when-ai-agents-pay-and-sellers-monetize-building-x402-apps-on-aws) (Anil Nadiminti, AWS, 20:41)