The 9 Stages of AI Agent Adoption, Explained

An AI shopping agent is like a customer with a credit card and no eyes: it evaluates product details, prices, shipping costs, and return policies rather than relying on visual design. Greg Eisenberg describes a trust ladder that runs from researching options, to preparing a purchase by filling a cart or requesting a quote, to paying. Many businesses accept agent research; payment is where authorization, fraud, and control concerns become sharper.

Businesses are not responding uniformly. Amazon has blocked Meta’s Muse from browsing and checking out since around September, while Delta says it works with zero outside agents for shopping or flight booking. Yelp allows agents through a paid data-licensing program, and Zillow welcomes agents acting for customers, though anti-scraping tools may still block them. Walmart said Muse’s failures on its site were accidental, and the retailer is a Muse partner. Some apparent resistance may be unintended: systems designed to stop scraping can mistake a customer’s assistant for a scraper, while traffic presenting itself as agent activity may actually seek prices or training data.

The commercial conflict is that agents can complete purchases while bypassing high-margin parts of the buying journey. The host points to Amazon sponsored products and airline offers for seats, baggage, and credit cards as examples of revenue an agent may skip. Amazon earns tens of billions of dollars a year from advertising, according to the host, and its strategy is to keep shopping within Amazon and its Rufus assistant. Google is adding agents and payments to Gemini while protecting an advertising business built on clicks. Advertising may change rather than disappear: ChatGPT has begun testing ads in AI search, while the host is less certain about Google’s approach.

The host’s adoption sequence begins with businesses blocking agents that bypass ads and upsells, followed by an arms race in which agents try humanlike workarounds and companies introduce new blocks. The market may then divide between businesses that welcome agents to win sales and data-focused businesses that charge for access. A trusted identity-and-payment layer could follow, a “Plaid moment” for personal agents. The host expects the stages to overlap and companies to progress at different speeds, rather than move together through a fixed timetable.

Several companies are trying to make agent commerce easier or safer. Meta is working with Walmart, Stripe, and others on an open standard; the host sees Walmart’s Muse partnership as a way for the number-two retailer to gain share from Amazon. OpenAI is adding checkout to ChatGPT, where users can buy from Etsy and Shopify stores, while Shopify’s agentic storefronts let merchants connect once to appear across AI channels. Visa and Mastercard are working with Cloudflare on ways to identify trusted agents, and Cloudflare is also developing ways for agents to pay for access. The host sees payment and security firms positioning themselves as intermediaries, much as Plaid bridged apps and banks after early conflicts over Mint.com.

The host argues that blocking every kind of agent activity may be a mistake when a business’s real concern is whether a payment is authorized. He predicts that agent-friendly businesses could win customers from blockers, and that companies will compete through prices and faster checkout designed for agents. Businesses can test real customer buying questions across several agents, check for accidental blocks, make total price, delivery, compatibility, and return terms easy to read, and set permissions separately for research, cart preparation, and payment. The host also sees potential benefits for specialist suppliers that are discoverable through agents, and opportunities in agent-readiness services, industry-specific identity and payment connections, insurance for agent mistakes, and analytics on what agents ask and where they leave.