My personal AI agents have helped me research and book flights, hotels, restaurants, rental cars and experiences. Too often, the difficulty came at the point of purchase: the business was unprepared to serve the agent I had sent.
Two years ago, I described an agent-to-agent (A2A) economy in which software would negotiate and transact for people and businesses. A second essay outlined three stages: agents taking over individual functions, representing whole stores or business units, and finally dealing directly with one another.
The next chapter of the A2A economy
With personal agents such as Instinct, Muse, and dots, customers are beginning to delegate the work of finding and buying what they want. Businesses now face a customer who can delegate the work of information gathering, negotiating, buying, and even making refunds.
Adobe reported that visitors referred by AI services to American retail websites in July 2026 converted at a rate 60% higher than visitors from other sources. That measures referred shoppers, not autonomous purchases. It nevertheless shows that AI is already directing valuable demand.
Businesses have adapted before. They acquired telephone numbers and listings in the Yellow Pages, built websites that search engines could find, and established profiles on social networks. Each became part of the cost of being reachable. I expect a customer AI agent to become another standard channel.
What a customer AI agent actually does
A personal agent works for the buyer. A customer AI agent represents the business, with knowledge of its products and policies and permission to act. It is a headless storefront: customers can reach it by voice or message, or send their own agent, without navigating a particular website.
Consider a hotel guest who wants to extend a stay. The guest’s agent needs someone who can check availability, quote the extra cost, obtain approval and amend the reservation. A link to the booking page leaves the work unfinished. A capable counterpart returns a confirmed booking.
Behind that counterpart may sit specialist agents, existing software and human employees. The guest should not have to navigate the hotel’s division of responsibility between reservations, billing and operations.
Some routine transactions will need only a reliable API/MCP. A customer AI agent earns its place by handling ambiguous requests, coordinating work across systems and knowing when an exception requires a person.
Businesses disagree on agent access
The response from businesses is far from uniform. Amazon has demanded that Perplexity remove it from Comet’s shopping experience, citing poor customer service and its right to decide whether to participate.
Shopify has chosen to open another sales channel. Its Agentic Storefronts help merchants sell through AI services, while its Meta integration supports purchases within Muse for eligible stores. “We’re making every Shopify store agent-ready by default,” says Tobi Lütke.
The incentives differ. Amazon has a marketplace to defend; Shopify benefits when its merchants sell through more channels. I think Lütke’s bet is well placed. By making its merchants easy for agents to buy from, Shopify can gain a distribution advantage as customers adopt new assistants.
Merchants are right to worry about fraudulent orders, misleading promises and unwanted automation. They also have a commercial interest in controlling the shopping experience. If a personal agent handles discovery, comparison and purchase, the merchant may fulfill the order while someone else controls the customer’s attention.
A business needs a supported way to admit authorized agents and enforce its terms. Refusing the channel altogether risks protecting its website while losing the sale.
AI agents will flow like water. Where acceptable alternatives exist, they will seek another route to what their customer wants. An exclusive product or a much-loved brand may justify the trouble of a difficult purchase. A retailer selling the same goods as its rivals has less room to obstruct it.
Customer AI agents should learn from the operation
Deploying these agents must also become easier. Businesses cannot be expected to rewrite every procedure inside an agent builder.
The agent should learn from the operation itself. With permission and in compliance with ever-evolving AI agent regulations, it should observe how employees handle requests, study internal knowledge and data, and work out how to carry a transaction through the company’s systems. It should propose actions, learn from corrections and take on more work as it proves reliable.
Watching employees is not enough: employees make mistakes, and habit or the status quo is not a coherent policy. The business must define what the agent may promise, which decisions need approval and how errors will be corrected. As the operation changes, the agent must keep learning.
Commerce between agents will need better ways to establish trust. Reputation services could help businesses assess an agent’s record. Authentication must establish whom it represents and what that person has authorized. Enterprise access controls should limit what it can see and do. Permission to amend one booking should confer access to that booking, not the guest database.
On October 20, Visa, BILL and Billtrust announced a pilot for trusted agents to complete approved invoice payments through supplier portals. The pilot uses authentication and payment controls to preserve the supplier’s oversight. It offers a concrete approach to admitting agents without surrendering control.
Business software will have to adapt as well. An agent using inventory, scheduling or payment systems needs clear permissions and dependable results. It must know whether a room was reserved, an order dispatched or a refund issued. Software designed for agents will be judged by how reliably it lets them finish the job.
Making customer AI agents easy to adopt
That is the task we have set for Delight: make it as easy as possible for a business to adopt a customer AI agent that learns its operations and represents it from discovery and purchase through service and repeat business. By helping people and their agents complete transactions as well as resolve problems, it should create a better customer experience and drive revenue.
Your next customer may never visit your website. You should still be able to earn their business.
