As personal AI agents such as Muse introduced by Meta spread, how they might change existing user experience has emerged as a point to watch.

In a recent post on social media X (Twitter), Decagon CEO Jesse Zhang (제시 장) said personal AI assistants will also bring major changes to corporate customer experience. "Amazon blocked Muse. Shopify did the opposite and opened a dedicated payment channel for Muse. Consumer-facing companies need to decide which side to stand on and think about a new UX," he stressed.

If a company is large like Amazon, a strategy of directly controlling the shopping experience and blocking external access could work. Even so, Zhang said, "In the end, Amazon will also back down. It is because what customers want is AI assistant integration."

According to him, use of personal agents such as Muse and Instinct is rising steeply. That is because these agents actually handle tasks. In Meta's case, Meta expects Muse to buy items directly using users' money and provides purchase protection of up to $1,000 per transaction.

Zhang stressed, "Every consumer company will face a customer segment it never considered. It is a robot with a wallet. It is a change as big as the shift from offline to online, and then to mobile apps."

He takes the view that transacting with agents is not something that ends with a single API and one page of documentation.

"If the customer is an agent, the existing UX is just baggage," he said. "Agents skip everything, from product displays and purchase-inducing flows to discount benefits that block cancellations hidden three screens deep. Agents do not look at screens. Rules that were enforced through screens now have to be enforced in other ways."

Zhang thinks agents should also take on that role. He said it is work that requires judgment, not rules. In this regard, he stressed five things.

First is verifying identity and authority. Companies must confirm whether an agent truly represents the customer and define the scope of its authority. Browsing and purchasing are different, and reordering $40 and booking $4,000 are also different.

Second is rules and procedures. Handling differs for flight changes, card reissuance and plan changes depending on customer type, laws and internal policies.

Third is negotiation limits. "If you cannot decide who gets a discount and when to stick to principles, agents will secure the terms most unfavorable to the company every time," Zhang said. "Agents do not get tired or embarrassed, and compare competitor offers in real time."

Fourth is exception handling. Even for requests past the refund deadline, a customer of 6 years and a customer who joined yesterday should be viewed differently.

Fifth is fraud happening at machine speed. While legitimate agents may buy goods in 0.2 seconds, malicious agents probe for policy loopholes 10,000 times an hour, so limiting request frequency alone cannot stop them.

The end state Zhang envisions is a structure in which AI customers and AI agents interact in transactions. People set what they want, such as "buy it at the lowest price," and companies set direction by saying, "protect margins and retain loyal customers." Transactions are handled machine to machine.

"A company AI that deals with customers must be able to deal with both people and AI assistants. If it cannot adapt, it will fall behind," he said. "If the effort required for a single conversation falls close to zero, conversations between consumers and companies will increase sharply, and economic potential that has not been realized will also come alive."