Travel companies are staring down a trust gap on the road to artificial intelligence (AI)-powered booking that is both technical and behavioral.
Following OpenAI’s recent relocation of its Instant Checkout, Walmart reported its Instant Checkout conversions in ChatGPT were three times worse than on its own website. The result indicates consumer reluctance to adapt to a new commerce structure, emphasizing a potential barrier to AI booking.
To enable agentic booking in the future, AI, travel and payments companies must close a gap that is as much about the plumbing of payments as it is about the psychology of the purchase.
“Trust doesn't transfer automatically through a new interface,” Matthew Mamet, a fractional chief product officer and chief growth officer wrote on LinkedIn. “You have to earn it in the new context, which takes time and usually means worse numbers before better ones.”
The takeaway isn’t that agentic commerce is dead, Mamet wrote. But it’s not going to become a standard method of transaction overnight either.
Trust gap
To move from AI planning to AI booking requires a double-sided evolution including technological guardrails and a shift in travelers’ buying habits.
Industry stakeholders have identified pain points in this evolution.
“Trust can only be accomplished if the underlying technology allows for a safe end-to-end transaction,” said Paul Beukers, chief commercial officer and co-founder of Katanox.
In a typical travel booking, customers visit the merchant’s website or app and input credit card details manually, according to Andrew Beckmann, travel and hospitality industry GTM lead for Stripe.
“The human is making … buying decisions,” Beckmann said during a March Destination AI virtual discussion titled “AI & The Next Booking Shift.”
That kind of transaction can’t be copied, pasted and handed to an agent, Beckmann said. There are too many fraud and liability risks such as a rogue agent booking more than it should.
The fundamentals of payment architecture are changing as players seek to accommodate agentic transactions, Beckmann said.
Beukers cautioned that trust also involves “how money moves,” and whether it’s moving in a compliant, secure and trustworthy way—no matter what interface a payment is made through.
“The biggest thing that we have to start to think about here is the operational correctness piece of it … how do we actually tie that into the interface,” Tyrone Millard, GTM for media and entertainment—previously travel and hospitality—at OpenAI, said at the Destination AI event.
“The technology to make that work matters,” Mamet said. “But the consumer behavior design around when and how to make the ask matters more.”
According to Mamet, the “trust transfer” is the barrier to agentic commerce success as evidenced by Walmart’s results.
“When a customer lands on Walmart.com, they're in Walmart's environment. The reviews are Walmart's reviews. The return policy is Walmart's,” Mamet wrote. “The whole context signals: this transaction is safe.”
Within ChatGPT, that setting melts away, he wrote. The customer is instead purchasing from the conversation. “The trust infrastructure that converts a browser into a buyer isn't in the checkout button. It's in everything around it.”
Mamet reflected on his time as VP of product for hotels at Tripadvisor when he led the Instant Booking effort in 2015. He told PhocusWire the security architecture was state of the art.
But that didn’t matter much for consumer adoption.
Trust doesn't transfer automatically through a new interface.
Matthew Mamet, fractional chief product officer and chief growth officer
What moved the needle was a “green checkmark” from Tripadvisor flagging the transaction was safe, according to Mamet. Travelers had already spent significant time using Tripadvisor to research hotel options. But asking them to book there was asking them to create a new habit, resulting in a snag.
“Those are two completely different problems, and the travel industry is going to confuse them again with agentic booking,” Mamet said.
Addressing the technology side
Travel and payments companies are taking steps to implement security layers around transactions to build technological trust.
Stripe’s Beckmann said that protocols and tokens could serve as building blocks for payments.
The first part is the protocol that would allow for an agent to act on a user’s behalf.
“We have one that we co-developed with OpenAI called the Agentic Commerce Protocol (ACP). Google's got one called Universal Commerce Protocol (UCP),” he said, adding it’s all open-source.
The second part Beckmann said is something Stripe has created called a “shared payment token” meant to address the risk of an agent going rogue.
The shared payment token is single-use, created when a transaction begins. It’s made with specific criteria and has guardrails around how much it can spend, how long it’s viable for and allowed use.
“It does all of that without exposing any sensitive credentials,” Beckmann said, adding that helps mitigate risk.
Hopper Technology Solutions (HTS) is separating what an AI agent reads from what it can write, according to Jo Lai, SVP of AI solutions for HTS.
“In most use cases, an agent can search for flights, surface prices, build a cart, etc.—those are read-only actions,” Lai said. “But when it needs to book something or process a payment, that’s a write operation and many of our partners are asking us for explicit human authorization at that step.”
Until things like shared payment tokens and ACP are more widely adopted, HTS is using device-verified payment flows, so that a user can confirm and complete payment outside of the AI conversation. Then, the authorization is passed back to the conversation.
That process allows HTS to abstract the payment from the large language model (LLM) and perceived risks.
“We work with financial institutions, and the industry is broadly converging on the idea that AI agents should not autonomously complete financial transactions without explicit human confirmation at some point in the experience,” Lai said.
Lai said HTS thinks of a checkout session as a “lifecycle.”
An AI agent can create and update a cart but transaction completion requires human authorization. That entails a payment link, stored credentials pre-approved by the customer or a checkout widget for agent handoff.
“The key architectural position we often take right now (and this can change quickly) is that the agent layer is separate from the payment action,” Lai said. “While capable of triggering real-world actions autonomously, an AI agent should only do so with the right authorization in place.”
Others see identification technology as the key.
Heather Dahl, CEO and co-founder of Indicio, said the missing piece is identity.
“Simply put, a customer must be identifiable in a way that provides assurance that they’re not a synthetic identity or a deepfake,” Dahl said. “The AI agent must be identifiable in a way that provides assurance that it is authentic—it’s not a spoof agent.”
Verifiable credentials can provide a foundation of trust for AI transactions, as they can show cryptographic proof that’s tied to a verified identity, according to Dahl.
Payment networks are working on building that on their side with agent payment protocols, she said, referencing Visa and Mastercard.
“But payment authorization is only half the equation,” Dahl said. “You also need to verify the human behind the agent, and you need to do it in a way that's privacy-preserving, portable across providers and grounded in standards that work globally."
Addressing behavioral holdup
The challenge with behavior is rooted in the fact that the travelers using AI are not immediately primed to purchase, according to Mamet.
Because they have been using AI for travel research, asking that audience to purchase inside ChatGPT right now is akin to asking someone to buy something at the library, Mamet said.
Trip planning is a high-effort, long-term activity, he said. There’s often one person that is coordinating elements including hotels, flights and experiences for a group. That person, a “high-value traveler,” is already using AI to start planning, according to Mamet.
“The mistake would be interrupting that planning phase with 'Book Now' calls to action before the traveler is ready,” he said.
Mamet theorized that if the booking option is presented outside of the “moment of readiness” then a traveler is likely to ignore it the same way they might ignore an advertisement.
But it can be compelling if it’s presented at the right moment under the right conditions: Once research is finished, with the best price available and with a plan if something goes wrong at check-in.
“Travel brands that lead with the technology story will miss the point,” Mamet said. “The ones that lead with the traveler's journey—and earn the booking moment rather than forcing it—will win.”