The Consumer Trust Gap

Between January 29 and March 20, 2026, hotels stopped talking about agentic distribution and started running it. Accor put its ALL app inside ChatGPT. Hyatt followed with a branded app. Lighthouse shipped an MCP-based ChatGPT app open to any hotel at a flat fee. Aven embedded Model Context Protocol across roughly 35,000 SynXis properties. OpenAI stepped back from in-chat checkout and left transactions to partner apps, while Google pushed ahead with tiered agentic booking under its Universal Commerce Protocol. The March 2026 Inflection reconstructs that six-week sequence; this article picks up where it ends.

The protocol layer is equally complete on paper. MCP (November 2024) handles discovery and tool calls. Google's Agent2Agent protocol (April 2025) describes agents negotiating with agents. Google's Agent Payments Protocol (September 2025) and OpenAI's Agentic Commerce Protocol with Stripe (September 2025) describe how an agent pays. UCP (January 2026) describes how it buys. Every layer an agent needs to find, price, and pay for a room now exists in some production or specification form.

The consumer has not moved at the same speed. Skift's March 2026 reporting on Skift Research and McKinsey survey data puts the share of travelers who use AI “extensively” for trip planning at 30%, up from 13% a year earlier — a 124% increase. That is a large number, and it is a planning number. Asking an assistant where to stay is not the same act as letting it choose, commit, and pay. Autonomous booking remains a trust fall most travelers are not yet willing to make. Skift's sharpest framing of the moment was a March 3 headline: travel brands are building AI agents for a consumer that does not exist.

The industry's default reading is that this is a maturity problem: better reasoning, longer memory, and finer preference capture will close the gap on their own. Some of it will. But the harder part of the gap is not in the model. It is in the arrangement around the transaction — who is accountable, what was promised, what the traveler can lose, and how they are made whole. Those are structural properties, and no amount of reasoning quality supplies them.

The abstract names three frictions. Each is easier to see in a concrete booking than in the abstract, so each is illustrated below with a scenario at a downtown Memphis property. The scenarios are illustrative, not case records.

Delegation anxiety

A traveler tells an assistant: find me a room downtown for the Friday of the festival weekend, under $250, and book it. The agent finds a queen at $149 a night and reaches the step that matters — confirm and pay. This is where most people stop. The charge would come from a merchant the traveler has never heard of, under a cancellation policy they have not read, initiated by software, with no one to call if it goes wrong. In card-network terms it is a card-not-present transaction started by a third party.

The familiar remedy does not fit. Card networks give consumers a chargeback window of up to 120 days from the transaction or the expected date of service, with reason codes, representment, and, since 2023, Visa's Compelling Evidence 3.0 rules for merchants. All of that was designed for stolen cards and goods that never arrived, not for “my agent misunderstood me.” Delegating a purchase is a relationship most people reserve for someone they know: an assistant, a long-standing travel agent, a spouse. Software has not earned that standing, and the payment system offers no substitute for it.

Preference opacity

The same traveler adds a constraint: something that feels like a boutique but isn't. Quiet but not isolated. The kind of place where the bar has good lighting. These are real decision inputs and they resist structured data. An MCP server can transmit search_availability, rates, room types, and a cancellation window. It cannot transmit atmosphere.

Structured fields also cut both ways. “0.2 miles from Beale Street” is a feature to one traveler and a defect to another, and the agent cannot know which without the conversation the traveler asked it to skip. The Dispute Resolution Working Group's own taxonomy treats subjective quality disputes as the hardest category to adjudicate for exactly this reason: the gap between marketing language and personal expectation. An agent that books on opaque preferences manufactures that category of dispute at scale.

Accountability ambiguity

Now suppose the agent, working from data it cached the day before, books a studio for three nights. The property's current rule requires four nights on festival weekends; the confirmation is generated anyway; the guest arrives to a re-priced folio and a front desk that did not write the rule the agent read. Who is responsible — the AI platform, the property's MCP server, the PMS vendor whose feed went stale, or the traveler who delegated? Today there is no forum in which that question is even well-formed.

ADAPT's dispute framework gives these cases a category of their own — Category G, agent-specific disputes: agent misrepresentation, authorization disputes (“I did not approve that booking”), protocol interpretation (“available: true” with an unsurfaced minimum stay), multi-agent conflicts over the same inventory, and stale data. None of these exist in an OTA resolution center or on a card network's reason-code list. Booking.com's Partner Hub and Expedia Partner Central mediate guest–property disputes; a card issuer mediates cardholder–merchant disputes; neither has a slot for “my agent.” Airbnb's Resolution Center and AirCover for Hosts show that a platform can build a remedy layer — for a marketplace it controls end to end. Nobody has built one for the open case.

Hospitality Net's Ira Vouk drew a useful line in February 2026 through what the industry lumps together as “agentic booking.” ADAPT's March 2026 briefing carries the same three-tier framework with labels tuned to hotel distribution. What matters here is not the technology in each tier but the trust each asks of the traveler.

The distance between the second row and the third is where the trust gap lives. It is the distance between reading a recommendation and handing over a card. Technology closes very little of it: the second row already works when the human approves. What the third row asks for is a change in the arrangement, not in the model.

Consumers already delegate payment to software at scale in two familiar places. Neither is hotel data, so treat what follows as analogy rather than evidence. It is still instructive, because the delegation happened for reasons that had little to do with how clever the software was.

Ride-hailing. The rider never negotiates the fare, never hands over a card at the curb, and never sees the driver's merchant account. The app shows a price before commitment, pre-authorizes the charge, settles it, and stands behind the trip with a refund flow measured in minutes and a rating that follows the driver from ride to ride. People delegated because accountability was legible: one counterparty, a bounded loss, a known remedy, portable reputation.

Marketplaces. Buyers pay strangers because the platform interposes itself: money-back guarantees, the ability to hold or reverse funds, published reputation, and a claims process that does not require a lawyer. Airbnb built the same layer for lodging — payout released to the host only after check-in, a Resolution Center, AirCover — and it was the remedy layer, not the quality of the search, that let people book a stranger's apartment.

Read across both, delegation required five conditions: a single accountable counterparty; the commitment shown before the charge; a loss that is bounded and, ideally, pre-funded; a fast remedy with a known process; and reputation that travels with the participant. The caveat is that a ride and a parcel are commodities. A hotel stay is experiential, its disputes are subjective, and the remedy layer therefore has to be hospitality-literate, not merely fast. That is the difference between a call-center script and a local arbiter.

Morgan Stanley's counter-thesis deserves a fair hearing. Its analysts argued in early 2026 that agentic AI is unlikely to disrupt Booking.com in the near term and may redirect discovery toward OTAs rather than away from them, because the OTAs already own the trust infrastructure travelers rely on at the moment of payment: reviews, guarantees, customer service, and a brand the traveler recognizes. A traveler who lets an assistant book “through Booking.com” is not trusting the agent. They are trusting a 2000–present institution standing behind it.

Operators should read that without resentment. OTAs built demand, they still deliver it, and they built the remedy layer most travelers have ever used. ADAPT's objection has never been to the existence of a trust layer; it is to the terms — a 15–25% commission and a gatekeeper position — and to resolution centers designed as customer-retention functions rather than adjudication. The Dispute Resolution Landscape documents that record in detail.

The counter-thesis is correct exactly as long as there is no alternative trust infrastructure. The question for independent operators is therefore not whether AI will make travelers trust hotels directly. It is whether a trust layer can exist that is open, local, and enforced by protocol rather than owned by a gatekeeper — and whether it can be made legible to a traveler at the confirm-and-pay step.

Each friction maps to a mechanism in the ADAPT stack, and each mechanism is a prerequisite for delegation rather than a feature added to it.

Localized arbiters answer accountability

The Dispute Resolution Working Group (ADAPT-WG-001) is drafting ADAPT-DRP around certified, localized arbiters rather than a central call center. The framework gives agent-mediated bookings what they lack today: a named category (Category G); an evidence standard that includes the property's listing as it stood at booking time and the agent's conversation log; a clock — 4 hours for safety and habitability, 24 hours for material impact on a stay, 72 hours for post-stay financial disputes, 7 business days for complex cases — and a person with local knowledge whose fee is the same whichever way the decision goes. Disputed funds sit in escrow from filing; the decision is the settlement instruction. Appeals go to a three-arbiter panel within 14 days. Decisions are published anonymized, and both properties and AI agents carry dispute profiles that other agents can read.

When an agent books the wrong room under this framework, the traveler can answer the accountability question before they delegate: there is a category, a standard, a deadline, and a named human. That is what a chargeback never offered and what an OTA offers only on its own terms.

Guest trust deposits answer delegation anxiety

ADAPT-DRP proposes a small, refundable, one-time guest trust deposit posted when a traveler joins the network, with three visible tiers: Verified (clean history, deposit posted), Bonded (deposit insured by a third party for a small fee), and New (no history yet, deposit required). The deposit is never touched unless a dispute is filed fraudulently. It is a two-way signal. The property sees an accountable guest and can waive its own deposits or offer better rates. The guest, in turn, is spending inside a bonded arrangement in which the disputed amount is already held and cannot be clawed back months later. For delegation specifically, the agent is no longer spending into the void: the loss on both sides is bounded, and it is bounded before the booking, not litigated after.

Transparent intermediation answers preference opacity

The manifesto floats AI-empowered local travel advisors earning 1–2% through the protocol. The advisor understands “quiet but not isolated” because they live in the destination; the agent handles logistics, the human handles judgment. This hybrid is probably the realistic path to delegated booking for anything more nuanced than a commodity room-night. Note the terminology: arbiters and advisors are intermediaries, and they remain — visibly, paid transparently, on open rails. They are not gatekeepers. That distinction is the whole design.

Portable identity lets reputation travel

The proposed WG-002, Guest Identity & Credentials, is the piece the marketplace analogy demands and the OTA model withholds: a portable, verifiable guest profile that travels with the guest by consent rather than living inside a platform account. A property should be able to confirm “Verified, twelve clean stays, no open disputes” without seeing the guest's dispute history or spending patterns, and the guest should be able to carry stated preferences — high floor, away from the elevator, late checkout on Mondays — from one direct booking to the next. Reputation that travels is what turned strangers into acceptable counterparties in every other delegated-payment market.

Better reasoning closes none of these rows. A more capable model can describe a room more accurately; it cannot make itself accountable, bound the traveler's loss, or hold a property to a snapshot of its own listing. Those are properties of the arrangement, and programmable settlement is the mechanism that supplies them: the terms travel with the transaction and are enforced by protocol, not by trust in an intermediary.

Delegation will be earned in roughly the order consumers granted it elsewhere: visibility first, approval loops second, bounded loss and accountable humans third, and only then autonomy. Sequenced for an independent operator, the path looks like this.

1. 1

   Be findable and honest

   Publish clean structured data — rooms, rates, policies, photos that match the room. Under ADAPT-DRP the listing is the contract, so accuracy is not marketing hygiene; it is the first line of dispute prevention.

2. 2

   Keep the human on the approve button

   Offer AI-assisted booking with the terms shown in plain language before payment: rate, taxes and fees, cancellation window, deposit. The approval loop is not a limitation to be engineered away. It is where trust is manufactured.

3. 3

   Bound the downside before removing the button

   Put the loss ceiling in the protocol: a dispute reserve held until checkout plus a filing window, escrow on any contested amount, and the guest trust deposit. Programmable settlement makes these terms of the transaction, not promises in a policy page.

4. 4

   Put an accountable local in the loop

   Stand up the arbiter network in your market through DR-WG certification, and let local advisors carry the judgment calls an agent cannot. A traveler who can name the person who decides is a traveler who can delegate.

5. 5

   Let reputation travel

   Recognize Verified and Bonded guests with waived deposits or better rates, and carry their stated preferences forward by consent under WG-002. Every clean stay should make the next delegation easier, wherever it happens.

6. 6

   Then delegate within limits

   Autonomous booking arrives last and inside stated boundaries — a maximum spend per booking, approved properties or trust tiers, a cancellation floor — with the limits enforced at settlement rather than requested in a prompt.

ADAPT's position is that the trust problem has to be solved before the technology problem, and that it has to be solved structurally — in the arrangement around the booking, where the traveler can inspect it — rather than asserted in marketing. Three working groups carry the work. The Dispute Resolution Working Group (WG-001) carries the arbiter model, the escrow flow, and Category G. The proposed WG-002, Guest Identity & Credentials, carries the trust tiers and portable, consent-based guest identity. WG-004, the Arbiter Certification Program, depends on DR-WG and will produce the training, examination, and fairness-scoring standards that make “certified local arbiter” mean something to a traveler who has never heard of one.

The hardest question is already on the agenda for the DR-WG's inaugural debate: would guests trust an arbiter they have never heard of over a brand like Booking.com? The honest answer today is no — and the analogies above suggest that brand was never the mechanism. Legible accountability, bounded loss, fast remedy, and portable reputation were. Build those into the protocol and the brand question answers itself over time; skip them and no improvement in AI reasoning will close the gap.