Hilton's AI Planner went live this morning. Here's the audit.

It is Tuesday morning at the Conrad in St James’s and I have been awake since six, partly because of the time difference and partly because the embargo email from Hilton went out at 02:14 London time and I wanted to be in the beta pool when the doors opened. They did, at 09:00 Eastern. The Hilton AI Planner is live on hilton.com as of this morning — generative-AI digital concierge, conversational interface, twenty-seven brands and over nine thousand properties piped into a single chat surface that wants to know where you are going and when. I had five test scenarios written into a notebook from the train down yesterday. By 10:40 I had run all five.

The contrarian thesis — and this is my read, not Hilton’s framing — is that the AI Planner is the first big-brand hospitality AI tool whose job is explicitly defined as narrowing a chain’s portfolio to one weekend’s worth of options, rather than inspiring a trip. Marriott’s RENAI agent is built like a search assistant. Hyatt’s ChatGPT app extension is built like a discovery toy. IHG’s Smart Planner is built like a recommendation engine. Accor’s AI Travel Companion is built like a brochure that talks back. Hilton’s Planner, in the five scenarios I ran this morning, behaves like a conversion funnel with a personality. That is a different product, and the OTA-bypass economics that follow from it are the real story, not the demo.

The cautions are real and I will get to them. The first independent hands-on reviews — Business Traveller, TheStreet, the European trade press — are due out later this month and the early read I am hearing from colleagues already in the beta pool is that the tool makes factual errors. I made a list of three I caught before lunch. The factual-error tax on a direct-channel AI tool is a different conversation from the factual-error tax on, say, a generic chatbot, because the disclosure standard for a chain’s own surface is higher. I will get to that too.

The five scenarios, and what the Planner did with them

I do this with every new AI surface. Five scenarios that map to the real bookings I have made in the last twelve months — two business, two leisure, one edge case — and a notebook open beside the screen. The Planner’s interface is a conversational pane on hilton.com, accessible via a Plan with AI CTA on the search bar. It opens with a greeting and a soft prompt: Tell me about your trip. There is no rigid form. You type the way you would email a concierge.

Scenario one: Frankfurt, Tuesday-to-Thursday, one client dinner, walking distance to the Messe. I typed roughly that, in those words. The Planner came back in about four seconds with three properties — a Hilton Frankfurt, a Canopy by Hilton Frankfurt City Centre, and a Hampton by Hilton Frankfurt Airport — and asked whether dinner was casual or formal and whether I cared about gym hours. I said formal, gym hours mattered less. It narrowed to the Hilton Frankfurt and the Canopy, surfaced a paragraph on each comparing meeting-room availability and the walk to the Messe entrance, and offered to check availability on my dates. The Canopy paragraph mentioned a restaurant on the property that, when I cross-checked against the property page, exists. The walking time it cited (twelve minutes to Messe Halle 1) is within a minute of Google Maps. Pass.

Scenario two: a long weekend in Lisbon with my partner, late April, walkable to Alfama, breakfast included, under £350 a night. This is the leisure brief that hotel chains historically lose to Booking.com. The Planner asked one clarifying question — do you want a boutique feel or something larger and more amenity-rich? — and on the second pass surfaced two properties: a Tapestry Collection by Hilton in Príncipe Real and a Curio Collection property near the river. Both have breakfast packages on hilton.com that I verified. The walking time to Alfama from the Príncipe Real property is fifteen-to-twenty minutes; the Planner said about twenty, which is fair. The under-£350 constraint was respected on the dates it offered. Pass, with a caveat: the Planner did not surface the Conrad Algarve as a “no, but consider” option even though my brief implied I cared about food and ambience. A human concierge might have said not Lisbon, but if you flexed south for two of the four nights, here is the case for it. The Planner did not flex.

Scenario three: a one-night stay in Edinburgh on a Friday, brand-flexible, near Waverley, expense-account. This is the business-traveller solo-night brief that should be the easiest case for any chain AI. The Planner returned two properties — a Waldorf Astoria and a DoubleTree — with a clean comparison paragraph on each. The Waldorf Astoria description mentioned a Michelin-starred restaurant on-site. This is one of my factual errors. The restaurant in question lost its star in the 2024 Michelin Guide and has not regained it. The property still has the kitchen, the chef, and a strong dining offer, but it is not currently Michelin-starred. I asked the Planner directly: Is the restaurant Michelin-starred? It repeated the claim. I asked: As of which Michelin Guide? It said the most recent Michelin Guide UK. That is incorrect. I noted it.

Scenario four: a family stay in Orlando over Easter week, two adults, two kids ages eight and twelve, theme-park access, two queen beds. The Planner returned three properties — a Signia by Hilton, a Waldorf Astoria, and an Embassy Suites — and asked whether we wanted parks-only or a balance of parks and resort downtime. I said balance. It narrowed to the Signia and Embassy Suites, with paragraph-level case-making on each. The Embassy Suites paragraph cited a complimentary cooked-to-order breakfast (true, brand-wide standard) and a pool with a waterslide (true, property-specific). The Signia paragraph cited a kids’ club. I cross-checked: the property has a recreation programme but the kids’ club framing is generous. Soft miss, not a factual error. It is the kind of phrasing a human concierge would correct themselves on.

Scenario five (the edge case): “I want to do New Year’s Eve somewhere warm with my parents, but my mother uses a wheelchair and we want a connecting-rooms setup.” Two real constraints — accessibility and connecting rooms — that are notoriously poorly served by chain search interfaces. The Planner asked three good clarifying questions: distance from a major airport with direct flights from London; pool accessibility; dining accessibility. On the second pass it surfaced a Hilton in the Caribbean with explicit accessible-room inventory on the dates I gave it and a note that connecting rooms with one accessible unit were available. This was the strongest answer of the five. I called the property to verify. The reservations agent confirmed the inventory and added one detail the Planner had not — a beach wheelchair available on request. The Planner had been correct on the harder facts and silent on the softer ones.

Across the five scenarios: four passes with caveats, one factual error of the kind that would embarrass a hotel brand if a guest had relied on it to book a celebratory dinner. The factual-error rate is real, but it is also concentrated in the property attribute layer — Michelin stars, kids’ clubs, restaurant names — and not in the availability and price layer. That distinction matters for what the Planner is actually for.

The competitive table

Here is the scoring I would put on the table, knowing this is hands-on against the Planner this morning and second-hand against the rivals based on testing I have done over the last six months. The scale is one-to-five, with five being best in class.

Tool	Conversational quality	Portfolio narrowing	Factual reliability	Conversion intent	Direct-channel integration
Hilton AI Planner (March 10 beta)	4	5	3	5	5
Marriott RENAI	4	3	4	3	4
Hyatt ChatGPT app extension	5	2	3	2	2
IHG Smart Planner	3	3	4	4	4
Accor AI Travel Companion	3	2	3	3	3

The Hyatt extension scores highest on pure conversational quality because it inherits ChatGPT’s model — but that is also its weakness, because it lives inside ChatGPT and the path from chat to booking is a handoff, not a continuation. The Hilton Planner scores lowest of the top three on factual reliability for the reasons the Edinburgh scenario surfaced, but scores top on the two columns that matter for direct-channel economics: portfolio narrowing and conversion intent.

The deeper Marriott AI stack — the property-level rollouts we walked through in the case study — is broader than the Hilton tool by surface area but does not have an analogue to the Planner today. RENAI is a search-front agent, not a portfolio-narrowing concierge. That gap is what the Planner is built into.

Why this is a conversion tool, not an inspiration tool

I keep coming back to this distinction because it is the design choice that makes the Planner different. Inspiration tools want to widen the funnel. Conversion tools want to narrow it. The Hilton Planner, in five scenarios this morning, narrowed every brief to two or three properties within two prompts. It did not surface every plausible Hilton in a city. It did not show me the long tail. It picked, and it picked with a point of view.

That is a design choice. It is also a structural admission about the Hilton portfolio. Twenty-seven brands is a lot. Nine thousand properties is a lot more. The classical chain-search interface — pick a city, see a list, filter — is exhausting in a portfolio that large, and the OTA value proposition has always partly rested on the fact that Booking.com’s interface is less exhausting than the chain’s own. A conversational tool that says given what you told me, here are the two you should consider is structurally an answer to the OTA interface, not a copy of it.

I am told — and this is interpretation, not anything in the corporate release — that the Planner was built with hilton.com’s conversion data in the loop. That is, the system has been trained on which paragraph framings, which comparison structures, and which clarifying-question sequences correlate with bookings on the existing site. If true, that explains the narrow fast behaviour: the Planner is not optimising for the breadth of the answer, it is optimising for the click. That is a defensible thing for a chain’s own site to do. It is also a thing the OTA channels have been doing for fifteen years.

The narrowing behaviour is what makes the OTA-bypass case real. An inspiration tool sends a guest back out to the broader market to compare. A conversion tool keeps them in the chain’s surface, with the chain’s pricing, and the chain’s loyalty hooks. The Planner is the latter.

The OTA-bypass economics

This is the part of the analysis the corporate release does not put numbers on, and where the operator stake is largest. Hilton books a significant share of its room nights through OTAs — Booking, Expedia, the smaller European players — and the channel cost of those bookings is the commission structure, broadly fifteen to twenty per cent depending on programme, plus the parity-and-rate-shopping overhead.

Let me put a back-of-envelope on it. Assume Hilton’s system-wide gross room revenue is roughly $40 billion at current scale — this is a public-disclosure-adjacent estimate, not a Hilton figure, and the actual number is in the annual report. Assume OTA-channel share of that revenue is in the thirty-per-cent range, which is the broadly cited industry figure for major chains. That is $12 billion of room revenue running through OTAs. Assume an average channel cost of seventeen per cent. That is roughly $2 billion a year in commissions across the system.

A one-per-cent shift from OTA to direct — that is, one out of every hundred guests who would have booked via Booking instead books via the Planner on hilton.com — is, in this back-of-envelope, $120 million of channel revenue retained at the chain level annually. Spread across nine thousand properties, that is on the order of $13,000 a property a year, which sounds modest until you compound it across a multi-year shift and remember that the channel-cost retained is high-margin revenue that does not require additional rooms, additional labour, or additional inventory. It is pure margin recovery.

That is the corporate-level math. The franchisee-level math is more interesting because the franchise contract structure means the direct-channel benefit flows differently to owners than to brand. I will not pretend to know how the Hilton management contracts treat AI-Planner-attributable bookings, but my read — interpretation — is that this will become a negotiation point at the next contract cycle. Franchisees will want the channel-cost saving to flow through to their P&L. Brand will want the loyalty data, the cross-property attachment, and the platform investment recouped. Watch this space.

There is a second-order economic effect worth naming. Loyalty enrolment. The Planner, when I asked it to hold availability on the Edinburgh property, prompted me — politely, once — to sign in to Hilton Honors or create an account. That is a soft loyalty hook embedded in the conversation flow, and it is the kind of CRM capture that an OTA channel cannot replicate because the OTA owns the customer relationship, not the chain. Every booking the Planner converts is, by structural default, a Honors-enrolled or Honors-recognised booking. Multiply the one-per-cent shift case by the lifetime-value uplift of a loyalty-enrolled guest relative to an anonymous OTA booking and the math compounds again. I would not be surprised, by year-end, to see Hilton disclose a Planner-attributable Honors-enrolment number in its earnings call. It is the kind of metric that proves the platform investment.

The cautions

The Edinburgh Michelin-star error is the one I want to come back to, because it is the canary in the trust mine. The Planner asserted a fact about a restaurant’s award status that was incorrect. When I asked it to verify, it repeated the assertion. When I asked it to specify the source — as of which Michelin Guide — it confabulated a source.

That sequence is the classical generative-AI failure mode, and the fact that the Planner exhibits it on launch day is not a surprise. The forthcoming independent reviews — Business Traveller’s hands-on is due in the next week or so and I expect it to call out exactly this category of error — will sharpen the public picture, and I would expect Hilton to ship a guardrail revision before April. The narrow-and-narrow-fast design is a strength on conversion and a liability on factual richness: the tool is generating paragraph-level case-making about properties, and every adjective it generates is a factual claim that could be wrong.

The cautions that follow are operator-side, not consumer-side. If you are a Hilton property with a featured attribute the Planner could misstate — a restaurant award, a renovation status, a pool feature, a kids’ programme — call your brand-management contact and ask how property-attribute data flows into the Planner’s context. If the answer is we are not sure, that is an answer. If you are a Hilton competitor thinking about deploying a similar tool, the Planner’s factual-error profile is the operational lesson. The tool is impressive at conversion-narrowing and the property-attribute layer is where it bleeds.

There is a second caution, structural rather than factual. The Planner’s conversion-funnel orientation is, by design, narrow. That is good for direct-channel revenue. It is potentially bad for the long tail of properties that do not surface in the Planner’s first-two-prompt narrowing. A Tapestry Collection property in a smaller market that the Planner does not surface in any of the obvious scenarios will see less direct demand than its bigger-brand siblings in the same city. The chain-level economics work; the property-level distribution effects need to be watched.

The third caution is jurisdictional. The Planner is live on hilton.com today in beta in the U.S. The EU rollout has not been confirmed, and as colleagues in the newsroom have been tracking, the EU AI Code of Practice on labelling AI-generated content is in its second draft with the comment window open through March 30. Article 50 obligations land August 2. A conversational AI tool on hilton.com in the EU will need to identify itself as AI at the start of the interaction, in clear language, in the language of the user, with a logged disclosure. The Planner I tested this morning opens with a friendly greeting that does not say I am an AI. That will need to change before any EU launch.

The fourth caution is the search-engine-shift exposure. Hilton’s direct-channel revenue today depends, in significant part, on paid search and the SEO long tail. If the Planner becomes the primary entry surface for hilton.com — that is, if guests start their trip planning in the AI Planner rather than via a Google query — the chain’s paid-search spend ought to fall, which is good, but the dependency on the Planner’s own discoverability rises. Whether the Planner is itself surfaced inside ChatGPT, Perplexity, and Google’s own AI surfaces becomes the next-order question. That is a partnership-and-distribution conversation, not a product conversation, and it is the conversation I expect to see at the next earnings call.

The bet

The bet is that direct-channel AI is the next ten years of the OTA-versus-chain conflict, and that the chain that ships the first credible conversion-narrowing tool gets a structural advantage in the platform race. Hilton, today, is that chain. Marriott will respond. IHG will respond. Hyatt will probably extend the ChatGPT integration into the hyatt.com surface. Accor will iterate the Travel Companion. The next twelve months will be a competitive build-cycle on this surface and the chain that compounds the most user-prompt data the fastest will harden the lead.

The risk is the trust risk. A conversion tool that gets the Michelin star wrong on Friday is a tool that loses the booking on Friday. A tool that loses the booking on Friday and confabulates the source when challenged on Saturday is a tool that loses the brand on Saturday. The Planner has to fix the factual-error rate faster than its competitors can match the narrowing behaviour. That is a race against two clocks at once.

I have one more read, and this is the one I will return to in the Voice Agent Maturity Curve piece coming in May. The portfolio-narrowing thesis applies to text today and will apply to voice within twelve months. A guest who can have this conversation in chat will, before the end of 2026, expect to have it on the phone. The hotel chain that builds the conversational concierge first — and gets the factual-error rate down to the floor — is the chain that owns the next generation of the booking relationship.

Hilton went live this morning. Five scenarios, four passes and one fail, one factual error that will need fixing before the independent reviews land. The audit is in. The bet is on. I will run the scenarios again at the end of the month and see what has shifted.

— Naomi covers hotel F&B and operator tech for TableTransfers. Tips: [email protected].