How Hilton shipped an Anthropic-powered travel planner to 9,100 hotels in 60 days

I have been sitting with the Hilton AI Planner build for a week now, trying to reconstruct it from the outside the way an operator would have to — press release on one screen, earnings transcript on another, a notebook of dates and named partners on the desk between them. I started Monday morning with the March 10 stories.hilton.com release on the left of the screen and the April 28 Q1 earnings call transcript on the right, and I worked backwards. The Planner was announced March 10. It was opened to all hilton.com visitors on March 17. By April 28, Chris Nassetta was on the earnings call naming Anthropic by name as the model partner behind it, sitting it inside a 41-use-case AI portfolio that also names Google and OpenAI, and framing the whole effort as part of a defense of what he called the only ones with that 25% of the market that can control rate inventory availability. The release was the news. The call was the strategic context. The 60 days between them are the case study.

The contrarian thesis — and this is my read, not Hilton’s framing — is that the Planner is not a customer-experience feature. It is a defense against agentic booking platforms taking the guest relationship. Every major hotel chain is staring at the same problem: a generation of AI-native search surfaces is being trained right now to negotiate travel on a customer’s behalf, and the chains that do not put their own conversational layer between the guest and the agent will end the decade being intermediated by something that is not Booking and not Expedia and looks more like a horizontal AI assistant with a payment instrument attached. Hilton went first. Hilton built the surface, picked Anthropic as the model partner, shipped beta-to-GA inside one quarter, and used the earnings stage to mark the territory. That is the build I want to walk through.

The interpretation work follows. The build details follow. I will get to all of them.

The 60-day timeline, in order

Let me put the dates in the order they happened, because the compression is the point. On March 10, Hilton issued a stories.hilton.com release introducing the AI Planner — a generative-AI digital concierge, conversational interface, twenty-seven brands and what the company described at the time as more than 9,000 properties piped into a single chat surface. The release quoted Michael Leidinger, Senior Vice President and Chief Information Officer, on the build philosophy. The Planner went into beta on hilton.com that morning, accessible behind a Plan with AI call-to-action on the search bar. I ran the launch-day audit on March 10 — five real travel scenarios, four against the Planner and one against the field — and the read I filed at the time was that this was the first big-brand AI tool that exists to narrow a portfolio, not inspire a trip.

A week later, on March 17, Hilton flipped the Planner from beta-cohort to all hilton.com visitors. The flip was quiet — there was no second press release, just a removal of the beta gating on the search page — and it is the date that does the most explanatory work in this story. Most enterprise AI deployments at this scale move from beta to general availability over a six-to-nine-month arc. Hilton did it in seven days. That is either a very confident product team or a very confident strategy team. By the April 28 earnings call, Nassetta would frame it as the latter.

In the six weeks between March 17 and April 28, the Planner ran live on hilton.com as a default surface for any visitor who clicked into the Plan with AI prompt. The first independent reviews landed in late March. The European trade press picked it up in early April. A small chorus of factual-error reports surfaced and got picked up by the consumer travel desks — I caught three in my own audit, the Edinburgh Michelin-star claim being the one I held onto longest — and Hilton’s response in conversation with operators was the same response any frontier-model-powered surface has to give: the model is improving, the guardrails are improving, the deployment is being iterated.

Then April 28. The Q1 print landed at 5:00 AM Pacific. RevPAR up 3.6% currency-neutral, adjusted EBITDA $901 million, full-year EBITDA guide raised. My Q1 read filed that morning captured the headline numbers. But the line that sat heaviest after the call was the AI section of the prepared remarks. Nassetta spent roughly four minutes on what he called the agentic-search and discovery layer, named the Planner explicitly, and in plain language attributed the model layer to Anthropic. The phrase he used, transcribed in the Motley Fool earnings transcript, was what we did with Anthropic and Claude. He named Google as a search-distribution partner and OpenAI as an integration partner. He put the total AI portfolio at 41 use cases in test across the enterprise.

That is the 60-day arc. Press release, beta-to-GA in seven days, six weeks of live operation, earnings stage to set the strategic frame. The build is the story.

The build: Leidinger’s team and what 90% cloud actually enables

The build question is where I keep getting stuck, because the public surface area is thin. Hilton has not published a technical post-mortem. Michael Leidinger’s stories.hilton.com quote is the single attributable piece of internal sourcing on the architectural decisions. The earnings call adds the model-partner name and the use-case count. The rest is inference from what is on the page, from the timing, and from what we know about Hilton’s tech-stack history.

Here is what is inferable. The Planner sits inside hilton.com — a property that runs on Hilton’s central reservation system, the brand’s loyalty and identity layer, and the rate and inventory feed that Nassetta on the April 28 call called the only ones with that 25% of the market that can control rate inventory availability, period, end of story, nobody can get it, unless we give it to them. The integration surface for the Planner has to talk to all three of those systems at conversational latency. The Planner answers questions about specific properties on specific dates with specific rates. That means it is not a brochure-style RAG-over-marketing-content build. It is wired into the live booking stack.

That is the piece that makes the 60-day timeline credible. The PhocusWire Q1 distribution write-up noted that Hilton reported approximately 90% of its enterprise technology footprint now sits on cloud infrastructure, against roughly 20% in 2020. That five-year cloud migration is the precondition. You cannot ship a conversational layer that reaches across the central reservation system, the loyalty platform, and the property-level rate feed in a single chat turn if the underlying systems are still living in mainframe or on-premise architectures that require batch updates and overnight reconciliation. Hilton spent 2020-2025 doing the unglamorous work of getting the booking stack into a place where a frontier-model surface could front it without falling over. The Planner is what 90% cloud lets you ship.

Leidinger’s team — the CIO function, working with what Hilton on the April 28 call described as cross-functional product, brand, and commercial groups — appears to have done the integration work in roughly two stages. Stage one, before the March 10 announcement, was the wiring: connecting Claude into the reservation, loyalty, and rate layers, designing the prompt scaffolding to keep the conversation inside the Hilton portfolio, and building the safety surface on top. Stage two, between March 10 and March 17, was the beta validation: opening the Planner to a cohort, watching the failure modes, and deciding whether the error rate was inside the band that justifies general availability.

The seven-day beta is the part that operators in other chains have been asking me about. The conventional answer would be that seven days is not enough beta time for a system that touches rate and inventory at scale. The Hilton answer, by demonstration, is that it is enough if you scoped the beta the right way and if you have the cloud-native instrumentation to watch the failure modes in real time rather than waiting for support tickets to aggregate. Whether the bet pays out is a question for Q2 and Q3. The bet itself is the case study.

The 41-use-case portfolio: where each partner sits

Nassetta’s call did one thing that no chain CEO had done before this cycle: he put a count on the internal AI portfolio. Forty-one use cases in test. The figure is unverifiable from outside, but the act of stating it is itself the news, because it tells the analyst desks and the operator community that Hilton wants to be read as a portfolio of AI bets rather than a single AI product. The Skift write-up of the same earnings call walked through several of the named use cases. The Hotel Dive piece on Hilton’s vendor partnerships from April 29 framed the portfolio as a multi-vendor strategy by design rather than a single-vendor commitment.

The way the named partners fit, as best I can reconstruct it from the public sourcing:

Anthropic and Claude sit at the customer-facing booking surface. That is the Planner. It is the use case where the model has to handle ambiguous natural-language briefs, hold a conversation across multiple turns, narrow a 9,100-property portfolio to a recommendation, and do so without hallucinating property facts. Anthropic’s framing of safety and steerability lines up with the kind of guardrail surface a hotel chain needs when the conversation has to terminate in a real reservation against a live rate. Interpretation: Hilton chose Claude for the Planner because the conversion-funnel use case rewards refusal to confabulate more than it rewards breadth of capability, and that is the regime Anthropic has been pricing into.

Google sits at the search-distribution layer. Nassetta named Google on the call as a search partner, which in operator language reads as the surface where Hilton properties show up when a traveller queries Google for a destination. The integration there is not a chat agent; it is the data layer that feeds Google’s travel surfaces — Google Hotels, Google Flights, the AI Overviews that surface above the blue links. The Google relationship is about distribution density on the surface that still owns the majority of travel-research traffic.

OpenAI and ChatGPT sit at the integration layer. The Skift piece covered this most directly — Hilton has been working on a presence inside the ChatGPT app environment, which is the agentic surface where travellers will, increasingly, do research and increasingly attempt to transact. That is the agentic booking platform threat in its most concrete form. Hilton’s integration there is about making sure that when a ChatGPT user asks the agent to find a hotel, Hilton properties surface with rate and inventory accuracy and the conversation can be handed off to a Hilton booking flow rather than to an OTA.

The three partners do three different jobs. Anthropic handles the conversation when the guest is already inside hilton.com. Google handles the visibility when the guest is searching upstream. OpenAI handles the handoff when the guest is inside someone else’s agent. The 41 use cases sit across that surface area, and the partner choice in each case appears to reflect the model strength against the specific job.

The other thirty-eight use cases are not enumerated publicly. From operator-side conversations and the named pieces in the prepared remarks, the categories appear to include: revenue management and dynamic pricing assistance, front-desk and concierge agent assist (the back-of-house counterpart to what I have been tracking in the voice-agent maturity work), housekeeping scheduling optimisation, F&B forecasting and procurement, and corporate-functions productivity (legal, finance, HR). The portfolio framing is the news. The Planner is the customer-facing case study inside it.

Nassetta’s “25% of the market” quote: the strategic frame

The quote is the one I keep returning to, because it is the line that explains why the 60-day arc happened on the schedule it happened. We are the only ones with that 25% of the market that can control rate inventory availability, period, end of story, nobody can get it, unless we give it to them. I have the full sentence written in the notebook in two colours of ink because I went back to the transcript twice to make sure I had it right. The shape of the sentence is the shape of the strategy.

What Nassetta is saying, in operator translation: Hilton (together with the other chain operators of similar scale) owns the rate and inventory at roughly a quarter of the U.S. hotel market. No agentic booking platform — no OpenAI, no Anthropic, no Google, no Perplexity, no future entrant — can transact a Hilton booking without going through Hilton’s distribution layer. The chain’s leverage in the AI distribution conversation is structural, not negotiated. The agents need the rate. Hilton owns the rate. The Planner, and the partner portfolio behind it, is how Hilton converts that structural leverage into a defensible position before the agentic-booking layer is fully built.

This is the contrarian read in concrete form. The Planner is not a customer-experience feature. It is a way of saying to the agentic-booking ecosystem: we will integrate with you on our terms, we will give you the rate feed and the booking flow you need, but we will do so via a surface that we control, with a model partner we picked, with a conversational layer that keeps the guest relationship inside the Hilton loyalty graph rather than handing it to your wallet. The Planner is the Hilton-controlled surface where that handshake happens.

The 25% framing also explains why Nassetta named partners explicitly rather than keeping them in the we are working with leading AI providers abstraction. Naming Anthropic, Google, and OpenAI on the earnings stage is a way of telling each of those partners that Hilton’s relationship is portfolio-style and not exclusive — which is the negotiating posture that maximises Hilton’s leverage. It also tells the buy side that the chain is not betting the booking surface on a single frontier model. The hedge is itself a strategic asset.

The 90% cloud-migration enabler

I have not seen enough operator coverage of the cloud-migration line, and it is the piece I want to mark for the record. Nassetta and the Hilton finance team referenced on the April 28 call that approximately 90% of enterprise technology now sits on cloud infrastructure, against approximately 20% in 2020. That is a five-year transformation, executed quietly through the post-pandemic recovery period, and it is the silent precondition for the AI Planner build.

The shape of what 90% cloud enables: real-time integration between systems that used to talk to each other in batches. Identity-and-loyalty data joined to rate-and-inventory data joined to property-level operational data, available to a conversational surface inside a single API call rather than an overnight ETL pipeline. The Planner can answer a question about whether the Hilton Frankfurt has rooms on a specific Tuesday under a specific rate plan because the cloud-native rate feed is queryable at chat-turn latency. That is not possible on the architecture Hilton had in 2020.

The operator takeaway here is the one that does not get enough column-inches: AI deployment at this scale is a tax on infrastructure debt. The chains that did the cloud migration through 2020-2025 can ship conversational layers in months. The chains that are still on mixed on-premise and legacy architectures will ship the same products in years, if at all. The Planner is not a software feature; it is the visible top layer of a five-year infrastructure investment. Operators evaluating their own AI roadmaps should read the 60-day timeline as a function of the 60-month migration that preceded it, not as a benchmark for what any chain can do on the current state of its stack.

Comparison to Marriott and Hyatt

The comparison set is forthcoming, in the sense that the field is moving even as I file this. Marriott’s Q1 print is forthcoming on May 6, and the read I am hearing from operator conversations in the back half of last week is that Marriott will use the call to disclose natural-language search capabilities — a different positioning than Hilton’s full conversational concierge surface, and one that suggests Marriott has chosen to put the AI layer one step higher in the booking funnel rather than at the conversational concierge layer. That is a defensible product decision and it lines up with Marriott’s RENAI work to date.

Hyatt printed Q1 last Thursday at +5.4% RevPAR, and the AI framing on that call was lighter than Hilton’s — a Wedding Guide product on the consumer side, the System of Hyatt rollout on the operational side, no named frontier-model partner on the prepared remarks. The cross-chain read I filed Thursday sat the three prints together: branded-bundle hotels compounding on the EBITDA line, Hilton having staked the agentic-discovery layer ahead of the field, Hyatt running a steadier and less-ambitious AI playbook. The Big 5 hotel AI scorecard is the comparison piece operators should read alongside this one — it tracked the relative position of each chain’s AI surface as of March 10, and the Hilton-Marriott-Hyatt ordering at that point was Hilton in the lead, Marriott close behind, Hyatt and IHG a step back, Accor running the Travel Companion brochure surface.

The forthcoming Marriott disclosure on May 6 will reset the comparison. If Marriott names a frontier-model partner on the call — and the operator-side conversations suggest a Microsoft-and-OpenAI orientation rather than Hilton’s Anthropic-centred one — the field will start to look like a two-vendor split at the top of the chain hierarchy, with each chain’s model partner choice reflecting a different read on what the conversational booking surface should do. Hilton bet on refusal-and-precision via Claude. If Marriott bets on breadth-and-search via GPT, the next two quarters will give us a clean comparison on which model regime wins the booking-conversion use case.

The bet for the next 12 months

Here is where I would put the markers, on a calendar I want to come back to. First, the next twelve months are about who owns the diner and guest relationship before agentic booking platforms arrive at scale. Hilton has staked the conversational concierge surface. Marriott is forthcoming. Hyatt has chosen a narrower path. The agentic booking platforms — OpenAI’s ChatGPT travel surfaces, Anthropic’s own consumer-facing efforts, Google’s AI Overviews, Perplexity, the smaller entrants — are still building. The window for chains to plant their own conversational surfaces inside their owned channels is open now and will probably begin closing inside twelve to eighteen months.

Second, the model-partner choice will get re-evaluated. Hilton picked Anthropic for the customer-facing Planner. If the use case proves out — measurable Planner-attributable bookings, measurable Honors enrollment lift, measurable rate-mix improvement — Anthropic will have a reference architecture inside one of the largest hotel chains in the world. If the factual-error rate stays in the band the consumer-travel desks have been flagging, the model layer will get swapped, and the swap will be the news. The 41-use-case portfolio framing gives Hilton the optionality to rotate vendors without rewriting the strategic story. That optionality is itself a product decision.

Third, the structural leverage Nassetta named on the call — the 25% rate-inventory control — is a finite asset. The agentic booking platforms can route around it if independent hotels and short-term-rental platforms aggregate enough alternative inventory. The chains have a window during which their distribution monopoly is binding, and the window is the runway on which the Planner has to convert. The 60-day timeline starts to make more sense when you read it as urgency about the runway rather than confidence about the product.

The Planner is the case study. The 60 days are the case study. The 25% quote is the strategic frame. The 90% cloud number is the precondition. The 41 use cases are the portfolio. The Anthropic name is the bet. Hilton went first. That is the news under the news.

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