How Hilton Shipped an Anthropic-Powered Travel Planner to 9,100 Hotels in 60 Days

I was sitting in a Hilton lobby in Midtown the morning the AI Planner went GA, drinking a coffee that cost slightly more than my self-respect and watching a woman two armchairs over try to plan a long weekend in Charleston on her laptop. She was on hilton.com. She typed a sentence into a chat box. She got back a list of properties, a rough itinerary, and — this is the part I want you to picture — she did not get up, walk to the front desk, and ask a human for help. She just kept typing.

That was March 17. I had flown in for a property visit and a coffee with a GM who has been an operator-client of mine for years, and the timing was an accident. Hilton had announced the AI Planner exactly one week earlier, on March 10, via stories.hilton.com. That release name-checked Michael Leidinger, Hilton’s SVP and Chief Information Officer, and described the tool in the kind of careful, undersold language that big-cap hospitality companies use when they don’t want to spook anyone: a “conversational travel planning experience” for hilton.com visitors. No model partner named. No adoption target. No “transformational” anywhere.

Six weeks later, on the Q1 earnings call on April 28, Chris Nassetta got specific. Asked about AI, he said — and I’m quoting from the Motley Fool transcript — “we did with Anthropic and Claude.” That sentence is the reason I’m writing this piece. Because what Hilton actually shipped between March 10 and March 17 is the most consequential hotel-AI deployment of Q1 2026, and almost nobody in our industry has correctly understood why.

It’s not the tool. The tool is fine. The tool is, frankly, a little boring. That is the entire point.

The contrarian read: unremarkable is the flex

Here is what I want operators reading this to internalize. In 2026, the differentiator at the top of the hotel pyramid is not what your AI can do. It’s how fast you can ship it, how cleanly it sits on your stack, and how many of them you can run in parallel without breaking your distribution math. Hilton’s AI Planner is unremarkable as a consumer experience. It is extraordinary as a deployment artifact — a thing that exists, on hilton.com, in front of guests, attached to a 9,100-property, 1.3-million-room, 143-country booking surface, six weeks after announcement and roughly two quarters after a credible “we are seriously building this” signal from leadership.

I have spent the last three years writing this column about operators trying to ship AI into hospitality. Most of what I’ve watched has been pilots — months-long beta programs scoped to a handful of properties, sometimes a single brand, occasionally a single market. (I wrote about Marriott’s deployment cadence in an earlier piece, and about the voice-agent timing curve here.) The Hilton AI Planner is the first time I have watched a top-three global brand take a guest-facing AI from press release to GA on the flagship booking surface inside a single quarter. The build-cadence is the news. The Claude integration is downstream of the build-cadence.

Let me reconstruct what actually happened, because the public story under-sells it.

The 60-day build the deck won’t show

The March 10 release is worth reading carefully, because it tells you what Hilton wanted you to notice and — by omission — what they didn’t. What they wanted you to notice: a conversational planner, embedded in hilton.com, that helps guests with destination ideas, itinerary suggestions, and property selection. Michael Leidinger’s quote in that release is the most operator-relevant piece of public text Hilton has produced about AI this year. It frames the planner as part of an ongoing program rather than a one-off product launch, and it positions the work as a team output rather than a single-leader bet. That framing matters. It’s the language of a company that intends to ship more of these.

What the release does not tell you: which model is under the hood, what the unit economics look like, how many guests had already used it in beta, or how the team got from internal demo to production in front of the entire hilton.com audience. We learned the model partner on April 28, when Nassetta confirmed Anthropic and Claude on the earnings call. This timing distinction is the most important reportorial point in this piece: if you read coverage that attributes the Anthropic partnership to the March 10 announcement, that coverage is wrong. The March 10 release does not name a model. The Anthropic attribution comes from the Q1 call six weeks later, which is itself a Hilton-controlled disclosure — a vendor name dropped onto an earnings transcript after the product had already been live for forty-two days.

If you’re keeping score, that’s a deliberate sequence: ship the product, watch it perform, then name the partner. Hilton did not need Anthropic’s logo to launch. They needed Anthropic’s logo to talk about the launch in the context of a 41-use-case AI portfolio. Which we’ll get to.

The build itself, as best as I can reconstruct it from conversations with three operators close to Hilton’s tech org and one former member of the IT leadership team, was scoped, engineered and tested inside what Leidinger’s organization calls the enterprise platforms group. The AI Planner sits on top of Hilton’s existing hilton.com search and property-detail infrastructure — it does not replace that surface, it overlays a conversational layer on it. From an engineering standpoint, that’s a meaningfully smaller surface area than a from-scratch agentic booking platform. It is, essentially, a retrieval-augmented chat experience pointed at the same inventory feed the legacy hilton.com search already consumes, with Claude doing the natural-language work on top.

That’s why 60 days was achievable. They did not rebuild the booking funnel. They wrapped it.

Why the model partner matters less than the cloud number

Here is the number from the April 28 call that I want every operator reading this to circle in red: roughly 90% of Hilton’s enterprise technology now runs on cloud, up from about 20% in 2020. That figure shows up in the earnings transcript and was independently flagged by Hotel Dive’s coverage. I’d argue it is more important than the Anthropic name-drop.

Why? Because the Anthropic partnership is a substitute good. If Claude weren’t the right fit, Hilton could have shipped on GPT-class models from OpenAI, or on Gemini from Google. Indeed, Nassetta name-checked all three vendors — Google, OpenAI, and Anthropic — as active partners on the same call. The model layer is, for a buyer of Hilton’s size, increasingly fungible. You pick the model that wins on the specific use case, and you re-pick next quarter if the leaderboard shifts.

The cloud migration is not fungible. It is the seven-year precondition that let the AI Planner ship in 60 days. In 2020, with 80% of the tech stack on-prem, the same internal team would have spent the build window arguing about VPC peering, data residency, and which legacy CRS system the planner needed to read from. In 2026, with 90% on cloud, the team spends the build window writing prompts and tuning retrieval. The unsexy infrastructure work that Leidinger and his predecessors have been grinding through since the start of the decade is what makes the sexy product velocity possible now.

Operators, this is the read: if your stack is not predominantly cloud, the AI roadmap you are writing for 2026 is not real. I don’t mean that as a hot take. I mean it as an engineering fact. The reason Hilton can run 41 AI use cases in parallel is that their data is reachable, their identity layer is unified, and their deployment pipeline can push to production without ticketing through three vendors. If you are a mid-scale operator with a hybrid stack and three property-management vendors, the cap on your AI velocity is not your AI budget. It’s your infrastructure.

This is also why I think the OpenTable piece I wrote last year is worth re-reading in this context. OpenTable’s AI integrations moved fast for the same reason: a clean, hosted data layer. The lesson is consistent across the sector.

Nassetta’s quiet quote on rate control

Let me pull a quote from the April 28 call that has been almost entirely under-covered by the trade press. Asked about agentic booking platforms and the threat of AI agents intermediating the guest relationship, Nassetta said: “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.”

Read that twice. Then read it a third time.

That sentence is not a comment about the AI Planner. It is a comment about why the AI Planner exists. The strategic backdrop for every AI deployment Hilton ships this year is the same backdrop Marriott, IHG, Accor, and Hyatt all face: a generation of AI-native travel agents — call them OpenAI’s ChatGPT shopping experiences, call them Google’s AI Overviews with booking intent, call them whatever Booking.com and Expedia ship next — are getting between the guest and the brand. The hotel companies that own significant chunks of US lodging supply have one structural advantage in that fight, and it is exactly what Nassetta named: they control rate, inventory, and availability. They get to decide whether the AI agent on the other end of the conversation sees a bookable room or a “check the hotel website” deflection.

The AI Planner is Hilton’s bet that you can keep the guest on hilton.com if the experience on hilton.com is good enough. It is, in that sense, a defensive product. It is not trying to be the best travel agent on the internet. It is trying to be a sufficiently good travel agent that a Hilton Honors member doesn’t tab over to ChatGPT to plan the same trip.

This framing also explains the Skift detail that the company is shipping into ChatGPT as well — a parallel surface where Hilton appears inside OpenAI’s agentic shopping experience. The two efforts are not contradictory. They are the same strategy from opposite ends. If the guest wants to start on hilton.com, the AI Planner meets them there. If the guest starts on ChatGPT, Hilton wants to be a first-class citizen in that flow too. Either way: Hilton’s inventory, Hilton’s rate, Hilton’s relationship.

The companies that don’t own rate-and-inventory control at scale will not get to play this game on both surfaces. That is what the quiet quote is really saying.

The 41-use-case portfolio, decoded

Skift and Phocuswire both surfaced the 41-use-case number from the call. Phocuswire’s writeup is the cleanest summary if you want the corporate framing; Skift’s piece is more skeptical and worth reading for the reporter’s-eye view.

Here is what 41 use cases actually means, decoded for operators.

It does not mean 41 products. It means 41 distinct AI workflows in test or production across the company — some guest-facing, most internal. From the transcript and the trade coverage, you can reasonably bucket them into four categories. Guest-facing search and planning: the AI Planner itself, the ChatGPT integration, and likely a Honors-app conversational layer. Operations and back-of-house: housekeeping optimization, demand forecasting, energy and labor scheduling at scale. Commercial: revenue management augmentation, group-sales enablement, marketing personalization. Corporate: internal copilots for finance, legal, and HR functions that don’t touch guests at all.

Most of the 41 are not the AI Planner. Most of the 41 are unsexy. That is, again, the point. A 9,100-property operator with a 527,000-room pipeline running across 143 countries does not win on a single hero product. It wins on portfolio velocity. If three of the 41 use cases each move RevPAR by twenty basis points, and five of them each move labor cost by a percentage point, you have moved Adjusted EBITDA — which Hilton reported at $901 million on the Q1 print, on RevPAR up 3.6% — by a number that matters. None of those individual use cases will get a press release. The portfolio gets the press release.

This is the model. This is the playbook. And it is only available to operators who have already done the cloud migration and the data unification. Everyone else is shipping one thing at a time and calling it a strategy.

What an independent or franchisee can copy

I get this question constantly, so let me handle it head-on. If you are reading this as an independent operator, or as a franchisee with one or two properties under a major flag, or as a small-portfolio group, what can you actually take from Hilton’s playbook?

You cannot copy the portfolio. You do not have 41 use cases worth of internal demand, and you do not have the engineering bench to ship them. But you can copy the sequencing, which is what actually matters.

The sequencing, as I read it: pick one guest-facing use case where you control the surface, ship it on a hosted retrieval-plus-LLM stack rather than rebuilding anything underneath, name the model partner only after the product is performing, and bundle the launch into a broader portfolio narrative on your next investor or owner update. That sequence is operationally accessible to a much smaller operator than Hilton. It does not require 9,100 properties. It requires one. The surface, for most independents, is the property website or the direct-booking widget. The retrieval stack is increasingly available as a managed service — and you do not need to commit to a model vendor on day one.

Two cautions. First, do not skip the cloud question. If your PMS, CRS, and CRM are not addressable from a single integration layer, your AI product will be brittle and you will spend more on engineering than on the model. Second, do not over-promise on adoption. Hilton has not disclosed active-user numbers for the AI Planner, and there is a reason: it’s early, the launch was March 17, and the company is sensibly letting the metric mature before talking about it. If a 9,100-property operator with a CIO-level sponsor is not yet publishing adoption data, you should not be promising adoption data to your owners. Ship the product. Measure it quietly. Talk about it when you have something honest to say.

The thing I would borrow most directly is the patience. The March 10 release did not over-promise. The April 28 framing did not over-promise. Hilton has been remarkably disciplined about under-stating what is, in fact, a real shipping cadence. That discipline is copyable, and it is more valuable than any specific technical pattern in the build.

Where Marriott will land differently

The Marriott Q1 print lands on May 6, the day after this piece publishes, and I want to be honest about what we are and are not going to learn from it.

We will learn how Marriott frames its AI portfolio relative to Hilton’s. We will probably get some version of a vendor-partner story — Marriott has been investing on this front and has its own set of named relationships. What we will likely not get is a shipped, guest-facing equivalent of the AI Planner on the marriott.com flagship surface, because Marriott’s deployment cadence on consumer-facing AI has been measurably more cautious. (I dug into the structural reasons in my earlier piece.) Marriott runs a meaningfully more complex brand portfolio than Hilton — more sub-brands, more loyalty surface, more international franchise variance — which means the same AI Planner concept, ported to marriott.com, is a harder engineering and brand problem. Not impossible. Harder.

The interesting tell on the May 6 call will not be whether Marriott names model partners. They will. The interesting tell will be whether Marriott can credibly cite a guest-facing AI surface that is in GA — not in beta, not in pilot, in GA — on marriott.com. If they can, the gap is closing. If they cannot, the read is that Hilton’s cloud-migration head start has converted into a durable shipping-velocity lead, and the rest of the cycle in 2026 is Marriott playing catch-up on the consumer-facing layer while compensating with internal-portfolio breadth.

I am genuinely undecided on which way it lands. I’ll write a follow-up next week.

Operator takeaways

A short list, the kind I send my operator-clients on Monday mornings:

• Cloud share is the leading indicator, not model choice. If you are below 70% cloud on your enterprise stack, your AI roadmap is fictional until that number moves. Hilton went from ~20% to ~90% in roughly six years. That is the timeline.
• Pick one surface, ship one product, name the partner later. Hilton announced March 10, went GA March 17, named Anthropic on April 28. That sequence is copyable at any scale.
• Do not promise adoption metrics you cannot defend. Hilton has not disclosed active users for the AI Planner. Neither should you, on day one, for whatever you ship.
• Defensive AI is real AI. The AI Planner exists to keep the guest on hilton.com, not to dazzle. A “boring” guest-facing AI that protects your direct channel is more valuable than a “transformational” one that ships in 18 months instead of 6.
• Read the rate-and-inventory quote in full. Nassetta’s “we are the only ones with that 25% of the market” line is the strategic context for every consumer-facing AI bet Hilton makes this year. If you do not own meaningful inventory, your defensive options are narrower and your offensive options are different. Plan accordingly.

What I’ll be watching

A few open threads as we move through Q2.

The first is whether Hilton starts publishing any kind of AI Planner usage data — even a soft signal like “millions of conversations” or “X% of hilton.com sessions.” Right now there is nothing public. The longer that silence holds, the more I’ll suspect adoption is below internal targets. The shorter it holds — say, a mention on the Q2 call — the more I’ll read it as a confidence signal.

The second is whether the ChatGPT integration story develops into a real distribution channel or stays a press-release artifact. Skift’s coverage was appropriately cautious on this. Agentic-booking-through-LLM is still an early surface; the unit economics for hotel companies remain unclear; and the question of whether OpenAI and Anthropic eventually compete with hotel brands for the guest relationship is unresolved.

The third is the Marriott print on May 6, which we will all read together.

And the fourth — the one I keep coming back to — is whether the 41-use-case portfolio model becomes the industry standard reporting unit for hospitality AI. Right now we report on hotel AI launch-by-launch. Hilton just reported on theirs portfolio-by-portfolio. If Marriott, IHG, and Accor follow suit on their next calls, the analyst frame for the sector shifts. That would be a small thing that ends up mattering a lot.

The woman in the lobby on March 17 is the part of this story I keep thinking about, though. She didn’t know she was using Claude. She didn’t know it was day one of GA. She didn’t know she was inside a 41-use-case portfolio narrative that would be reframed seven weeks later on an earnings call. She just typed a sentence and got a useful answer, and she stayed on hilton.com.

That’s the whole game, in 2026. The operators who can deliver that experience, on their own surface, on their own inventory, at speed — those are the ones who keep the guest. Hilton just demonstrated they can. The rest of the industry has until roughly the next print to demonstrate the same.

— Priya files The Operator. Tips: [email protected].