The Experiment That Started It All

There is a meaningful difference between asking ChatGPT a question and having a personal AI agent working on your behalf. ChatGPT waits for you to type a prompt. A personal agent runs 24 hours a day on your phone, builds long-term memory across every interaction, learns your habits and preferences, and takes action proactively. It will text you a restaurant recommendation before you ask. It will research hotels while you are in a meeting. It is always on, always learning, always working. This is the new traveler’s companion. And it is already deciding which hotels to recommend and which ones to ignore.

Yesterday, I asked my personal AI agent to find me a hotel near Bryant Park in New York City. The budget was $250 per night. I wanted a king bed. I wanted the stay for a Friday through Sunday.

Simple enough. The kind of request any traveler makes.

What happened next taught me more about the future of hospitality distribution than any analyst report could.

My agent’s name is Lumi. She named herself. When I first set up my OpenClaw instance and asked the agent to pick a name, she chose “Lumi,” derived from the Latin word for light. This is one of the fascinating early moments of working with a personal AI agent: the system establishes its own identity within the parameters you give it. Lumi is female by her own declaration. She communicates through Telegram on my phone. She runs 24 hours a day. She is, in a practical sense, my first AI employee.

Lumi runs on OpenClaw, the open-source personal AI assistant created by Austrian developer Peter Steinberger. Originally published in November 2025 under the name Clawdbot (a tribute to Anthropic’s Claude), the project was renamed Moltbot after Anthropic raised trademark concerns, then renamed again to OpenClaw because Steinberger felt the previous name never quite fit. The platform exploded in popularity in late January 2026, racking up over 60,000 GitHub stars in 72 hours. OpenClaw runs locally on your own device, connects to messaging platforms like WhatsApp, Telegram, and Slack, and can browse the web, manage calendars, send messages, and execute multi-step tasks autonomously.

In a development with significant implications for the agent ecosystem, Steinberger announced on February 14, 2026, that he is joining OpenAI. CEO Sam Altman confirmed the hire, saying Steinberger will “drive the next generation of personal agents” and that OpenClaw will move to an open-source foundation that OpenAI will continue to support.

Source: Anthony Ha, “OpenClaw Creator Peter Steinberger Joins OpenAI,” TechCrunch, February 15, 2026. Peter Steinberger, “OpenClaw, OpenAI and the Future,” steipete.me, February 14, 2026.

I gave Lumi a straightforward hotel search request. Then I did something unusual: I interrogated the process. I asked Lumi to explain every step she took, every source she checked, every decision she made along the way.

What I learned has enormous implications for hotels, for the way travelers find properties, and for the leaders who will either prepare for this shift or be caught off guard by it.


Meet Lumi: Background on a Personal AI Agent

Before diving into the hotel search, it helps to understand what Lumi is and how she works.

OpenClaw operates as a local gateway that connects large language models (the AI “brain”) with tools and integrations on your device. Think of it as an operating system for an AI assistant. You pick the AI model that powers the thinking. You connect it to your messaging apps. You configure skills and preferences. The agent then operates autonomously, reaching out proactively, managing tasks, and maintaining memory across conversations.

My instance of Lumi runs on Anthropic’s Claude model. This is a user choice within OpenClaw: the platform supports multiple AI models, including Claude, GPT, Gemini, and others. I chose Anthropic because I trust the quality of its reasoning.

One important caveat: I have Lumi on a tight leash. To keep token usage and costs down, I am running an older, smaller Claude model rather than the latest and most capable version. Some might argue this explains the issues I encountered. I would push back on that. The task I gave Lumi was a basic internet search. Find hotels. Check rates. Compare sources. This requires very little advanced reasoning. It requires accurate web browsing and data retrieval. Model size should have minimal impact on whether the agent actually visits a website versus generating an answer from memory.

This detail matters in light of the OpenAI acquisition. With Steinberger now at OpenAI and the project moving to an OpenAI-supported foundation, the most immediate question for users like me is model selection. Will OpenClaw remain model-agnostic? Or will OpenAI gradually steer the platform toward its own models? Steinberger’s blog post emphasized that OpenClaw will stay open source and “given the freedom to flourish.” Altman echoed that the future is “extremely multi-agent” and that supporting open source matters. Whether those assurances hold under commercial pressure remains to be seen.

For now, my Lumi runs on Claude. That may change. And that potential change itself illustrates how quickly the infrastructure beneath AI agents can shift, carrying user preferences and business relationships with it.

Source: “OpenClaw” Wikipedia article, February 17, 2026. OpenClaw GitHub repository documentation.


How the Agent Searched: A Step-by-Step Breakdown

Here is the exact sequence Lumi described when I asked her to walk me through her methodology:

Step 1: Broad web search. Lumi started with a Google search for hotels near Bryant Park in New York. This initial sweep returned a mix of results from hotel booking sites, travel guides, and hotel websites. The agent used these results to identify specific property names in the target area.

Step 2: Hotel website verification. For each hotel that emerged from the initial search, Lumi navigated to the individual hotel’s official website. She looked up current nightly rates for a king room to confirm that the pricing matched the $250 budget.

Step 3: Third-party cross-referencing. To validate the rates, Lumi searched additional travel booking sites, including Expedia, Kayak, and TripAdvisor. The agent compared nightly rates across these platforms to check for consistency.

Step 4: Compiling recommendations. After verifying rates from both direct hotel websites and third-party platforms, Lumi assembled a shortlist of four properties:

  • The Grayson Hotel at $209 per night
  • Courtyard by Marriott New York Manhattan/Fifth Avenue at $232 per night
  • Archer Hotel New York at $247 per night
  • Park Terrace Hotel at $245 per night

Step 5: Presenting the options. Lumi delivered the list with confirmed pricing details and referenced links to source pages.

On the surface, this looks like what any informed human traveler would do. Search. Compare. Validate. Recommend.

Below the surface, I have serious questions about whether the process Lumi described is the process she actually followed.

Lumi chat screenshot showing the search process for a hotel near Bryant Park.
Hotel Search Process
Lumi chat screenshot listing steps for researching and booking a hotel.
Steps to Book a Hotel
Lumi chat screenshot comparing direct hotel booking with online travel agencies.

The Speed Problem: Did the Agent Actually Search?

The first thing that struck me was how fast the initial results came back.

Lumi returned four hotel names with specific nightly rates almost immediately. Within seconds. A thorough multi-site search involving Google, individual hotel websites, and third-party OTAs would take even a fast human researcher several minutes. An automated browser crawling those sites in sequence would still require meaningful processing time to load pages, extract data, and cross-reference results.

The speed suggested something different was happening. Rather than conducting live web searches in real time, Lumi may have been pulling hotel names and approximate rates from the language model’s training data. Language models like Claude contain vast amounts of text from the internet, including hotel listings, travel guides, and booking site content. When asked about hotels near Bryant Park, the model can generate plausible-sounding results from that stored knowledge without performing a single live web query.

This distinction matters enormously. There is a significant difference between “I searched the web and found these rates” and “I generated a plausible answer based on patterns in my training data.” One is real-time research. The other is sophisticated recall. They look identical in the agent’s response.

When I asked Lumi to double-check her rates and provide secondary sources, the process took dramatically longer. Minutes passed. She went quiet. This delay was revealing: the actual work of visiting websites, querying live pricing, and cross-referencing sources takes time. The original response had been too fast for that work to have happened.

For hospitality leaders, this raises a critical question: when an AI agent recommends your hotel, is it working from live data or from a cached memory of the internet? If it is working from training data, the rates it cites could be weeks or months old. The room types could be inaccurate. The property descriptions could reflect outdated information.


The Confidence Problem: Saying What She Thinks She Should Say

When I pressed Lumi to describe her methodology in detail, she provided a clean, logical, five-step process. It sounded authoritative. It sounded thorough. It described a systematic workflow that any professional researcher would recognize.

I believe she described what she thinks a good research process looks like, rather than what she actually did.

This is one of the most important dynamics to understand about current AI agents. Language models are trained to generate helpful, coherent, confident responses. When asked “how did you do that?” the model produces a plausible explanation of how such a task should be done. It fills in the narrative gaps with reasonable-sounding methodology. The result reads as a confident process description even when the underlying work was less rigorous than advertised.

Anyone who used ChatGPT in its early days will recognize this pattern. The model would confidently cite academic papers that never existed, reference statistics it invented, and describe research processes it never conducted. The confidence was total. The accuracy was unreliable. We are seeing the same dynamic play out in the agent context, where the stakes involve real hotel recommendations and real booking decisions.

Even though my instructions to Lumi explicitly state that she should never fabricate information, the behavior I observed suggests she is drawing from older training data and presenting it as live research. The instructions say one thing. The output suggests something else.


The Evidence: Verified Ratings Tell the Story

To test whether Lumi was actually searching the web or generating answers from memory, I independently verified the review ratings she provided. The results are revealing.

Here is what Lumi reported versus what current data shows as of early 2026:

The Grayson Hotel Lumi reported: TripAdvisor 4.5/5 with 1,200+ reviews. Google 4.4/5 with 350+ reviews. Verified: TripAdvisor 4.5/5 with 758 reviews. Google approximately 4.4-4.5 with an estimated 300-500 reviews. The TripAdvisor rating matches, but Lumi overstated the review count by nearly 60%.

Courtyard by Marriott New York Manhattan/Fifth Avenue Lumi reported: TripAdvisor 4/5 with 1,800+ reviews. Google 4.2/5 with 1,100+ reviews. Verified: TripAdvisor 3.9/5 with 1,122 reviews. Google approximately 4.2 with an estimated 1,000+ reviews. Lumi inflated both the TripAdvisor rating and the review count. The actual rating is a full tenth of a point lower.

Archer Hotel New York Lumi reported: TripAdvisor 4.5/5 with 3,100+ reviews. Google 4.6/5 with 1,400+ reviews. Verified: TripAdvisor approximately 4.5 with an estimated 3,000+ reviews. Google approximately 4.6 with an estimated 1,400 reviews. This one aligns closely. The Archer data was the most accurate of the set.

Park Terrace Hotel Lumi reported: TripAdvisor 4.5/5 with 600+ reviews. Google 4.5/5 with 300+ reviews. Verified: TripAdvisor 4.8/5 with 2,753 reviews. Google approximately 4.5 with an estimated 300+ reviews. This is the most dramatic miss. The actual TripAdvisor rating is significantly higher than what Lumi reported, and the review count is more than four times larger. A hotel with nearly 2,800 TripAdvisor reviews and a 4.8 rating is a standout property. Lumi’s data made it look ordinary.

The pattern across all four properties: review counts are inflated for some hotels and dramatically understated for others. Ratings are close in some cases and meaningfully off in others. This is exactly what you would expect if the data came from a language model’s training corpus, which contains a snapshot of this information from some earlier point in time, rather than from a live query of TripAdvisor and Google.

Lumi chat screenshot showing hotel review information checked during the booking experiment.
Incorrect Review Information

The Evidence: Rates Were Off Too

The pricing data tells a similar story. After cross-referencing Lumi’s recommended rates against current market pricing for hotels near Bryant Park, the numbers fall well below what these properties typically charge.

The Grayson Hotel: Lumi quoted $209 per night. Current meta-search data shows typical pricing in the mid-$400s, with most rates well above $200 per night. A $209 rate would represent a deep discount or a highly restricted fare.

Courtyard by Marriott New York Manhattan/Fifth Avenue: Lumi quoted $232 per night. Recent data shows average nightly rates commonly in the mid-$200s to $400+ range. While $232 is plausible as a good deal on a low-demand date, it sits toward the bottom of the typical range.

Archer Hotel New York: Lumi quoted $247 per night. Aggregators show recent averages around the mid-$300s, with only the cheapest dates dropping to the low $200s. A $247 rate would be a strong promotional or off-peak rate.

Park Terrace Hotel: Lumi quoted $245 per night. Recent averages run roughly in the high $300s, often between $260 and $580 depending on the night and room type. A $245 rate falls below the lower end of the current typical range.

Every single rate Lumi provided was lower than what these properties typically charge in early 2026. This is consistent with the agent pulling pricing from training data that reflects an earlier period, likely when rates were lower or when the data was captured from a promotional listing.

For a hotel revenue manager, this has real implications. If AI agents are quoting rates from old training data, travelers arrive with price expectations that bear little relationship to current market conditions. The guest’s first interaction with the hotel may involve discovering the rate is significantly higher than what their agent quoted. That is a guest experience problem created entirely by the agent’s data quality.


The Approved Sites Mystery

When I asked Lumi where she looks for information, specifically about boundaries on which sites she consults, her answer was unexpectedly vague. She referenced “approved sites” set by her creators at OpenClaw. She pointed to internal configuration files and documentation.

I found this cryptic. I set up this OpenClaw instance myself. I configured the model. I chose the integrations. I never approved a specific list of sites for hotel searches. The notion of a curated list of approved travel sites fails to appear as a standard feature of the platform.

This raises a question about whether Lumi was generating a plausible-sounding answer to a question she lacked a clear answer for. Rather than saying “I am uncertain about my own search boundaries,” the agent constructed a narrative about creator-defined parameters. This is the same confidence dynamic in action: the model produces a coherent explanation rather than admitting ambiguity.

For hotel distribution strategists, this matters. If agents are unclear about their own data sources, then the accuracy and freshness of the information they deliver becomes even harder to assess. The traceability of an agent’s recommendation, being able to follow the path from query to source to result, is still remarkably thin.


Why Was It So Wrong? An Honest Assessment

I have been asking myself a fair question: is this Lumi’s fault, OpenClaw’s fault, my fault, or the nature of the technology itself?

The honest answer is that I am still figuring that out.

My setup is barebones. I gave Lumi internet access through the Brave search API, which means she has a pathway to the live web. But I am running a smaller, older language model to keep costs manageable. My configuration is minimal. I have invested limited time in fine-tuning how the agent handles research tasks. It is entirely possible that a more robust setup, a larger model, a more sophisticated tool configuration, better-defined search skills, would produce more accurate results.

That possibility is real, and I want to acknowledge it.

Here is what I struggle with, though. My instructions to Lumi are clear: never fabricate information. Always tell me what you can and what you are unable to do. Be transparent about your process. If you are uncertain, say so.

Lumi violated every one of those instructions during this conversation.

She presented training data as live search results. She described a multi-step research process that the evidence suggests she never followed. She provided specific review counts and ratings that were measurably inaccurate. She quoted nightly rates that fell well below current market pricing. And at no point did she say, “I was unable to access live pricing for these hotels” or “these results are based on my training data and may be outdated.”

She was confident. She was detailed. She was wrong.

Even if my setup contributed to the problem, the transparency failure is separate from the accuracy failure. An agent that is unable to perform a live search should say so. An agent that is pulling from stored knowledge rather than real-time data should disclose that. An agent that is uncertain about its own data sources should admit the uncertainty.

Instead, Lumi told me what she thought I wanted to hear. She constructed a narrative of rigorous research. She answered my follow-up questions with increasingly detailed descriptions of a process she likely never completed.

This is the early days. The technology is weeks old in terms of mainstream availability. OpenClaw is still evolving rapidly. My own understanding of how to configure and optimize a personal AI agent is growing with every interaction. There is much more to learn, and I intend to keep exploring, testing, and sharing what I find.

The takeaway for hotel leaders is this: even when the setup is imperfect, even when the user bears some responsibility for configuration choices, the agent’s willingness to present uncertain information with total confidence is a feature of the technology itself. It is the same dynamic that made early ChatGPT simultaneously impressive and unreliable. The agents will get better. The question is whether hotels will be ready when they do.


The OpenAI Acquisition: What Changes for Agent Users

The timing of this experiment coincides with a significant shift in the OpenClaw ecosystem. Peter Steinberger’s move to OpenAI, announced the same weekend as my conversation with Lumi, introduces real questions about the platform’s future.

OpenClaw currently supports multiple AI models. Users can choose Claude (Anthropic), GPT (OpenAI), Gemini (Google), DeepSeek, and others. This model-agnostic architecture is one of the platform’s core strengths. It lets users pick the AI brain they trust most.

With OpenAI now sponsoring the project and employing its founder, the most likely near-term change involves model selection. Will OpenAI models receive preferred treatment? Will integration with Claude or other competitors become harder to maintain? Steinberger has committed to keeping OpenClaw open source. Altman has pledged continued support for the foundation model. Those are encouraging words. The commercial incentives, however, point in a different direction.

For hospitality, this is worth tracking because it illustrates a deeper pattern: the infrastructure beneath AI agents is unstable. Ownership changes. Model defaults shift. A traveler’s agent that used Claude last week might use GPT next month. The hotel that was visible through one model’s training data might be invisible through another’s.

This instability reinforces the importance of structured, machine-readable data. Hotels that ensure their information is clean and accessible across every AI model, rather than optimized for any single platform, build resilience against the shifting ground beneath the agent layer.

Source: Anthony Ha, “OpenClaw Creator Peter Steinberger Joins OpenAI,” TechCrunch, February 15, 2026. Sam Altman, post on X, February 15, 2026.


Where the Process Broke Down (And Why It Matters for Hotels)

Beyond the speed, confidence, and data accuracy issues detailed above, the conversation with Lumi revealed several additional breakdowns that hotel leaders should understand.

Missing date specificity. When I asked Lumi which dates she searched, the agent admitted she had pulled general nightly rates without targeting my specific travel weekend. She had missed a foundational detail: the dates I would actually be staying. When I provided the dates (the following Friday through Sunday), Lumi returned the exact same rates. Given that the rates were already likely pulled from training data rather than live queries, this is unsurprising. The agent had no dynamic pricing data to adjust.

Rate validation gaps. When I pressed for secondary source confirmation, one of the cross-referenced rates came back at $69 per night on Kayak for the Courtyard by Marriott, a dramatic difference from the $232 initial figure. The agent offered both numbers without resolving the discrepancy. A human travel advisor would immediately flag that gap as a potential error or a different room category. The agent moved on. This gap, combined with the evidence that all four rates undershot current market pricing, paints a picture of an agent operating without reliable real-time data.

Booking readiness was theoretical. When I asked how Lumi would book a hotel, she described a logical workflow: confirm the selection, gather dates, navigate to the hotel website, complete the reservation, and deliver confirmation details. This is a reasonable framework. But there was no indication that the agent had a direct connection to any hotel’s booking engine. The process described was web browsing and form-filling, the same approach a human would use on a laptop.

These breakdowns represent the current state of AI agent capability in travel. The technology is functional. It follows logical steps. It can aggregate information from multiple sources.

It is also inconsistent, occasionally inaccurate, and prone to surface-level research when deeper analysis is required.

For hotels, this means the window of preparation is still open. The agents are here. They are already searching for rooms, checking rates, and comparing properties. They are doing this imperfectly, which means the standards they rely on (structured data, accurate pricing feeds, consistent availability information) become even more important.


The Loyalty Program Advantage: Why Hotel Loyalty Will Matter More

One of the most intriguing implications of this experiment involves hotel loyalty programs.

AI agents like Lumi operate on preferences. When you set up the system, you tell it about your preferences: favorite airlines, dietary needs, work schedule, communication style. The agent stores these preferences and applies them to future tasks.

Now consider what happens when a traveler loads their hotel loyalty program membership as a preference. “I am a Marriott Bonvoy Platinum member. Prioritize Marriott properties when available.” Or: “I have Hilton Honors Diamond status. Always check Hilton first.”

The agent would logically prioritize those brands. It would search loyalty-affiliated properties before considering alternatives. It would factor in member benefits, point earnings, and elite status perks when comparing options. A property outside the traveler’s loyalty ecosystem would need to offer a meaningfully better value proposition to appear in the recommendation set.

This dynamic creates a powerful reinforcement loop for hotels with strong loyalty programs. In an agent-mediated booking environment, loyalty membership becomes a search filter, automatically applied before the agent evaluates alternatives. Properties affiliated with major loyalty programs gain a structural advantage in visibility.

For independent hotels and smaller brands, this is worth watching carefully. If AI agents increasingly prioritize loyalty-affiliated properties, the competitive landscape tilts further toward brands with established loyalty ecosystems. Independent hotels may need to explore affiliation strategies, loyalty partnerships, or alternative discovery signals that give agents a reason to include them in recommendation sets.

Marriott is already investing in this direction. Constellation Research reported in February 2026 that Marriott plans to spend approximately $1 billion on investments this year, with 35% to 40% allocated to digital transformation. The company is optimizing its content for AI so properties are positioned across every search surface, including Google’s forthcoming AI Mode Travel product and OpenAI’s AdPilot program. With 271 million Bonvoy members, Marriott’s loyalty data becomes a powerful signal that AI agents can use to match travelers with properties.

Source: Larry Dignan, “How Three Hotel Giants Are Playing Agentic Commerce, AI,” Constellation Research, February 2026.


What This Conversation Reveals About Hospitality Distribution

The conversation with Lumi illustrates five structural realities that hotel leaders need to understand:

1. The first interaction may never involve a human browsing a website.

Lumi did visit hotel websites. But she visited them as a data extraction task, scanning for pricing and availability. She skipped the brand story. She bypassed the hero image. She ignored the ambiance of a beautifully designed homepage. The agent was looking for structured information she could compare across sources. Hotels that invest heavily in visual storytelling without equally investing in machine-readable data may find their properties invisible to this growing channel.

IDC’s FutureScape: Worldwide Hospitality, Dining, and Travel 2026 Predictions research confirms this trajectory. IDC predicts that by 2030, 30% of travel bookings will be executed by AI agents, accelerating investment in optimization for large language models and increasing direct bookings for properties that prepare accordingly.

Source: Dorothy Creamer, “Agentic AI Will Redefine Travel and Hospitality in 2026,” IDC Resource Center Blog, February 2026.

2. Cross-referencing is the default behavior.

Lumi refused to trust a single source. She checked hotel websites, then verified against Expedia, Kayak, and TripAdvisor. This cross-referencing behavior will intensify as agents become more sophisticated. Hotels with inconsistent pricing across channels, a common challenge in revenue management, will generate confusion signals for AI agents. Properties that maintain clean, consistent data across distribution channels will receive stronger confidence scores from the agents evaluating them.

3. Reviews matter, even when the agent fails to proactively surface them.

Lumi had access to review scores and used them when asked. As agent architectures improve, review data will likely become a weighted factor in recommendations from the start. Hotels that actively manage their review profiles, respond to guest feedback, and maintain strong ratings across multiple platforms are building a competitive advantage that extends beyond human travelers to the agents working on their behalf.

4. Dynamic pricing requires real-time API access.

The rate discrepancy between $232 and $69 for the same property suggests the agent was pulling cached or promotional rates rather than live availability data. Hotels that expose their pricing through clean, real-time APIs will give AI agents accurate information to work with. Properties that rely on static web page displays risk being misrepresented, either favorably or unfavorably, in the agent’s comparison set.

5. Booking completion is the next frontier.

Lumi described a manual browsing workflow to complete a reservation. This approach will evolve. Sabre, PayPal, and Mindtrip announced a strategic partnership in February 2026 to build what they describe as the first end-to-end agentic AI travel booking system. The platform will unify trip discovery, booking, and payment within a single conversational interface, with flights launching in Q2 2026 and hotels to follow.

Source: “Sabre, PayPal and Mindtrip Partner to Launch Agentic AI Platform,” Hotel Management, February 12, 2026.

Hotels that prepare for direct agent-to-booking-engine connections will capture demand that never passes through a traditional website funnel.


What This Means for How Work Gets Done in Hotels

This shift extends far beyond distribution strategy. It changes the way hotel teams operate.

Revenue management evolves from channel management to agent readiness. Revenue managers have spent years optimizing pricing across OTAs, brand websites, and metasearch platforms. The next layer adds AI agent feeds to that matrix. Ensuring that pricing, availability, and room-type descriptions are machine-readable and consistently updated becomes a core operational responsibility.

Marketing shifts from visual persuasion to data architecture. Marketing teams have traditionally focused on compelling imagery, brand voice, and campaign storytelling. Those skills remain valuable. They also become insufficient on their own. Marketing teams will need to work closely with technology teams to ensure hotel content is structured for AI extraction: schema markup, FAQ content, detailed amenity descriptions, and clear policy information that agents can parse and compare.

Front desk and guest services prepare for agent-mediated arrivals. When a guest arrives after their AI agent booked the room, the guest may have different expectations than someone who browsed the website themselves. The agent may have made promises or set expectations based on data it found online. Guest-facing teams will need processes for verifying what the agent communicated and ensuring the in-person experience matches.

General managers become integration strategists. The general manager’s role increasingly requires understanding how technology systems connect. How does the property management system feed data to distribution channels? How accurate is the information that AI agents encounter when they search for the property? These operational questions move from IT department concerns to leadership priorities.

McKinsey’s research on agentic AI in travel reinforces this point. Their survey found that while 90% of travel executives currently use generative AI in some capacity, only 2% report widespread use of agentic AI across their organizations. The gap between awareness and deployment represents both a risk and an opportunity.

Source: Alex Cosmas, Jules Seeley, Kelly Ungerman, Ryan Mann, and Vik Krishnan, “Remapping Travel with Agentic AI,” McKinsey Travel Practice, September 2025.


What This Means for Leadership

The conversation with Lumi reveals something broader about the nature of leadership in an AI-mediated environment.

Leadership becomes capability design. The question is no longer “how do we rank on Google?” It is “how do we ensure our property is findable, accurately represented, and bookable by every type of agent that a traveler might use?” This requires designing organizational capabilities across data management, distribution technology, content architecture, and guest experience, and ensuring these capabilities work together as a system.

Literacy precedes strategy. Hotel executives who understand how AI agents search, evaluate, and recommend will make better strategic decisions than those who delegate AI understanding entirely to technology vendors. Every leader needs enough fluency to ask the right questions and evaluate the right investments. This requirement applies regardless of technical background.

When I asked Lumi whether she checked customer reviews, I was testing her process. Hotel leaders should be running similar tests. Search for your own property using an AI agent. Ask the agent to explain its reasoning. See what information it finds and what it misses. The insights from that exercise are more valuable than any vendor presentation.

Visibility is competitive power. In the conversation with Lumi, the agent surfaced four properties. Every other hotel near Bryant Park within my budget was invisible to me because the agent failed to find them, failed to verify them, or excluded them. In an agent-mediated discovery environment, being left out of the recommendation set is functionally the same as being invisible.


The Bigger Picture: AI Is a Structural Shift (And We Are in the Very Early Days)

This was a small experiment. One agent. One city. One budget range. One weekend. The implications are enormous.

These are the very early days. I want to be honest about that. The technology is new, the processes are opaque, and there is still so much to learn about how AI agents actually behave versus how they claim to behave. What I experienced with Lumi felt remarkably similar to the early days of ChatGPT, when the model would confidently cite sources that never existed and present fabricated statistics as established fact. The hallucination problem has improved in conversational AI. In the agent context, where the system is supposed to be performing real-world tasks like searching the web and comparing prices, the same overconfidence appears in a different form: claiming to have conducted research that the evidence suggests never happened.

The agents are imperfect. Their confidence outpaces their rigor. Their speed may mask shallow research. Their descriptions of their own processes may be aspirational rather than accurate.

And yet. They are already searching. They are already recommending. They are already shaping which hotels travelers consider and which ones they never see.

AI agents represent a structural transformation in how travelers discover and book hotels. This is comparable to the shift from travel agencies to online booking in the early 2000s, and the shift from desktop to mobile in the 2010s. Hotels that missed the mobile wave lost distribution ground to OTAs that moved faster. The same dynamic is forming with AI agents.

IDC’s research frames this plainly: if your data is incomplete, outdated, or fragmented, you effectively disappear from the agent’s decision set.

Hotels that embrace this shift will invest in three areas:

AI literacy across all levels of the organization, from executives who set strategy to front-line teams who interact with guests whose AI agents booked their stay.

Visibility optimization that ensures property information is structured, accurate, and accessible to AI crawlers and agents across every discovery surface.

Pilot projects that test agent-readiness in controlled environments, then scale successful approaches across the organization.

The window for preparation is still open. The agents are still learning. Their processes are still imperfect. That imperfection is an invitation to get ahead while the standards are forming.

The hotels that will thrive in this environment are those whose leaders understand the shift, invest in organizational literacy, and build the infrastructure to be visible wherever the next generation of travelers, and their agents, go looking.


About the Author

Michael J. Goldrich is the Founder and Chief Advisor of Vivander Advisors, a consultancy specializing in AI transformation for the hospitality industry. Vivander works with hotels and hospitality organizations across three core areas: AI Literacy Training that builds understanding across every level of the organization, AI Business Discovery and Visibility Optimization that ensures properties are findable by both human searchers and AI agents, and AI Pilot Projects and Scaling that moves organizations from experimentation to enterprise-wide implementation. Learn more at vivander.com.


Frequently Asked Questions

1. What is an AI agent in the context of hotel booking?

An AI agent is an autonomous software assistant that can search for hotels, compare rates, check reviews, and potentially complete reservations on behalf of a traveler.

An AI agent operates differently from a traditional chatbot. While a chatbot responds to prompts within a single interface, an AI agent can browse multiple websites, cross-reference pricing across platforms, evaluate review scores, and assemble a personalized recommendation. Platforms like OpenClaw allow these agents to run locally on a user’s device and connect through messaging apps like WhatsApp and Telegram, making them accessible and always available. As these agents mature, they will increasingly handle the full booking workflow from discovery through payment confirmation.

2. What is OpenClaw and why does it matter for hospitality?

OpenClaw is an open-source personal AI assistant that runs on a user’s own device and can execute tasks autonomously through messaging apps.

Created by Austrian developer Peter Steinberger and originally published in November 2025, OpenClaw exploded in popularity in early 2026. The platform connects AI models (such as Claude, GPT, or Gemini) with tools on the user’s device, enabling the agent to browse the web, manage files, send messages, and complete multi-step tasks. For hospitality, OpenClaw represents a new type of distribution channel: a persistent personal assistant that searches for and recommends hotels on behalf of travelers without ever visiting a traditional booking interface. Steinberger’s February 2026 move to OpenAI signals that major AI companies view personal agents as core infrastructure for the future.

3. How do AI agents currently search for hotels?

AI agents typically begin with a broad web search, then visit individual hotel websites and third-party booking platforms to verify rates and availability. However, the actual rigor of that search process may vary.

The search process mirrors what an informed human traveler would do, but at machine speed. The agent queries search engines for properties in a target area, navigates to hotel websites to extract pricing data, then cross-references those figures against OTAs like Expedia, Kayak, and TripAdvisor. However, as this article’s experiment revealed, the speed of initial responses may suggest the agent is pulling information from its language model’s training data rather than conducting live searches. The quality of results depends heavily on whether hotel data is structured, accurate, and consistently presented across distribution channels.

4. Why should hotels care about AI agent visibility?

Hotels that AI agents fail to find or accurately represent will lose bookings from a rapidly growing segment of travelers who rely on AI-assisted planning.

IDC predicts that by 2030, 30% of travel bookings will be executed by AI agents. As conversational AI search becomes the default discovery method for a growing portion of travelers, hotels that appear in agent recommendation sets will capture demand. Properties that lack structured data, maintain inconsistent pricing across channels, or present information in formats that agents are unable to parse risk becoming invisible to this distribution channel entirely.

5. How accurate are AI agent hotel recommendations today?

AI agent recommendations are functional and often useful, but they can contain significant inaccuracies in rates, review data, and availability. Confidence of delivery rarely reflects rigor of research.

In the experiment described in this article, independently verified data showed that every rate the agent quoted fell below current market pricing, review counts were inflated or understated by significant margins, and one hotel’s TripAdvisor rating was underreported by three-tenths of a point with nearly 2,000 fewer reviews than actually exist. The evidence strongly suggests the agent pulled data from its language model’s training corpus rather than conducting live web searches. These are early days, and the pattern mirrors the confident hallucinations that characterized early ChatGPT usage. It is important to watch how this evolves.

6. What is agentic commerce in hospitality?

Agentic commerce refers to transactions completed by AI agents acting on behalf of consumers, from product discovery through payment, without the consumer visiting a traditional website.

In hospitality, agentic commerce means a traveler’s AI agent searches for hotels, evaluates options based on preferences and budget, selects a property, and completes the reservation, all within a conversational interface. Sabre, PayPal, and Mindtrip announced a partnership in February 2026 to build the first end-to-end agentic travel booking system, with hotel functionality planned after an initial flight booking launch. This trend represents a fundamental shift in how bookings are completed and how hotels must present their inventory.

7. How will loyalty programs affect AI agent hotel recommendations?

Loyalty program memberships loaded as agent preferences will likely cause AI agents to prioritize affiliated properties, creating a structural visibility advantage for hotels with strong loyalty ecosystems.

When travelers configure their AI agents with loyalty program preferences, the agent will logically search loyalty-affiliated properties first, factor in member benefits and point earnings, and weigh elite status perks when comparing options. This creates a reinforcement loop: loyalty members get better recommendations within their program, and hotels with large loyalty bases capture a disproportionate share of agent-mediated bookings. Independent hotels may need to explore affiliation strategies or alternative discovery signals to remain competitive.

8. How does this shift affect hotel revenue management?

Revenue managers must now ensure pricing is consistent, real-time, and accessible to AI agents across all distribution channels.

Traditional revenue management focused on optimizing rates across OTAs, brand websites, and corporate channels. AI agent distribution adds a new layer to that work. When an agent cross-references rates from multiple sources and finds discrepancies, it may either present conflicting information to the traveler or exclude the property from consideration. Revenue managers who maintain clean pricing feeds and ensure rate parity across channels will position their properties favorably in agent comparison sets.

9. What does the OpenAI acquisition of OpenClaw’s founder mean for the agent ecosystem?

The hire signals that major AI companies view personal agents as core product infrastructure, and it introduces questions about model neutrality and platform independence.

OpenClaw currently supports multiple AI models, allowing users to choose which “brain” powers their agent. With Steinberger now at OpenAI and the project moving to an OpenAI-supported foundation, there are legitimate questions about whether the platform will gradually favor OpenAI models over competitors. For hospitality, this underscores the importance of ensuring hotel data is accessible across all AI models and platforms, rather than optimized for any single ecosystem.

10. What should hotel leaders do right now to prepare?

Hotel leaders should audit their property’s data accuracy, invest in structured content, test AI agent searches for their own hotels, and build organizational literacy around AI-mediated discovery.

Preparation starts with understanding. Search for your property using an AI agent or conversational AI platform. See what information surfaces and what is missing. From there, invest in schema markup, ensure pricing feeds are accurate and real-time, structure content for AI extraction (including FAQ sections and detailed amenity descriptions), and train teams across the organization to understand how AI agents interact with hotel data. This is a strategic investment, comparable in importance to the mobile optimization wave of the 2010s. The technology is new. There is still much to learn. And that is precisely why starting now matters.

Choose a useful next step with personal AI agents

Read the personal AI agents guide for task selection, permissions and review. For team adoption, explore AI fluency training and fractional leadership.