Sort the AI a hotel can buy by a single question, how much of the work it does on its own and three frameworks separate out.
- Co-pilots: A co-pilot works beside your experts and makes them faster, with a person still making every call.
- Agents: Agents work inside your systems and carry whole tasks to completion on their own.
- FDE: Forward-deployed engineering sends a vendor's engineers in to build a custom system around your own data.
None of the three is more advanced than the others and none is the right answer by itself. Each hands a different amount of work to the software, asks for a different amount of trust and earns its keep on a different kind of problem. What matters is which one fits the team you have and the work you need done.
Handing more work to the software saves more time and frees more people, but it asks for more trust and more setup. Handing over less keeps you in control but leaves more on human shoulders. So fit depends on the situation: a hotel sharpening a strong team needs something different from one buried in repetitive volume, or one facing a hard, specific problem that no off-the-shelf tool touches.
The Vertical Co-Pilot Makes Your Best People Faster
The most familiar form of hotel AI is the co-pilot, a tool that sits beside an expert and makes them faster and better informed. Lighthouse's Ernest is a clear example, giving a commercial team one place to ask across revenue, marketing, sales and distribution and to act within controls that keep a person in the loop. The human still makes the call; the co-pilot makes the call faster and easier to get right.
Its great strength is that it amplifies the people you already trust, with very little risk, because nothing happens without a person reviewing it. It is quick to adopt, it fits the way skilled people already work and its benefits tend to show up almost immediately. It also does more for labor than it is usually given credit for. A revenue manager who can do in an hour what used to take a morning can cover more hotels and a corporate team at a management company can run more properties per person. That is real efficiency, the kind that lets a group grow its portfolio without growing its overhead at the same pace.
Its limits are the mirror image of its strengths. A co-pilot is only as valuable as the people using it, so it rewards hotels that have skilled operators and good habits and it does less where no one is staffed to ask it anything. Its payoff also depends on adoption, since a capable co-pilot that nobody opens changes nothing. It is a way to get more out of the good people you already have.
A hotel can put a co-pilot to work anywhere judgment lives, which is to say almost everywhere: pricing and forecasting in revenue, campaign and content decisions in marketing, group and corporate negotiations in sales, budgeting in finance and problem-solving on the operations floor. The impact is more and better output per person, sharper decisions and the ability to stretch a strong team and a growing portfolio, further than it could go on its own.
OS-Native Agents Do The Work Inside Your Systems
The second model hands over more: instead of advising, it does the work. These agents live inside the systems that already run the hotel and carry tasks from start to finish, within the guardrails a hotel sets. Canary's Hospitality AI Agent Studio is the clearest example in hotels: a builder where a hotel stands up its own agents, starting from templates for the front desk, concierge, reservations and post-stay feedback or assembling custom ones from a library of skills and runs them across its whole tech stack and every department. Other vendors approach the same model from different angles: Guesty in the vacation rental space recently launched a coordinated set of agents into its PMS and Apaleo runs a more open version where third-party agents act across connected systems like a marketplace of specialists.
The appeal is straightforward. Agents take repetitive work off a team's plate and scale without adding headcount and because they run inside systems a hotel already uses, they extend the existing stack rather than replacing it; in the CFO's office, more than three-quarters say they prefer AI layered onto their system of record over a rip-and-replace. The approaches mainly differ in who assembles the agents and how far they reach: a studio like Canary's gives a hotel the most control and the broadest span across the stack, a coordinated suite from a PMS vendor gives a set that already works together under one owner and an open marketplace gives the widest selection of specialists.
The tradeoff is trust. Because an agent acts on its own, a hotel has to be comfortable with what it does unattended, which is why this model is strongest where the work is high in volume and the rules are clear and why it deserves hard questions about guardrails, ownership and the audit trail when something goes wrong. Those are answerable questions and the precedent next door is encouraging: in housing, EliseAI built much the same model and now handles more than 90 percent of resident interactions with no person involved, scaling to roughly 10 percent of the U.S. apartment market.
Agents can handle guest messaging and reservations, push rate and inventory updates in revenue, coordinate housekeeping, respond to reviews in marketing, or reconcile invoices in finance. The impact, applied where the volume is, is genuine labor savings and the ability to grow the work without growing the team to match.
Forward-Deployed Engineers Build AI Around Your Business
The third model does not hand you a product off the shelf at all. Instead, the vendor embeds its own engineers to build a system around your data and your specific problems. Actabl does this in hotels, putting engineers inside hotel companies to build on the financial, labor, service and asset data it has already unified, so that even a hotel's hardest and highest-stakes decisions can be handled by something built to fit them exactly.
This is the model behind some of the most capable enterprise AI anywhere. It is how Palantir turned tangled internal data into action a company could trust and the frontier labs have stood up their own embedded-engineering teams to do the same for large enterprises. Its strength is depth. When a problem is genuinely specific to your business, a system built around your own data will fit it in a way no packaged tool can and it can take on work too important or too unusual to trust to a generic agent.
The tradeoffs are cost and commitment. This model asks for scale, clean-enough data and budget and it makes the most sense when the problem really is unique to you; if it is a common problem that an off-the-shelf tool already solves, that tool will be cheaper and faster. The buyer question worth asking is whether the vendor turns each engagement into a reusable product, so the cost of the next deployment falls rather than repeating, which is the difference between buying software and funding an open-ended consulting engagement. The firms that do this model well, as the finance world has learned, are the ones that build the data work into the product rather than billing it forever.
The same approach can build a bespoke revenue decision, a custom guest-experience workflow, a labor-optimization model, or an asset-level analysis. The impact, when the fit is right, is AI shaped to your business on the problems that matter most to it, with a depth that generic tools cannot reach.
How A Hotel Should Decide Where To Focus
None of this is a contest with a single winner and most hotels will end up using more than one of these models. The useful work is matching each to the job it does best. If you have strong people and want them to do more, start with a co-pilot. If you have high-volume, repetitive work you would rather not staff against, look hard at agents. If you have a specific, high-stakes problem and the data and budget to support it, forward-deployed engineering is built for exactly that. The real question is which problem to attack first.
A few things hold true across all three. Judge them on results in production rather than in a demo, because the gap between the two is wide; in finance, where adoption is further along, fewer than one in 20 AI pilots clears even a fifty-percent success rate, so ask for references from hotels actually running the tool. Ask who is accountable when the tool gets something wrong and what the audit trail looks like. And treat clean, connected data as the cover charge for all three, because none of them works well without it.
These three models are the shape of AI in hotels for the next several years. The hotels that get the most from them will be the ones that learned each model well enough to put it where it pays off.
Jordan Hollander is the co-founder of HotelTechReport.com, a hotel industry's app store. He is also an angel investor in boutique hotels and hospitality tech startups. He was previously on the global partnerships team at Starwood Hotels & Resorts.