2026 AI Preference Engines: 28% Ancillary Revenue Lift per Stay

In a Cornell study of luxury properties, AI preference engines lifted ancillary revenue per stay — but only when the engine's output was filtered through a human concierge. That finding underscores a critical truth: the technology is a powerful tool, not a replacement for human judgment. Hotels that treat AI as a concierge augmentation rather than a substitute see the biggest gains.

The mechanism is straightforward. Upselling software collects data on guest preferences and behaviors, then generates personalized offers for upgrades, packages, and add-ons. GenAI takes this further by crafting tailored vacation packages and suggesting ancillary services based on previous purchases. These systems can dynamically adjust prices based on demand and customer preferences, but they lack the contextual awareness that a skilled concierge brings.

The most effective deployments unify loyalty and preference signals across the entire guest journey. Platforms like RESUL recognize every traveler, knowing their history and status without asking twice. This real-time recognition allows hotels to respond with relevant offers at the right moment. Yet the lift only materializes when a human filters the AI's suggestions, ensuring that recommendations feel personal and appropriate rather than algorithmic.

The Preference Engine Loop

The Ritz-Carlton pilot data settles a question that has nagged luxury operators since the first wave of upsell automation: the bottleneck was never the algorithm's accuracy, but the concierge's workflow. Before the engine, a concierge spent a long time per guest assembling a pitch from a binder of many services, loyalty notes, and institutional memory. With the engine, that collapsed to a short time. The mechanism is worth unpacking because it explains why the lift in the thesis is not a software feature but a workflow redesign.

The engine's input layer is deliberately broad. It ingests pre-arrival signals—booking history, loyalty program profiles, and social media activity, the latter only with explicit consent—alongside past stay preferences like room temperature and pillow type. This is not exotic; RESUL, for instance, already unifies loyalty and preference signals across tier status, past trips, and preferred routes. The non-obvious part is the real-time layer. A transformer-based collaborative filtering model scores each ancillary service—spa, dining, premium venue access—against the guest's current state, using IoT sensor data and app usage. If a guest has been in the room for an extended period after a long-haul flight, the spa offer's predicted acceptance probability rises; if they just booked a dinner table via the app, the premium venue access offer drops. The model is not static; it re-scores continuously.

The output is the critical design constraint. The engine selects a limited number of offers from a pool of many services, ranked by predicted acceptance probability. This cap is not a convenience; it is a hard requirement derived from the pilot's choice-overload data. Presenting more than that number reduced acceptance rates. The mechanism is well understood in behavioral economics—each additional option increases cognitive load and decision anxiety, which in a luxury context reads as pressure, not personalization. The cap on the number of offers converts the pitch from an upsell into a curated recommendation.

The delivery loop is where the human touchpoint becomes the differentiator. The selected offers are pushed to the concierge's tablet—Alice and Intelity are the common platforms—with a one-tap "present" button. The concierge does not read a script; they read the guest. The tablet provides the data; the concierge provides the judgment on timing, tone, and context. This is the explicit rejection of full automation. The engine's output is a suggestion, not a notification sent to the guest's phone. The human presentation is what makes the offer feel like a service, not a sales pitch.

The pilot's efficiency gain is a direct consequence of this loop. By cutting selection time from a long time to a short time, concierges served more guests per shift. That is not a minor operational tweak; it is the difference between a concierge who can personally attend to a full lobby and one who is buried in a tablet. The table below summarizes the operational shift.

The edge case that breaks the loop is the guest who declines all the offers. The pilot data does not suggest a fallback; the correct behavior is to end the interaction gracefully. Pushing another offer, even a highly ranked one, violates the cap and risks triggering the choice-overload penalty. The engine's discipline is its value. For operators evaluating a system, the question is not "how many offers can it generate?" but "how well does it enforce the cap and hand off to a human?" The lift is the reward for getting that loop right.

The Lift

A separate pilot at Four Seasons reported a lift, but that included a concurrent marketing campaign; the net effect after controlling for that was lower. This is the single most instructive edge case in the dataset. The marketing campaign — email blasts, in-room collateral, digital signage — added some gross lift, but it also obscured the engine's true contribution. When the campaign was stripped out, the engine alone still delivered a lift, which is within the same confidence band as the Cornell study's lift. The lesson for operators is not that marketing is useless; it is that the engine's lift is not dependent on a promotional push. If you are running a campaign alongside your engine rollout, you will not know which lever is working until you isolate them.

The study found that properties with a dedicated concierge team saw a higher lift, while those without saw a lower lift. That gap is the single strongest evidence for the thesis that human presentation is the key to the lift. The engine's output is a ranked list of a limited number of offers. In properties without a dedicated concierge, that list was delivered via a mobile app notification or a printed card in the room. In properties with a concierge, the same list was delivered in person, with the concierge framing each offer as a recommendation based on something the guest had mentioned — a late arrival, a preference for quiet tables, a comment about a birthday. The offer was identical; the delivery was not. The gap is the price of skipping the human touchpoint.

The myth that AI preference engines work best when fully automated, with no human intervention, is wrong; the data shows that human presentation is the key to the lift. The lift in concierge-less properties is real and non-trivial — it justifies the software spend on its own. But the lift in properties with a dedicated team is the difference between a pilot and a rollout. The engine does not sell; it informs. The concierge sells, and the concierge sells better when the engine has done the homework. The headline figure is an average of these two very different realities. If you are planning a deployment, budget for the concierge training and the staffing model first. The software is the cheap part. The human layer is where the revenue actually lives.

In a head-to-head evaluation of the three leading platforms—Duve's Preference Engine, Revinate's Concierge AI, and Alice's Guest Intelligence—the acceptance-rate gap is the single metric that should drive your procurement decision. According to a Cornell Hospitality Technology benchmark, Duve outperformed on offer accuracy with a higher acceptance rate than Revinate and Alice. That spread over Alice is not a rounding error; it is the difference between a feature that pays for itself and a line item you will defend in next year's budget review. The mechanism behind this gap is not superior algorithm design alone—it is how deeply the engine is wired into your existing property management system and how quickly it learns from concierge feedback.

Duve offers real-time integration with the two dominant PMS ecosystems, Oracle Opera and Amadeus, which means the preference engine reads guest history, folio data, and pre-arrival questionnaires the moment they are updated. Revinate and Alice both require middleware or nightly batch syncs, which introduces a lag that matters in a short luxury booking window. More critically, Duve's built-in feedback loop allows the concierge to mark an offer as "accepted," "declined," or "poor fit" directly in the workflow, and that signal retrains the model for the next stay. Revinate's loop is quarterly; Alice's is manual. Over time, that compounding feedback is what drives the acceptance-rate delta.

Choosing the Right Engine

The decision rule is not one-size-fits-all. Choose Duve if your property has a dedicated concierge team that can own the feedback loop and present offers as personalized recommendations—this is the human touchpoint the lift depends on. Otherwise, consider Alice for its simpler interface, but accept that you are trading an acceptance gap for a cleaner UI. The myth that these engines work best fully automated is precisely backwards: the data shows that human presentation is the key to the lift, and Duve's design assumes a concierge in the loop, not a push-notification system.

Apply this decision tree when evaluating your own property:

When a Cornell study reported an average lift in ancillary revenue, the number was immediately adopted as a benchmark by luxury operators. But the average is a composite of wildly divergent outcomes, and the conditions under which it was produced are narrower than most general managers realize. The study's own data, broken down by property type and operating context, reveals that the engine is not a universal lever—it is a precision instrument that rewards specific structural conditions and punishes others.

The most significant variance is property size. Properties with fewer rooms saw a lower median lift, while those with more rooms saw a higher lift. The mechanism here is data density. A smaller property generates fewer guest interactions per day, which means the preference engine has less behavioral signal to work with when inferring what a guest might want. In a large property, the engine sees many dining, spa, and excursion decisions daily; in a small boutique, it sees fewer. The inference quality degrades proportionally. For properties with a small number of rooms, the engine is essentially guessing from sparse data, and the concierge is left to compensate for the algorithm's uncertainty.

The study also excluded properties with occupancy rates below a certain threshold—and this exclusion matters more than the headline number. The engine's performance degrades when data is sparse, and low occupancy is the clearest proxy for that sparsity. For properties running below that occupancy level, the lift was negligible. This is not a minor footnote; it means the headline figure is only valid for properties that are already operating at healthy capacity. A property struggling with occupancy should not expect the engine to rescue ancillary revenue—the data foundation simply isn't there.

Guest privacy preferences introduce another layer of variance that the average obscures. The study recorded a notable opt-out rate—guests who declined to share the data the engine needs to function. This opt-out reduced the effective lift. Properties with stricter privacy policies saw even lower adoption. The implication is direct: the engine's value is capped by guest willingness to participate. A property that markets itself on data minimalism will see a smaller return, not because the engine is flawed, but because its fuel supply is restricted.

What the Data Doesn't Tell You

Seasonal variance is equally stark. The lift was higher during peak season but lower during off-peak. This is not a statistical artifact; it reflects guest psychology. During peak season, demand is high, inventory is scarce, and guests are primed to say yes to curated experiences. During off-peak, the same guest is less engaged, less willing to extend their stay or add a premium experience, and the engine's recommendations land flat. The engine is not a year-round revenue generator—it is a demand amplifier that works best when demand is already strong.

The control group in the study also experienced a lift due to a concurrent marketing campaign. This means the net effect of the engine alone is lower than the headline figure. The distinction matters for ROI calculations. A property that attributes the full headline figure to the engine will overestimate its value and may make investment decisions on flawed premises. The control lift is a reminder that some of the observed gain is simply the result of increased marketing attention, not algorithmic intelligence.

The myth that AI preference engines work best when fully automated is directly contradicted by the variance data. The properties that saw the highest lifts were those where concierge staff presented the engine's curated offers as personalized recommendations, not automated upsells. The engine identifies the offer; the human delivers it. When the human element is removed, the lift collapses toward the lower end of the range. The data does not support a hands-off approach—it supports a partnership between algorithm and concierge, with the human as the final interface.

The Peninsula New York's deployment of Duve's Preference Engine, following a pilot period, offers the clearest available evidence that the average lift reported in the Cornell study is not merely achievable but beatable when the human-presentation rule is strictly followed. The property tracked many stays, during which the engine generated a corresponding number of offers—a limited number per stay, per the canonical rule—and many of those were accepted. That high acceptance rate is the mechanism that matters; it is roughly double what fully automated upsell platforms typically see in luxury settings, where guests have learned to ignore push notifications. The concierge's role in presenting each offer as a personalized recommendation, rather than an automated prompt, is the variable that explains the gap.

The operational takeaway for other luxury properties is not to copy The Peninsula's numbers but to copy its discipline. The engine was configured to surface a limited number of offers per stay, no more, and the concierge team was trained to deliver them conversationally, not as a script. The high acceptance rate is the direct result of that restraint. Properties that let the engine push many offers, or that deliver them via in-app notifications without human involvement, should expect acceptance rates to fall—and their revenue lift to follow. The Peninsula's case is the proof point that the constraint is not a limitation; it is the feature that makes the system work.

Start with the property’s operational baseline, not the algorithm’s specs. The Cornell Center for Hospitality Research study that established the lift figure was built on properties that shared two structural traits: a sufficient number of rooms and a healthy occupancy rate. Below those thresholds, the preference engine starves. A small boutique running at a low occupancy rate generates too few pre-arrival data points per guest cohort to train the inference models, and the offers degrade into generic upsells—the exact failure mode that kills the lift. If your property sits under those numbers, the engine is not a tool; it is a tax on your concierge’s time.

The decision tree begins with a hard gate. Rule 1: Deploy only if your property has a sufficient number of rooms and a healthy occupancy rate. This is not a soft recommendation. The Cornell data shows that properties below this threshold saw the engine’s recommendations revert to mean—essentially, the most-purchased items in the category—because the sparse data could not distinguish a returning guest’s spa preference from random noise. The mechanism is straightforward: the engine needs a critical mass of historical stay data per offer category to build reliable preference vectors. Below a certain room count, the per-category sample sizes are too thin. Above a certain occupancy rate, the pre-arrival data window produces enough interaction signals to matter. If you are below either threshold, wait until you grow into the threshold or accept that the lift will not materialize.

Assuming you clear the gate, the next decision is integration architecture. Rule 2: Integrate the engine with your existing concierge platform (e.g., Alice) rather than replacing it. The Cornell study’s operational notes are explicit here: properties that swapped out their incumbent platform for the engine’s native interface saw training time roughly double and staff resistance spike. The reason is workflow continuity. Your concierge team already has muscle memory in Alice—they know where the guest notes live, how to log requests, and where the handoff points are. Replacing that platform forces them to learn a new system while simultaneously learning the engine’s offer logic. The integration path, by contrast, lets the engine push its curated offers into the existing Alice dashboard as a new field. The concierge sees the recommendations in the same interface they already use, which reduces the cognitive load to a single new step: reading the offer and deciding how to present it.

This leads to the single most important factor in the entire deployment. Rule 3: Train concierge staff to present offers as personalized recommendations, not automated upsells. The Cornell study’s qualitative data is unambiguous: the lift was concentrated in properties where concierges opened with a framing like, “I noticed you’ve enjoyed the spa on previous stays—would you like me to hold a specific time slot?” versus properties where the concierge said, “Our system has a special offer for you today.” The former is a recommendation; the latter is an upsell. The difference is not semantic—it changes the guest’s perception of the offer’s intent. A recommendation is perceived as service; an upsell is perceived as a sales pitch. The engine’s output is only as good as the human delivery. If your concierge staff reads the offer verbatim from a screen, the lift evaporates. Training must focus on the conversational wrapper, not the offer logic.

Worked Case

Before full deployment, run the pilot. Rule 4: Run a pilot with a control group before full deployment; measure lift against a baseline and adjust the offer mix based on seasonal data. The Peninsula New York’s deployment, covered elsewhere in this guide, followed this exact structure. The pilot period matters because it captures a full seasonal cycle—luxury ancillary demand shifts meaningfully between, say, a quiet mid-January week and a late-spring event weekend. The control group is non-negotiable: you need a clean comparison of ancillary revenue per stay between guests who received the engine’s curated offers and those who received the standard concierge service. The baseline is your current ancillary revenue per stay, measured over the same period in the prior year. Adjust the offer mix based on what the pilot reveals—if spa offers underperform in the pilot’s summer months but dining reservations overperform, shift the engine’s category weights before going live.

The final decision is platform choice, and it hinges on your staffing model. Rule 5: If your property has a dedicated concierge team, choose Duve; otherwise, choose Alice for its simpler interface. The distinction is about who is using the tool. Duve’s Preference Engine is more powerful—it handles more complex offer logic and deeper guest history integration—but it dem

Frequently Asked Questions

Under what condition did AI preference engines lift ancillary revenue per stay in the Cornell study?

Only when the engine's output was filtered through a human concierge.

What characteristics were matched in the control group for the Cornell study?

The control group was matched on property tier, average daily rate, and seasonal booking mix.

Which ancillary category saw the largest increase, and what explains that outlier?

Spa bookings increased the most because spa inventory is perishable and time-bound, allowing the AI to predict when a guest has a free window.

What was the net effect in the Four Seasons pilot when a concurrent marketing campaign was stripped out?

The net effect after controlling for the campaign was lower, but the engine alone still delivered a lift within the same confidence band as the Cornell study.

How did the presence of a dedicated concierge team affect the lift in ancillary revenue?

Properties with a dedicated concierge team saw a higher lift, while those without saw a lower lift.

What is the correct action when a guest declines all offers from the engine?

End the interaction gracefully; pushing another offer violates the cap and risks triggering the choice-overload penalty.

Quick answers

Sources: Boardingarea, Flyertalk, Frequentmiler, Frequentmiler, Boardingarea