Who Controls the Decision When Hotel AI Recommends It?

A Saturday morning staffing call

It's 6:40 on a Saturday morning. The housekeeping schedule is already posted, but the hotel's labor-planning tool now projects that the property can run with one fewer room attendant than the manager thinks it needs. The live question is whether to call in extra help.

The occupancy forecast is soft, and the corporate labor target doesn't leave much room to spend more. On the other side, the front desk is holding a long list of early check-in requests, and engineering has warned that a guest elevator may be down for part of the morning.

Neither clue settles it. Early check-in requests don't always turn into early arrivals. The elevator may be back before housekeeping hits its busiest stretch. Last month, the manager added a shift because a group arrival looked difficult and ended the day with more unused labor hours than she could defend. On another weekend, the forecast looked safe until late room changes made it obsolete.

She can stay with the planned coverage or request more help. But what does either choice actually mean? Did the forecast include the early-arrival signals? Can she adjust the labor target, or was it fixed somewhere upstream? If she calls someone in, who has to approve the extra cost? If she follows the tool and rooms are late, will the postmortem look at what she reasonably knew at 6:40, or only at the outcome?

The scenario is hypothetical, and it has no predetermined right answer. Research from actual hotels makes the simple choice between trusting the algorithm and trusting the manager hard to defend. A study of more than 20 million hotel occupancy forecasts found that frequent human adjustments were associated with better forecasting for groups and special events but worse results for transient demand (Koupriouchina, van der Rest, & Schwartz, 2023). An experiment with 57 experienced hotel revenue managers found that changing the interface changed how much people altered an identical forecast (Schwartz & Cohen, 2004). A study of housekeeping room attendants reported differences in how an algorithmic assignment app was configured across properties (Spektor et al., 2023).

These studies do not establish that hotel managers are routinely punished for an algorithm's mistakes. They do force a narrower and more useful question. When a manager signs off on a recommendation, who else has already shaped the choices in front of her?

The first Hotel Decision Lab article, When Hotel Managers Override the Algorithm, asked when a manager's review actually improves an AI-supported decision. It argued that review is most likely to help when the manager contributes information the system lacks, can recognize consequential errors, and works inside a governance structure that supports effective intervention. It ended with a recommendation: assign accountability according to actual control.

Article 04 examines that third condition, governance, and what it means to assign accountability according to the control managers actually have. Article 01 separated decision authority (the right to decide), operational responsibility (the work assigned to people), and answerability (who has to explain the result). Here those questions are split further, into formal authority, practical control, responsibility, blame, and accountability. Two questions follow. When the organization later judges a manager's AI-supported decision, does that judgment match what she could actually see and change? And what about the reverse, when she has plenty of control and nobody checks how she uses it?

Whether a manager improves a decision and whether she's judged fairly for it are separate questions, and the answers can come apart. A manager can add little to a forecast and still answer for all of it. Another can hold wide discretion that nobody ever checks. This article calls the arrangement behind both cases a hotel's accountability architecture: who sets the objective, who controls the data and the settings, who may change the recommendation, and how the outcome gets judged afterward. Some of its most important parts are decided before any recommendation appears.

The decision before the decision

A labor recommendation doesn't just appear. Someone decided what the tool should optimize. Someone chose the data inputs and constraints. Someone determined whether the manager may change a forecast, amend a schedule, request more labor, or escalate a recommendation she disputes.

Those choices may sit with different people. A vendor can build a feature without deciding whether a hotel switches it on. Corporate teams may set labor targets without controlling every property-level exception. A general manager may keep more discretion than the department head who actually works in the interface. And a product that supports overrides doesn't mean every manager or employee can use them.

In hotel housekeeping, Spektor and colleagues (2023) studied the experiences of unionized room attendants using an algorithmic assignment system. Participants described differences in how the same product was configured across properties, including whether features that let workers sequence their own work were available. That grounds the idea that settings matter. The researchers listened to frontline workers, though, not a representative sample of hotel managers, and the study cannot establish how managers are disciplined for these decisions.

Another hotel-focused study came at the question from the manager's side. Jianu, Ashton, and Lugosi (2025) used a three-stage Delphi approach with managers and academic experts to explore what algorithmic management means for frontline managers. The study anticipated more work translating system requirements into operational and interpersonal practice. That's informed expert judgment about an evolving role, not direct measurement of whether software has taken managers' authority away.

The design-time point also has a theoretical foundation. Grote, Parker, and Crowston (2026) argue that organizations should align control with accountability across the people who develop AI systems and the people who use them. Their work is conceptual. It helps frame the institutional question but supplies no hotel-specific prevalence figures.

Accountability is doing too much work

In many discussions of AI, that one word covers situations that have little in common. A manager can feel responsible for a decision nobody formally assigned to her. A guest can blame the visible manager without knowing how the schedule was configured. An employer can assess the manager against labor cost and service targets without ever asking whether the recommendation was sensible when it appeared. Legal liability is a different question again, and this review doesn't evaluate it.

Five terms keep these apart in the rest of the article:

• Authority is the formal right to make or change a decision.

• Control is the practical ability to change the outcome, given the information, time, and permissions a person actually has.

• Responsibility concerns an actor's assigned duties or contribution to a decision. Perceived responsibility is a separate judgment about who caused an outcome or should have prevented it, and the decision-maker, an observer, a guest, and an employer can each attribute it differently.

• Blame is a retrospective negative judgment. It shouldn't be equated with formal responsibility.

• Accountability is the obligation to explain a decision, and accept its consequences, to someone with standing to evaluate it.

Hotel Decision Lab reads the two failure modes in this article as mismatches between these terms. In the first, accountability runs ahead of control: a manager answers for a decision she couldn't practically change. In the second, control runs ahead of accountability: a manager can change decisions freely, but nobody checks whether the changes help. That reading builds on Grote, Parker, and Crowston's (2026) argument that control and accountability should be aligned. A misalignment can run in either direction.

The distinctions matter because the research in this area measures different judgments, and the results don't all point the same way.

• Outside observers. Arnestad and colleagues (2024) asked people to judge a person after an automated system made a mistake. Observers assigned more blame when a manual override had been available, even if it went unused.

• The public judging a decision-maker. Ozer, Waggoner, and Kennedy (2024) asked US survey respondents to judge a judge whose decision preceded a bad outcome. Following an algorithm's advice did not reduce the blame placed on the judge, and blame was similar whether the advice came from an algorithm or from a person.

• The decision-maker judging themself. Kares and colleagues (2026) studied 261 participants who selected job applicants for interviews with an AI system's assistance, under different levels of automation and system reliability. Participants in the higher-automation condition, which combined constraints on inspecting applicant information and editing scores, assigned less responsibility and post-error blame to themselves and more to the system. The experiment does not isolate edit permission from information access.

A fourth line of work looks at decisions made under accountability pressure. Kotzian, Bauch, and Weißenberger (2025) ran an online experiment on advice-taking in a managerial forecasting task. They varied whether advice came from a human or an algorithm, whether a rationale came with it, and whether participants had to justify their judgment to a superior. For human advice, a rationale mattered more when participants had to justify themselves. For algorithmic advice, a rationale mattered either way. The study measured stated intentions to rely on advice, not forecast accuracy or what happened to anyone afterward.

None of these studies took place in hotels, and none measured how an employer actually evaluated a manager. They're easier to read together once you notice who is doing the judging. In Arnestad et al. (2024), an automated system acted and a person could have intervened manually. In Ozer et al. (2024), a human decision-maker acted with advice. In Kares et al. (2026), people judged their own AI-assisted choice. A reasonable hotel implication is that the same staffing outcome could be judged one way by the manager, another way by a guest, and a third way by the manager's boss, depending on how the system was set up.

The missing piece is specific. None of the reviewed studies follows a hotel labor recommendation from system settings through the manager's decision, guest and employee outcomes, and the employer's later accountability decision. That is a bounded gap in the sources reviewed, not proof that such events never happen.

Failure mode one: approval without practical control

Keeping a human in the process can make responsibility look simple. The manager approved the recommendation, so the manager owns the outcome.

Formal authority and practical control are not the same thing. A reviewer may lack the information to spot an error, the time to investigate it, or the right to change a binding organizational target. An override can exist in the software and still be expensive, slow, or politically difficult to use.

Research outside hotels offers a warning. Arnestad, Meyers, Gray, and Bigman (2024) found that observers assigned more blame to a person after AI mistakes when a manual-control option was available, even if the person didn't use it. The experiments covered automated driving, medical decisions, and investing. In a final study, the authors examined perceived causation and counterfactual thinking as explanations for the effect: observers could imagine that the person should have intervened.

That is evidence about observers' moral judgments in experimental settings. It is not evidence that hotel employers routinely sanction general managers, or that every managerial approval button is illusory.

The hotel evidence closest to restricted control concerns the people doing the work, not their managers. Spektor and colleagues (2023) documented room attendants' reported constraints under algorithmic room assignments. Wang and Kunasekaran (2026) reported associations between perceived algorithmic scheduling exposure, lower control over work time, and work-family conflict among surveyed hotel employees in China. That study is cross-sectional and does not establish causation.

On managers specifically, the limited hotel evidence is more mixed. In Jianu, Ashton, and Lugosi's (2025) three-stage Delphi study, the statement that managerial control remains unless algorithmic management is fully automatic received 73.34% agreement among respondents to that item, which the authors categorized as weaker consensus. That figure reports expert expectations, not observed hotel-manager control or a representative industry estimate. The study also identified managerial translation, communication, and coaching as important emerging functions. The Spektor study, for its part, shows why local product settings and the experiences of frontline room attendants matter. It does not independently establish which override rights a particular hotel's supervisor had.

So the manager can be caught in the same architecture, but the reviewed hotel research has not measured the proposed manager-level blame mechanism directly.

Failure mode two: real discretion used poorly

More human control isn't automatically the fix.

Sometimes a manager has good information the model missed. Sometimes the manager adds noise, or a familiar bias, to a capable forecast. Both possibilities have shown up in hospitality research.

Koupriouchina, van der Rest, and Schwartz (2023) analyzed 20,081,973 hotel occupancy forecasts from 1,752 properties across multiple forecast horizons. Higher override frequency was associated with better accuracy for group demand and around special events, but worse accuracy for transient demand. A single rule to override more, or less, would miss that split. The study is about revenue-management forecasts, and its findings shouldn't be passed off as direct evidence on housekeeping labor decisions.

Schwartz and Cohen (2004) experimentally varied the interface through which 57 experienced hotel revenue managers received the same forecast. The presentation affected how much they changed it. That matters because part of what looks like managerial judgment can come from interface design, before the manager has weighed any new local information.

Transferred field evidence points the same way. Kim and colleagues (2024) examined restaurant inspections where public inspectors kept final discretion. Inspectors were less likely to visit restaurants selected only by an algorithm, even though those restaurants had more recorded violations when inspected. The authors did not measure blame, and a public inspector's objective differs from a hotel's competing labor, guest, and employee priorities. Still, the study challenges the assumption that final human discretion necessarily improves a system's decisions.

Together, the studies caution against assuming that a manager's opportunity to intervene necessarily improves the result. Hotel override frequency was associated with better or worse forecast accuracy depending on demand segment; an interface experiment changed adjustment behavior; and inspectors with real discretion underused more predictive recommendations. They do not establish why an individual manager chose a particular action.

One organizational possibility, and it is an HDL inference rather than a finding, is that overrides never get compared with what the forecast would have done. A recurring pattern then stays invisible to the people who could correct it. The manager holds real control, and nobody systematically reviews how it's used.

The two failure modes don't force a choice between more human control and less. Both lead to the same question: which decision rights help on which task, under what operating conditions?

Configuration connects the two failures

Picture two hotels using a similar staffing product. At one, a department head sees the forecast's uncertainty, can inspect the relevant inputs, and can change a recommendation within defined bounds. At the other, the department head sees a number to approve and needs remote authorization to change the target.

Now picture a third property that allows unlimited adjustment but never compares those adjustments with what the forecast would have achieved.

All three could say they have a human in the loop. They don't have the same accountability architecture.

The three hotels are a proposed illustration, not an observed comparison. They show why the approval click can't be the whole unit of analysis. An accountability architecture can be described at three moments:

• Before the recommendation: who set the objective, which data the system can see, what the manager is allowed to change, and how hard it is to escalate.

• At the decision: what the manager knew that the system didn't, what she did, and why.

• After the outcome: what the result gets compared against, and who is asked to explain it.

The first property is set up well before the recommendation. The second isn't, and that's where the first failure mode lives. The third falls short after the outcome, which is where the second failure mode lives. The audit later in this article follows the same order.

An experienced operator might push back here. Hotel leaders are judged on results for good reasons. Owners pay for outcomes, outcomes are visible, and reviewing the reasoning behind every decision is slow and can turn into excuse-making. That's a fair objection. The audit below keeps outcome accountability and adds just enough of a record to separate three explanations for a bad day: the setup limited the options, the manager misjudged, or the day was unlucky.

A researcher might push back differently: none of this is unique to AI. The Ozer study supports that point, since the source of the advice made little difference to blame. A reasonable hotel implication is that AI tools change where the choices sit and what gets recorded. More of the decision gets set upstream, in configuration. A recommendation system can also keep a record of what it recommended, which makes it possible to compare the manager's call with the alternative. Studies like Koupriouchina et al. (2023) depend on exactly that kind of record: the system's forecast and the human change, both on file.

The organizing proposition is conditional. Accountability can become misaligned when the authority someone appears to hold differs from what they could practically know and do. Generous discretion can become a problem too, when repeated overrides make decisions worse. The existing evidence supports studying both. It does not support declaring which is more common among hotel managers.

What the hotel evidence does not yet settle

The evidence for the two failure modes is not equal.

• Approval without practical control: Preliminary. The mechanism appears in experiments outside hospitality, and hotel research documents constrained control among room attendants. No reviewed study directly measures this accountability failure mode for hotel managers.

• Real discretion used poorly: Hotel revenue-forecasting studies found that override frequency was associated with better accuracy in some demand segments and worse accuracy in others, while an experiment found that interface design affected how much managers adjusted forecasts. They do not directly measure whether managerial discretion is appropriately reviewed or assigned accountability. Evidence for that accountability failure mode in hotel labor decisions remains preliminary.

Hotel research on algorithms and human judgment isn't thin. Beyond the studies above, Ivanov and Webster (2024) surveyed hoteliers about their preferred degree of automation across decisions, which captures attitudes rather than observed override behavior. Li, Zheng, and Lin (2026) describe a demand-prediction and staff-scheduling optimization system for one hotel. Using the hotel's historical data, the authors compared a system-generated September 2023 schedule against the hotel's manual schedule and actual September demand. Under their stated cost assumptions, they reported 11.35% lower staffing cost and 81.95% lower schedule-adjustment cost. This was a retrospective evaluation, not a reported live deployment, and it did not measure changes in managers' accountability.

What this reviewed set lacks is a joined record: the system recommendation, the relevant inputs, the manager's real intervention rights, the reasons for accepting or overriding, the effects on employees and guests, and how the organization later allocated responsibility. Without that chain, the statement that managers are becoming AI's scapegoats would be stronger than the data.

An operator audit worth trying

Hotel Decision Lab proposes a practical way to make the problem visible. Take one recurring, consequential labor decision, such as next-day housekeeping coverage, and reconstruct it end to end with five questions.

1. Who owns the target? Is the tool optimizing a service constraint, labor expense, workload fairness, forecast accuracy, or some combination? If those goals conflict, who decides the tradeoff?
2. What did the system know at the time? Which arrivals, room conditions, labor requirements, and exceptions were included? Which facts were visible only to someone on property? Does the user see forecast uncertainty, or just a recommendation?
3. What could the manager actually do? Can a manager modify the forecast, change the schedule, pause a recommendation, request extra coverage, or escalate? Are those rights direct, bounded, delayed, or contingent on someone else's approval? Are relevant product features switched off?
4. Is the reason on record? Record reasons for consequential decisions, whether the manager follows the tool or departs from it. Start with a sample of decisions, or with high-stakes days, rather than every shift. A useful reason names a specific fact the system didn't have, or an identifiable concern about the forecast. Neither "the algorithm said so" nor "I know my hotel" is enough on its own.
5. How do comparable decisions look over time? Review forecast error by demand segment, the operational outcomes that mattered, and what information was actually available when the call was made. Evaluate avoidable overrides as carefully as unquestioning acceptance. Don't treat a bad outcome as proof that a reasonable decision was bad when it was made.

Run the Saturday call through these questions. Before 6:40, someone had already decided whether the forecast could see the early check-in requests, whether the housekeeping manager could change the labor target, and whether adding an attendant needed someone else's approval. After the day ends, someone will decide whether to judge her on late rooms, on labor hours, or on what she could reasonably have known that morning. Those choices won't tell her what to do at 6:40. They do decide whether anyone can later tell a bad call from a bad setup.

The audit is an HDL proposal, not a demonstrated intervention, and it can be tested. If the first failure mode is real, post-incident reviews should hold approving managers responsible at similar rates whether or not they could see the relevant data or change the target. If the second is real, properties that never compare overrides with the forecast should keep making the same segment-specific override errors, while properties that do compare should see those errors shrink. A proper test would compare properties or decision regimes with different information access and authority, track behavior at the recommendation level, and examine service, labor, employee, and financial consequences rather than only subjective satisfaction.

Bad call, bad setup, or bad luck

Back to 6:40 on Saturday.

The housekeeping manager may be the last person to decide whether to add coverage to that schedule. That doesn't mean she designed the options on her screen. It doesn't mean wider discretion would necessarily have made her call better, either.

The question worth keeping is whether the hotel has built a decision process where authority, information, intervention rights, and later accountability make sense together. That design work happens before the morning staffing call, when the hotel decides what the system should optimize, who can challenge it, and what evidence will count when people disagree.

A hotel's last bad labor day is a fair place to test it. From what was recorded at the time, could anyone tell whether that day was a bad call, a bad setup, or bad luck?

The next research task is to observe that chain in actual hotel labor decisions. Article 05 can examine candidate decision-right structures. Before calling any structure a solution, it will need to show that the structure helps the hotel make better decisions, not merely assign responsibility more neatly.

How this review was done

This article is a research-informed conceptual essay based on a structured integrative review. It is not a systematic review.

Sources came from targeted searches of hospitality, management, and human-automation research on algorithmic recommendations, override behavior, and responsibility attribution, plus the reference lists of the studies those searches found. Hotel studies were prioritized. Research from other settings is labeled as transferred.

AI research assistants supported source discovery, reading, and claim checks under my editorial direction. The research record documents full-text reviews of six sources: Jianu et al. (2025), Kares et al. (2026), Kim et al. (2024), Ozer et al. (2024), Spektor et al. (2023), and Wang and Kunasekaran (2026). A later independent audit also reviewed the full texts of Kim, Kotzian, Li, Ozer, Spektor, and Wang. Access and verification levels differ by source and by reviewer. The Jianu passage on managerial control was independently checked again, and selected other claims were cross-checked separately, but I did not personally repeat all six full-text reviews. The remaining sources were evaluated from available abstracts and publisher sections, and the claims drawn from them are limited accordingly. I am responsible for the article's interpretations and conclusions.

When the article says the reviewed evidence does not show something, that statement applies to this source set, not to all research.

References