When Hotel Managers Override the Algorithm

At 2:30 p.m., a hotel’s labor system recommends sending two front-desk employees home early.

The recommendation appears reasonable. Occupancy is below forecast. Check-in volume is lighter than expected. The property is already running above its labor target.

The manager knows something the system may not.

A youth sports group is arriving three hours late. One of the remaining agents is still in training. The lobby bar is short-staffed and has already asked the desk for help. A group billing problem is likely to require managerial attention during the evening rush.

This is a hypothetical example, but the underlying decision is familiar.

Should the manager accept the recommendation, modify it, reject it, pause to investigate, or escalate the decision?

Hotel companies often respond with a reassuring phrase:

Keep a human in the loop.

That phrase sounds responsible because it preserves visible human control. But it says very little about what the human is expected to know, detect, or do.

A manager may be required to approve a recommendation without understanding the system’s limitations. The approval may be perfunctory. Disagreement may create administrative or political friction. The manager may possess relevant local information, or may simply feel confident that experience is superior to the model.

All of these situations contain a human in the loop. They do not create the same decision process.

Defining the problem

Several terms need to be defined before evaluating that argument.

The goal is not maximum trust, maximum acceptance, or maximum automation. It is calibrated reliance based on comparative capability (Lee & See, 2004).

Why hotels are a useful setting for this question

Hotels are not merely a convenient industry in which to apply general AI research.

Hotel labor decisions combine characteristics that make human-AI decision rights unusually difficult:

• Service capacity is perishable.
• Demand can change quickly.
• Operations continue around the clock.
• Departments are interdependent.
• Service failures become visible immediately.
• Staffing decisions affect employee hours and income.
• Property conditions vary locally.
• Managers often have limited time to decide.
• Corporate goals and property realities may conflict.
• Authority is distributed across owners, brands, management companies, regional leaders, property managers, and technology providers.

A forecasting error in another industry may become visible weeks later. A poor hotel staffing decision may produce a line at the desk within thirty minutes.

That combination of volatility, immediacy, human consequences, and distributed authority makes hotels a valuable setting for studying when human judgment complements algorithmic recommendation and when it undermines it.

There is not one hotel labor decision

Broad discussions of AI-supported management often treat forecasting, scheduling, and real-time personnel actions as though they were one decision. They are not.

Prediction decisions

These include forecasting occupancy, predicting arrival patterns, estimating housekeeping workload, and forecasting required labor hours. The central question is comparative predictive capability: which method produces the strongest forecast under relevant conditions?

Resource-allocation decisions

These include building a weekly schedule, assigning rooms or workloads, allocating departmental coverage, and distributing overtime. The questions extend beyond prediction to operational constraints, fairness, employee preferences, legal requirements, and the objective being optimized.

Real-time personnel decisions

These include sending an employee home, calling someone in, extending a shift, closing an outlet, or moving an employee between departments. They incorporate current operating conditions, employee consequences, service risk, reversibility, and limited time.

A model may produce an accurate demand forecast while the staffing recommendation built from it is poor. The recommendation may be reasonable while the property executes it badly. A manager may correctly recognize an exception but choose the wrong response.

The decision process should therefore distinguish among:

1. The prediction
2. The translation of that prediction into a recommended allocation
3. The final operating action
4. The execution of that action
5. The outcome

Without this separation, organizations may blame “the AI” for a scheduling-policy failure or credit “manager judgment” for an outcome that resulted from luck.

What the evidence establishes

The Hotel Decision Lab review identified 25 high-value sources. Seventeen full texts or accepted manuscripts were reviewed, and 15 of the most relevant received detailed evidence dossiers. The review separated hotel-specific evidence from research transferred from forecasting, human-factors, consumer, retail, and organizational settings.

Across that evidence, algorithm aversion was not a universal response. Expectations, expertise, autonomy, incentives, interface design, and competing ideas about rationality all influenced whether people accepted or rejected algorithmic advice. The relevant behavior was discriminatory use, not blanket trust or rejection (Burton, Stein, & Jensen, 2020).

Control increases adoption more reliably than decision quality

Dietvorst, Simmons, and Massey found that people were substantially more willing to use an imperfect algorithm when they were permitted to modify its output.

In one experiment, algorithm use increased from 32 percent without modification rights to 73 percent with bounded modification. Broader adjustment authority produced a similar adoption rate.

The performance effects differed by the amount of control. Bounded changes preserved much of the model’s advantage, while unrestricted changes harmed accuracy. Control increased satisfaction and willingness to use the system without making every intervention beneficial (Dietvorst, Simmons, & Massey, 2018).

Sele and Chugunova reported a related contradiction. Allowing participants to monitor and adjust algorithmic recommendations increased algorithm uptake and participant confidence. Yet final decisions were less accurate in the human-in-the-loop condition.

Participants were also least likely to correct the largest recommendation errors, where effective intervention should have mattered most (Sele & Chugunova, 2024).

These studies did not involve hotel staffing, employee income, or guest service. They should not be treated as direct evidence that manager review harms hotel decisions.

Giving someone authority to intervene does not establish that the person can identify when intervention will improve the decision.

Human judgment adds more value when it adds information

Human knowledge is most valuable when it contributes a signal missing from the model.

Nair and Huchzermeier compared expert judgment, machine-learning forecasts, and an integrated approach using demand data from a bicycle manufacturer. Machine learning outperformed judgment alone. The integrated approach performed better than either one separately.

Expert information added more value for innovative and new products, where historical patterns were less useful. Historical data contributed more for stable, carried-over products. The experts also demonstrated systematic optimism in some settings (Nair & Huchzermeier, 2024).

The hotel implication is conditional. A manager’s knowledge is more likely to improve a recommendation when the manager can identify a current fact that is absent from the system, material to the decision, and not already captured indirectly in the model.

The broader forecasting evidence does not support unrestricted discretion or the assumption that every adjustment based on experience improves the forecast. Small habitual adjustments, optimism, and intervention in stable settings can introduce noise.

Expertise can produce signal or confidence-driven discounting

Hotel organizations understandably value experience. Experienced managers may understand property conditions that cannot be seen from a corporate dashboard. But experience is not equivalent to incremental information.

Logg, Minson, and Moore found that nonexpert participants often placed more weight on advice when it was described as algorithmic. Domain experts relied less on advice overall and discounted algorithmic advice more heavily.

In one expert comparison, the experts’ resulting probabilistic judgments were less accurate than those of nonexperts receiving the same algorithmic advice (Logg, Minson, & Moore, 2019).

This does not prove that hotel managers are inferior to algorithms. It does show why title, tenure, and confidence should not automatically determine override authority. Experience may increase access to useful local information. It may also increase confidence in patterns that are no longer predictive.

Explanation does not necessarily improve error detection

Explanation is often presented as the solution to weak trust or poor adoption.

Bansal and colleagues tested whether AI explanations improved complementary human-AI performance. Explanations increased reliance when the AI was correct, but also increased reliance when it was wrong. They did not significantly improve performance beyond recommendation and confidence information, producing little net gain in the tested tasks (Bansal et al., 2021).

The relevant design question is not simply whether the manager understands why the recommendation was made. It is whether the explanation helps the manager recognize missing data, weak model coverage, unfamiliar operating conditions, high uncertainty, known failure modes, or a mismatch between the system’s objective and the hotel’s actual priorities.

A persuasive explanation and a diagnostically useful explanation are not the same thing.

Visible errors can create excessive rejection

Dietvorst, Simmons, and Massey found that people lost confidence in algorithms more quickly after seeing them make errors, even when the algorithm still outperformed human forecasting. Visible imperfection caused people to reject a comparatively stronger method (Dietvorst, Simmons, & Massey, 2015).

That does not mean every negative reaction to a system failure is irrational. A hotel labor system may be more accurate on average while producing rare failures with unacceptable consequences. An average-performance advantage does not automatically justify automation when errors are difficult to detect, severe, or irreversible.

The correct comparison is not, “Did the system make a mistake?” It is:

How does the system perform relative to available alternatives under comparable conditions, and what are the consequences of its errors?

Acceptance is an inadequate measure of success

Research on AI reliance distinguishes following behavior from signal interpretation, belief formation, and expected decision value. Common measures such as accepting correct recommendations or rejecting incorrect ones can produce conflicting conclusions about reliance quality (Guo et al., 2024).

A high acceptance rate in a hotel could reflect:

• Strong recommendations
• Automation bias
• Organizational pressure
• Superficial approval
• Administrative friction attached to disagreement
• A belief that overriding the system is politically unsafe

Acceptance should not be treated as the primary success measure.

What the hotel-specific evidence shows

Hotel-specific evidence

Hotel research supports the relevance of the problem but does not resolve it.

Ivanov and Webster surveyed 130 hotel managers across 23 decisions in eight functions. Managers were generally positive toward AI but preferred to retain control. Human-out-of-the-loop arrangements were least preferred, especially for decisions perceived to require emotional intelligence.

The study establishes stated preferences. It does not show that retained control produces better decisions, nor does it measure live override behavior, labor outcomes, employee consequences, or guest outcomes (Ivanov & Webster, 2024).

A systematic review of 80 hospitality employee studies found that employee responses to AI were associated with usefulness, ease of use, trust, risk, uncertainty, organizational support, AI awareness, and anticipated work consequences. Much of the reviewed literature focused on service robots and Asian contexts, leaving limited direct evidence about algorithmic staffing or manager overrides (Kumawat et al., 2025).

Hotel-specific technical evidence also cautions against assuming superior capability from greater complexity. Kozlovskis and colleagues compared multiple machine-learning approaches with an econometric benchmark using occupancy data from one hotel. The strongest machine-learning specification did not outperform the benchmark (Kozlovskis et al., 2023).

Within the reviewed set, no full-text hotel study directly measured managers accepting, modifying, overriding, pausing, investigating, or escalating an AI-supported labor recommendation and then compared downstream operating outcomes.

That is a result of the current search, not proof that such research has never been conducted.

Three proposed conditions for evaluating human review

These conditions are not independent. A manager may possess excellent local information but lack the authority to act on it. Another may have authority but lack enough information about system capability to diagnose a failure. A third may have both, but operate under incentives that reward labor reduction regardless of service or employee consequences.

Taken together, the three conditions provide a proposed basis for evaluating whether review may function as a safeguard. Their joint effect has not been validated in hotel operations.

When human review can be a real safeguard

Human review is more likely to improve an AI-supported labor decision when:

• The manager possesses information unavailable to the system.
• That information is material to the specific decision.
• The system’s important errors are observable to the manager.
• The manager understands the decision, not merely the software interface.
• The consequences are serious enough to justify review but intervention remains possible.
• The manager has a defined action beyond approval or rejection.
• Incentives do not punish responsible disagreement.
• The result of the intervention can be evaluated.
• Useful overrides are incorporated into future learning.

Under these conditions, human review may identify missing context, limit harm, and improve the combined decision. Without them, review may increase confidence, diffuse responsibility, or create unmanaged discretion.

Power and incentives are part of the decision system

Human-AI decisions do not occur in an organizational vacuum.

Bader and Kaiser’s qualitative case study of an AI-supported call-center system found that employees possessed conversational context missing from the system, but professional identity and organizational incentives also contributed to workarounds and data manipulation. Human involvement was shaped by the interface and organizational setting rather than simply added on top of the technology (Bader & Kaiser, 2019).

This matters in hotels.

A property manager may follow a labor recommendation they distrust because ownership has made labor reduction the dominant priority. Another manager may override because the system threatens professional autonomy. A third may approve the recommendation because disagreement creates political risk or additional work.

These behaviors cannot be explained by trust alone.

The organization should ask:

• Is the manager genuinely permitted to disagree?
• Which outcomes are rewarded?
• Who bears the consequences of a bad decision?
• Does the approval process create meaningful review or merely transfer blame?
• Does the system encourage managers to continue gathering local information?

Formal theory suggests that allocating authority to AI may also change the human incentive to acquire private information. If local knowledge no longer affects the decision, managers may invest less effort in producing it. This effect has not been demonstrated in hotels, but it is an important theoretical risk (Athey, Bryan, & Gans, 2020).

Accountability is not one thing

Organizations frequently say that the manager remains accountable. That statement may hide more than it clarifies.

A responsible governance model should distinguish:

• Decision authority: Who is permitted to make the choice?
• Operational responsibility: Who must execute it?
• Answerability: Who must explain the result?
• Liability: Who bears formal consequences?
• Model ownership: Who is responsible for capability, data quality, and maintenance?
• Governance ownership: Who determines how authority is allocated?

A manager may approve a system-generated schedule without controlling the objective function, forecast model, labor standard, input data, corporate target, or interface design.

Calling that manager fully accountable may create the appearance of governance while leaving the actual decision structure untouched. Accountability should follow control.

A proposed Hotel AI Decision Rights Process Model

The following model is a conceptual synthesis of the reviewed evidence and hotel operating logic. It has not been empirically validated.

The purpose of recording overrides is not simply to monitor compliance. It is to learn whether the system missed information, the manager introduced value, or the intervention reduced decision quality.

Design-time authority and run-time response

Decision authority should be established before the live decision occurs.

Conceptual research distinguishes among full AI delegation, sequential human-AI arrangements, and aggregated decision structures. The appropriate structure may depend on task specificity, interpretability, speed, repeatability, and the number of available alternatives (Shrestha, Ben-Menahem, & von Krogh, 2019).

At the design level, a hotel organization might assign one of four proposed modes:

1. Human-led: The manager owns the decision and uses system information as an input.
2. AI-recommended: The system proposes an action, but the manager decides.
3. AI-generated with required review: The system creates the intended action, but implementation requires a defined human review.
4. Automated with human monitoring: The system acts unless a threshold, exception, or later review triggers intervention.

These categories are a proposed vocabulary rather than a validated framework. At run time, the person should still have clearly defined response options. “Human in the loop” does not explain whether the manager can adjust, override, pause, investigate, or escalate.

Six article-level propositions for future research

The current evidence supports propositions that can guide future hotel research without pretending that the answers have already been established.

What hotel companies should do now

These propositions are not validated hotel rules, but they provide a defensible starting point.

1. Separate prediction, allocation, and personnel-action decisions.
2. Test systems against relevant baselines before expanding their authority.
3. Define what information justifies an override.
4. Evaluate whether reviewers can detect consequential errors.
5. Use bounded modification where unrestricted intervention is unlikely to add value.
6. Preserve pause and escalation as legitimate responses.
7. Measure outcomes rather than acceptance alone.
8. Record structured override reasons.
9. Review both system failures and manager interventions.
10. Assign accountability according to actual control.

Technology providers should design for diagnosis rather than persuasion. Managers need to know what the system sees, what it does not see, where it performs well, and where intervention may be warranted.

The real standard

The hotel industry does not need managers who always follow the algorithm.

It also does not need systems that defer to every assertion of local experience.

It needs a decision process that compares the capabilities and limitations of both.

The system may be better at identifying patterns across thousands of historical observations. The manager may be better at recognizing a local condition that has not yet appeared in the data. Either one may introduce error.

Human review may add value when it contributes missing information, detects a meaningful failure, and occurs within a governance structure that permits useful intervention. Current evidence does not establish this as a validated rule for hotel labor operations.

Without those conditions, keeping a human in the loop may improve comfort, adoption, or political defensibility without improving the decision.

The question is not simply whether a hotel manager should trust the algorithm. The better questions are:

What does each party know?

Which errors can each party detect?

Who has meaningful authority?

Who bears the consequences?

How will the organization learn from the result?

Putting a human in the loop is not the end of responsible design.

It is the beginning of a decision-rights problem.