Stakeholder perspectives on integrating explainable artificial intelligence into hotel revenue management system

Black-box revenue management systems

Since RM was adapted from airlines to hotels in the 1970s and 1980s, technology has played a central role in supporting economic decision-making—market analysis, demand forecasting, price-inventory optimisation, and performance evaluation (Denizci Guillet and Mohammed, 2015). Over time, this led to the development of RMSs—initially simple software—now transformed into advanced, data-driven tools powered by algorithms, machine learning, and predictive analytics (Nam et al., 2021). As complexity increased, RMSs began operating as “black boxes,” generating outputs that lack transparency and clarity. This has led to mistrust (Yeoman, 2019), frequent overrides (Mohammed and Denizci Guillet, 2025b), and algorithmic system aversion (Mohseni et al., 2021). In response, industry leaders like Richard Ratliff of Sabre Labs have advocated for more transparent systems, urging the shift from black boxes to “glass boxes” (Walson, 2022: para 20).

Glass boxing RMSs and XAI

A glass (or white or grey) box is a term used to describe a transparent system or a system that is not a black box (Rai, 2020) and the process of transforming a black box into a glass box is commonly called glass boxing (Mohseni et al., 2021). Generally, a system is deemed a black box when its construction, internal functions, logic, and parameters are opaque, making it difficult to understand and explain (Zhang et al., 2025). This opacity has led to interest in XAI as a possible solution to enhance transparency and explainability (Mohseni et al., 2021). Although the notion of explainable expert systems dates to the 1970s (Miller, 2019), the term explainable AI was coined in 2004 by van Lent et al. as AI systems that can explain their rationale to a human user, characterise their strengths and weaknesses, and convey an understanding of how they will behave in the future (Gunning and Aha, 2019).

Explanations

In XAI literature, Arrieta et al. (2020) state that, “given a certain audience, an explanation refers to the details and reasons a model gives to make its functioning clear or easy to understand” (p. 85). Based on the interpretation scale, Mohseni et al. (2021) distinguish two types of explanations: global versus local. Global (or model) explanations thoroughly describe how a model works, while local (or instance) explanations elucidate the reasoning behind a particular prediction (Mohseni et al., 2021; Doshi-Velez et al., 2019). Miller (2019) categorises explanations based on the type of questions they answer: “what” (associative reasoning), “how” (interventionist reasoning), and “why” (counterfactual reasoning). Expanding on Miller’ (2019) categorisation for XAI, Mohseni et al. (2021) generalise explanations into six question-driven types: how, why, why-not, what if, how-to, and what else, which are utilised in this study.

RM professionals’ desires to understand black-box RMSs – what should be explained

RM professionals expressed varying levels of desire for explanations about black-box RMSs, with clear differences between end-users and system providers. Most end-users wanted specific, situational explanations, such as demand forecasts or price recommendations for a certain day or period, especially when system outputs deviated from expectations. One participant said, “I want to know why the system increases the room night rates for specific months and not the entire period.” (P18). Another added, “I don’t want to know the algorithm or the calculation behind the whole system. Instead, what I want is very simple explanations of the algorithm’s results and input data.” (P12).

Some end-users also wanted a broader understanding of how the system functions overall. As one explained, “I want to understand more about how it [the system] generates its recommendations and the steps it takes to generate them.” (P14). Another echoed, “I want to understand how the system works and what it generates.” (P2). In contrast, system providers preferred general explanations to identify exceptions, validate outputs, and debug. One noted, “I need explanations for exceptional patterns so that I can investigate them to see if the system is making the right decision or not.” (P27). Another added, “I wish the system can spot outliers and alert me where there are data concerns for interventions.” (P7).

Despite these differing expectations, there was agreement on the need for improved understanding. However, a tension exists between user demands for clarity and provider concerns about revealing proprietary information. One provider commented, “We need to understand the amount of detail that people want, need, and should get because those three things aren’t the same.” (P19). As a result, end-users often felt their questions were met with insufficient answers, despite acknowledging, “Nobody will be able to give you 100% of the information because of proprietary reasons.” (P1).

Beyond proprietary issues, competition in the hotel industry also discouraged transparency. One end-user observed, “The competitive nature of the hotel industry makes full transparency a double-edged sword; it can promote deeper understanding and collective learning, just as it can compromise competition and encourage price fixing.” (P2). System providers justified limited transparency as necessary for operational protection. One explained, “Sometimes, we do see that the lack of transparency is a protection for the hotels’ operations.” (P6). These statements suggest that the lack of transparency in RMS could be deliberate.

XAI’s capabilities and professionals’ beliefs

XAI is known to offer both global (model-level) and local (output-specific) explanations (Mohseni et al., 2021), as well as static or interactive formats (Arya et al., 2019). These capabilities suggest that XAI can meet RM professionals’ task requirements. However, whether this alignment will support adoption remains uncertain. Two insights emerged. First, many users lacked awareness of XAI’s capabilities and questioned its effectiveness. One asked, “If the human agent cannot offer satisfactory explanations, how can we have a technology capable of doing it right?” (P1).

Second, system providers doubted whether end-users could understand the complexity of XAI’s explanations. One stated, “Imagine if you have a prediction with thousands of variables, how can XAI explain that to a user to grasp fully?” (P7). Another added, “I don’t think revenue management systems’ users are going to understand how a machine-learning system comes up with its recommendations even if the system could explain itself.” (P20). An end-user agreed: “Being able to explain is one thing, and understanding is another.” (P26).

Simplification and brevity

Participants stressed the need for concise, easily digestible explanations. One said, “I prefer it in a bullet-point summary of no more than 250 to 300 words” (P3). Another quizzed, “If I am a third grader and it has all these numbers and variables, will I be able to understand it without being overwhelmed?” (P20).

Contextualisation and personalisation

Participants found current RMS explanations too generic and static. They called for context-specific and user-adapted outputs. As one put it, “If I want to drill down further, I am unable to do so because the answers are standard and not tailored to my needs” (P12). Another added, “We get generic explanations from them. XAI’s explanations should be contextualised, providing a little bit more specifics on how it works” (P23).

Interactive conversation and gamification

Participants favoured conversational, human-like explanations delivered in natural language. One explained, “If the XAI can explain to me in a natural language using a human-like voice, that will be great” (P13). Another supported a “more natural language way” (P8). Gamified elements were also suggested to boost engagement. One participant proposed, “One way to achieve an engaging explanation is to gamify it so that a little endorphin is kicked in when people interact with the system” (P17).

Technology

The identified technological factors were grouped into four thematic domains: compatibility with existing systems and adaptability, availability and access to quality data, vulnerability and risks, and dependability of the new system. Each domain had two sub-themes, with mentions ranging from 2 to 11. Privacy (11) and security (9), under vulnerability and risks, were the most frequently cited, followed by compatibility (8). Availability and access to quality data was also seen as a major barrier, as XAI depends on clean, reliable data. One respondent noted, “For XAI to work well with RMS, you need good quality clean data, which is a challenge.” (P10). Another added, “It is hard to find the right data. Hotel companies do not give third party access to their data because it reduces their value proposition.” (P15). Although dependability of the new system had the fewest mentions, concerns centred on the validity and reliability of XAI explanations. This may reflect limited awareness and a need for broader organisational training.

Organisation

Participants identified four organisational factors affecting XAI adoption: size and resources (8), culture of adoption and change (4), ownership structure and management (9), and RM executives’ readiness (7). Resources and ownership were the most cited, reflecting concerns about limited funding or lack of interest from hotel owners. One participant noted, “It will be a little bit hard for the hospitality group, like [Name withheld], to do this because of the R&D cost.” Another added, “It will be driven by the personalities at the top and the day-to-day demands from the revenue manager.” (P19).

At the operational and individual level, views ranged from fear to optimism. One respondent reflected, “Whenever I hear sentiments like that, they sound a bit like fear mongering… a bit like, people are nervous about their jobs being taken away… However, as hotel revenue managers, we can be XAI enhanced, not replaced.” (P17). Generational differences also shaped adoption views. As one participant put it, “Different generations have different viewpoints on whether technology is good or bad. So, these perceptions will influence the adoption.” (P15).

Environment

The analysis revealed three thematic domains of environmental factors: regulatory lapses and deficiencies, competitive pressures, and green concerns and the planet. The regulatory lapses and deficiencies, with three sub-themes (ethics, regulatory silos and responsibility), was mentioned most frequently. On the sub-theme of ethics, one respondent noted, “The unregulated use of IP [intellectual property] to train the machine and generate content is a serious concern.” (P17). Another lamented the lack of global Internet regulations and mechanisms of seeking protection. She stated, “…. there is no global Internet regulation. The same goes for AI which is of great concern.” (P21). These perspectives highlight the need for stiffer regulations, and governments can lead the role.

Enhancing transparency and understanding

The participants agreed that integrating XAI into RMSs can enhance transparency and contribute to a better understanding of the systems’ outputs and recommendations. One participant expressed this benefit as follows: “The enhanced transparency will allow RM executives to understand why certain prices are recommended, enabling them to effectively communicate these prices to customers and other stakeholders” (P9).

Promoting human-system interaction

Given the capabilities of XAI to explain its outputs and how it works, the participants believed that when integrated into RMS, it could promote human-system interaction and engagement to reduce overrides and biases. In support of this, one participant remarked, “It will make users feel more involved and less intimidated by the technology.” (P17).

Building stakeholder trust

The participants argued that having an interactive XAI in RMS can foster stakeholder trust since the interaction will align users’ expectations with systems’ output, which is crucial for adoption and use. This point was made by two participants as follows: “Knowing how it works will make me trust the system” (P8) and “In order to trust the system, I need to know how it works” (P16).

Increasing user confidence

Considering the human-centredness of XAI, the participants articulated that integrating it into RMS can increase user confidence. By explaining questions like why-not, users can obtain counterfactual evidence to boost their confidence in the system’s recommendations vis a vis their expectations. One respondent stated, “If the XAI can explain to me why the RMS recommends a particular price instead of another, it will boost my confidence to accept or override its recommendation” (P9).

Improving decision quality and efficiency

The participants noted that since XAI can provide insights to help developers fine-tune its results or debug any operational failures, the systems’ output and decision quality can improve. This conviction was expressed by a participant, stating, “It takes a lot of time to figure out small issues or problems, sometimes, an hour to find an issue with a rate. So, if I have a tool that could help me to understand where the issue is, then I think I will be more efficient.” (P13).

Growing adoption and use

The participants argued that when the integration of XAI improves decision quality and performance, large hotel companies and RMS providers are bound to increase their R&D and investment in XAI-enabled RMS. This could lead to a widespread adoption and use of these technologies. One respondent noted, “People need to see the wins before they decide. A series of small wins can make a huge difference.” (P26).