An AI system recommends cutting an evening shift. Your department manager clicks approve. If service suffers later, who owns the decision?
The name beside the approval button doesn't tell you the whole story. One team sets the labor target. Another maintains the forecast data. A property manager may adjust shifts but can't change company rules or system assumptions. Formal authority and actual control aren't always the same.
Operational responsibility means carrying out an action. Accountability means answering for decisions and controls within a role's authority. Blame is a judgment about fault, not another word for a disappointing result.
A useful principle is that accountability should follow what each person could actually control, not merely the last person to approve the recommendation.
Understanding is not error detection
A manager can understand why the system expects a quiet evening and still miss an error. The demand forecast might be accurate while the recommended coverage overlooks a required qualification or an employee committed elsewhere. A sound recommendation can also fail in execution.
Research outside hospitality gives reason for caution. In controlled tasks, AI explanations increased acceptance of both correct and incorrect advice without significantly improving human-AI team performance beyond confidence information (Bansal et al., 2021). Repeating the system's reasoning isn't the same as recognizing a bad assumption.
The useful question is: what could make this recommendation wrong here, today? Look for missing inputs, outdated coverage information, unenforced constraints, or an objective that values labor savings while overlooking service consequences.
Set the boundary before the recommendation
These are proposed operating choices, not a proven seven-part hotel method. The exact split of authority will vary by property and organization. For each recurring staffing action, settle seven things before the recommendation arrives.
- Authority. Who can accept, modify, or override the recommendation? Building a draft schedule and sending an employee home don't need identical permissions.
- Modification limits. Set bounds on hours, headcount, skills, and department coverage, and identify which changes need higher approval. Those limits must match the property's policies.
- Pause triggers. Stop for stale data, missing qualified coverage, an unusually large change, a disputed restriction, or a hard-to-reverse personnel action.
- Escalation. Name the next decision-maker, a backup if that person is unavailable, and the temporary operating choice while the issue is unresolved. Silence isn't approval.
- Override evidence. Record the current fact, its source, its timing, and why it changes the decision. Check whether the system already represents it. "I know this property" isn't enough.
- Feedback. Keep the prediction, recommendation, authorized change, and executed action distinct. Route recurring missing inputs to the data owner, broken constraints to the policy owner, and forecast problems to the system owner.
- Monitoring and accountability. Check service, labor, employee effects, repeat exceptions, missed errors, and unnecessary pauses. Assign corrective work to whoever can actually change the process involved.
A Saturday staffing decision (hypothetical)
Say a hotel's Saturday arrival forecast looks light, and the system recommends releasing a front-desk agent at 6 p.m. The front office manager has a revised agenda for a conference next door. The event will now finish earlier than planned, so attendees may check in together. The forecast doesn't reflect the timing change.
Under a hypothetical preapproved boundary, the manager can extend the agent's shift by two hours, document the revised agenda, and notify the team. If the change exceeds the manager's hour limit, affects a required rest or pay condition, or moves an employee into a role they can't cover, the decision pauses and goes to the named approver.
The next day, the hotel checks arrivals, lines, actual hours, employee impacts, and service problems. It also looks into why the revised event timing was missing.
Even if the evening went well, that alone can't prove the override was the best choice. The alternative was never observed under identical conditions.
What the evidence can and cannot establish
An override is most defensible when it adds timely, material information the system is missing, rather than ceremonial approval or unsupported confidence.
The studies cited here come from outside hospitality. A field forecasting study at a bicycle company found value in structured expert input in some settings, alongside systematic optimism in others (Nair & Huchzermeier, 2024). Experiments on algorithm adjustment found that giving people control can increase use without reliably improving decisions (Dietvorst et al., 2018; Sele & Chugunova, 2024). None of them validates a hotel staffing override rule.
In the evidence Hotel Decision Lab has reviewed, hotel-specific research establishes relevant applications and manager preferences but hasn't validated these decision-rights practices against downstream hotel outcomes. Much of the transferred evidence is experimental. Some is observational or field-based, from other industries.
Hypothetical observer judgments about blame, even if measured, wouldn't establish actual organizational consequences, discipline, or liability.
Forecast accuracy isn't decision quality either. A staffing action carries service, employee, financial, and execution consequences beyond the forecast. Responsibility for acting, blame for a mistake, and formal accountability should be examined separately.
Start with one decision
The goal is a decision process whose limits and results can be evaluated, not more approval clicks or maximum automation.
Pick one recurring action, such as reducing evening front-desk coverage. Write down who can change it, when it should pause, who resolves a disagreement, and where the outcome gets recorded. Then review accepted recommendations and overrides alike.
Drew Potter founded Hotel Decision Lab, which studies decision authority in AI-supported hotel operations. hoteldecisionlab.com