AI hotel reporting: plain-language questions on your own data

AI hotel reporting means asking a question in plain language and getting an answer built from the hotel's own records, with the report it read, the filters it applied and the date of the data shown beside it. The software retrieves and summarizes. It does not predict, and a person checks anything that moves money.

For example, a general manager types "how many arrivals tomorrow are unpaid?" and gets a number back. The useful question is the next one: which report did that number come from, what counted as unpaid, and was it read before or after night audit. This article walks through what happens between the question and the answer, so you can judge any plain-language reporting tool, including the one you may already be sold, by what it shows beside its answer rather than by how confident it sounds.

What happens between a manager's question and the answer?#

The question is turned into a defined report or query over the hotel's own records, narrowed by filters and a date range, and returned with those choices visible. Nothing is remembered from other hotels and nothing is estimated. The chain is mechanical: the question is matched to a report or table the hotel already has, filters are applied, rows are counted or summed, and the result is written in one sentence.

Ambiguity is the normal failure, not the exception. "Unpaid arrivals" needs a definition: a balance above zero at arrival, a deposit not received, or a payment not yet matched. A good system states the definition it used, or asks which one you meant. A bad one picks silently.

The moment of capture changes the number. A figure read at 07:00 differs from the same figure read at 23:30, because night audit closes the business day and moves charges, no-shows and room status. The answer should therefore be reproducible: the same question with the same filters on the same business date should return the same number, and a person should be able to re-run the underlying report by hand and compare.

A plain-language answer is a summary of rows, not a new source of truth. If the summary and the report disagree, the report wins and the summary is a bug. Ask for the count behind the answer: "eleven" is checkable, "a few" is not.

Why must every answer show its source, filters and date?#

Because an answer you cannot trace cannot be checked, and an unchecked number becomes a decision. Source, filters and date are what turn a sentence into a figure a finance manager can defend. Three things to demand every time: which report or record was read, which filters were applied, and when the data was captured relative to night audit.

A number without a definition starts an argument between two people who are both right about different things. Stating the definition ends the argument before it starts. Traceability then makes the answer auditable: someone can re-run the same report with the same filters and either match the figure or find the difference.

Corrections should be visible, not silent. A system that voids and reverses a wrong charge rather than deleting it lets a later re-run show the correction instead of a changed past. If the assistant cannot name its source, treat the output as a draft sentence, not a figure, and do not paste it into a board pack. This is the same discipline that applies to any AI output in the hotel, covered more broadly in AI in hotels: a practical guide to uses, limits and safeguards.

What should a morning briefing written from your own numbers contain?#

A short set of figures the hotel already reports, in plain language, each with its comparison and its business date. It is a rendering of existing reports for the morning meeting, not a new source of truth. The briefing must carry the business date and whether it was written before or after night audit, because the same figures move across that boundary.

Every line needs its comparison: yesterday, the same weekday last week, the same period last year, or the budget if the hotel keeps one. A number alone tells a manager nothing. The briefing must not add interpretation it cannot support: "revenue is behind because of the weather" is not in the records; "revenue is 8 percent behind the same weekday last month" is, in an example where the records say so.

Keep it short enough to read standing up. A briefing that repeats the whole report library gets skimmed and then ignored. What belongs on the page line by line is covered in what a GM's morning page should contain; this article covers where the numbers come from. The briefing opens the morning meeting. It does not replace the conversation about what the team will do that day.

How do alerts work when a number moves outside its usual range?#

The system compares today's figure with the range that same figure has occupied in the hotel's own history for comparable days, and flags it when it falls outside. It describes what is unusual for this hotel. It does not predict what the number should have been.

The mechanism has four choices: the metric, the comparison set (same weekday, same season, the last several weeks), the band (the observed minimum and maximum, or a percentile band), and the threshold for raising an alert. The comparison set is a decision the hotel makes and must be able to see. A band built on four weeks of data is thin; a band built across all seasons is meaningless.

Seasonality breaks naive bands. A range learned in February will fire every day in July. Compare like with like: same weekday, same month, same holiday pattern where the hotel has the history. Alerts are description, not prediction: nothing here forecasts demand, recommends a rate or estimates tomorrow. It says only that a number sits outside its own recent range.

Alert fatigue is the real risk. Start with a handful of metrics the team already watches, and add more only when the first set is acted on. Every alert should open the report behind it, so the reader can see the rows rather than trust the flag.

What should the assistant say when the data is missing or wrong?#

It should name what is missing, which records are incomplete and what it can answer instead. A confident guess is the failure mode, not the fallback. Türkiye's Personal Data Protection Authority, in its guide on generative AI and personal data (opens in a new tab) published 24 November 2025, describes hallucination as generated content that does not match reality but often looks plausible and convincing, and asks for checking, regular monitoring and human oversight where needed. NIST's Generative AI Profile (NIST AI 600-1, July 2024) (opens in a new tab) calls the same behavior confabulation, the production of confidently stated but erroneous content, and lists over-reliance on such output among generative AI risks.

In hotel terms, the common gaps are unglamorous: reservations without a rate code, rooms not checked out, restaurant charges posted to a house account, deposits received but not matched to a booking. A good refusal is specific. It says how many records are affected and which step closes the gap, so the answer becomes a task rather than a dead end.

Practical checks before acting: does the answer name a report, does the total reconcile with a figure you already know, and is the date range the one you asked for. When the records are wrong, fix the records. Asking the same question again will not produce a better answer.

Who may ask what?#

Access should follow the person's job, not their curiosity: the assistant should answer within the same permissions the person already holds in the hotel's systems. Map the roles before switching anything on. Front desk needs today's arrivals and room status; finance needs billing and invoices; revenue needs rates and pickup; housekeeping needs room and task status.

A plain-language answer is still a disclosure. Guest identity and contact details are personal data under KVKK (kişisel verilerin korunması kanunu), and a friendly sentence does not make sharing them acceptable. Separate aggregates from records: a count of unpaid arrivals is not the same as a list of guest names and balances, and the two should not travel together by default.

Keep a record of who asked what. If an answer later turns out to be wrong, the question and the report behind it should be reviewable. Access ends with the job: someone who has left should not be able to ask questions through an assistant any more than through the front desk system. This is general practice, not a description of any particular product. Confirm your own setup with your software provider and your legal adviser.

Which questions should a report never answer?#

Questions about cause, intent and what to do next. A report can show what happened in the hotel's records; it cannot tell you why, and it should not choose for you. "Why are we behind last year?" needs evidence from outside the system: a group cancellation, a market shift, a road closure, a competitor's rate. The system can supply the size of the gap, not the reason.

"Should we drop the rate on Tuesday?" is a judgment with money at stake. The assistant can show occupancy, pickup and the rate history; the decision and its signature belong to a person. NIST AI 600-1 (July 2024) warns that human-AI arrangements can lead to automation bias and over-reliance, and a tidy, well-formatted answer invites exactly that. The tidier the output, the more deliberate the check.

Use a decision rule: if the answer would change money, a contract or a guest's stay, a person decides and a person signs. The same line runs through guest-facing AI, where the assistant prepares and the person decides, as covered in AI at the front desk: which tasks change and which stay human.

What to do next#

Pick one question your team asks every morning, write down the report it should come from, the filters it needs and the business date, and check whether today's answer can show all three. That single test tells you whether your reporting is ready for plain-language questions, and it is the same test you should apply to any assistant you are shown, including ours. The PMS assistant of AI Otel is built: it answers questions about a hotel's own property in plain language, carries out routine actions, and a person confirms anything sensitive. Beside it sit a plain-language morning briefing written from the hotel's own numbers, and an automatic alert when a number moves outside its normal range. If you want to see how these answers look against a live question of yours, request a briefing.