Hotel data quality before AI: what to clean up and why

Hotel data quality is the accuracy, completeness and consistency of a hotel's own records: room types, rate plans, guest profiles, segment codes and folios. Before any AI assistant reads them, fix the records it will answer from, in this order: inventory, rates, guest profiles, codes, open balances. Then keep them clean monthly.

Ask an assistant how many rooms of a given type you can sell next Tuesday and it will answer in one clear sentence. If your system holds two room types that describe the same physical rooms, that sentence will be confident and wrong, and nobody will see the contradiction the way they would see an empty column in a report. This article gives you an order for the cleanup, a check for each record type, and a monthly routine that keeps the work from undoing itself. For a broader view of where assistants fit in a hotel, see AI in hotels: a practical guide to uses, limits and safeguards.

Why does AI make bad hotel data visible faster?#

Because an assistant writes fluent sentences from whatever the records contain, a gap or a contradiction comes back as a stated fact instead of a blank field that a person would stop and question.

A report shows its working: an empty column, a zero, a total that does not add up. An assistant hides the working and states the conclusion, so the reader has nothing to check against. The cost is asymmetric. A blank field invites a question; a wrong sentence invites a decision.

This is not a hypothetical failure mode. Türkiye's Personal Data Protection Authority describes hallucination as generated content that does not match reality but often looks quite plausible and convincing, notes that answers can be wrong even when they read coherently, and stresses regular monitoring of outputs supported by human oversight where needed (Kişisel Verileri Koruma Kurumu, Üretken Yapay Zekâ ve Kişisel Verilerin Korunması Rehberi (15 Soruda) (opens in a new tab), 24 November 2025). The regulator's point applies with extra force when the assistant is answering from your own records.

Bad records take three shapes. Duplicates: one guest, several profiles. Contradictions: two room types covering the same rooms, or two rate plans with the same rules. Orphans: a rate plan attached to no channel, a folio attached to no reservation. Each shape has its own fix, and the sections below give them in order.

Order the work by how often a record is read, not by how easy it is to fix. Inventory and rates are read on every availability question; a mailing list is read rarely. And do not look for a public figure on how often assistants answer wrongly: none exists, and a vendor's number would not describe your property anyway. Sample twenty answers a month against the underlying records and you have your own measure.

Room types and inventory: how do you get one list with one meaning?#

One room type should mean one physical product with one code, one bed setup, one size and one view, and the count of rooms in the system should match the count of doors on the floor.

Start with a physical count. Walk the floors and count rooms per type, then compare with the system. Every difference is either a data error or a room that has quietly changed use, and both need a decision. This is the check that most often finds a real contradiction, and it takes an afternoon.

Retire, do not delete. A room type that has past bookings stays in the system as inactive, so old reservations and invoices still read correctly. Name for the reader, not for the staff: a guest, a receptionist and an assistant should all understand the same name, and internal abbreviations survive only in the code field.

Stop inventing a new type for a one-off request. Sea view, high floor and connecting door are attributes of a room, not new products, and types multiply when attributes are missing. Keep the mapping between your type codes and the codes used by channels in one place, and check it whenever a channel changes its list.

Give the list one owner. Two people creating room types is how a hotel ends up with a long list of types and a set of physical differences nobody can reconcile, for example eleven types and nine physical differences.

Rate plans: which ones should you retire?#

Retire any rate plan that has produced no booking in a defined window and any plan that duplicates another, and name the survivors by what they actually do.

Set the window before you look at the list, for example twelve months, so the decision is a rule rather than an argument about a favourite plan. Name by function: breakfast included, non-refundable, member rate, long stay. A plan named after a year, a person or a campaign nobody remembers will be misread by staff and by any assistant.

Check each surviving plan for three attachments: a cancellation policy, a meal plan and at least one channel or portal where it is sold. A plan missing any of the three is an orphan. Two plans that differ only in name produce two different answers to the same question; merge them into one and keep the other as inactive.

Retire in stages: stop selling first, keep the plan readable for past reservations, then archive it once your retention decision allows. Pricing stays a person's decision throughout. The cleanup is about which plans exist and what they mean, not about what they should charge.

Guest profiles: how do you handle duplicates, consent and retention?#

Merge duplicates into one surviving profile with the stay history linked, keep the purposes of processing unchanged by the merge, and send every retention question to your legal adviser.

Match on stable identifiers: last name, first name and one of phone or e-mail. Matching on room number or on a shortened name creates false merges that are harder to undo than duplicates. Merge, do not delete: keep one surviving record, log who merged what and when, and carry the stay history across so the guest's past is not lost.

A merge must not widen the purpose. Article 10 of Law No. 6698 requires you to inform people, when collecting their data, of the controller's identity, the purpose of processing, to whom and for what purpose data may be transferred, the method and legal basis of collection, and their rights under Article 11 (Mevzuat Bilgi Sistemi, Law No. 6698, Article 10 (opens in a new tab), consolidated text). Merging two profiles collected for different purposes quietly breaks that promise.

Accuracy and retention are duties, not preferences. Article 4 requires personal data to be accurate and, where necessary, up to date, and kept only as long as the law requires or the purpose needs; Article 7 says that once the reasons for processing no longer exist, the controller deletes, destroys or anonymizes the data on its own initiative or at the person's request (Mevzuat Bilgi Sistemi, Law No. 6698, consolidated text (opens in a new tab)).

Identity records are a separate obligation from marketing consent. Under Article 2 of the Kimlik Bildirme Kanunu (Law No. 1774), the responsible operators of hotels, motels, pensions, holiday villages and every other kind of lodging place must keep identity and arrival and departure records of everyone they accommodate, domestic or foreign, paid or free, day by day in the prescribed form, ready for inspection by the general law enforcement at any time (Mevzuat Bilgi Sistemi, Law No. 1774, consolidated text (opens in a new tab)). You cannot delete your way out of that duty, and you should not mix it up with consent for marketing.

Security covers the merged profile too. Article 12 requires all necessary technical and administrative measures for an appropriate level of security against unlawful processing and unlawful access, and makes the controller jointly responsible with anyone processing the data on its behalf (Mevzuat Bilgi Sistemi, Law No. 6698, Article 12). Do not write a retention period into your procedure because a blog said so. Your accountant knows the fiscal side, your legal adviser knows the personal data side, and the answer depends on your obligations and your stated purposes. For how this plays out at the desk, see AI at the front desk: which tasks change and which stay human.

Segment and source codes: what is the shortest check worth doing?#

Take twenty recent reservations, have three people assign segment and source independently, and treat every disagreement as a finding about the codes rather than about the people.

One code, one meaning. If a code can be read two ways, two reports and two answers will disagree and both will look defensible. The check is cheap and repeatable: twenty reservations, three people, one hour, a list of disagreements. Do it before any redesign, so you know which codes actually cause trouble.

Free-text notes are not codes. Anything you want counted, compared or asked about has to live in a field with a fixed list of values. Keep the mapping between your internal codes and each channel's codes in one document, and review it when a channel changes its list.

Do not redesign the scheme in the middle of a cleanup. Code design is its own job with its own rules, and a half-finished redesign is worse than a messy but stable list. Also keep source and segment separate: source answers where the booking came from, segment answers what kind of business it is. Codes that blur the two make both answers unreliable. If you decide the scheme needs work, start from Hotel market segmentation: how to set up segment codes.

Why do open folios break reports and AI answers?#

A folio left open past departure puts revenue in the wrong business day and leaves a balance that every later report and every later question will read literally.

Night audit closes the business day. Anything not posted before it moves to the next day, and the daily figures stop matching what happened at the desk. Run three lists before night audit: departures with a non-zero balance, folios with no posting for several days, and payments with no fiscal document linked.

Correct by voiding and reversing, never by deleting. This is ordinary accounting practice, and your accountant should confirm how it is applied in your system and your documents. An assistant asked who owes money, or how much revenue yesterday produced, reads these balances as they stand. It has no way to know that a folio is open because a guest is still deciding.

Close the loop with the front office. A folio that stays open usually points at a step missing from the departure routine, not at a careless person. Check the same three lists monthly, not only in busy weeks; the failures cluster where the process is thinnest, and that is usually a shift nobody watches. For what plain-language questions on your own data look like in practice, see AI hotel reporting: plain-language questions on your own data.

What does a monthly data-quality routine look like, and who owns it?#

One named owner, one hour a month, five fixed checks recorded in the same place each time, and a quarterly review of retention and consent with your legal adviser.

Name the owner and a deputy. In most hotels the front office manager or the revenue manager is the right owner, because both already touch inventory, rates and balances daily. Keep the routine to five checks so it survives a busy month, and record the result in the same file every month. A number that moves tells you where a process is failing; a one-off cleanup tells you nothing about next month.

Fix at the source. If the same check fails three months running, the cause is a step in a routine or a missing field, not the person doing the check. Keep a short change log: who changed a room type, a rate plan or a merge, and when. Without it, the next cleanup repeats the last one.

Once a quarter, sit with your legal adviser on retention, deletion and consent wording, and with your accountant on documents and corrections. The monthly routine keeps records accurate; they decide how long the records may stay.

What to do next#

Start this week with the physical room count, because it is the check that most often finds a real contradiction and it takes an afternoon. Then work down the order this article gives: rates, guest profiles, segment codes, open folios. The monthly routine is what keeps the work from undoing itself, and the retention questions belong with your legal adviser, not with the person running the checks.