Systems and records
Hotel archive data qualification: which records make a complete AI training case
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A hotel record becomes a complete AI training case when one reservation decision can be replayed end to end: the booking state, the rate, cancellation and deposit rule in force when it was sold, dependent records such as group blocks, folio routing and agency contracts, who changed what and when, and the verified end state. Reservation exports alone are fragments. Inventory linkage and rights per record class before extracting anything.
System export
Hotel archive data
Key takeaways
- A complete case is one decision with its rule, its dependencies, its approvals and its verified outcome, not a row from a reservation export.
- Freeze the cancellation and deposit rule in force on the booking date; today's configuration may differ.
- Group blocks, folio routing and agency contracts often explain decisions that look arbitrary on the reservation alone.
- Label cases complete, reconstructed or fragment, and report counts per label instead of one completeness score.
- Guest data, card data, payer identity and contract terms need transformation or exclusion, and ownership must be settled separately from privacy.
Which hotel operating records form complete AI training cases?
A complete hotel training case is one reservation decision that can be replayed from start to finish: the booking as it stood, the rate and cancellation rule in force when it was sold, every dependent record that constrained the change (group block, routing to a company folio, travel agency contract), who changed what and when, and the verified end state on the reservation and the bill. A reservation export alone is a fragment, because it shows the result without the rule or the dependencies that produced it.
That definition matters for hotel archive data qualification because most property management system (PMS) exports are flat tables of reservations. They answer what happened to a booking. They rarely answer why the property was allowed to do it, which is the part an AI system for revenue, front office or reservations work actually has to learn.
The practical test is simple. Pick a decision, such as a late cancellation, a no-show charge, a group room release or a relocation during a disruption, and ask whether someone outside the hotel could reconstruct the decision using only the linked records. If they need to ask the front office manager what the policy was that week, the case is not complete yet.
Where each fact lives in a hotel archive
Each fact in a case has one system that should be treated as its source of truth, and the worksheet below only works if you name that system per fact. The PMS holds reservation state and timestamps; rate and policy configuration holds the purchase-time rules; the billing module holds folios and payments; the sales module or contracts register holds group and agency terms.
Field names differ by vendor. The examples here use the public Mews Connector API documentation because it is openly published, and Oracle OPERA Cloud documentation for deposit rules and routing. Treat them as reference vocabularies for what to look for, not as a claim about what your own system exports.
| Fact in the case | Usual source | Example field or object | Why the case needs it |
|---|---|---|---|
| Reservation lifecycle | PMS reservations | State (Inquired, Optional, Requested, Confirmed, Started, Processed, Canceled), CreatedUtc, UpdatedUtc, CancelledUtc | Shows where the booking was in its lifecycle when the decision was made |
| Planned versus actual stay | PMS reservations | ScheduledStartUtc and ActualStartUtc, ScheduledEndUtc and ActualEndUtc | Separates early check-outs, late arrivals and no-shows from the original plan |
| Why a booking ended | PMS reservations | CancellationReason (for example NoShow, ForceMajeure, BookedElsewhere, InputError) | Turns a cancellation into a labeled outcome instead of an unexplained state change |
| Rule sold with the rate | Rate and policy configuration | Cancellation policy tied to a rate group, with ApplicabilityOffset and fee settings | The rule that decides whether a fee applies must be the one in force at booking |
| Group dependency | Sales and catering or block module | AvailabilityBlockId on the reservation; block State and ReleaseStrategy | Group releases and attrition decisions depend on the block, not the single room |
| Who pays what | Billing and folios | Bill AccountId; routing instructions to folio windows | Disputes and adjustments are decided on the folio, often a company's, not the guest's |
| Channel and agency context | Channel manager and contracts | ChannelNumber, TravelAgencyId, travel agency contract | Explains distribution terms that limit what the hotel could change |
The record linkage and rights inventory worksheet
Fill in one row per record class before anything is extracted. The worksheet does two jobs at once: it proves each record can be joined to the reservation it belongs to, and it records who controls the record and what has to change before it can leave the property. A record that links but cannot be shared, or can be shared but does not link, does not belong in the case yet.
Score each row Keep, Transform or Exclude. Keep means the record links reliably and raises no rights or confidentiality question. Transform means it links but needs removal or tokenization of specific fields first. Exclude means it fails linkage or carries terms that the hotel cannot release.
| Record class | Link key to the reservation | Completeness test | Rights and confidentiality question | Default decision |
|---|---|---|---|---|
| Reservation and change log | Reservation ID and confirmation number | Every state change has a timestamp and the profile that made it | Does the management agreement let the operator license PMS history, or does the owner or brand control it? | Keep, after the ownership question is answered in writing |
| Rate and cancellation policy versions | Rate ID and rate group, plus policy update time | The policy in force on the booking date can be identified, not just today's | Usually property-controlled configuration; brand standard policies may sit under the franchise agreement | Keep if versions are recoverable; otherwise flag the case as reconstructed |
| Deposit rules and schedules | Reservation ID; rate code or reservation type | The deposit rule actually applied to the booking is recorded | Low sensitivity on its own; payment details attached to it are not | Keep the rule, transform any payment details |
| Group or availability block | Block ID on each reservation | Block state, release strategy and release time are present for the stay dates | Group contracts often carry confidentiality clauses that cover the account and its terms | Transform: keep block mechanics, remove the group account identity |
| Folios, bills and routing | Bill ID, account ID and routing instruction | Every charge on the decision can be traced to a folio window and payer | Company payers are third parties whose billing terms may be confidential | Transform: keep charge types and timing, tokenize payer identity |
| Payments and card records | Payment ID linked to the bill | Payment state and timing exist without any card security data | Card security codes must not be kept after authorization; card numbers stay out of training sets | Transform or exclude: keep payment type, state and timing only |
| Guest profiles | Account ID on the reservation | Profile can be replaced by a stable token without breaking joins | Guest personal data, including birth date, nationality, email and internal staff notes | Transform: tokenize, drop free-text notes unless separately reviewed |
| Travel agency and channel contracts | Travel agency ID and channel reservation number | The contract active on the booking date is identifiable | Commercial terms with agencies and channels are usually confidential to both parties | Exclude the terms; keep a category label such as agency booking |
| Disruption and relocation notes | Reservation ID, date and staff profile | Each note can be tied to a specific decision and time | Free text can name guests, staff and other properties | Transform after a reviewer reads a sample |
Which purchase-time rules have to be frozen with each booking?
Freeze the cancellation rule, the deposit rule and the rate that applied on the day the booking was made, because the current configuration may say something different. Hotels edit policies constantly, and a case labeled with today's policy teaches a model a rule the guest never agreed to.
In Mews, a cancellation policy belongs to a rate group and carries its own created and updated timestamps, an applicability offset from the start of the stay, and absolute or relative fee settings, and the API can return the policies attached to specific reservations. That structure is helpful, but the updated timestamp also tells you the policy can change after a booking is sold. If your PMS does not keep prior versions, record the case as reconstructed rather than complete.
OPERA Cloud documentation describes a precedence order worth copying into your inventory: only one deposit rule applies to a reservation at a time, with a rule on the rate code taking precedence over one on the reservation type, followed by a rule attached to the reservation itself. A case should record which of those three levels supplied the rule, because the same booking can look inconsistent if you only see the manual rule on the reservation.
- Rate identifier and rate group on the booking date
- Cancellation rule: deadline offset, fee basis and which nights it covers
- Deposit rule and the level it came from: rate code, reservation type or manual
- Any manual override, the profile that applied it and the timestamp
- Property-specific variations, such as event-period policies, recorded as separate rule versions
Which dependent bookings and approvals limit a change?
The constraint on a hotel decision is often another record: the group block the room belongs to, the company folio the charges route to, or the agency contract the booking came through. A case that omits these will make legitimate decisions look arbitrary.
Group blocks are the clearest example. In Mews, a block has its own state, and only a confirmed block can have reservations assigned to it; it can be released at a fixed time, on a rolling offset from each day, or not automatically at all. A room cancelled inside a block before its release date is a different decision from the same cancellation after release, and the training case has to carry the block state and release settings to show which one happened.
Routing is the second. OPERA Cloud lets charges be routed to separate folio windows or to another reservation, with unrouted charges staying on the primary guest's window, and routing can be set up automatically from company, agent, source or contact profiles. When a dispute is resolved by moving a charge, the case needs the routing instruction in force, or the adjustment will look like an error.
Approvals are the third. Record the profile that made each change and, where your PMS supports it, the reason code. Mews records both the creator and the last updater of a reservation, which is a start; a full authorization trail usually needs the PMS audit log as well.
Illustrative scenario: a weather disruption across a group block
The following is an illustrative example, not a real property. A conference hotel holds a group block for a three-night event with a rolling release offset. A regional storm closes the airport on the first arrival day. Over two days the reservations team cancels part of the block, moves some guests to later dates and relocates others to a sister property.
The archive team pulls every reservation tied to the block and runs the worksheet. The counts below are illustrative only.
| Decision type | Cases pulled (illustrative) | Complete | Missing piece in the rest | Label |
|---|---|---|---|---|
| Cancellation inside block, fee waived | 18 | 15 | Policy version on booking date not recoverable for 3 cases | Complete for 15; reconstructed rule for 3 |
| Date change within block | 9 | 9 | None | Complete |
| Relocation to sister property | 6 | 2 | Relocation recorded only in free-text notes for 4 cases | Fragment until notes are linked and reviewed |
| Charge moved to company folio | 5 | 5 | None, routing instruction present | Complete after payer tokenization |
| No-show charged despite disruption | 2 | 0 | No record of who approved the charge | Exclude pending approval evidence |
Which assumption most often flips a case label, and how to report the doubt
The assumption that changes most outcomes is that the policy in force at booking can be recovered. If it cannot, every fee or waiver decision becomes a guess about the rule, and the case should be labeled reconstructed with the source of the reconstruction noted, such as a dated rate sheet or a policy email.
Three other assumptions deserve a column in the worksheet. Time: Mews returns reservation times in UTC, while staff and guests reason in property local time, so convert once and record the property time zone with the case. Vocabulary: state names change across systems and even across API versions of the same system, for example an older Mews version lists Enquired where the current one says Inquired, so map every state to one controlled list. Manual work: decisions made by phone or email and recorded later in notes will carry later timestamps than the decision itself.
Show uncertainty as counts per label rather than a single completeness percentage. A buyer or internal team should be able to see how many cases are complete, how many rest on a reconstructed rule, and how many are fragments, with the reason for each.
What has to come out before hotel records leave the property?
Before any extract, remove or tokenize guest identity, card data, payer identity and confidential contract terms, and settle who has the right to license the history at all. These are separate questions. System access does not grant a right to share, privacy review does not settle ownership, and de-identification does not create a license right that the operator never had.
Guest profiles in a PMS hold more than names. The Mews customer object includes fields such as birth date, nationality code, email and internal notes about the customer; free-text notes are the fields most likely to contain something sensitive. Replace account identifiers with stable tokens so joins still work, and drop or separately review free text.
Card data needs a hard rule. The PCI Security Standards Council states that card verification codes may not be kept once the transaction they were collected for has been authorized, even for card-on-file use and even with the cardholder's permission. A training extract should carry payment type, state and timing, never card numbers or security codes.
Ownership is the question hotel groups most often skip. In managed and franchised hotels, the operator, the owner and the brand may each claim parts of the PMS history under the management or franchise agreement. Travel agency contracts, which in Mews can include commission and channel adjustment settings, are commercial terms the counterparty may expect to stay confidential. Counsel should read the agreements before the inventory is marked Keep.
When a hotel archive is ready to discuss with a licensing partner
An archive is ready to discuss when the worksheet is filled in for each record class, the ownership question is answered in writing, and you can state how many decision cases are complete versus reconstructed. You do not need a finished extract to start that conversation.
SourceX helps hotel companies license selected operating records to AI developers, handling qualification, rights review, contracting, preparation and delivery, and its first step is a short company assessment that asks for no data uploads. For hotel archives the rough bar is consistent history in one system over two or more years and tens of thousands of records, with smaller sets able to qualify when the work is rare, so a multi-year reservation history with linked policies and folios is worth assessing. PMS records leave the property only after the hotel company has approved the scope and terms and signed the agreement.
Frequently asked questions
Is a PMS reservation export enough to train a hotel AI model?
Usually not on its own. A reservation export shows states and dates, but not the cancellation or deposit rule in force when the booking was sold, the group block or routing that constrained changes, or who approved them. Without those links, the model sees outcomes without the reasons that produced them.
What should the record linkage and rights inventory worksheet include?
One row per record class, with the key that joins it to the reservation, a completeness test, the rights or confidentiality question it raises, and a Keep, Transform or Exclude decision. Fill it in before extraction so linkage gaps and ownership questions surface early rather than after data has moved.
How do we handle policies that changed after a booking was made?
Use the policy version in force on the booking date. If your system keeps only the current configuration, rebuild the earlier rule from dated evidence such as rate sheets and label those cases as reconstructed, with the evidence noted, instead of mixing them with fully documented cases.
Can card data stay in a training extract if it is encrypted?
No for security codes, and it should not be needed for anything else. PCI guidance says card verification codes cannot be kept after authorization, and encryption does not change that. Keep payment type, state and timing, and leave card numbers out entirely.
Who owns the operating history of a managed or franchised hotel?
It depends on the management and franchise agreements, which may give the owner, the operator and the brand different rights over guest and operating data. Confirm in writing who can authorize licensing before the inventory marks any record class as Keep. Counsel should review the specific agreements.
Sources
- The Mews Connector API documents reservation states, scheduled and actual stay times, cancellation reasons, channel and travel agency identifiers, availability block links, and creator and updater profiles, with times in UTC. Source
- Mews cancellation policies belong to a rate group, carry created and updated timestamps, an applicability offset and absolute or relative fee settings, and can be retrieved for specific reservations. Source
- Mews availability blocks have their own states, only confirmed blocks can have reservations assigned, and release can be fixed, rolling or not automatic. Source
- Mews bills are issued to a customer or company account and have open or closed states. Source
- Mews travel agency contracts can record whether commission is included in the rate and channel manager price adjustments. Source
- Mews customer records include personal fields such as birth date, nationality code, email and internal notes. Source
- OPERA Cloud applies only one deposit rule to a reservation at a time, with the rate code rule taking precedence over the reservation type rule and then the rule attached to the reservation. Source
- OPERA Cloud routes charges to folio windows 2 through 8 or to another reservation, keeps unrouted charges on window 1 for the primary profile, and can apply routing from company, agent, source or contact profiles. Source
- The PCI Security Standards Council says card verification codes may not be retained after the transaction is authorized, including for card-on-file or recurring use and even with the customer's approval. Source
See if your company qualifies
A short company assessment. No data uploads are needed.