AI for reservation paperwork: booking emails, rooming lists and agency vouchers

AI data extraction from documents turns a booking email, a PDF voucher or an Excel rooming list into a table of named fields: guest name, arrival date, room type, board. The tool proposes the values; a person compares them with the contract and the existing booking before anything is typed into the hotel system.

A rooming list arrives for the third time, the subject line says "final", and one surname, one night and one board type have changed since the version already in the system. Someone now has to find those three differences by eye. This guide covers where retyping happens in a reservations office, what extraction can and cannot do, which documents suit it, the field list to pull, the three checks before entry, and how to count errors before and after the change.

Where does retyping happen in a reservations office?#

Retyping happens wherever a booking arrives as a document rather than as a structured message: an agency email, a PDF voucher, an Excel rooming list, or a reply that changes one line of an earlier one.

The same fact lives in three places at once: the agency's email, the voucher attached to it, and the hotel's own reservation record. Every hand copy is a chance for the three to disagree. The typical shapes are a confirmation written into the body of an email, a one-page PDF voucher per reservation, an Excel rooming list for a group, and a short reply that moves one guest to a different room type.

Group changes are the hardest case, because the document arrives repeatedly. Version two and version three look identical at a glance, and nothing on the file says which one the hotel acted on. The error is invisible at entry and visible at the desk: a misspelled surname, a stay one night short, or a board type read as half board when the contract says bed and breakfast all look normal in the system until the guest is standing in front of you.

The time cost is not the typing. It is the re-reading: an agent checks the same line against the source two or three times, because there is no single place where source and record sit side by side. The full request-to-arrival flow, including where the reservation record is created and who owns each step, is set out in our guide to how a hotel reservations department works.

What does AI data extraction from a booking document actually do?#

It reads the layout of the document, finds the values that match a field list you set, and returns them as a table with the source text next to each value.

Three steps happen in order. The tool reads the page as text and as layout, it matches what it finds to the fields you named, and it returns a table. Nothing in that sequence understands what a booking is.

Layout tolerance is the real gain. The same reservation written as a paragraph, as a two-column voucher or as a spreadsheet row can all be reduced to the same field list. What the tool cannot do is judge. It does not know whether the rate exists in the contract, whether the room is held, or whether the person named is the guest or the payer.

Good tools mark which values they read cleanly and which they inferred. Treat an unmarked confidence signal as a warning: you cannot check what you cannot see. Ask for a fixed output: the same columns every time, plus a column holding the original text for that field. A table you can compare line by line is worth more than a tidy summary.

The output belongs in a review screen, not in the reservation. Extraction shortens the checking, it does not replace it.

Which booking documents suit extraction, and which do not?#

Documents with a stable, repeated layout suit extraction; documents carrying identity data, or arriving as photographs of handwriting, do not.

The first filter is legal, not technical. Passports, identity cards and guest identity notifications belong to the check-in flow, never to a document extraction tool. The second filter is volume: if a document type reaches you a few times a week, retyping it is cheaper than building and maintaining a field list for it. The third filter is versioning: a document that arrives repeatedly with changes is only safe to extract if you also record which version the hotel acted on.

In Türkiye, every stay involves a payment, a fiscal document (an invoice or e-document), a guest identity notification (KBS, misafir kimlik bildirimi) and personal data (KVKK, the personal data protection law). Booking documents carry the commercial terms and the payment; identity data arrives at the desk, in person. Scanned or photographed vouchers are the borderline case: they can work when the scan is straight and clean, and fail quietly when it is not. Start with the digital originals.

Which fields should you pull from a voucher or rooming list?#

Pull the fields you would otherwise type into the reservation, and nothing else: the ones that, if wrong, change the guest's stay or the bill.

Keep names exactly as the agency printed them, including the spelling you believe is wrong. Correcting at extraction time destroys the trail back to the source document. Board type is where meaning drifts: half board, HB and yarım pansiyon must map to one value in your system, and the mapping is your decision, not the tool's.

Dates need arithmetic the tool will not do for you: nights between arrival and departure, and whether the departure date is the last night or the morning after. Never extract a promise: a request for a sea view or a connecting room is a request, and it stays as text until a person allocates it. Record the document's issue date and version alongside the fields, so a later dispute can be settled by pointing at the paper that was acted on.

Group blocks, the rooming list itself and the master folio are covered in our article on hotel group bookings: block, rooming list and master folio.

How do you check extracted data against the contract, the allotment and the existing booking?#

Three comparisons, in this order: the contract says what the rate and terms should be, the allotment says whether the rooms are held, and the existing booking says whether this is a new reservation or a change.

Contract check: rate code, board, cancellation terms and commission basis. A rate that does not exist in the loaded contract is a stop, not a query to resolve later. Allotment check: the rooms held for that agency on those dates. A voucher asking for more rooms than the block holds is a request to the agency, not a confirmed booking.

Existing booking check: match on the agency reference first, then on surname plus dates. A duplicate created by a fast extraction is worse than a slow entry, because it consumes inventory twice. Change against new: rooming list version three replaces version two, it does not add to it. The version number is the only thing that tells the two apart.

Arithmetic stays with the person: nights multiplied by rate, children's pricing, extras and the deposit already received. The tool has read numbers, it has not calculated anything. Contract loading, including how rate codes and cancellation policies should be entered so that a check is possible at all, is covered in our contract loading checklist for agency rates.

Why must a person confirm before anything enters the PMS?#

Because the tool has read a document, not checked a booking, and the confirmation step is where the hotel takes responsibility for the record.

The failure mode is quiet. A wrong date produces a reservation that looks entirely normal in the system and becomes a problem only when the guest arrives a day earlier than the record says. Confirmation should be one screen with the source document beside the proposed values. If the reviewer has to open the PDF in another window and re-key the fields to compare, the gain has disappeared.

Some fields deserve a second pair of eyes: price, dates, and anything that alters an existing reservation rather than creating a new one. Log who confirmed, when, and from which document version. That log is what you produce when an agency disputes a booking six weeks later.

AI Otel's PMS assistant is built to answer questions about a hotel's own property in plain language, to carry out routine actions, and to wait for a person's confirmation on anything sensitive. It does not read booking emails or documents, and nothing in this workflow writes into reservations unattended. The general pattern, one task, a defined set of permitted actions and a named person who approves, is set out in our guide to setting up an AI agent for one hotel task.

How do you count reservation data entry errors before and after?#

Pick one definition of an error, count it by hand for a month, then count it again after the change using the same definition and the same sample.

Define the error narrowly: a field in the reservation that differs from the source document, found before or at arrival. Anything vaguer cannot be counted twice in the same way. Count in one place: the arrivals list each morning, checked against the voucher for that day, gives a repeatable sample without extra work.

Record the type as well as the count: name, date, room type, board, rate. The type tells you which field to fix in the template or the mapping. Keep a denominator, errors per hundred reservations entered, so that a quiet week does not look like an improvement.

Count the checking time too. If reviewing the extracted table takes as long as retyping the document, the gain is not there yet, whatever the error count says. Do not adopt an accuracy figure from a vendor, including from AI Otel. Layout, language and scan quality are local, so the only number that means anything is the one measured on your own documents.

What to do next#

Take one document type, the PDF voucher, write the field list, run twenty real vouchers through the tool and check every value by hand against the source. Count the mismatches and the minutes spent checking. That small test tells you whether extraction earns a place in your reservations workflow, and it produces the field list and the review screen you will need if it does. If you want to see how AI Otel approaches human confirmation in hotel software, you can request a briefing.