AI in hotel purchasing: comparing quotes, reading supplier emails and what buyers must still check

AI in procurement means using a language model to read supplier quotes and emails, extract the same fields from each, and normalize units, currency and VAT into one table. It drafts and compares. It does not choose the supplier, approve the order or replace the check at receiving.

Three quotes for the same item arrive as three PDFs: one priced per kilogram, one per case of twelve, one per litre, with different VAT treatment, different payment terms and validity dates six weeks apart. Somebody has to turn that into one number a manager can sign. This guide shows how a language model helps with the reading, the normalizing and the drafting, and where the buyer's own work begins.

Where does hotel purchasing time actually go?#

Most of it goes to reading and re-keying documents, not to choosing suppliers. The judgment takes minutes; the preparation around it takes the afternoon.

The same item arrives described three different ways: per kilogram from one supplier, per case from another, per litre from a third, with the pack size buried inside the product description. Quotes go stale. When prices move monthly, a buyer re-requests instead of comparing, and the comparison is postponed again. Chasing is constant: confirming a delivery date, asking for a corrected invoice, reminding a supplier about a missing document, checking whether a promised substitution was actually delivered.

The comparison itself is short work once the numbers sit in the same unit. That is exactly the part a machine can do. So split the work explicitly. Extraction, normalization and drafting go to the tool. Judgment, approval and the physical check at the door stay with people.

Write the split down before you buy anything. A team that has not decided what it will never delegate will delegate the wrong thing first.

How does AI turn three PDF quotes into one comparable table?#

It reads each document, pulls the same fields out of all three, and converts them to one unit and one currency so the numbers can stand in the same column. The output is a draft table with its working shown, not a verdict.

The fields to extract are always the same: item, unit of measure, pack size, unit price, VAT rate and whether the price includes it, currency, payment term, validity date, delivery terms and minimum order quantity. Normalization is the real work: cases to kilograms, litres to millilitres, foreign currency to lira at a stated daily central-bank rate with the date written next to it, VAT-exclusive and VAT-inclusive prices kept in separate columns.

Demand the working, not just the result. Every normalized figure should carry the source line it came from and the arithmetic that produced it, so a controller can follow it in ten seconds. Ask the tool to mark its own uncertainty: any line where it guessed a unit or could not read a figure should be flagged, not silently filled.

The six fields that break every comparison#

What a normalized table looks like#

Expect specific failures. A line reading "1 koli (12 x 1 L)" hides the pack size; a discount row at the bottom of the page changes the effective unit price; an item sold in 5 kg bags but priced per kilogram needs a conversion the tool may skip. Text-based PDFs extract far more reliably than scans and phone photographs. Treat anything that came in as an image as unverified until a person has read it.

Why should you recompute every total in your own sheet?#

Because an extraction error is invisible in a tidy table. A wrong total looks exactly like a right one, and the tidier the table, the less anyone checks it.

Recompute unit price times quantity, then apply VAT, then compare your figure with the tool's. Investigate every difference rather than accepting the closer-looking number. Spot-check by value, not by row count: the five largest lines and every line the tool flagged as uncertain. Keep the source document next to the row. A number with no page reference cannot be checked by the person who signs it, or by an auditor a year later.

Know which differences are legitimate. Rounding and a different VAT treatment produce small gaps that are fine; a gap that scales with quantity is an error. Rebuild the table by hand once, on a real comparison, before you trust the automated version. You will learn which fields your suppliers actually get wrong. If you want to go deeper on checking an AI-built table against your own spreadsheet formulas, the same rules apply here.

Treat the spreadsheet as the record and the AI output as an input to it. That order matters when someone asks who checked what.

How do you see what moved since the last order?#

Match each new quote line to the last order for the same item, in the same unit, and flag every change with its size and its date. The comparison is against your own history, not against the other suppliers.

Match on your internal item code, never on the supplier's description. Suppliers rename products, merge lines and change pack sizes without saying so. Flag four things separately: unit price, pack size, VAT treatment and payment term. A pack that shrinks while the price holds is an increase, and it hides well. Record the validity date and re-check it before the order goes out. Where prices move every month, a quote is a snapshot with an expiry, not a standing price; price validity, contracts and stock levels when prices rise every month deserve their own routine.

Separate a real increase from a unit artifact. The same product at the same price per case is cheaper per kilogram if the case got heavier. Set a threshold above which a person must look at the change before anything is ordered, and note everything below it. Keep the flagged changes in one place through the year. That list is what you take into the next contract negotiation.

Can AI draft supplier emails and reminders?#

Yes, drafting is the strongest fit in the whole purchasing workflow. The assistant writes the request, the correction note or the reminder, and a person reads it before it leaves the building.

The useful drafts are routine: a quote request in a fixed format, a request for a corrected invoice, a delivery-date confirmation, a reminder about a missing document, a note asking a supplier to confirm a substitution. Use the drafting to fix the problem at the source. Send suppliers a standard quote template with the fields you need, and the next comparison starts from a document that extracts cleanly.

Language and register matter. A Turkish supplier email is written in formal address and plain sentences; the assistant can draft in Turkish and English, and a person checks names, quantities, dates and item codes. Make the draft reference the quote number and its date. That single line is what keeps a thread auditable when three suppliers quote the same item.

Never let a draft carry a commitment. A price agreement, an order confirmation or a delivery acceptance is a person's message, sent by a person. Read the draft as if a supplier will forward it to their manager. Anything ambiguous in it will be read in the supplier's favor.

What must a buyer still decide and sign off?#

The supplier choice, the approval, the quality at receiving and the written reason for the decision. None of these move to a tool, whatever the table looks like.

Choice is not the lowest number. Delivery reliability, substitution policy, payment terms, contract conditions and the cost of a failed delivery all sit outside the price column. Scoring suppliers on delivery, substitution and terms is a separate exercise from ranking prices, and the two should be kept visibly separate. Approval limits come from the hotel's own internal policy, not from software. The tool can present the numbers; the signature and the limit behind it belong to a named person.

Receiving is physical and cannot be delegated to a document: count, weight, temperature, condition, and whether the delivery note matches the order and the invoice. The receiving procedure from order to storeroom is where a good comparison either holds or falls apart. Write one line on why this supplier at this price. It is what you need at the next renewal, and it is what an auditor will ask for.

AI output is never an approval record. It is a draft that a person has read, corrected and signed, or it is nothing. Keep the approval trail in your own systems. A conversation with a tool is not an audit trail, however well it is worded.

Which purchasing data may go into an AI tool?#

Only what the hotel has approved for that specific tool. Supplier price lists, contracts and personal data need a decision before they are pasted anywhere, and a public chat window is not an approved tool.

Confidentiality first: supplier contracts commonly restrict disclosure of prices and terms, and pasting a price list into a public tool is disclosure. Personal data has its own rule. Supplier contact names, bank details and staff data fall under Turkish personal data law, so minimize what you send and use a tool the hotel has approved. If the hotel has no approved tool list, ask for one before the first quote is uploaded. That request is easier to make now than after an incident.

Reading a contract with AI is possible and useful for orientation, but it misses clauses and misreads cross-references. A person and a legal adviser confirm anything that binds the hotel; what AI finds and misses when reading supplier contracts is worth knowing before you rely on it. Keep a short internal log of what was sent to which tool and when. It costs a line and answers the question nobody wants to answer from memory.

When in doubt, send the structure and not the content: ask the tool to normalize a table you built, rather than uploading the supplier's document.

What to do next#

Take the last three quotes you compared and rebuild them into one normalized table by hand, once. You will see exactly which fields your suppliers get wrong, which conversions you keep redoing, and which of those steps you are willing to hand to a tool. Everything above that line, the choice, the approval, the count at the door and the reason written down, stays where it is.