AI clauses in hotel software contracts: data use, model changes and wrong answers
An AI feature adds three contract questions a standard template misses: what data it reads, what it produces, and who can change it. Find each clause and ask for a written answer.
Contents
Key takeaways
- An AI feature adds three contract questions a standard template misses: what data it reads, what it produces, and who may change it.
- The permission to run the service and the permission to improve the service are two separate permissions; ask for them separately and get any opt-out in writing.
- Ask for advance notice of model changes and the right to retest with the hotel's own questions before a change reaches guests.
- Logs are both the evidence you need after a wrong answer and a store of guest words; agree a retention period for each record type.
- Exit terms should name what leaves with the hotel: its operational data, the instructions it wrote, and the AI's records.
An AI contract clause is a term that decides three things: what the vendor may do with the data the AI reads, how you are told when the model or feature changes, and who answers when an output reaches a guest. Find each clause, ask for a written answer, and let counsel read it.
The contract is signed, the assistant is answering guest messages, and then a reply goes out that no one at the hotel would have written. The first question is who fixes it. The second is what the contract actually said. This article names the AI contract clauses to look for, the question to ask about each, and what a clear answer sounds like, so you can mark the gaps before you sign rather than after a guest sees them. For the clauses every hotel software contract needs, with or without AI, start with the companion piece on hotel software contracts: clauses to read before you sign.
Why does AI add clauses a standard software contract does not cover?
Because the product now reads the hotel's own content and writes sentences a guest may act on, while the engine behind those sentences can change without the hotel doing anything. A standard hotel software template is built around uptime, support, backups, price and data ownership. It assumes the software does the same thing tomorrow as it does today.
An AI feature breaks that assumption in three ways. Data goes in: guest messages, reservation notes, rate and occupancy figures. Free text comes out instead of a fixed report. And the model underneath is replaced on the vendor's schedule, not yours.
Keep the AI terms where you can find them. They may sit in the main contract, a data processing annex, an acceptable use policy, or a page on the vendor's website that the contract refers to. A term that lives on a web page can change without your signature. Written answers given during the sales process are worth keeping as a dated annex; they are the part a later conversation can point to when memories differ. This article names clauses and questions. Whether a clause is enforceable, and what it means in your case, is a question for counsel.
The three questions that find the clause
The practical test for any AI feature: what does it read, what does it produce, and who can change it. Each answer points to a clause you should be able to find. The table below extends that test to the five areas where AI contract clauses most often hide.
| What the AI touches | Question to ask | What a clear answer looks like |
|---|---|---|
| Hotel data used to improve the service | Which of our data may you use to improve or train, and can we refuse in writing? | A named list of data types, a stated opt-out, and the consequence of using it |
| Model or feature change | How much notice do we get, and may we retest before it reaches guests? | A notice period, a change log, a retest window, a switch-off right |
| A wrong answer to a guest | Who confirms before sending, and what record do we get afterwards? | A named confirmation step, a record of prompt and output, a response time |
| Logs and subcontractors | What is stored, for how long, and who else processes it? | Retention per record type, a subcontractor list, a notice duty when it changes |
| Exit | What do we take with us, in what format, and what is deleted? | A format, a test export, a deletion confirmation |
May the vendor use our hotel's data to improve its service?
Only to the extent the contract says, and the permission to improve the service is a different permission from the one to run it. Two permissions hide behind one sentence. Processing to deliver the service, such as reading a guest message to draft a reply, is necessary for the product to work. Processing to improve, train, benchmark or build new features is separate, and it is negotiable.
Ask what improvement covers: staff prompts typed into the assistant, guest messages, uploaded documents, reservation and rate data, complaint texts. Ask whether improvement happens per hotel or across all customers, and whether anything that identifies the hotel or the guest survives the process. "Aggregated and anonymized" is a phrase; ask who decides and how the process is described.
Get any opt-out in writing, in the contract or a signed annex, not in a settings page the vendor controls and can change without telling you. Ask what changes if you opt out: a feature that stops working, a different price, slower support. An opt-out with an unstated cost is not a decision you can make.
Guest data is a second layer. The hotel is the party the guest deals with, so the vendor's use has to fit what the hotel told the guest. Which role each side holds is a question for counsel, and the privacy side of it is covered in AI assistants and guest data: KVKK questions for your adviser. If the vendor keeps a copy for improvement, deletion at the end of the contract may not reach it. Ask where improvement copies live and when they go.
ChecklistBefore signing, you should be able to tick every line below
- A named list of the data types the AI reads, not the phrase "your data"
- A separate sentence on improvement and training, distinct from the sentence on delivering the service
- A written, signed opt-out with its consequence stated
- A statement of whether improvement is per hotel or across customers
- A statement of where improvement copies are kept and when they are deleted
- A named contact for data questions and a response time
What happens when the model or the feature changes under us?
Something will change without the hotel's involvement, so the contract should decide how you hear about it, what you may test, and what you may switch off. Models are replaced, prompts are tuned, features are renamed or withdrawn. The front desk does not pause while that happens.
Notice: ask for advance notice of changes that can alter what the assistant says or does, not a changelog published after the fact. Ask who in the hotel receives the notice. Retesting: build a small set of questions with the answers the hotel considers correct, and ask for the right to run them after a change and before guests meet it. Ask who pays for the retest and how long the window stays open.
Switch-off: the right to turn one feature off without ending the contract, plus a written description of what the workflow looks like without it. Version pinning: some vendors let a customer stay on a named version for a period; ask whether it exists, for how long, and what ends it. Price: a model change can change the vendor's cost, so check whether that reaches your bill and on what notice, a topic covered in how AI tools are priced: seats, usage and what drives the bill. Change log: ask for a dated record of changes that is sent to you, so a later disagreement has a timeline rather than a memory.
TipBuild the test set before you sign
Ten questions the hotel's own staff would ask, with the answers they would accept, written down and dated. It costs an afternoon and it turns "the answers got worse" into something you can show.
Who is responsible when a wrong answer reaches a guest?
The hotel answers to the guest; the contract decides how far the vendor stands behind the output and what the hotel must do first. Three questions get mixed together here: who checks before a message is sent, who is answerable to the guest afterwards, and who carries the cost. Separate them in the negotiation.
Confirmation: ask which actions and messages require a person's approval, whether the hotel can change that list, and whether a change takes effect immediately. The irreversible step: a message already sent to a guest cannot be recalled. Ask what the system does in the seconds before sending, and what it records. Evidence: after a wrong answer the hotel needs the exact prompt, the output and the data the system read. Ask how quickly that record is produced and in what form. Complaint path: who the guest complains to, who at the vendor responds, and in what time.
Liability wording is not something to negotiate yourself. Ask counsel what the wording means against a realistic worst case, and keep the vendor's own marketing claims in the file. Staff instructions are part of the answer too: a wrong reply is more likely when staff paste guest data into a tool that was never meant to hold it, so the contract should say what staff may put in. The buying-stage version of these questions is in questions to ask an AI vendor before buying for your hotel.
Watch outA wrong answer is not only a vendor problem
If the hotel's staff sent the message without reading it, the guest will not distinguish between the two. Agree the confirmation step in writing, then train to it.
How long are prompts and outputs kept, and who else processes them?
Every AI feature leaves a record, and the contract should say what is stored, for how long, who can read it and who else handles it. Logs cut both ways: they are how a wrong answer is investigated, and they are a store of guest words. Both sides belong in the clause.
Ask for retention per record type: staff prompts, guest messages, outputs, generated documents, error reports. Access: which of the vendor's staff can read the content, and whether reading is itself recorded. Subcontractors: the model provider, the hosting provider, a speech or translation service. Ask for a list, a duty to tell you when it changes, and a right to object.
Where processing happens: ask in general terms and let counsel read the answer. A sentence in an annex is not the same as a commitment you can act on. Deletion at the end: what is deleted, in what time, and what evidence you receive. Ask whether backups are included in that answer.
NoteAsk the subcontractor question in writing and keep the reply
A list that arrives by email, dated, is more useful than a link to a page that changes.
What should the exit terms say?
They should name what leaves with the hotel: its operational data, the instructions it wrote, and the records of what the AI did. Three things to take: the operational data (reservations, guest records, bills), the configuration (the instructions, rules, approved answers and offer texts the hotel wrote), and the AI's records.
Ask for the export format before signing, and run a test export during onboarding rather than discovering the format on the last day. Deletion after export: what remains, for how long, and what proof you get. Continuity: what happens to messages and offers in flight when the contract ends.
Describe the exit as a sequence of steps with a time for each, not a single date. Any fee for export or transition should be asked up front and written into the contract.
What to do next
Pull the contract you have and the vendor's data annex, and mark which of the five questions in the table they already answer. Send the rest to the vendor in writing, ask for dated replies, and take those replies to counsel together with the contract. If the base clauses are still unclear, start with the general contract checklist and come back to the AI ones.
Frequently asked questions
Do we need a separate AI contract, or can it be an annex?
Either works. What matters is that the AI terms are written, dated and referenced from the main contract, and that you keep the version that applied when you signed. An annex is easier to update when the model changes, which is also why you should ask how updates to it are notified.
Can we stop a vendor using our data for improvement after signing?
Ask, and get the answer in writing as a signed annex or an amendment. Ask what changes if you take the opt-out, so the decision has a stated cost. A toggle in a settings page the vendor controls is not the same as a contract term.
Who owns what the AI writes?
Read the clause rather than assuming. The hotel's own operational data and the instructions it wrote are the hotel's, and that should be stated. Ownership of generated text is a separate question, and counsel should read the wording before you sign.
What do we do when staff say the answers got worse after an update?
Run the test set you built before signing and compare it with the dated answers. Send the difference to the vendor in writing, use the notice and retest rights in the contract, and switch the feature off while it is fixed if the contract allows it.
Is a wrong AI answer to a guest our problem or the vendor's?
The guest deals with the hotel, so the hotel answers first. The contract decides how far the vendor stands behind the output and who carries the cost. Keep the record of prompt, output and source data, because that record is what any later conversation rests on.