Ten things to check in a generative-AI or API contract
In short
- The first thing to settle in an AI contract is not copyright in the output. It is where the data goes and what the upstream model provider does with it.
- A company that builds a product on someone else's API is a provider to its customers and a user to its model vendor. What it promises downstream cannot exceed what it has secured upstream.
- Input data, training use, personal data and cross-border transfer, output rights, third-party IP claims, model changes, logs and liability caps have to be read as one set, not clause by clause.
- Where the vendor's standard terms cannot be negotiated, enterprise settings, opt-outs, a DPA, an SLA and internal usage rules can still move the risk.
"We are adopting a generative-AI tool and want the terms reviewed." "We want to build OpenAI's API into our own SaaS." "We want to build an agent that works on customer data." These have become ordinary requests in Japanese corporate practice.
When people read an AI contract, their eye goes to intellectual property and the liability cap. In practice there is something to settle before that: who puts which data in, how far it travels, what the vendor uses it for, which model processes it, and where the output and the logs end up.
This note sets out ten things worth checking before signing, written for a company either adopting an AI service in Japan or building one on top of someone else's model.
Start by drawing the value chain
Before reading a single clause, draw the chain.
A typical arrangement has a foundation-model developer, a company offering that model through an API, a SaaS company that embeds the API in its product, the customer company that subscribes, and the employee who types into it. Personal data, trade secrets and customer records move along that chain, and so does responsibility.
The position that causes the most trouble is the middle one. A SaaS company is a provider to its customers and a user to its model vendor at the same time. If its customer contract promises that customer data is never used for training, that promise has to be backed by the upstream terms. It usually is not, unless someone checked.
So the first question is not "is this clause acceptable" but "does what we are about to promise downstream match what we have actually secured upstream."
1. Service, model, role and intended use
Identify what is being contracted for: which service, which model or model family, whether the vendor may substitute models, and what the service is meant to be used for.
Intended use matters more than it looks. A tool sold as producing draft text for a human to check is a different legal object from one sold as removing the human check. That difference runs through everything below — the performance you can expect, the review you owe, and who answers when the output is wrong. It is covered separately in who is liable when generative AI causes loss.
2. Input data — is it allowed to go in at all?
Before asking what the vendor may do with your inputs, ask whether those inputs may leave your company.
Customer personal data, employee data, trade secrets, information received under an NDA, source code, and material licensed from a third party each have their own constraints. An NDA with a client will often restrict disclosure to third parties, and an AI vendor is a third party.
In practice the control that works is not a clause but a rule: decide which categories of data may be entered into which tool, write it down, and make it visible where people actually work.
3. Will the vendor train on your inputs?
This is the clause people ask about, and the answer is usually configurable rather than fixed.
Check whether inputs and outputs are used to train or improve the vendor's models, whether that applies to the plan you are on, whether there is an opt-out and whether it is on by default, whether it covers both the API and the web interface, and whether sub-processors have the same restriction.
Note that "we do not train on your data" and "we retain your data for 30 days for abuse monitoring" are different statements, and both can be true. Retention, access by vendor staff, and deletion on termination are separate points from training.
4. Personal data, cross-border transfer and the processing relationship
If personal data goes in, Japan's Act on the Protection of Personal Information applies on top of the contract.
Two questions decide the analysis. First, is the vendor handling the data as a processor on your instructions, or for its own purposes? If it is improving its own models with your customers' data, the second characterisation is hard to avoid, and consent or another basis becomes the issue. Second, where is the data processed? Overseas processing engages the cross-border transfer rules, which are set out in using overseas SaaS and cross-border transfers.
The 2026 amendment to the Act tightens the rules on processors and on the information used to reach a specific individual, and introduces a surcharge for serious violations. What changes is summarised in the 2026 amendment to Japan's privacy law.
5. Security, sub-processors and incidents
Ask who else touches the data. Most AI vendors use cloud infrastructure and some use further model providers, and each is a sub-processor.
Then ask what happens when something goes wrong: the notification period, what information you receive, whether you get enough to run your own investigation and meet your own notification duties, and whether you can audit. A contract that promises notification "promptly" without specifying content leaves you unable to tell your own customers anything useful.
6. Output: what you may do with it, and who defends a claim
Two separate questions hide in the word "output."
The first is permission: may you use the output commercially, modify it, redistribute it, and does the vendor claim any rights in it? Most commercial terms now assign or license output to the customer, but the scope varies.
The second is exposure: if the output infringes a third party's copyright, trade mark or other right, who defends the claim and who pays? Several vendors offer an indemnity, and those indemnities are conditioned — on using current models, on not disabling filters, on not supplying infringing input. Read the conditions, because they describe how your engineers must actually configure the product.
7. Performance, hallucination and the human in the loop
Vendors do not warrant accuracy, and that is not going to change. The useful work is elsewhere.
Where the service matters, consider measurable performance indicators, the conditions under which accuracy is known to fall, out-of-scope uses, and what happens to performance when the model is updated. Then make your own documentation consistent with it: if your contract says the customer makes the final decision while your sales deck says human review is unnecessary, the second one is what a court will hear about.
8. Model changes, deprecation and continuity
An AI service is not stable in the way other software is. Models are updated, deprecated and withdrawn, and behaviour changes without any change to your code.
Ask for notice of material changes, a window to test, access to the prior version for a period, and clarity on how long a model you have built around will remain available. For regulated or high-stakes uses, the ability to pin a version is worth more than most price concessions.
9. Logs, audit and being able to explain
When something goes wrong, you will be asked what the system was told and what it answered.
Decide who keeps the input, the output, the model version, the settings and the timestamps, for how long, and whether you can obtain them. A vendor that deletes inputs after 30 days for privacy reasons may also have deleted the only evidence of what happened. These two goals pull against each other, and the resolution should be deliberate.
10. Liability, indemnity, termination and governing law
Caps are usually expressed as a multiple of fees paid, which on a modest subscription is a small number against the loss an AI failure can cause.
Rather than arguing about the headline cap, consider whether particular categories sit outside it — IP infringement, breach of confidentiality, data breach — and where the cap should carve out your own exposure to third parties. Then settle termination: return and deletion of data, transition assistance, and what happens to fine-tuned models or embeddings built from your data.
For a Japanese company contracting with a foreign vendor, governing law and forum are not boilerplate. A clause nominating a foreign court can make a modest claim uneconomic to bring.
When the standard terms cannot be changed
Often they cannot. A small customer does not negotiate a large model vendor's terms.
Things that still move the risk: choosing the enterprise or business tier where training is off by default, signing the vendor's data processing addendum, turning on the available retention and region controls, restricting which data categories may be entered, keeping your own logs, and writing an internal AI usage rule that people can follow.
The point is that risk allocation is not only contractual. Configuration and internal process carry a large part of it.
If you are the one providing the AI service
Everything above applies in reverse. Your terms of service, your privacy policy, your DPA and your sales material all describe the same system, and they have to agree with each other and with the upstream contract.
The common failure is a confident statement to customers — about training, retention, accuracy or indemnity — that the upstream vendor has not given you. Map each downstream promise to the upstream clause that supports it before the first enterprise customer asks.
The three I read first
Given a limited amount of time with an AI contract, I start with:
- Training and retention of inputs — because it decides what data may be used at all.
- The gap between downstream promises and upstream rights — because that gap is carried by the company in the middle.
- Model change and continuity — because it is the clause companies forget until the model they built on is withdrawn.
In short
An AI contract is not primarily an intellectual property document. It is a description of where data goes, what is done with it, how the system is expected to be used, and who carries the consequences when it is wrong.
Read it alongside the configuration you will actually deploy and the promises you will actually make to your own customers. Where those three disagree, the contract is not the thing that will protect you.