Contract Clauses for an AI Project: Ten Points a Standard Software Contract Misses
November 11, 2026 · 7 min read · Etzal Group Team
A company signs its usual software development contract for an AI assistant that will answer customer questions from its policy documents. Six months after launch the model provider retires the version the system was built on. Answers change. Some get worse. The client says the system no longer meets the specification. The vendor says the code is unchanged and the specification never mentioned model versions. Both are reading the contract correctly, because the contract says nothing about it. Good contract clauses for an AI project exist to settle such questions before they arise.
This article describes what to cover and why. It is not legal advice. Have a lawyer qualified in your jurisdiction draft and review the actual terms.
1. Data: what the vendor and its suppliers may do with it
The question: what happens to the data we provide and the data users enter?
Cover:
- Purpose. Data is used only to deliver this project.
- Training. Neither the vendor nor any model provider it uses may train general models on your data, unless you agree in writing.
- Retention. How long prompts, outputs and logs are kept, by whom, and how they are deleted.
- Location. Where data is processed and stored, and rules for transfer across borders.
- Sub-processors. A list of third parties that receive data, including the model provider, with notice before any change.
- Personal data. Roles and duties under the applicable law, such as Indonesia's personal data protection law or Singapore's PDPA, including breach notification between the parties.
Ask to see the model provider's terms for the specific service and plan used. The vendor's promise is only as good as what its own supplier has agreed.
2. The model: which one, and what happens when it changes
The question: what are we actually buying, when the central component belongs to someone else?
Cover:
- The model and version at launch.
- Notice of provider deprecations as soon as the vendor learns of them.
- An obligation to re-test against the acceptance set before moving to a new version.
- Who pays for migration work when a version is retired. This will happen, so price it or agree a rule.
- The client's right to approve a change of model provider.
3. Acceptance: quality as a measured level
The question: how do we know it is done, when no AI system is right every time?
Replace "the system answers correctly" with:
- A test set agreed by both sides: representative questions or documents with expected results, including difficult and out-of-scope cases.
- Metrics and thresholds. For example, at least 90 percent of answers rated correct, no more than 2 percent rated seriously wrong, and all out-of-scope questions declined or handed over.
- Who rates, and how disagreements are settled.
- Re-measurement after any change to model, prompts or knowledge.
The figures here are illustrations. The right thresholds depend on the use and should be set from a trial, not guessed.
4. Ongoing quality: after go-live
The question: what if quality falls later?
Cover:
- Periodic re-measurement on the test set, for example quarterly.
- Monitoring of live quality by sampling.
- A remedy if quality falls below the agreed level: a fix within a stated time, service credits, or the right to terminate.
- Who maintains the knowledge base and who is responsible when outdated content causes wrong answers. Usually the client owns content, and the contract should say so.
5. Intellectual property: who owns what
The question: what do we own at the end?
Sort the assets:
- Client data and documents: the client.
- Prompts, configurations, workflows and evaluation sets created for the project: state it. A client that paid for them usually expects to own them.
- Fine-tuned model weights or embeddings built from client data: state it, and state where they are held.
- Vendor's pre-existing tools and libraries: the vendor, with a licence for the client to keep using what the system needs.
- Outputs generated by the system: the client, as far as the law recognises rights in them.
6. Third-party content and infringement
The question: who bears the risk if output reproduces someone else's protected material?
Cover whether the vendor passes through any indemnity offered by the model provider, what conditions apply to it, and what each side must do to stay within those conditions. Be realistic: vendors can rarely offer more than their provider gives them.
7. Running costs
The question: what will this cost each month, and who carries the risk of usage growth?
Cover:
- Whether model usage fees are billed by the provider directly to the client, or through the vendor with or without a margin.
- An estimate at stated volumes, and the assumptions behind it.
- Alerts and caps on usage.
- What happens when the provider changes its prices.
A fixed monthly fee that includes usage is simple and hides a risk for one side. A pass-through at cost with a cap is usually clearer.
8. Security and misuse
The question: what protects the system from being manipulated?
Cover:
- Access controls and the principle that the system has only the permissions it needs.
- Protection against instructions hidden in content the system reads.
- Logging of inputs, outputs and actions, with retention periods.
- Security testing before launch, including attempts to make the system reveal data or act outside its limits.
- Incident notification: how fast the vendor must tell you, and what it must provide.
9. Human oversight and limits of use
The question: which decisions may the system take alone?
State what the system is for, what it must not be used for, and where a person must approve. If the system can act, such as sending messages or changing records, list the actions and their limits. This protects both sides: the client from uncontrolled automation, and the vendor from liability for uses it never agreed to support.
10. Liability and exit
The question: who pays when it goes wrong, and how do we leave?
Liability. Vendors will not accept unlimited liability for the output of a probabilistic system, and clients should not accept none. Common ground: the vendor is responsible for building to the agreed quality level, for security and for following instructions, and the client is responsible for the decisions it makes using the output and for its content. Caps and exclusions follow from that division.
Exit. On termination, the client receives its data, prompts, configurations, evaluation sets, documentation and any model artefacts it owns, in usable formats, with a period of assistance. The vendor deletes client data and confirms it. Without this clause, leaving means starting again.
A short checklist before signing
- Do we know which model is used and what happens when it changes?
- Is acceptance defined by a test set and thresholds?
- Is our data excluded from training, with retention and location stated?
- Do we own the prompts, configurations and evaluation sets?
- Are running costs estimated, with caps?
- Is there a remedy if quality drops?
- Can we leave with everything needed to continue elsewhere?
Seven yes answers do not guarantee a good project. A no on any of them is a known gap.
What not to expect
- A guarantee of correctness. Agree a measured level and a remedy.
- A promise that the model will never change. Agree what happens when it does.
- Terms better than the model provider gives the vendor. Read the provider's terms.
We raise these points ourselves when proposing AI work, because a contract that is silent on them serves neither side when the first model deprecation arrives. See how we approach projects in AI Consulting, or contact us to discuss a project you are scoping.
Summary
An AI contract should cover what standard terms miss: use, retention and location of data; the model version and who pays when it changes; acceptance and ongoing quality as measured levels with remedies; ownership of prompts, configurations and model artefacts; running costs with caps; security and incident notice; limits of use and human oversight; a realistic division of liability; and an exit that leaves the client able to continue. Have a qualified lawyer write the words.
Frequently asked questions
Why does an AI project need different contract terms?
Three things differ from conventional software. Output is probabilistic, so "works as specified" has to be defined as a measured quality level. The system usually depends on a model from a third party that can change. And your data flows through it, which raises questions of use, retention and location. Standard clauses are silent on all three.
Who owns the prompts and the fine-tuned model?
Whatever the contract says, so it should say. A common fair position is that the client owns prompts, configurations, evaluation sets and any model weights trained on its data for the project, while the vendor keeps its pre-existing tools and general know-how. Unstated, this becomes a dispute at exit.
Can a vendor guarantee that an AI system will be accurate?
Not absolutely, and a vendor who promises it should be questioned. What can be agreed is a measured level on an agreed test set, for example a minimum share of correct answers and a maximum share of serious errors, with a method for re-measuring after changes and a remedy if quality falls below the level.
Related service
WhatsApp AI Chatbot
We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.
