Disclaimer
This guide is not legal advice. It is general information for education and discussion, and reading it does not create a lawyer and client relationship. Get advice from a qualified lawyer before you act on it, because it does not take account of your circumstances, your contract or the law of your jurisdiction. The guide is not specific to any one jurisdiction. It draws generally on common law principles of commercial contracting, so the position may differ in a civil law country or where local statute overrides the contract. The author accepts no liability for any loss that arises from reliance on it. Vendor terms are quoted from public versions as at the update date at the top of this page, and vendors change them often. The clauses and dialogues are illustrative and do not describe any client matter, and the views are the author’s own and not those of any employer or client.
Overview
Purpose of the guide
This guide explains how contracts for AI systems are negotiated, and why each party takes the positions it does.
AI regulation differs by country. Some binding rules already apply, such as parts of the EU AI Act, while elsewhere most rules remain soft law. Even so, the parties to an AI deal must decide now who bears each risk. The right position usually depends on the AI system, the data, the deal size and the bargaining power.
This guide is for lawyers and commercial teams on either side who know contracts but not AI. It covers the two kinds of AI deal in the box below. Many checklist issues apply to both, although issues such as agents and pricing arise mainly in AI as a service, and consent to the use of AI arises mainly in AI in service delivery.
The Foundations explain how an AI system works and what shapes the contract. The Checklist covers 18 issues where AI changes the usual position, each with a simulated negotiation, and Interconnecting issues shows how they connect. Hover over an underlined term for its definition.
Two kinds of AI deal
The customer buys access to an AI system itself, such as a model through an API, an AI application or an agent.
A supplier uses AI to deliver another service, such as outsourcing, consulting or creative work, and the customer buys the result rather than the AI system.
Both kinds of deal call for the same core analysis. In each, the parties need to follow the data, work out who controls each part of the system and place each risk with the party best able to manage it, so the issues on data use, retention, output quality, IP and liability apply in much the same way.
The deals differ in what the customer is buying. While a customer buying AI as a service negotiates the terms of the AI system itself, such as service levels, model changes and price, a customer buying a service delivered with AI usually leaves those terms to the supplier and focuses on the result. That customer instead asks whether the supplier must disclose its use of AI or obtain consent for it, how the supplier checks the output before delivery and who owns the deliverables. Although the supplier carries the delivery risk in that case, its ability to stand behind the service still depends on the terms it obtained from its own AI vendor.
Foundations
The core foundations
Summary of the four principles
- Understand the picture. Start by understanding the components of the AI system and where the parties sit in the broader AI value chain, so that you can see what issues are at stake.
- Understand the bargaining power game. Where a party sits in the AI value chain typically sets how much leverage it has to assert its positions and interests.
- Follow the data. Track how data moves through the AI system, from input to output, and this will help you uncover issues around consent, permitted use, infringement and security.
- Understand the interplay between control and risk. Once you have followed the data and worked out the contentious issues, how each party addresses them comes down to an equation between control and risk, where each risk is best allocated to the party that controls the part of the system where it arises.
Principle 1
Understand the picture
The small picture
To understand the various moving parts in an AI contract, it is important to understand the parts of a single AI system and how one request moves through them.
- Inputs. The customer supplies most of the inputs, which are the prompts that its users type and the files and records that it connects. The contract therefore needs to say who owns the inputs, what the vendor may do with them and whether the customer has the right to supply them.
- Application. The application adds its own instructions, retrieves relevant documents and applies guardrails, which are controls that block unsafe requests and outputs. Because the application vendor controls this part, it is the part that the vendor can most readily warrant and support with service levels.
- Model. The model differs from ordinary software, because it produces each output by prediction and the same input can produce different outputs. Although the application vendor sells the AI system, it often buys access to the model from a model provider. This explains why vendors resist a warranty of accuracy, and why the contract deals with third-party models and model changes.
- Output. The output is the text, code or recommendation that the system returns to the customer. The contract needs to say who owns the output, who is liable if it is wrong and who answers a claim that it infringes another person’s rights.
- Storage. The system also stores data that the customer may not see, because the application vendor and the model provider may keep logs of prompts and outputs. Either of them may want to use that data to improve its product, so the contract sets retention periods and limits on training.
A further note on agentic AI
An agent is an AI system that works towards a goal in several steps and can act in other systems.
- Input. The customer gives the agent a goal and a set of permissions in place of a single prompt. The permissions limit what the agent may access and do, so the contract should record them and say who is responsible for setting them.
- Agent. The agent asks the model to plan a step, carries it out, reads the result and repeats until the task is done. Because the agent chooses its own steps, the contract needs a definition of an agent action and a rule on who is liable for one.
- Tools. The agent sends data to tools and connectors that third parties often build and run. The data therefore travels further than in a chat-based AI system, and the contract should say which party answers for each tool and which terms apply to it.
- Action. The result is an action in a real system, such as a payment or a changed record. While a person can check an output before relying on it, an agent may act before anyone has checked, so the liability cap and the exclusions matter more.
- Human approval. A person can approve or block an action before it happens where the AI system requires it. The contract usually allocates this oversight duty to the customer, although the duty is fair only for actions that the customer can in fact see and stop.
The big picture
In addition to understanding the components of AI / agentic system, it is also important to understand where the relevant AI system fits within the bigger picture of the AI value chain.
The layers of the value chain now compete with each other to sell direct to the enterprise customer.
- Cloud platforms offer their own models alongside other companies’ models, and they sell services for building agents on top of them.
- Model providers sell enterprise subscriptions and their own agents direct to the customer, so they now compete with the application vendors that buy access to their models.
- Application vendors try to reduce their dependence on one model provider by using several models or by training their own.
This competition matters for the contract. Although buying direct from a cloud platform or a model provider removes a layer from the chain, the customer then usually buys on that party’s standard terms and has less room to negotiate. While an application vendor tends to negotiate more readily, it can promise only what it receives from the parties above it. The same AI system can therefore reach the customer through several routes, and each route comes with different terms.
Pick a layer to read a profile of that party.
Principle 2
Understand the bargaining power game
Each side’s bargaining power comes from two things.
- Its alternative if the deal fails. This is commonly called the BATNA (best alternative to a negotiated agreement). A customer that can switch to another model has a strong alternative, and so does a vendor with many buyers waiting.
- How much each side needs the other. The vendor’s need depends on the customer’s spend, brand and use case, while the customer’s need depends on how unique the AI system is and how deeply it is embedded.
Each new category of AI system starts with few vendors, so vendors hold more power at first. As more vendors enter, customers gain power until the next category arrives. Although customers had gained ground on chatbot terms by 2025, vendors of agentic AI systems appear to hold more power in 2026.
Place your deal on the chart
Gauging the bargaining power of each party early tells you which positions are realistic before you spend effort on them. A party with little leverage should focus on the few issues where it is most exposed and accept standard terms on the rest, while a party with strong leverage can ask for protections that suit its own needs. Either way, the next step is to work out which issues matter most in the deal. In my view, following the data through the AI system is the most reliable way to find them, because most contentious issues in an AI contract arise where data is collected, used, stored or reused.
Principle 3
Follow the data
Data is the core focus of AI contracting because an AI system runs on data at every stage. The customer’s prompts and files go in, the model was trained on data that the customer did not supply, and the system creates new data in its outputs and logs. Each of these raises a question that ordinary software contracts rarely face, such as whether the vendor may train on the customer’s content, who owns the output and whether either could infringe another person’s rights. While the legal issues can look scattered, they tend to arise at the points where data changes hands, so tracking the data flow leads to the clauses that matter most.
Start where the data is collected or uploaded, follow it through the application and the model, and finish at the output and at whatever is stored afterwards. At each stage, ask four questions. The table shows which question matters most at each stage.
- Who consented?
- What use is allowed?
- Could anything infringe?
- Is the data secure?
This exercise shows where the parties tend to disagree. Whereas the vendor and the model provider want to use prompts and outputs to improve their products, the customer wants to control each use.
Most data clauses in an AI contract, such as training restrictions, retention limits and ownership of output, answer one of the questions in the table.
| Stage | What happens to the data | Questions to ask |
|---|---|---|
| 1. Upload | Users type prompts and the customer connects files and records. | Who consented, and does the customer have the right to send it? |
| 2. Application | The application adds instructions and retrieved documents. | Where is the data processed, and who can see it? |
| 3. Model | The model provider’s model reads the prompt and the context. | Does the provider keep the data or train on it? |
| 4. Tools | An agent sends data to tools and connected systems. | Which third parties receive it, and on what terms? |
| 5. Output | The system returns text, code, a recommendation or an action. | Who owns it, and could it infringe or cause harm? |
| 6. Storage | Logs, stored prompts and agent memory remain after the task. | How long is the data kept, and can the customer delete it? |
Principle 4
Understand the interplay between control and risk
A long-standing principle of contract risk allocation says that a risk should be allocated to the party best able to control it. The party that controls a risk can prevent the loss at the lowest cost, so giving it the liability gives it the incentive to do so.
Control over the model tends to fall at each step away from the model provider, while willingness to negotiate tends to rise, because suppliers become smaller and compete more. The party most willing to agree to a clause is therefore often the party least able to perform it.
The table below takes a more granular look at what this control dynamic means for the contract. It lists the parts of an AI system that the parties usually negotiate over, and it marks the party that typically controls each one. While the customer controls how it deploys and uses the AI system, the parties further up the value chain control the application, the model and the infrastructure, so the contract tends to place each risk with the party marked in that row.
Vendor control and data sensitivity
The chart below depicts this interplay. The complexity of a negotiation typically turns on two factors, which are the vendor’s control over its model and the sensitivity of the customer’s data. Here the vendor is whichever party supplies the AI system to the customer, so it may be a cloud platform, a model provider, an application vendor or an integrator, depending on where the deal sits in the value chain. A vendor that built and trained its own model knows what went into it and can stand behind it, whereas a vendor that builds on another company’s model can usually promise only what its own supplier promised to it. On the customer’s side, personal information, confidential information and know-how need more protection than public material. While a deal in the bottom right of the chart is easy to close, a deal in the top left is hard, because the customer that needs the most protection is dealing with the vendor least able to give it.
Place your deal on the chart
Guidance
Unlock the rest of the guide with a paid subscription
The Overview and Foundations are free to read. The practical guidance that builds on them is available on any paid plan.
- ✓Checklist of issues. 18 issues where AI changes the usual contract position, with the context and the market practice for each.
- ✓Simulated negotiations. A redline for each issue, played round by round between the customer and the vendor.
- ✓Interconnecting issues. A clause map that shows how a position on one issue changes what a party needs on another.
- ✓Glossary. Definitions of 43 key terms.