AI Agent Intellectual Property: Who Owns the Code, Model, Data and Outputs?

Map ownership, access and control separately

Three questions should be kept distinct. Ownership asks who holds the relevant legal or contractual right. Access asks whether a party can technically reach or use the component. Control asks who can change, suspend, terminate or constrain that access. An agent may be fully controlled by the seller at the application layer while remaining dependent on an external model API that the seller does not own and cannot guarantee indefinitely.

Copyright in source code starts with authorship and chain of title

Human-authored software can attract copyright protection where the jurisdictional requirements are satisfied. For commercialization, the practical issue is often not whether source code can be protected, but whether the company can prove that it owns the relevant rights. Employment agreements, founder assignments, contractor agreements, repository history and third-party contributions should be reviewed before licensing or acquisition.

Patent rights can protect qualifying technical inventions

An AI-agent product can include technical features that merit patent analysis, such as orchestration, persistent state, secure tool execution, identity and authorization, distributed agent coordination, device control or resource optimization. Patentability depends on the actual claims and jurisdiction. The dedicated AI-agent patents page should own the detailed international comparison rather than repeating it here.

Read the AI Agent Patents guide →

Inputs and customer content require their own rights analysis

A customer may own information supplied to the agent, but the seller still needs the contractual authority to process that information through its systems and relevant providers. OpenAI and Anthropic both place responsibility on business customers to possess the rights or permissions needed for supplied content. That requirement does not establish that every downstream use of the data is permitted.

Data ownership, processing, training and reuse should not be collapsed into one concept

For each important dataset or knowledge source, ask who owns or controls it, whether the seller may process it, whether a model or service provider may retain or use it for training, and whether outputs or retrieved material may be redistributed commercially. Microsoft Azure AI documentation is useful because it distinguishes customer data storage and service processing from use of customer material for foundation-model training.

Microsoft: Azure AI documentation states that customer fine-tuning data is not used to train foundation models without the customer's permission or instruction and explains customer data handling across relevant services. Microsoft Azure AI data privacy guidance.

Open-source and open-weight components can create different obligations

Public availability is not the same as unrestricted commercial use. Model repositories and software ecosystems contain permissive, copyleft, research-only, non-commercial and other licenses. Each material component should be reviewed individually before the seller promises unrestricted enterprise deployment or sublicensing.

License taxonomy: Hugging Face lists a broad range of software, model and data licenses, including MIT, Apache, GPL-family, OpenRAIL and Creative Commons variants, showing why individual license review is necessary. Hugging Face repository licenses.

Trade-secret protection can preserve valuable non-public implementation knowledge

Where technical know-how is not easily observable from the product and disclosure is not required for patent strategy, confidential orchestration logic, internal evaluation methods, operational tuning, deployment know-how or customer-specific implementation methods may have trade-secret value subject to applicable law and reasonable secrecy controls. Patent and trade-secret strategy should be coordinated because patent filing requires disclosure while trade-secret protection depends on secrecy.

Trademarks protect market identity, not the underlying agent functionality

The product name, logo or other source identifier may be protectable as a trademark if legal requirements are met. Trademark rights should therefore be considered separately from rights in software, models, patents and data. A company acquiring an AI-agent business may need both the technical assets and the market identity that customers recognize.

Acquisition diligence should test whether the important rights can actually transfer

When an AI-agent company or product is sold, the transaction may need to address patents and applications, source code, domains, trademarks, confidential know-how, customer contracts, datasets, model-provider arrangements, APIs, marketplace accounts and third-party licenses. Some rights can be assigned freely, some require consent, and others may be non-transferable. The buyer's economic value can therefore depend on contract portability as much as nominal ownership.

AI Agent IP Chain-of-Title Checklist

• Confirm founder, employee and contractor assignments for source code and inventions.
• Identify all patent applications and determine the correct applicant, assignee and inventors.
• Review model-provider agreements and distinguish service access from model ownership.
• Identify all external APIs, MCP tools and integrations that are material to the product.
• Review open-source and open-weight licenses used in production.
• Confirm rights in training data, retrieval corpora and customer-provided content.
• Document provider and customer positions on generated outputs.
• Identify trade secrets and verify reasonable confidentiality controls.
• Check trademark ownership and clearance for the product identity.
• Confirm that material customer, marketplace and vendor contracts can survive investment or acquisition.

Frequently asked questions

Who owns an AI agent built on OpenAI, Anthropic or Azure?

The application developer may own its own code and other proprietary assets, while the external provider retains rights in its own model or service. The exact position depends on the relevant contracts, contributions and applicable law.

Does owning AI output mean the output is copyrighted?

Not necessarily. A provider contract can allocate rights between provider and customer, while statutory copyright protection is determined separately under applicable law. In the United States, human authorship remains central to copyright protection.

Who owns a fine-tuned AI model?

The answer depends on the technical implementation and provider terms. Fine-tuning can involve customer-provided data and configuration layered onto provider-controlled technology, so ownership and control should be checked rather than assumed.

Can contractors retain rights in AI-agent code?

Potentially, depending on the contract and applicable law. Companies should use written agreements that clearly address ownership and assignment of relevant work product and inventions.

Can an AI-agent business be sold if it relies on third-party models?

Potentially, but buyers will examine whether the relevant provider, API, data, marketplace and customer contracts can continue or transfer after the transaction. Dependency portability can affect valuation and deal structure.