An agent’s autonomy is not the same as your company’s ownership
An AI agent can generate code, text and analysis, query archives and trigger tools. The commercial value of the result, however, does not by itself prove that an exclusive right exists.
When an AI agent produces a deliverable, the question “who owns it?” bundles together several distinct issues: whether the result is protectable at all; who has acquired any resulting rights; whether the company can use it without infringing the rights of others; and who answers for the actions carried out through the system.
Conflating these layers is particularly risky in cross-border transactions. A clause assigning outputs to the customer may govern the relationship with the vendor, but it does not create a copyright where the law does not recognise one, and it does not automatically bind third parties.
The strategic question is not only “who owns the output?” but “which rights can we prove, enforce and transfer in the result, and in the elements that make its production possible?”
Seven checks before deployment
Output
Is the result protectable? What human creative contribution can be documented?
Data
Which rights, licences and legal bases allow the use of documents, datasets and personal data?
Models
Which terms govern the model, fine-tuning, configurations and open-source components?
Vendors
Which rights are granted, and which warranties, exclusions and caps accompany the service?
People
How are the contributions of employees, consultants, developers and partners regulated?
Accountability
Who authorises actions, verifies outputs and handles errors, disputes and incidents?
Evidence
Can sources, versions, human reviews, approvals and contractual terms be reconstructed?
One workflow, three levels of verification
The technology flow should be read together with the provenance of inputs, the protectability of the result and the authorisation of operational actions. In the diagram, the highlight moves from one step to the next.
On a smartphone, scroll the diagram horizontally: the text is not compressed to fit every node on the screen.
Italy: human contribution remains decisive
Article 25 of Italian Law no. 132 of 23 September 2025 amended Article 1 of Law no. 633 of 22 April 1941 (the Italian Copyright Act). The provision refers to works of human ingenuity of a creative nature and extends to those created with the aid of artificial intelligence tools, provided they result from the author’s intellectual work.
Using AI therefore does not, in itself, preclude protection. But owning the account, funding the workflow or issuing a command is not enough: what matters is the creative expression that can concretely be attributed to a human being. Selection, coordination and reworking of materials must be assessed case by case, and not every formal review or corrective edit demonstrates a protectable creative contribution.
The same provision introduced Article 70-septies into the Copyright Act, which cross-refers to Articles 70-ter and 70-quater for reproductions and extractions of text and data carried out through AI models and systems. It also introduced a specific criminal offence, linked to Article 171, for unlawful extractions. For companies feeding an agent with third-party content, the exposure is not merely civil.
Protectability and acquisition of rights remain separate questions. For software and databases created by an employee in the performance of their duties, Article 12-bis of Law 633/1941 is relevant, for example; for other categories of works, and for consultants, the contract and the applicable rules must be examined.
An assignment transfers rights that exist and are available. It does not turn a result lacking the legal requirements into a protected work, and it does not eliminate third-party rights.
European Union: four regimes not to be confused
Copyright and text and data mining
Articles 3 and 4 of Directive (EU) 2019/790 provide specific exceptions for text and data mining, with different conditions and beneficiaries.
Article 4 allows rightsholders to reserve their rights, including by appropriate machine-readable means for content made publicly available online.
The fact that content is accessible on the web does not authorise every use. Training, indexing, retrieval, reproduction in outputs and publication each require separate analysis.
GPAI models and Article 53 of the AI Act
Article 53 of Regulation (EU) 2024/1689 requires providers of general-purpose AI models to maintain a policy for compliance with Union copyright law and to publish a sufficiently detailed summary of the content used for training.
The obligations in Chapter V apply in general from 2 August 2025. For models placed on the market before that date, Article 111(3) provides for compliance by 2 August 2027.
A customer that integrates a model does not automatically become the provider of that model. Qualification requires an examination of the activities actually performed.
Transparency: roles and content determine the obligations
Article 50 of the AI Act distinguishes the obligations of providers from those of deployers. Technical marking of synthetic outputs is not the same as informing people that they are interacting with an AI system, or disclosing the artificial nature of certain content.
Paragraph 2 concerns machine-readable marking and detectability of synthetic content, within the limits and exceptions provided. Paragraph 4 specifically addresses deepfakes and certain texts published to inform the public on matters of public interest.
There is no single, undifferentiated duty to attach the same visible label to every corporate output. Role, content, use and exceptions must be checked, together with the application dates currently in force: the AI Act timetable is the subject of simplification measures at EU level and should be verified at the time of assessment.
Personal data, trade secrets and confidential information
If the workflow processes personal data, the GDPR requires a separate analysis: purpose, legal basis, data minimisation, roles, security, any impact assessment and transfers to third countries.
A licence to use an output does not demonstrate that the processing of data contained in the input or the result was lawful. Trade secrets and confidential information also require dedicated safeguards.
The international dimension: a contract does not create a global right
The international framework does not grant a uniform right in agentic outputs. Intellectual property treaties do not replace the analysis of national laws and of the rules applicable to each individual dispute.
In the United States, the Copyright Office requires human authorship: AI-generated material is not protected as such, while human creative contributions meeting the requirements may be. Registration applications must correctly describe the human contribution and the AI-generated material. The Office’s 2025 report on copyrightability confirms that prompts alone are generally not sufficient.
Other legal systems reason differently. The United Kingdom has a specific regime for computer-generated works, currently under debate and review, while in China some decisions have recognised protection for images produced with AI where there was a significant human contribution. The choice of law governing a contract does not by itself determine protectability in every country and does not neutralise third-party rights.
| Level | Core issue | Operational consequence |
|---|---|---|
| Italy | Human ingenuity and the author’s intellectual work, including with AI assistance; TDM rules. | Document the creative contribution and verify the acquisition of economic rights. |
| European Union | Copyright, TDM, GPAI, transparency and personal data follow distinct regimes. | Do not equate AI compliance with ownership of the output. |
| United States | Human authorship and accurate disclosure of AI material in registration. | Separate generated elements from human contributions and keep evidence. |
| United Kingdom and China | Different approaches to computer-generated works and human contribution. | Check the target market before promising exclusivity. |
| Cross-border transactions | Territorial protection and contractual regulation do not overlap. | Coordinate contracts, markets, dispute handling and evidence. |
Data is not a single category of property
Speaking generically of data ownership can conceal very different situations: copyright in content, the sui generis right of database makers (Articles 102-bis et seq. of Law 633/1941), trade secrets, confidentiality, contractual access rights and personal data protection.
A company should map not only who holds a dataset, but who can access it, for what purposes, with which tools, for how long, and with what rights of reuse or transfer.
The chain of title of the agentic company
Due diligence should not stop at the formula “outputs belong to the customer”. It must reconstruct the contributions and distinguish pre-existing rights from those possibly developed in the project.
| Asset | Verification | Documentary safeguard |
|---|---|---|
| Code and orchestration | Ownership of contributions and compatibility of component licences. | Development agreements, IP provisions, dependencies and version control. |
| Prompts and workflows | Possible protectability and protection as confidential information or know-how. | Restricted access, confidentiality and reuse rules. |
| Knowledge base | Content, databases, personal data and confidential information. | Sources, licences, purposes, authorisations and retention. |
| Output | Human contribution, third-party rights and vendor terms. | Reviews, approvals and allocation of existing rights. |
| Feedback and improvements | Vendor reuse and use for training or service development. | Clauses on training, confidentiality, licences, deletion and exit. |
What to ask your model vendor
Negotiation with the vendor is where a company can make the points above concrete. Some clauses deserve priority attention:
- Output and training: clarify whether inputs and outputs may be used to train or improve the service, and which opt-out options exist.
- Indemnity and warranties: check whether protection against third-party claims over outputs is offered, under what conditions and within what liability caps.
- Exit and continuity: provide for export of data and configurations, verifiable deletion and notice of material changes to the service.
Accountability: software does not break the legal chain
An agent does not become a legally responsible party because it executes a plan autonomously. The analysis concerns the people and organisations involved and depends on obligations, conduct, powers, causation and damage.
Not every error automatically falls on the director, the employee or the vendor. Contractual, tortious and regulatory liability must be assessed separately, together with any other relevant aspects.
- Before the action: define permissions, spending thresholds, prohibited operations, approvals and credential management.
- During the workflow: log the necessary events, control exceptions and allow human intervention.
- After an incident: preserve evidence, contain the effects and assess remedies, communications and notification duties.
Traceability must be proportionate: it does not justify indiscriminate retention of personal data, trade secrets or credentials.
Red flags that reduce the value of the asset
Outputs presented as exclusive without verifying protectability and human contribution.
Customer or partner archives used without analysing authorisations and purposes.
Contracts that generically transfer “all the AI” without distinguishing software, models, data and deliverables.
External actions with no limits and no way to reconstruct who approved an operation.
Vendor dependence with no workable plan for export, continuity and deletion.
Documentary readiness checklist
A guidance tool for organising the review. Completing it does not certify compliance or absence of risk.
0 of 8 measures selected
Selections stay on the current page: the script does not save or send them.
Legal and institutional sources
- Italian Law no. 132 of 23 September 2025 — Article 25, amendments to copyright law (Italian text).
- Italian Law no. 633 of 22 April 1941 — Italian Copyright Act (Arts. 1, 12-bis, 70-ter, 70-quater, 70-septies, 102-bis, 171; Italian text).
- Regulation (EU) 2024/1689 — AI Act — Articles 50, 53, 111 and 113.
- Directive (EU) 2019/790 — Articles 3 and 4, text and data mining.
- Regulation (EU) 2016/679 — GDPR — processing of personal data.
- U.S. Copyright Office — Works Containing Material Generated by Artificial Intelligence — registration guidance on human authorship; see also the Office’s series of reports on copyright and AI.
- WIPO — Intellectual Property Policy and Artificial Intelligence — discussion document, not binding.
Closing considerations and reader notice
Assessing an agentic workflow requires connecting technology assets, intellectual property, data and accountability: from verifying the chain of title to contracting with vendors and developers, through to the conditions for deploying the system internationally. The goal is not to attribute to the company an abstract ownership of AI, but to identify provable rights, sustainable controls and terms of use consistent with the business model.
This article is for general information only and does not constitute legal advice or a certification of compliance. Conclusions depend on the use case, human contributions, contracts, categories of data, sector and jurisdictions involved. Specific transactions and decisions require a dedicated analysis of the framework applicable at the time of assessment.
// Commenti