Who Owns AI Agents Built by Employees?

Reading Time: 12 minutes

Quick Answer: Under the default rule in both the UK and the US, an employer owns the work an employee creates in the course of their job. But that rule assumes a copyright exists to hand over in the first place, and where an AI agent is built mostly by prompting a model with little human input, current law in both countries treats the result as something that may not be protectable by copyright at all. What a company can reliably own is the human-authored part built around it: the prompts, the workflow design and the configuration choices that shaped the final agent.

In March 2026, the UK government recommended scrapping the copyright rule that has applied to computer-generated work for the last 35 years, right around the same time employees across marketing, operations and customer service teams started building AI agents on their own initiative, often without anyone in IT or legal knowing.

Most of what gets written about this question treats it as an access problem: who can run the agent, who can switch it off, what happens to its login when someone leaves. That is a real question, but a different one from who actually owns the agent as intellectual property. The access question is about control; the ownership question is about whether there is anything legally protectable at all, and who holds the rights to it if so. That second question is the one most businesses have not worked through properly, and it has a different answer depending on which side of the Atlantic the business sits.

Why This Question Didn’t Used to Matter

Large language models have turned employees far outside engineering and product teams into builders, with no code involved at any point:

• A customer service lead puts together an automation that drafts and routes tickets.
• An operations manager designs a prompt library and wires up an internal agent that summarises supplier calls.
• A marketing associate builds a tool that turns a brief into a first draft campaign.

This is agentic transformation in practice: not one company-wide initiative, but dozens of small agents built for everyday tasks, appearing across teams at once.

Ten years ago, this question mostly came up in R&D departments that already used invention assignment agreements. AI has turned who owns an AI agent into a company-wide question, since valuable internal tools are now built by people whose job descriptions never mentioned software.

The Default Rule for Who Owns an AI Agent

What “Created in the Course of Employment” Covers

The starting position is broadly the same on both sides of the Atlantic. In the UK, section 11(2) of the Copyright, Designs and Patents Act 1988 gives the employer first ownership of copyright works an employee creates in the course of their employment, unless there is an agreement to the contrary. In the US, the equivalent is the work-for-hire doctrine: work an employee creates within the scope of their job is treated as authored by the employer from the outset.

This rule has handled ordinary work product cleanly for decades. A report, a spreadsheet model or a piece of marketing copy is straightforwardly owned by the employer the moment it is created, with no separate assignment needed. Who owns an AI agent an employee builds at work is a harder case, because the rule was never written with agents in mind.

Where It Runs Into a Wall

Both rules assume there is a copyright to hand over. That holds for human-authored work, but gets shakier once an agent is built mostly by prompting a model and accepting the result.

The US Copyright Office has taken the position, most recently in its January 2025 report, that purely AI-generated material without meaningful human creative input is not eligible for copyright protection at all. That position has now been tested and confirmed at the highest level available: in March 2025, the DC Circuit affirmed in Thaler v. Perlmutter that copyright requires a human author, and in March 2026 the Supreme Court declined further review, leaving that ruling as settled law. If an AI agent’s design and behaviour came almost entirely from the model rather than from meaningful human authorship, the work-for-hire doctrine has nothing to assign, because there is no copyright underneath it for anyone to own.

The UK Position Specifically

UK law has a feature the US does not: a provision written for exactly this situation, decades before generative AI existed. Section 9(3) of the CDPA addresses a “computer-generated work”, defined as one produced by a computer in circumstances where there is no human author.

• Authorship is assigned to the person who undertook the arrangements necessary for the work’s creation, usually read as whoever wrote the prompt.
• The protection this creates is narrower than ordinary copyright: 50 years from creation rather than the life of the author plus 70 years, and no moral rights attach.
• On 18 March 2026, the UK government published its Report on Copyright and AI, recommending that section 9(3) protection be removed altogether, on the basis that it is unclear and rarely relied on in practice.
• No legislation has been tabled yet and the government has said it will not act without firmer evidence, but the direction of travel is clear enough that section 9(3) should not be treated as a dependable, long-term answer to who owns an AI-built agent.

Who Owns an AI Agent: How the UK and US Compare

Put side by side, the two systems answer who owns an AI agent built by an employee in similar ways, by different routes. Both converge on the same idea: ownership follows meaningful human input, not the act of pressing a button.

	UK	US
Default rule for employee work	CDPA section 11(2): employer owns	Work-for-hire doctrine: employer owns
Purely AI-generated output, minimal human input	Narrow, separate protection may apply (section 9(3))	No copyright protection at all
Who counts as “author” of AI-only output	Person who made the arrangements, usually the prompt-writer	No one; the output is not copyrightable
Direction of travel	Government recommended removing this protection (March 2026)	Settled: Supreme Court declined to revisit the human-authorship rule (March 2026)

In practice, ownership is not settled just because an employee built the agent on company time. What is protectable, in either country, is the part a person actually shaped.

Who Owns the AI Agent When Someone Leaves

All of this stays theoretical until someone hands in their notice, at which point it stops being a matter for a law firm’s blog and starts being a matter for whoever has to explain why a team’s invoice-processing agent has stopped working.

Much of the exposure here has nothing to do with malice. An employee builds something useful in a personal AI account because it is faster than waiting for an approved tool, and never mentions it to anyone. Nobody has looked at who owns it, because nobody knew it existed. That is a related but distinct problem, covered fully in our piece on shadow AI risk.

Whether ownership transfers to the employer, and whether the agent survives the departure, depends heavily on where and how it was built. The table below sets out who owns the AI agent across common scenarios, though the specific facts will always matter more than a general rule.

Scenario	Likely ownership position
Employee builds it in a company-provided AI tool, during work hours	Employer, under the default employment rule
Employee builds it in a personal AI account, for work purposes	Genuinely contested; depends on employment contract wording
Independent contractor builds it, no written agreement in place	Contractor may retain rights; company may hold only a licence
Built almost entirely by prompting, with minimal human editing	May not be protected by copyright for anyone to own

Only the first row is close to settled. Everything below it depends on facts the business often has not documented, which is exactly the gap worth closing before who owns the agent becomes a dispute rather than a policy decision.

What Actually Protects a Business

None of this means a business is defenceless, only that the default rule does not settle who owns an AI agent on its own. A few practical steps close most of the gap.

• Document the human contribution as it happens. The prompts, edits and choices about what to keep are what make an agent protectable, and a record of that contribution settles who owns it far more cheaply than reconstructing it after a dispute starts.
• Treat prompt libraries and agent configurations as confidential. Trade secret protection does not depend on the copyright question being resolved first, and it protects exactly the material most likely to walk out the door with a departing employee.
• Update employment contracts and AI use policies. Name which tools are authorised and state plainly that resulting work belongs to the company. Most employment agreements were drafted before generative AI existed and simply do not address it.
• Get contractor assignments in writing before work starts. Use present tense assignment language rather than a promise to assign later. This gap is often larger than expected, since protections that apply automatically to employees do not extend to contractors.

Some of this exposure only exists because agents get built in personal environments in the first place. A business that never allows that has less to dispute later.

Closing the Gap by Design

Ownership disputes and knowledge loss are two symptoms of the same problem: automation built ad hoc, by one person, outside any central system. The same agent that raises the question of who owns it is often also the agent nobody else in the business can find or reuse once its builder moves on.

The alternative is to remove the grey period rather than resolve it after the fact. Interviewing every team to surface the work worth automating, then building and hosting the resulting agents centrally from the start, means an agent is never created inside one person’s personal account to begin with. It is company infrastructure from the moment it exists, and the question of who owns the AI agent never has room to come up.

That also solves the continuity problem directly. When an agent lives in a shared, company-owned system rather than an individual’s account, the business does not depend on that person staying to keep using it. The agent stays. Who happened to build it becomes a far less important question.

If it would help to see how this looks in practice, GrowthNation can show you the system: map a single team, see the scorecard, and hand over a working agent, built and owned centrally from day one. No lengthy deck involved.

FAQ

Who owns an AI agent that an employee builds using a company-approved AI tool?

Under the default employment rule in both the UK and the US, the employer owns work an employee creates in the course of their job, which answers who owns an AI agent built using a company-approved tool, provided the agent involved enough human input to be protectable.

Does it matter if the employee used a personal AI account instead of an approved one?

Yes. A personal account moves the situation into contested territory, because the automatic employment ownership rule assumes company resources and company time, not tools an employee subscribed to individually. Most employment contracts do not address this scenario at all.

Who Owns an AI-Generated Agent, Exactly?

Only the parts a person meaningfully shaped. Current law in the UK and the US does not protect output produced almost entirely by an AI model with minimal human input. What remains protectable is the human-authored scaffolding around it: the prompts, the architecture and the editorial choices behind it.

Is the UK’s rule on computer-generated works (section 9(3)) still valid?

Yes, for now. Section 9(3) of the CDPA still assigns authorship of a computer-generated work to whoever made the arrangements necessary for its creation. The UK government’s March 2026 report recommended removing this protection, so it should not be treated as a durable basis for ownership.

What happens to an AI agent when the employee who built it leaves?

It depends on where and how the agent was built. An agent built inside a company-approved, centrally managed system typically continues to belong to the business. An agent built inside a personal AI account is far less certain to transfer, and may not even be accessible once that person leaves.

What should a company do now to protect its position?

Document human contributions as they happen, treat prompt libraries as confidential regardless of copyright status, update employment contracts and AI use policies to name authorised tools, and put written assignment agreements in place for contractors before work begins.